Upload 194 files

FluxMusic_jupyter.ipynb

@@ -0,0 +1,190 @@
+{
+    "cells": [
+        {
+            "cell_type": "code",
+            "execution_count": null,
+            "metadata": {
+                "id": "VjYy0F2gZIPR"
+            },
+            "outputs": [],
+            "source": [
+                "%cd /content\n",
+                "!git clone -b dev https://github.com/camenduru/FluxMusic\n",
+                "%cd C:/Users/Curt/Developer/AItools/AIaudio/AudioCreation/FluxMusicJupyter/FluxMusic\n",
+                "\n",
+                "!apt -y install -qq aria2\n",
+                "!aria2c --console-log-level=error -c -x 16 -s 16 -k 1M https://huggingface.co/audo/FluxMusic/resolve/main/musicflow_b.pt -d C:/Users/Curt/Developer/AItools/AIaudio/AudioCreation/FluxMusicJupyter/FluxMusic -o musicflow_b.pt\n",
+                "\n",
+                "!pip install transformers diffusers accelerate einops soundfile progressbar unidecode phonemizer torchlibrosa ftfy pandas timm matplotlib numpy==1.26.4 thop flash-attn==2.6.3 sentencepiece"
+            ]
+        },
+        {
+            "cell_type": "code",
+            "execution_count": null,
+            "metadata": {
+                "id": "NoTEt9Wto70D"
+            },
+            "outputs": [],
+            "source": [
+                "%cd C:/Users/Curt/Developer/AItools/AIaudio/AudioCreation/FluxMusicJupyter\n",
+                "\n",
+                "import os\n",
+                "import torch\n",
+                "import argparse\n",
+                "import math\n",
+                "from einops import rearrange, repeat\n",
+                "from PIL import Image\n",
+                "from diffusers import AutoencoderKL\n",
+                "from transformers import SpeechT5HifiGan\n",
+                "\n",
+                "from utils import load_t5, load_clap, load_ae\n",
+                "from train import RF\n",
+                "from constants import build_model\n",
+                "\n",
+                "def prepare(t5, clip, img, prompt):\n",
+                "    bs, c, h, w = img.shape\n",
+                "    if bs == 1 and not isinstance(prompt, str):\n",
+                "        bs = len(prompt)\n",
+                "\n",
+                "    img = rearrange(img, \"b c (h ph) (w pw) -> b (h w) (c ph pw)\", ph=2, pw=2)\n",
+                "    if img.shape[0] == 1 and bs > 1:\n",
+                "        img = repeat(img, \"1 ... -> bs ...\", bs=bs)\n",
+                "\n",
+                "    img_ids = torch.zeros(h // 2, w // 2, 3)\n",
+                "    img_ids[..., 1] = img_ids[..., 1] + torch.arange(h // 2)[:, None]\n",
+                "    img_ids[..., 2] = img_ids[..., 2] + torch.arange(w // 2)[None, :]\n",
+                "    img_ids = repeat(img_ids, \"h w c -> b (h w) c\", b=bs)\n",
+                "\n",
+                "    if isinstance(prompt, str):\n",
+                "        prompt = [prompt]\n",
+                "    txt = t5(prompt)\n",
+                "    if txt.shape[0] == 1 and bs > 1:\n",
+                "        txt = repeat(txt, \"1 ... -> bs ...\", bs=bs)\n",
+                "    txt_ids = torch.zeros(bs, txt.shape[1], 3)\n",
+                "\n",
+                "    vec = clip(prompt)\n",
+                "    if vec.shape[0] == 1 and bs > 1:\n",
+                "        vec = repeat(vec, \"1 ... -> bs ...\", bs=bs)\n",
+                "\n",
+                "    print(img_ids.size(), txt.size(), vec.size())\n",
+                "    return img, {\n",
+                "        \"img_ids\": img_ids.to(img.device),\n",
+                "        \"txt\": txt.to(img.device),\n",
+                "        \"txt_ids\": txt_ids.to(img.device),\n",
+                "        \"y\": vec.to(img.device),\n",
+                "    }\n",
+                "\n",
+                "version=\"base\"\n",
+                "seed=2024\n",
+                "prompt_file=\"C:/Users/Curt/Developer/AItools/AIaudio/AudioCreation/FluxMusicJupyter/config/example.txt\"\n",
+                "\n",
+                "print('generate with MusicFlux')\n",
+                "torch.manual_seed(seed)\n",
+                "torch.set_grad_enabled(False)\n",
+                "device = \"cuda\" if torch.cuda.is_available() else \"cpu\"\n",
+                "\n",
+                "latent_size = (256, 16)\n",
+                "\n",
+                "model = build_model(version).to(device)\n",
+                "local_path = 'C:/Users/Curt/Developer/AItools/AIaudio/AudioCreation/FluxMusicJupyter/musicflow_b.pt'\n",
+                "state_dict = torch.load(local_path, map_location=lambda storage, loc: storage, weights_only=True)\n",
+                "model.load_state_dict(state_dict['ema'])\n",
+                "model.eval()  # important!\n",
+                "diffusion = RF()"
+            ]
+        },
+        {
+            "cell_type": "code",
+            "execution_count": null,
+            "metadata": {
+                "id": "B5ebyTmto70D"
+            },
+            "outputs": [],
+            "source": [
+                "t5 = load_t5(device, max_length=256)\n",
+                "clap = load_clap(device, max_length=256)\n",
+                "\n",
+                "vae = AutoencoderKL.from_pretrained('cvssp/audioldm2', subfolder=\"vae\").to(device)\n",
+                "vocoder = SpeechT5HifiGan.from_pretrained('cvssp/audioldm2', subfolder=\"vocoder\").to(device)"
+            ]
+        },
+        {
+            "cell_type": "code",
+            "execution_count": null,
+            "metadata": {
+                "id": "xqG8Px6xo70D"
+            },
+            "outputs": [],
+            "source": [
+                "prompt_file=\"C:/Users/Curt/Developer/AItools/AIaudio/AudioCreation/FluxMusicJupyter/config/example.txt\"\n",
+                "\n",
+                "with open(prompt_file, 'r') as f:\n",
+                "    conds_txt = f.readlines()\n",
+                "L = len(conds_txt)\n",
+                "unconds_txt = [\"low quality, gentle\"] * L\n",
+                "print(L, conds_txt, unconds_txt)\n",
+                "\n",
+                "init_noise = torch.randn(L, 8, latent_size[0], latent_size[1]).cuda()\n",
+                "\n",
+                "STEPSIZE = 50\n",
+                "img, conds = prepare(t5, clap, init_noise, conds_txt)\n",
+                "_, unconds = prepare(t5, clap, init_noise, unconds_txt)\n",
+                "with torch.autocast(device_type='cuda'):\n",
+                "    images = diffusion.sample_with_xps(model, img, conds=conds, null_cond=unconds, sample_steps = STEPSIZE, cfg = 7.0)\n",
+                "\n",
+                "print(images[-1].size(), )\n",
+                "\n",
+                "images = rearrange(\n",
+                "    images[-1],\n",
+                "    \"b (h w) (c ph pw) -> b c (h ph) (w pw)\",\n",
+                "    h=128,\n",
+                "    w=8,\n",
+                "    ph=2,\n",
+                "    pw=2,)\n",
+                "# print(images.size())\n",
+                "latents = 1 / vae.config.scaling_factor * images\n",
+                "mel_spectrogram = vae.decode(latents).sample\n",
+                "print(mel_spectrogram.size())"
+            ]
+        },
+        {
+            "cell_type": "code",
+            "execution_count": null,
+            "metadata": {
+                "id": "ytAXlAEdo70D"
+            },
+            "outputs": [],
+            "source": [
+                "!mkdir C:/Users/Curt/Developer/AItools/AIaudio/AudioCreation/FluxMusicJupyter/FluxMusic/b_output\n",
+                "\n",
+                "for i in range(L):\n",
+                "    x_i = mel_spectrogram[i]\n",
+                "    if x_i.dim() == 4:\n",
+                "        x_i = x_i.squeeze(1)\n",
+                "    waveform = vocoder(x_i)\n",
+                "    waveform = waveform[0].cpu().float().detach().numpy()\n",
+                "    print(waveform.shape)\n",
+                "    # import soundfile as sf\n",
+                "    # sf.write('reconstruct.wav', waveform, samplerate=16000)\n",
+                "    from  scipy.io import wavfile\n",
+                "    wavfile.write('C:/Users/Curt/Developer/AItools/AIaudio/AudioCreation/FluxMusicJupyter/FluxMusic/b_output/sample_' + str(i) + '.wav', 16000, waveform)"
+            ]
+        }
+    ],
+    "metadata": {
+        "accelerator": "GPU",
+        "colab": {
+            "gpuType": "T4",
+            "provenance": []
+        },
+        "kernelspec": {
+            "display_name": "Python 3",
+            "name": "python3"
+        },
+        "language_info": {
+            "name": "python"
+        }
+    },
+    "nbformat": 4,
+    "nbformat_minor": 0
+}

README.md

@@ -1,12 +1,59 @@
----
-title: FluxMusicGUI
-emoji: 🏢
-colorFrom: pink
-colorTo: purple
-sdk: gradio
-sdk_version: 4.42.0
-app_file: app.py
-pinned: false
----
-
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+## FluxMusic: Text-to-Music Generation with Rectified Flow Transformer <br><sub>GUI Implementation</sub>
+
+<a href="https://arxiv.org/abs/2409.00587"><img src="https://img.shields.io/static/v1?label=Paper&message=FluxMusic&color=purple&logo=arxiv"></a> &ensp;
+<a href="https://huggingface.co/feizhengcong/fluxmusic"><img src="https://img.shields.io/static/v1?label=Models&message=HuggingFace&color=yellow"></a> &ensp;
+
+This repo contains a Graphical User Interface (GUI) implementation of the FluxMusic model, based on the paper *Flux that plays music*. It explores a simple extension of diffusion-based rectified flow Transformers for text-to-music generation.
+
+### FluxMusic GUI
+
+We have created a user-friendly GUI for FluxMusic using Gradio. This interface allows users to easily generate music based on text prompts without needing to interact with command-line interfaces.
+
+#### Features:
+
+1. **Model Selection**: Users can choose from different FluxMusic models (small, base, large, giant) via a dropdown menu.
+
+2. **Text Prompt**: Enter your desired text prompt to guide the music generation.
+
+3. **Sliders and Inputs**:
+   - **Seed**: Set a seed for reproducibility (0 for random).
+   - **CFG Scale**: Adjust the Classifier-Free Guidance scale (1-40).
+   - **Steps**: Set the number of diffusion steps (10-200).
+   - **Duration**: Specify the length of the generated audio in seconds (10-300).
+
+4. **File Management**:
+   - **Models Folder**: Place your FluxMusic model files (`.pt`) in the `models` folder.
+   - **Generations Folder**: Generated audio files are saved in the `generations` folder.
+
+5. **File Naming System**: Generated files are named using the format: `[prompt]_[seed]_[model]_[counter].wav`
+
+### Setup and Running
+
+1. Install the required dependencies:
+   ```
+   pip install -r requirements.txt
+   ```
+
+2. Place your FluxMusic model files in the `models` folder.
+
+3. Run the GUI:
+   ```
+   python fluxGUI.py
+   ```
+
+4. Use the interface to generate music based on your prompts and preferences.
+
+### Model Information
+
+FluxMusic comes in four sizes: Small, Base, Large, and Giant. You can download these models from the following links:
+
+|  Model | Url |
+|---------------|------------------|
+| FluxMusic-Small | [link](https://huggingface.co/feizhengcong/FluxMusic/blob/main/musicflow_s.pt) |
+| FluxMusic-Base  | [link](https://huggingface.co/feizhengcong/FluxMusic/blob/main/musicflow_b.pt) |
+| FluxMusic-Large | [link](https://huggingface.co/feizhengcong/FluxMusic/blob/main/musicflow_l.pt) |
+| FluxMusic-Giant | [link](https://huggingface.co/feizhengcong/FluxMusic/blob/main/musicflow_g.pt) |
+
+### Acknowledgments
+
+The codebase is based on the awesome [Flux](https://github.com/black-forest-labs/flux) and [AudioLDM2](https://github.com/haoheliu/AudioLDM2) repos.

audioldm2/__init__.py

@@ -0,0 +1,2 @@
+from .utils import seed_everything, save_wave, get_time, get_duration, read_list
+from .pipeline import *

audioldm2/__main__.py

@@ -0,0 +1,183 @@
+#!D:\GitDownload\SupThirdParty\audioldm2\venv\Scripts\python.exe
+import os
+import torch
+import logging
+from audioldm2 import text_to_audio, build_model, save_wave, get_time, read_list
+import argparse
+
+os.environ["TOKENIZERS_PARALLELISM"] = "true"
+matplotlib_logger = logging.getLogger('matplotlib')
+matplotlib_logger.setLevel(logging.WARNING)
+
+parser = argparse.ArgumentParser()
+
+parser.add_argument(
+	"-t",
+	"--text",
+	type=str,
+	required=False,
+	default="",
+	help="Text prompt to the model for audio generation",
+)
+
+parser.add_argument(
+	"--transcription",
+	type=str,
+	required=False,
+	default="",
+	help="Transcription for Text-to-Speech",
+)
+
+parser.add_argument(
+	"-tl",
+	"--text_list",
+	type=str,
+	required=False,
+	default="",
+	help="A file that contains text prompt to the model for audio generation",
+)
+
+parser.add_argument(
+	"-s",
+	"--save_path",
+	type=str,
+	required=False,
+	help="The path to save model output",
+	default="./output",
+)
+
+parser.add_argument(
+	"--model_name",
+	type=str,
+	required=False,
+	help="The checkpoint you gonna use",
+	default="audioldm_48k",
+	choices=["audioldm_48k", "audioldm_16k_crossattn_t5", "audioldm2-full", "audioldm2-music-665k",
+	         "audioldm2-full-large-1150k", "audioldm2-speech-ljspeech", "audioldm2-speech-gigaspeech"]
+)
+
+parser.add_argument(
+	"-d",
+	"--device",
+	type=str,
+	required=False,
+	help="The device for computation. If not specified, the script will automatically choose the device based on your environment.",
+	default="auto",
+)
+
+parser.add_argument(
+	"-b",
+	"--batchsize",
+	type=int,
+	required=False,
+	default=1,
+	help="Generate how many samples at the same time",
+)
+
+parser.add_argument(
+	"--ddim_steps",
+	type=int,
+	required=False,
+	default=200,
+	help="The sampling step for DDIM",
+)
+
+parser.add_argument(
+	"-gs",
+	"--guidance_scale",
+	type=float,
+	required=False,
+	default=3.5,
+	help="Guidance scale (Large => better quality and relavancy to text; Small => better diversity)",
+)
+
+parser.add_argument(
+	"-dur",
+	"--duration",
+	type=float,
+	required=False,
+	default=10.0,
+	help="The duration of the samples",
+)
+
+parser.add_argument(
+	"-n",
+	"--n_candidate_gen_per_text",
+	type=int,
+	required=False,
+	default=3,
+	help="Automatic quality control. This number control the number of candidates (e.g., generate three audios and choose the best to show you). A Larger value usually lead to better quality with heavier computation",
+)
+
+parser.add_argument(
+	"--seed",
+	type=int,
+	required=False,
+	default=0,
+	help="Change this value (any integer number) will lead to a different generation result.",
+)
+
+args = parser.parse_args()
+
+torch.set_float32_matmul_precision("high")
+
+save_path = os.path.join(args.save_path, get_time())
+
+text = args.text
+random_seed = args.seed
+duration = args.duration
+sample_rate = 16000
+
+if ("audioldm2" in args.model_name):
+	print(
+		"Warning: For AudioLDM2 we currently only support 10s of generation. Please use audioldm_48k or audioldm_16k_crossattn_t5 if you want a different duration.")
+	duration = 10
+if ("48k" in args.model_name):
+	sample_rate = 48000
+
+guidance_scale = args.guidance_scale
+n_candidate_gen_per_text = args.n_candidate_gen_per_text
+transcription = args.transcription
+
+if (transcription):
+	if "speech" not in args.model_name:
+		print(
+			"Warning: You choose to perform Text-to-Speech by providing the transcription.However you do not choose the correct model name (audioldm2-speech-gigaspeech or audioldm2-speech-ljspeech).")
+		print("Warning: We will use audioldm2-speech-gigaspeech by default")
+		args.model_name = "audioldm2-speech-gigaspeech"
+	if (not text):
+		print(
+			"Warning: You should provide text as a input to describe the speaker. Use default (A male reporter is speaking)")
+		text = "A female reporter is speaking full of emotion"
+
+os.makedirs(save_path, exist_ok=True)
+audioldm2 = build_model(model_name=args.model_name, device=args.device)
+
+if (args.text_list):
+	print("Generate audio based on the text prompts in %s" % args.text_list)
+	prompt_todo = read_list(args.text_list)
+else:
+	prompt_todo = [text]
+
+for text in prompt_todo:
+	if ("|" in text):
+		text, name = text.split("|")
+	else:
+		name = text[:128]
+
+	if (transcription):
+		name += "-TTS-%s" % transcription
+
+	waveform = text_to_audio(
+		audioldm2,
+		text,
+		transcription=transcription,  # To avoid the model to ignore the last vocab
+		seed=random_seed,
+		duration=duration,
+		guidance_scale=guidance_scale,
+		ddim_steps=args.ddim_steps,
+		n_candidate_gen_per_text=n_candidate_gen_per_text,
+		batchsize=args.batchsize,
+	)
+
+	save_wave(waveform, save_path, name=name, samplerate=sample_rate)

audioldm2/audiomae_gen/__init__.py

@@ -0,0 +1 @@
+from .sequence_input import Sequence2AudioMAE

audioldm2/audiomae_gen/sequence_input.py

@@ -0,0 +1,429 @@
+import torch
+import torch.nn as nn
+from audioldm2.latent_diffusion.util import (
+    instantiate_from_config,
+)
+
+# from latent_diffusion.modules.encoders.modules import CLAPAudioEmbeddingClassifierFreev2
+from transformers import GPT2Config, GPT2Model
+import torch.optim.lr_scheduler as lr_scheduler
+
+class Sequence2AudioMAE(nn.Module):
+    def __init__(
+        self,
+        base_learning_rate,
+        sequence_gen_length,
+        sequence_input_key,
+        sequence_input_embed_dim,
+        cond_stage_config,
+        optimizer_type="AdamW",
+        use_warmup=True,
+        use_ar_gen_loss=False,
+        use_audiomae_linear=False,
+        target_tokens_mask_ratio=0.0,
+        random_mask_ratio=False,
+        **kwargs
+    ):
+        super().__init__()
+        assert use_audiomae_linear == False
+        self.random_mask_ratio = random_mask_ratio
+        self.learning_rate = base_learning_rate
+        self.cond_stage_config = cond_stage_config
+        self.use_audiomae_linear = use_audiomae_linear
+        self.optimizer_type = optimizer_type
+        self.use_warmup = use_warmup
+        self.use_ar_gen_loss = use_ar_gen_loss
+        # Even though the LDM can be conditioned on mutliple pooling rate
+        # Our model always predict the higest pooling rate
+
+        # self.time_pool = max(self.cond_stage_config["crossattn_audiomae_pooled"]["params"]["time_pooling_factors"])
+        # self.freq_pool = max(self.cond_stage_config["crossattn_audiomae_pooled"]["params"]["freq_pooling_factors"])
+        # self.mae_token_num = int(512/(self.time_pool*self.freq_pool))
+
+        self.mae_token_num = sequence_gen_length
+        self.sequence_input_key = sequence_input_key
+        self.sequence_input_embed_dim = sequence_input_embed_dim
+        self.target_tokens_mask_ratio = target_tokens_mask_ratio
+
+        self.start_of_sequence_tokens = nn.Embedding(32, 768)
+        self.end_of_sequence_tokens = nn.Embedding(32, 768)
+
+        self.input_sequence_embed_linear = nn.ModuleList([])
+        self.initial_learning_rate = None
+
+        for dim in self.sequence_input_embed_dim:
+            self.input_sequence_embed_linear.append(nn.Linear(dim, 768))
+
+        self.cond_stage_models = nn.ModuleList([])
+        self.instantiate_cond_stage(cond_stage_config)
+        self.initialize_param_check_toolkit()
+
+        # configuration = GPT2Config(n_layer=1) # TODO
+        # self.model=GPT2Model(configuration)
+        ###################
+        # self.model=nn.Linear(768,768, bias=False) # TODO change the model
+        # with torch.no_grad():
+        #     self.model.weight.copy_(torch.eye(768))
+        ###################
+        self.model = GPT2Model(GPT2Config.from_pretrained("gpt2"))
+        ###################
+        # self.model = nn.LSTM(input_size=768, hidden_size=768, num_layers=1,bias=False) # TODO
+
+        # self.loss_fn = nn.MSELoss()
+        self.loss_fn = nn.L1Loss()
+
+        self.logger_save_dir = None
+        self.logger_exp_name = None
+        self.logger_exp_group_name = None
+        self.logger_version = None
+
+    def set_log_dir(self, save_dir, exp_group_name, exp_name):
+        self.logger_save_dir = save_dir
+        self.logger_exp_group_name = exp_group_name
+        self.logger_exp_name = exp_name
+
+    def cfg_uncond(self, batch_size):
+        unconditional_conditioning = {}
+        for key in self.cond_stage_model_metadata:
+            model_idx = self.cond_stage_model_metadata[key]["model_idx"]
+            unconditional_conditioning[key] = self.cond_stage_models[
+                model_idx
+            ].get_unconditional_condition(batch_size)
+        assert (
+            "crossattn_audiomae_pooled" in unconditional_conditioning.keys()
+        ), "The module is not initialized with AudioMAE"
+        unconditional_conditioning[
+            "crossattn_clap_to_audiomae_feature"
+        ] = unconditional_conditioning["crossattn_audiomae_pooled"]
+        return unconditional_conditioning
+
+    def configure_optimizers(self):
+        lr = float(self.learning_rate)
+        # params = list(self.model.parameters()) + list(self.input_sequence_embed_linear.parameters())
+        params = list(self.parameters())
+
+        # opt = torch.optim.Adam(params, lr=lr, betas=(0.9, 0.98), eps=1e-9)
+        opt = eval(self.optimizer_type)(params, lr=lr)
+        scheduler = lr_scheduler.StepLR(opt, step_size=10, gamma=0.8)
+        return [opt], [scheduler]
+
+    def add_sos_eos_tokens(self, _id, sequence, attn_mask):
+        batchsize = sequence.size(0)
+
+        new_attn_mask_step = torch.ones((batchsize, 1)).to(sequence.device)
+        key_id = torch.tensor([_id]).to(sequence.device)
+
+        # Add two more steps to attn mask
+        new_attn_mask = torch.cat(
+            [new_attn_mask_step, attn_mask, new_attn_mask_step], dim=1
+        )
+
+        # Add two more tokens in the sequence
+        sos_token = self.start_of_sequence_tokens(key_id).expand(batchsize, 1, -1)
+        eos_token = self.end_of_sequence_tokens(key_id).expand(batchsize, 1, -1)
+        new_sequence = torch.cat([sos_token, sequence, eos_token], dim=1)
+        return new_sequence, new_attn_mask
+
+    def truncate_sequence_and_mask(self, sequence, mask, max_len=512):
+        if sequence.size(1) > max_len:
+            print(
+                "The input sequence length to GPT-2 model is too long:",
+                sequence.size(1),
+            )
+            return sequence[:, :max_len], mask[:, :max_len]
+        else:
+            return sequence, mask
+
+    def get_input_sequence_and_mask(self, cond_dict):
+        input_embeds = None
+        input_embeds_attn_mask = None
+        for _id, sequence_key in enumerate(self.sequence_input_key):
+            assert sequence_key in cond_dict.keys(), (
+                "Invalid sequence key %s" % sequence_key
+            )
+            cond_embed = cond_dict[sequence_key]
+            if isinstance(cond_embed, list):
+                assert (
+                    len(cond_embed) == 2
+                ), "The crossattn returned list should have length 2, including embed and attn_mask"
+                item_input_embeds, item_attn_mask = cond_embed
+
+                item_input_embeds = self.input_sequence_embed_linear[_id](
+                    item_input_embeds
+                )
+
+                item_input_embeds, item_attn_mask = self.add_sos_eos_tokens(
+                    _id, item_input_embeds, item_attn_mask
+                )
+
+                if input_embeds is None and input_embeds_attn_mask is None:
+                    input_embeds, input_embeds_attn_mask = (
+                        item_input_embeds,
+                        item_attn_mask,
+                    )
+                else:
+                    input_embeds = torch.cat(
+                        [input_embeds, item_input_embeds], dim=1
+                    )  # The 1-st dimension is time steps
+                    input_embeds_attn_mask = torch.cat(
+                        [input_embeds_attn_mask, item_attn_mask], dim=1
+                    )  # The 1-st dimension is time steps
+            else:
+                assert isinstance(cond_embed, torch.Tensor)
+                cond_embed = self.input_sequence_embed_linear[_id](cond_embed)
+                attn_mask = torch.ones((cond_embed.size(0), cond_embed.size(1))).to(
+                    cond_embed.device
+                )
+
+                item_input_embeds, item_attn_mask = self.add_sos_eos_tokens(
+                    _id, cond_embed, attn_mask
+                )
+
+                if input_embeds is None and input_embeds_attn_mask is None:
+                    input_embeds, input_embeds_attn_mask = (
+                        item_input_embeds,
+                        item_attn_mask,
+                    )
+                else:
+                    input_embeds, input_embeds_attn_mask = torch.cat(
+                        [input_embeds, item_input_embeds], dim=1
+                    ), torch.cat([input_embeds_attn_mask, item_attn_mask], dim=1)
+
+        assert input_embeds is not None and input_embeds_attn_mask is not None
+
+        input_embeds, input_embeds_attn_mask = self.truncate_sequence_and_mask(
+            input_embeds, input_embeds_attn_mask, int(1024 - self.mae_token_num)
+        )
+        cond_sequence_end_time_idx = input_embeds.size(
+            1
+        )  # The index that we start to collect the output embeds
+
+        return input_embeds, input_embeds_attn_mask, cond_sequence_end_time_idx
+
+    def warmup_step(self):
+        if self.initial_learning_rate is None:
+            self.initial_learning_rate = float(self.learning_rate)
+
+        # Only the first parameter group
+        if self.global_step <= 1000:
+            if self.global_step == 0:
+                print(
+                    "Warming up learning rate start with %s"
+                    % self.initial_learning_rate
+                )
+            self.trainer.optimizers[0].param_groups[0]["lr"] = (
+                self.global_step / 1000
+            ) * self.initial_learning_rate
+        else:
+            # TODO set learning rate here
+            self.trainer.optimizers[0].param_groups[0][
+                "lr"
+            ] = self.initial_learning_rate
+
+    def mask_target_sequence(self, target_embeds, target_embeds_attn_mask):
+        time_seq_mask = None
+        if self.target_tokens_mask_ratio > 1e-4:
+            batchsize, time_seq_len, embed_dim = target_embeds.size()
+            _, time_seq_len = target_embeds_attn_mask.size()
+            # Generate random mask
+            if self.random_mask_ratio:
+                mask_ratio = torch.rand(1).item() * self.target_tokens_mask_ratio
+            else:
+                mask_ratio = self.target_tokens_mask_ratio
+
+            time_seq_mask = (torch.rand((batchsize, time_seq_len)) > mask_ratio).to(
+                target_embeds.device
+            )
+            # Mask the target embedding
+            target_embeds = target_embeds * time_seq_mask.unsqueeze(-1)
+            target_embeds_attn_mask = target_embeds_attn_mask * time_seq_mask
+        return target_embeds, target_embeds_attn_mask, time_seq_mask
+
+    def generate_partial(self, batch, cond_dict=None, no_grad=False):
+        if cond_dict is None:
+            cond_dict = self.get_input(batch)
+
+        print("Generate partially prompted audio with in-context learning")
+        # self.model.train()
+        # assert self.model.training==True
+
+        target_embeds, target_embeds_attn_mask = (
+            cond_dict["crossattn_audiomae_pooled"][0],
+            cond_dict["crossattn_audiomae_pooled"][1],
+        )
+
+        target_time_steps = target_embeds.size(1)
+
+        (
+            input_embeds,
+            input_embeds_attn_mask,
+            cond_sequence_end_time_idx,
+        ) = self.get_input_sequence_and_mask(cond_dict)
+
+        model_input = torch.cat(
+            [input_embeds, target_embeds[:, : target_time_steps // 4, :]], dim=1
+        )
+        model_input_mask = torch.cat(
+            [
+                input_embeds_attn_mask,
+                target_embeds_attn_mask[:, : target_time_steps // 4],
+            ],
+            dim=1,
+        )
+
+        steps = self.mae_token_num
+
+        for _ in range(3 * steps // 4):
+            output = self.model(
+                inputs_embeds=model_input, attention_mask=model_input_mask
+            )["last_hidden_state"]
+            # Update the model input
+            model_input = torch.cat([model_input, output[:, -1:, :]], dim=1)
+            # Update the attention mask
+            attention_mask_new_step = torch.ones((model_input_mask.size(0), 1)).to(
+                model_input.device
+            )
+            model_input_mask = torch.cat(
+                [model_input_mask, attention_mask_new_step], dim=1
+            )
+
+        output = model_input[:, cond_sequence_end_time_idx:]
+
+        return output, cond_dict
+
+    def generate(self, batch, cond_dict=None, no_grad=False):
+        if cond_dict is None:
+            cond_dict = self.get_input(batch)
+
+        # self.model.train()
+        # print("!!!!!!!!!!!!!train")
+
+        (
+            input_embeds,
+            input_embeds_attn_mask,
+            cond_sequence_end_time_idx,
+        ) = self.get_input_sequence_and_mask(cond_dict)
+        model_input = input_embeds
+        model_input_mask = input_embeds_attn_mask
+
+        steps = self.mae_token_num
+
+        for _ in range(steps):
+            output = self.model(
+                inputs_embeds=model_input, attention_mask=model_input_mask
+            )["last_hidden_state"]
+            # Update the model input
+            model_input = torch.cat([model_input, output[:, -1:, :]], dim=1)
+            # Update the attention mask
+            attention_mask_new_step = torch.ones((model_input_mask.size(0), 1)).to(
+                model_input.device
+            )
+            model_input_mask = torch.cat(
+                [model_input_mask, attention_mask_new_step], dim=1
+            )
+
+        return model_input[:, cond_sequence_end_time_idx:], cond_dict
+
+    def get_input_item(self, batch, k):
+        fname, text, waveform, stft, fbank = (
+            batch["fname"],
+            batch["text"],
+            batch["waveform"],
+            batch["stft"],
+            batch["log_mel_spec"],
+        )
+        ret = {}
+
+        ret["fbank"] = (
+            fbank.unsqueeze(1).to(memory_format=torch.contiguous_format).float()
+        )
+        ret["stft"] = stft.to(memory_format=torch.contiguous_format).float()
+        # ret["clip_label"] = clip_label.to(memory_format=torch.contiguous_format).float()
+        ret["waveform"] = waveform.to(memory_format=torch.contiguous_format).float()
+        ret["text"] = list(text)
+        ret["fname"] = fname
+
+        for key in batch.keys():
+            if key not in ret.keys():
+                ret[key] = batch[key]
+
+        return ret[k]
+
+    def get_input(self, batch):
+        cond_dict = {}
+        if len(self.cond_stage_model_metadata.keys()) > 0:
+            unconditional_cfg = False
+
+            for cond_model_key in self.cond_stage_model_metadata.keys():
+                cond_stage_key = self.cond_stage_model_metadata[cond_model_key][
+                    "cond_stage_key"
+                ]
+
+                # if(not self.training):
+                #     if(isinstance(self.cond_stage_models[self.cond_stage_model_metadata[cond_model_key]["model_idx"]], CLAPAudioEmbeddingClassifierFreev2)):
+                #         assert cond_stage_key == "text" # CLAP model should use text for evaluation
+
+                # The original data for conditioning
+                xc = self.get_input_item(batch, cond_stage_key)
+                if type(xc) == torch.Tensor:
+                    xc = xc.to(self.device)
+
+                c = self.get_learned_conditioning(
+                    xc, key=cond_model_key, unconditional_cfg=unconditional_cfg
+                )
+                cond_dict[cond_model_key] = c
+
+        return cond_dict
+
+    def instantiate_cond_stage(self, config):
+        self.cond_stage_model_metadata = {}
+
+        for i, cond_model_key in enumerate(config.keys()):
+            model = instantiate_from_config(config[cond_model_key])
+            self.cond_stage_models.append(model)
+            self.cond_stage_model_metadata[cond_model_key] = {
+                "model_idx": i,
+                "cond_stage_key": config[cond_model_key]["cond_stage_key"],
+                "conditioning_key": config[cond_model_key]["conditioning_key"],
+            }
+
+    def get_learned_conditioning(self, c, key, unconditional_cfg):
+        assert key in self.cond_stage_model_metadata.keys()
+
+        # Classifier-free guidance
+        if not unconditional_cfg:
+            c = self.cond_stage_models[
+                self.cond_stage_model_metadata[key]["model_idx"]
+            ](c)
+        else:
+            if isinstance(c, torch.Tensor):
+                batchsize = c.size(0)
+            elif isinstance(c, list):
+                batchsize = len(c)
+            else:
+                raise NotImplementedError()
+            c = self.cond_stage_models[
+                self.cond_stage_model_metadata[key]["model_idx"]
+            ].get_unconditional_condition(batchsize)
+
+        return c
+
+    def initialize_param_check_toolkit(self):
+        self.tracked_steps = 0
+        self.param_dict = {}
+
+    def statistic_require_grad_tensor_number(self, module, name=None):
+        requires_grad_num = 0
+        total_num = 0
+        require_grad_tensor = None
+        for p in module.parameters():
+            if p.requires_grad:
+                requires_grad_num += 1
+                if require_grad_tensor is None:
+                    require_grad_tensor = p
+            total_num += 1
+        print(
+            "Module: [%s] have %s trainable parameters out of %s total parameters (%.2f)"
+            % (name, requires_grad_num, total_num, requires_grad_num / total_num)
+        )
+        return require_grad_tensor

audioldm2/audiomae_gen/utils.py

@@ -0,0 +1,27 @@
+import torch.nn as nn
+
+
+class Prenet(nn.Module):
+    def __init__(self, in_dim, sizes=[256, 128], dropout_rate=0.5):
+        super(Prenet, self).__init__()
+        in_sizes = [in_dim] + sizes[:-1]
+        self.layers = nn.ModuleList(
+            [
+                nn.Linear(in_size, out_size)
+                for (in_size, out_size) in zip(in_sizes, sizes)
+            ]
+        )
+        self.relu = nn.ReLU()
+        self.dropout = nn.Dropout(dropout_rate)
+
+    def forward(self, inputs):
+        for linear in self.layers:
+            inputs = self.dropout(self.relu(linear(inputs)))
+        return inputs
+
+
+if __name__ == "__main__":
+    model = Prenet(in_dim=128, sizes=[256, 256, 128])
+    import ipdb
+
+    ipdb.set_trace()

audioldm2/clap/open_clip/__init__.py

@@ -0,0 +1,25 @@
+from .factory import (
+    list_models,
+    create_model,
+    create_model_and_transforms,
+    add_model_config,
+)
+from .loss import ClipLoss, gather_features, LPLoss, lp_gather_features, LPMetrics
+from .model import (
+    CLAP,
+    CLAPTextCfg,
+    CLAPVisionCfg,
+    CLAPAudioCfp,
+    convert_weights_to_fp16,
+    trace_model,
+)
+from .openai import load_openai_model, list_openai_models
+from .pretrained import (
+    list_pretrained,
+    list_pretrained_tag_models,
+    list_pretrained_model_tags,
+    get_pretrained_url,
+    download_pretrained,
+)
+from .tokenizer import SimpleTokenizer, tokenize
+from .transform import image_transform

audioldm2/clap/open_clip/bpe_simple_vocab_16e6.txt.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:924691ac288e54409236115652ad4aa250f48203de50a9e4722a6ecd48d6804a
+size 1356917

audioldm2/clap/open_clip/factory.py

@@ -0,0 +1,276 @@
+import json
+import logging
+import os
+import re
+from copy import deepcopy
+from pathlib import Path
+
+import torch
+
+from .model import CLAP, convert_weights_to_fp16
+from .openai import load_openai_model
+from .pretrained import get_pretrained_url, download_pretrained
+from .transform import image_transform
+
+_MODEL_CONFIG_PATHS = [Path(__file__).parent / f"model_configs/"]
+_MODEL_CONFIGS = {}  # directory (model_name: config) of model architecture configs
+
+
+def _natural_key(string_):
+    return [int(s) if s.isdigit() else s for s in re.split(r"(\d+)", string_.lower())]
+
+
+def _rescan_model_configs():
+    global _MODEL_CONFIGS
+
+    config_ext = (".json",)
+    config_files = []
+    for config_path in _MODEL_CONFIG_PATHS:
+        if config_path.is_file() and config_path.suffix in config_ext:
+            config_files.append(config_path)
+        elif config_path.is_dir():
+            for ext in config_ext:
+                config_files.extend(config_path.glob(f"*{ext}"))
+
+    for cf in config_files:
+        if os.path.basename(cf)[0] == ".":
+            continue  # Ignore hidden files
+
+        with open(cf, "r") as f:
+            model_cfg = json.load(f)
+            if all(a in model_cfg for a in ("embed_dim", "audio_cfg", "text_cfg")):
+                _MODEL_CONFIGS[cf.stem] = model_cfg
+
+    _MODEL_CONFIGS = {
+        k: v
+        for k, v in sorted(_MODEL_CONFIGS.items(), key=lambda x: _natural_key(x[0]))
+    }
+
+
+_rescan_model_configs()  # initial populate of model config registry
+
+
+def load_state_dict(checkpoint_path: str, map_location="cpu", skip_params=True):
+    checkpoint = torch.load(checkpoint_path, map_location=map_location)
+    if isinstance(checkpoint, dict) and "state_dict" in checkpoint:
+        state_dict = checkpoint["state_dict"]
+    else:
+        state_dict = checkpoint
+    if skip_params:
+        if next(iter(state_dict.items()))[0].startswith("module"):
+            state_dict = {k[7:]: v for k, v in state_dict.items()}
+    # for k in state_dict:
+    #     if k.startswith('transformer'):
+    #         v = state_dict.pop(k)
+    #         state_dict['text_branch.' + k[12:]] = v
+    return state_dict
+
+
+def create_model(
+    amodel_name: str,
+    tmodel_name: str,
+    pretrained: str = "",
+    precision: str = "fp32",
+    device: torch.device = torch.device("cpu"),
+    jit: bool = False,
+    force_quick_gelu: bool = False,
+    openai_model_cache_dir: str = os.path.expanduser("~/.cache/clip"),
+    skip_params=True,
+    pretrained_audio: str = "",
+    pretrained_text: str = "",
+    enable_fusion: bool = False,
+    fusion_type: str = "None"
+    # pretrained_image: bool = False,
+):
+    amodel_name = amodel_name.replace(
+        "/", "-"
+    )  # for callers using old naming with / in ViT names
+    pretrained_orig = pretrained
+    pretrained = pretrained.lower()
+    if pretrained == "openai":
+        if amodel_name in _MODEL_CONFIGS:
+            logging.info(f"Loading {amodel_name} model config.")
+            model_cfg = deepcopy(_MODEL_CONFIGS[amodel_name])
+        else:
+            logging.error(
+                f"Model config for {amodel_name} not found; available models {list_models()}."
+            )
+            raise RuntimeError(f"Model config for {amodel_name} not found.")
+
+        logging.info(f"Loading pretrained ViT-B-16 text encoder from OpenAI.")
+        # Hard Code in model name
+        model_cfg["text_cfg"]["model_type"] = tmodel_name
+        model = load_openai_model(
+            "ViT-B-16",
+            model_cfg,
+            device=device,
+            jit=jit,
+            cache_dir=openai_model_cache_dir,
+            enable_fusion=enable_fusion,
+            fusion_type=fusion_type,
+        )
+        # See https://discuss.pytorch.org/t/valueerror-attemting-to-unscale-fp16-gradients/81372
+        if precision == "amp" or precision == "fp32":
+            model = model.float()
+    else:
+        if amodel_name in _MODEL_CONFIGS:
+            logging.info(f"Loading {amodel_name} model config.")
+            model_cfg = deepcopy(_MODEL_CONFIGS[amodel_name])
+        else:
+            logging.error(
+                f"Model config for {amodel_name} not found; available models {list_models()}."
+            )
+            raise RuntimeError(f"Model config for {amodel_name} not found.")
+
+        if force_quick_gelu:
+            # override for use of QuickGELU on non-OpenAI transformer models
+            model_cfg["quick_gelu"] = True
+
+        # if pretrained_image:
+        #     if 'timm_amodel_name' in model_cfg.get('vision_cfg', {}):
+        #         # pretrained weight loading for timm models set via vision_cfg
+        #         model_cfg['vision_cfg']['timm_model_pretrained'] = True
+        #     else:
+        #         assert False, 'pretrained image towers currently only supported for timm models'
+        model_cfg["text_cfg"]["model_type"] = tmodel_name
+        model_cfg["enable_fusion"] = enable_fusion
+        model_cfg["fusion_type"] = fusion_type
+        model = CLAP(**model_cfg)
+
+        if pretrained:
+            checkpoint_path = ""
+            url = get_pretrained_url(amodel_name, pretrained)
+            if url:
+                checkpoint_path = download_pretrained(url, root=openai_model_cache_dir)
+            elif os.path.exists(pretrained_orig):
+                checkpoint_path = pretrained_orig
+            if checkpoint_path:
+                logging.info(
+                    f"Loading pretrained {amodel_name}-{tmodel_name} weights ({pretrained})."
+                )
+                ckpt = load_state_dict(checkpoint_path, skip_params=True)
+                model.load_state_dict(ckpt)
+                param_names = [n for n, p in model.named_parameters()]
+                # for n in param_names:
+                #     print(n, "\t", "Loaded" if n in ckpt else "Unloaded")
+            else:
+                logging.warning(
+                    f"Pretrained weights ({pretrained}) not found for model {amodel_name}."
+                )
+                raise RuntimeError(
+                    f"Pretrained weights ({pretrained}) not found for model {amodel_name}."
+                )
+
+        if pretrained_audio:
+            if amodel_name.startswith("PANN"):
+                if "Cnn14_mAP" in pretrained_audio:  # official checkpoint
+                    audio_ckpt = torch.load(pretrained_audio, map_location="cpu")
+                    audio_ckpt = audio_ckpt["model"]
+                    keys = list(audio_ckpt.keys())
+                    for key in keys:
+                        if (
+                            "spectrogram_extractor" not in key
+                            and "logmel_extractor" not in key
+                        ):
+                            v = audio_ckpt.pop(key)
+                            audio_ckpt["audio_branch." + key] = v
+                elif os.path.basename(pretrained_audio).startswith(
+                    "PANN"
+                ):  # checkpoint trained via HTSAT codebase
+                    audio_ckpt = torch.load(pretrained_audio, map_location="cpu")
+                    audio_ckpt = audio_ckpt["state_dict"]
+                    keys = list(audio_ckpt.keys())
+                    for key in keys:
+                        if key.startswith("sed_model"):
+                            v = audio_ckpt.pop(key)
+                            audio_ckpt["audio_branch." + key[10:]] = v
+                elif os.path.basename(pretrained_audio).startswith(
+                    "finetuned"
+                ):  # checkpoint trained via linear probe codebase
+                    audio_ckpt = torch.load(pretrained_audio, map_location="cpu")
+                else:
+                    raise ValueError("Unknown audio checkpoint")
+            elif amodel_name.startswith("HTSAT"):
+                if "HTSAT_AudioSet_Saved" in pretrained_audio:  # official checkpoint
+                    audio_ckpt = torch.load(pretrained_audio, map_location="cpu")
+                    audio_ckpt = audio_ckpt["state_dict"]
+                    keys = list(audio_ckpt.keys())
+                    for key in keys:
+                        if key.startswith("sed_model") and (
+                            "spectrogram_extractor" not in key
+                            and "logmel_extractor" not in key
+                        ):
+                            v = audio_ckpt.pop(key)
+                            audio_ckpt["audio_branch." + key[10:]] = v
+                elif os.path.basename(pretrained_audio).startswith(
+                    "HTSAT"
+                ):  # checkpoint trained via HTSAT codebase
+                    audio_ckpt = torch.load(pretrained_audio, map_location="cpu")
+                    audio_ckpt = audio_ckpt["state_dict"]
+                    keys = list(audio_ckpt.keys())
+                    for key in keys:
+                        if key.startswith("sed_model"):
+                            v = audio_ckpt.pop(key)
+                            audio_ckpt["audio_branch." + key[10:]] = v
+                elif os.path.basename(pretrained_audio).startswith(
+                    "finetuned"
+                ):  # checkpoint trained via linear probe codebase
+                    audio_ckpt = torch.load(pretrained_audio, map_location="cpu")
+                else:
+                    raise ValueError("Unknown audio checkpoint")
+            else:
+                raise f"this audio encoder pretrained checkpoint is not support"
+
+            model.load_state_dict(audio_ckpt, strict=False)
+            logging.info(
+                f"Loading pretrained {amodel_name} weights ({pretrained_audio})."
+            )
+            param_names = [n for n, p in model.named_parameters()]
+            for n in param_names:
+                print(n, "\t", "Loaded" if n in audio_ckpt else "Unloaded")
+
+        model.to(device=device)
+        if precision == "fp16":
+            assert device.type != "cpu"
+            convert_weights_to_fp16(model)
+
+        if jit:
+            model = torch.jit.script(model)
+
+    return model, model_cfg
+
+
+def create_model_and_transforms(
+    model_name: str,
+    pretrained: str = "",
+    precision: str = "fp32",
+    device: torch.device = torch.device("cpu"),
+    jit: bool = False,
+    force_quick_gelu: bool = False,
+    # pretrained_image: bool = False,
+):
+    model = create_model(
+        model_name,
+        pretrained,
+        precision,
+        device,
+        jit,
+        force_quick_gelu=force_quick_gelu,
+        # pretrained_image=pretrained_image
+    )
+    preprocess_train = image_transform(model.visual.image_size, is_train=True)
+    preprocess_val = image_transform(model.visual.image_size, is_train=False)
+    return model, preprocess_train, preprocess_val
+
+
+def list_models():
+    """enumerate available model architectures based on config files"""
+    return list(_MODEL_CONFIGS.keys())
+
+
+def add_model_config(path):
+    """add model config path or file and update registry"""
+    if not isinstance(path, Path):
+        path = Path(path)
+    _MODEL_CONFIG_PATHS.append(path)
+    _rescan_model_configs()

audioldm2/clap/open_clip/feature_fusion.py

@@ -0,0 +1,192 @@
+"""
+Feature Fusion for Varible-Length Data Processing
+AFF/iAFF is referred and modified from https://github.com/YimianDai/open-aff/blob/master/aff_pytorch/aff_net/fusion.py
+According to the paper: Yimian Dai et al, Attentional Feature Fusion, IEEE Winter Conference on Applications of Computer Vision, WACV 2021
+"""
+
+import torch
+import torch.nn as nn
+
+
+class DAF(nn.Module):
+    """
+    直接相加 DirectAddFuse
+    """
+
+    def __init__(self):
+        super(DAF, self).__init__()
+
+    def forward(self, x, residual):
+        return x + residual
+
+
+class iAFF(nn.Module):
+    """
+    多特征融合 iAFF
+    """
+
+    def __init__(self, channels=64, r=4, type="2D"):
+        super(iAFF, self).__init__()
+        inter_channels = int(channels // r)
+
+        if type == "1D":
+            # 本地注意力
+            self.local_att = nn.Sequential(
+                nn.Conv1d(channels, inter_channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm1d(inter_channels),
+                nn.ReLU(inplace=True),
+                nn.Conv1d(inter_channels, channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm1d(channels),
+            )
+
+            # 全局注意力
+            self.global_att = nn.Sequential(
+                nn.AdaptiveAvgPool1d(1),
+                nn.Conv1d(channels, inter_channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm1d(inter_channels),
+                nn.ReLU(inplace=True),
+                nn.Conv1d(inter_channels, channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm1d(channels),
+            )
+
+            # 第二次本地注意力
+            self.local_att2 = nn.Sequential(
+                nn.Conv1d(channels, inter_channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm1d(inter_channels),
+                nn.ReLU(inplace=True),
+                nn.Conv1d(inter_channels, channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm1d(channels),
+            )
+            # 第二次全局注意力
+            self.global_att2 = nn.Sequential(
+                nn.AdaptiveAvgPool1d(1),
+                nn.Conv1d(channels, inter_channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm1d(inter_channels),
+                nn.ReLU(inplace=True),
+                nn.Conv1d(inter_channels, channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm1d(channels),
+            )
+        elif type == "2D":
+            # 本地注意力
+            self.local_att = nn.Sequential(
+                nn.Conv2d(channels, inter_channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm2d(inter_channels),
+                nn.ReLU(inplace=True),
+                nn.Conv2d(inter_channels, channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm2d(channels),
+            )
+
+            # 全局注意力
+            self.global_att = nn.Sequential(
+                nn.AdaptiveAvgPool2d(1),
+                nn.Conv2d(channels, inter_channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm2d(inter_channels),
+                nn.ReLU(inplace=True),
+                nn.Conv2d(inter_channels, channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm2d(channels),
+            )
+
+            # 第二次本地注意力
+            self.local_att2 = nn.Sequential(
+                nn.Conv2d(channels, inter_channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm2d(inter_channels),
+                nn.ReLU(inplace=True),
+                nn.Conv2d(inter_channels, channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm2d(channels),
+            )
+            # 第二次全局注意力
+            self.global_att2 = nn.Sequential(
+                nn.AdaptiveAvgPool2d(1),
+                nn.Conv2d(channels, inter_channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm2d(inter_channels),
+                nn.ReLU(inplace=True),
+                nn.Conv2d(inter_channels, channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm2d(channels),
+            )
+        else:
+            raise f"the type is not supported"
+
+        self.sigmoid = nn.Sigmoid()
+
+    def forward(self, x, residual):
+        flag = False
+        xa = x + residual
+        if xa.size(0) == 1:
+            xa = torch.cat([xa, xa], dim=0)
+            flag = True
+        xl = self.local_att(xa)
+        xg = self.global_att(xa)
+        xlg = xl + xg
+        wei = self.sigmoid(xlg)
+        xi = x * wei + residual * (1 - wei)
+
+        xl2 = self.local_att2(xi)
+        xg2 = self.global_att(xi)
+        xlg2 = xl2 + xg2
+        wei2 = self.sigmoid(xlg2)
+        xo = x * wei2 + residual * (1 - wei2)
+        if flag:
+            xo = xo[0].unsqueeze(0)
+        return xo
+
+
+class AFF(nn.Module):
+    """
+    多特征融合 AFF
+    """
+
+    def __init__(self, channels=64, r=4, type="2D"):
+        super(AFF, self).__init__()
+        inter_channels = int(channels // r)
+
+        if type == "1D":
+            self.local_att = nn.Sequential(
+                nn.Conv1d(channels, inter_channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm1d(inter_channels),
+                nn.ReLU(inplace=True),
+                nn.Conv1d(inter_channels, channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm1d(channels),
+            )
+            self.global_att = nn.Sequential(
+                nn.AdaptiveAvgPool1d(1),
+                nn.Conv1d(channels, inter_channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm1d(inter_channels),
+                nn.ReLU(inplace=True),
+                nn.Conv1d(inter_channels, channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm1d(channels),
+            )
+        elif type == "2D":
+            self.local_att = nn.Sequential(
+                nn.Conv2d(channels, inter_channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm2d(inter_channels),
+                nn.ReLU(inplace=True),
+                nn.Conv2d(inter_channels, channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm2d(channels),
+            )
+            self.global_att = nn.Sequential(
+                nn.AdaptiveAvgPool2d(1),
+                nn.Conv2d(channels, inter_channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm2d(inter_channels),
+                nn.ReLU(inplace=True),
+                nn.Conv2d(inter_channels, channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm2d(channels),
+            )
+        else:
+            raise f"the type is not supported."
+
+        self.sigmoid = nn.Sigmoid()
+
+    def forward(self, x, residual):
+        flag = False
+        xa = x + residual
+        if xa.size(0) == 1:
+            xa = torch.cat([xa, xa], dim=0)
+            flag = True
+        xl = self.local_att(xa)
+        xg = self.global_att(xa)
+        xlg = xl + xg
+        wei = self.sigmoid(xlg)
+        xo = 2 * x * wei + 2 * residual * (1 - wei)
+        if flag:
+            xo = xo[0].unsqueeze(0)
+        return xo

audioldm2/clap/open_clip/htsat.py

@@ -0,0 +1,1304 @@
+# Ke Chen
+# knutchen@ucsd.edu
+# HTS-AT: A HIERARCHICAL TOKEN-SEMANTIC AUDIO TRANSFORMER FOR SOUND CLASSIFICATION AND DETECTION
+# Some layers designed on the model
+# below codes are based and referred from https://github.com/microsoft/Swin-Transformer
+# Swin Transformer for Computer Vision: https://arxiv.org/pdf/2103.14030.pdf
+
+import torch
+import torch.nn as nn
+from itertools import repeat
+import collections.abc
+import math
+import warnings
+
+from torch.nn.init import _calculate_fan_in_and_fan_out
+import torch.utils.checkpoint as checkpoint
+
+import random
+
+from torchlibrosa.stft import Spectrogram, LogmelFilterBank
+from torchlibrosa.augmentation import SpecAugmentation
+
+from itertools import repeat
+from .utils import do_mixup, interpolate
+
+from .feature_fusion import iAFF, AFF, DAF
+
+
+# from PyTorch internals
+def _ntuple(n):
+    def parse(x):
+        if isinstance(x, collections.abc.Iterable):
+            return x
+        return tuple(repeat(x, n))
+
+    return parse
+
+
+to_1tuple = _ntuple(1)
+to_2tuple = _ntuple(2)
+to_3tuple = _ntuple(3)
+to_4tuple = _ntuple(4)
+to_ntuple = _ntuple
+
+
+def drop_path(x, drop_prob: float = 0.0, training: bool = False):
+    """Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).
+    This is the same as the DropConnect impl I created for EfficientNet, etc networks, however,
+    the original name is misleading as 'Drop Connect' is a different form of dropout in a separate paper...
+    See discussion: https://github.com/tensorflow/tpu/issues/494#issuecomment-532968956 ... I've opted for
+    changing the layer and argument names to 'drop path' rather than mix DropConnect as a layer name and use
+    'survival rate' as the argument.
+    """
+    if drop_prob == 0.0 or not training:
+        return x
+    keep_prob = 1 - drop_prob
+    shape = (x.shape[0],) + (1,) * (
+        x.ndim - 1
+    )  # work with diff dim tensors, not just 2D ConvNets
+    random_tensor = keep_prob + torch.rand(shape, dtype=x.dtype, device=x.device)
+    random_tensor.floor_()  # binarize
+    output = x.div(keep_prob) * random_tensor
+    return output
+
+
+class DropPath(nn.Module):
+    """Drop paths (Stochastic Depth) per sample  (when applied in main path of residual blocks)."""
+
+    def __init__(self, drop_prob=None):
+        super(DropPath, self).__init__()
+        self.drop_prob = drop_prob
+
+    def forward(self, x):
+        return drop_path(x, self.drop_prob, self.training)
+
+
+class PatchEmbed(nn.Module):
+    """2D Image to Patch Embedding"""
+
+    def __init__(
+        self,
+        img_size=224,
+        patch_size=16,
+        in_chans=3,
+        embed_dim=768,
+        norm_layer=None,
+        flatten=True,
+        patch_stride=16,
+        enable_fusion=False,
+        fusion_type="None",
+    ):
+        super().__init__()
+        img_size = to_2tuple(img_size)
+        patch_size = to_2tuple(patch_size)
+        patch_stride = to_2tuple(patch_stride)
+        self.img_size = img_size
+        self.patch_size = patch_size
+        self.patch_stride = patch_stride
+        self.grid_size = (
+            img_size[0] // patch_stride[0],
+            img_size[1] // patch_stride[1],
+        )
+        self.num_patches = self.grid_size[0] * self.grid_size[1]
+        self.flatten = flatten
+        self.in_chans = in_chans
+        self.embed_dim = embed_dim
+
+        self.enable_fusion = enable_fusion
+        self.fusion_type = fusion_type
+
+        padding = (
+            (patch_size[0] - patch_stride[0]) // 2,
+            (patch_size[1] - patch_stride[1]) // 2,
+        )
+
+        if (self.enable_fusion) and (self.fusion_type == "channel_map"):
+            self.proj = nn.Conv2d(
+                in_chans * 4,
+                embed_dim,
+                kernel_size=patch_size,
+                stride=patch_stride,
+                padding=padding,
+            )
+        else:
+            self.proj = nn.Conv2d(
+                in_chans,
+                embed_dim,
+                kernel_size=patch_size,
+                stride=patch_stride,
+                padding=padding,
+            )
+        self.norm = norm_layer(embed_dim) if norm_layer else nn.Identity()
+
+        if (self.enable_fusion) and (
+            self.fusion_type in ["daf_2d", "aff_2d", "iaff_2d"]
+        ):
+            self.mel_conv2d = nn.Conv2d(
+                in_chans,
+                embed_dim,
+                kernel_size=(patch_size[0], patch_size[1] * 3),
+                stride=(patch_stride[0], patch_stride[1] * 3),
+                padding=padding,
+            )
+            if self.fusion_type == "daf_2d":
+                self.fusion_model = DAF()
+            elif self.fusion_type == "aff_2d":
+                self.fusion_model = AFF(channels=embed_dim, type="2D")
+            elif self.fusion_type == "iaff_2d":
+                self.fusion_model = iAFF(channels=embed_dim, type="2D")
+
+    def forward(self, x, longer_idx=None):
+        if (self.enable_fusion) and (
+            self.fusion_type in ["daf_2d", "aff_2d", "iaff_2d"]
+        ):
+            global_x = x[:, 0:1, :, :]
+
+            # global processing
+            B, C, H, W = global_x.shape
+            assert (
+                H == self.img_size[0] and W == self.img_size[1]
+            ), f"Input image size ({H}*{W}) doesn't match model ({self.img_size[0]}*{self.img_size[1]})."
+            global_x = self.proj(global_x)
+            TW = global_x.size(-1)
+            if len(longer_idx) > 0:
+                # local processing
+                local_x = x[longer_idx, 1:, :, :].contiguous()
+                B, C, H, W = local_x.shape
+                local_x = local_x.view(B * C, 1, H, W)
+                local_x = self.mel_conv2d(local_x)
+                local_x = local_x.view(
+                    B, C, local_x.size(1), local_x.size(2), local_x.size(3)
+                )
+                local_x = local_x.permute((0, 2, 3, 1, 4)).contiguous().flatten(3)
+                TB, TC, TH, _ = local_x.size()
+                if local_x.size(-1) < TW:
+                    local_x = torch.cat(
+                        [
+                            local_x,
+                            torch.zeros(
+                                (TB, TC, TH, TW - local_x.size(-1)),
+                                device=global_x.device,
+                            ),
+                        ],
+                        dim=-1,
+                    )
+                else:
+                    local_x = local_x[:, :, :, :TW]
+
+                global_x[longer_idx] = self.fusion_model(global_x[longer_idx], local_x)
+            x = global_x
+        else:
+            B, C, H, W = x.shape
+            assert (
+                H == self.img_size[0] and W == self.img_size[1]
+            ), f"Input image size ({H}*{W}) doesn't match model ({self.img_size[0]}*{self.img_size[1]})."
+            x = self.proj(x)
+
+        if self.flatten:
+            x = x.flatten(2).transpose(1, 2)  # BCHW -> BNC
+        x = self.norm(x)
+        return x
+
+
+class Mlp(nn.Module):
+    """MLP as used in Vision Transformer, MLP-Mixer and related networks"""
+
+    def __init__(
+        self,
+        in_features,
+        hidden_features=None,
+        out_features=None,
+        act_layer=nn.GELU,
+        drop=0.0,
+    ):
+        super().__init__()
+        out_features = out_features or in_features
+        hidden_features = hidden_features or in_features
+        self.fc1 = nn.Linear(in_features, hidden_features)
+        self.act = act_layer()
+        self.fc2 = nn.Linear(hidden_features, out_features)
+        self.drop = nn.Dropout(drop)
+
+    def forward(self, x):
+        x = self.fc1(x)
+        x = self.act(x)
+        x = self.drop(x)
+        x = self.fc2(x)
+        x = self.drop(x)
+        return x
+
+
+def _no_grad_trunc_normal_(tensor, mean, std, a, b):
+    # Cut & paste from PyTorch official master until it's in a few official releases - RW
+    # Method based on https://people.sc.fsu.edu/~jburkardt/presentations/truncated_normal.pdf
+    def norm_cdf(x):
+        # Computes standard normal cumulative distribution function
+        return (1.0 + math.erf(x / math.sqrt(2.0))) / 2.0
+
+    if (mean < a - 2 * std) or (mean > b + 2 * std):
+        warnings.warn(
+            "mean is more than 2 std from [a, b] in nn.init.trunc_normal_. "
+            "The distribution of values may be incorrect.",
+            stacklevel=2,
+        )
+
+    with torch.no_grad():
+        # Values are generated by using a truncated uniform distribution and
+        # then using the inverse CDF for the normal distribution.
+        # Get upper and lower cdf values
+        l = norm_cdf((a - mean) / std)
+        u = norm_cdf((b - mean) / std)
+
+        # Uniformly fill tensor with values from [l, u], then translate to
+        # [2l-1, 2u-1].
+        tensor.uniform_(2 * l - 1, 2 * u - 1)
+
+        # Use inverse cdf transform for normal distribution to get truncated
+        # standard normal
+        tensor.erfinv_()
+
+        # Transform to proper mean, std
+        tensor.mul_(std * math.sqrt(2.0))
+        tensor.add_(mean)
+
+        # Clamp to ensure it's in the proper range
+        tensor.clamp_(min=a, max=b)
+        return tensor
+
+
+def trunc_normal_(tensor, mean=0.0, std=1.0, a=-2.0, b=2.0):
+    # type: (Tensor, float, float, float, float) -> Tensor
+    r"""Fills the input Tensor with values drawn from a truncated
+    normal distribution. The values are effectively drawn from the
+    normal distribution :math:`\mathcal{N}(\text{mean}, \text{std}^2)`
+    with values outside :math:`[a, b]` redrawn until they are within
+    the bounds. The method used for generating the random values works
+    best when :math:`a \leq \text{mean} \leq b`.
+    Args:
+        tensor: an n-dimensional `torch.Tensor`
+        mean: the mean of the normal distribution
+        std: the standard deviation of the normal distribution
+        a: the minimum cutoff value
+        b: the maximum cutoff value
+    Examples:
+        >>> w = torch.empty(3, 5)
+        >>> nn.init.trunc_normal_(w)
+    """
+    return _no_grad_trunc_normal_(tensor, mean, std, a, b)
+
+
+def variance_scaling_(tensor, scale=1.0, mode="fan_in", distribution="normal"):
+    fan_in, fan_out = _calculate_fan_in_and_fan_out(tensor)
+    if mode == "fan_in":
+        denom = fan_in
+    elif mode == "fan_out":
+        denom = fan_out
+    elif mode == "fan_avg":
+        denom = (fan_in + fan_out) / 2
+
+    variance = scale / denom
+
+    if distribution == "truncated_normal":
+        # constant is stddev of standard normal truncated to (-2, 2)
+        trunc_normal_(tensor, std=math.sqrt(variance) / 0.87962566103423978)
+    elif distribution == "normal":
+        tensor.normal_(std=math.sqrt(variance))
+    elif distribution == "uniform":
+        bound = math.sqrt(3 * variance)
+        tensor.uniform_(-bound, bound)
+    else:
+        raise ValueError(f"invalid distribution {distribution}")
+
+
+def lecun_normal_(tensor):
+    variance_scaling_(tensor, mode="fan_in", distribution="truncated_normal")
+
+
+def window_partition(x, window_size):
+    """
+    Args:
+        x: (B, H, W, C)
+        window_size (int): window size
+    Returns:
+        windows: (num_windows*B, window_size, window_size, C)
+    """
+    B, H, W, C = x.shape
+    x = x.view(B, H // window_size, window_size, W // window_size, window_size, C)
+    windows = (
+        x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size, window_size, C)
+    )
+    return windows
+
+
+def window_reverse(windows, window_size, H, W):
+    """
+    Args:
+        windows: (num_windows*B, window_size, window_size, C)
+        window_size (int): Window size
+        H (int): Height of image
+        W (int): Width of image
+    Returns:
+        x: (B, H, W, C)
+    """
+    B = int(windows.shape[0] / (H * W / window_size / window_size))
+    x = windows.view(
+        B, H // window_size, W // window_size, window_size, window_size, -1
+    )
+    x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(B, H, W, -1)
+    return x
+
+
+class WindowAttention(nn.Module):
+    r"""Window based multi-head self attention (W-MSA) module with relative position bias.
+    It supports both of shifted and non-shifted window.
+    Args:
+        dim (int): Number of input channels.
+        window_size (tuple[int]): The height and width of the window.
+        num_heads (int): Number of attention heads.
+        qkv_bias (bool, optional):  If True, add a learnable bias to query, key, value. Default: True
+        qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set
+        attn_drop (float, optional): Dropout ratio of attention weight. Default: 0.0
+        proj_drop (float, optional): Dropout ratio of output. Default: 0.0
+    """
+
+    def __init__(
+        self,
+        dim,
+        window_size,
+        num_heads,
+        qkv_bias=True,
+        qk_scale=None,
+        attn_drop=0.0,
+        proj_drop=0.0,
+    ):
+        super().__init__()
+        self.dim = dim
+        self.window_size = window_size  # Wh, Ww
+        self.num_heads = num_heads
+        head_dim = dim // num_heads
+        self.scale = qk_scale or head_dim**-0.5
+
+        # define a parameter table of relative position bias
+        self.relative_position_bias_table = nn.Parameter(
+            torch.zeros((2 * window_size[0] - 1) * (2 * window_size[1] - 1), num_heads)
+        )  # 2*Wh-1 * 2*Ww-1, nH
+
+        # get pair-wise relative position index for each token inside the window
+        coords_h = torch.arange(self.window_size[0])
+        coords_w = torch.arange(self.window_size[1])
+        coords = torch.stack(torch.meshgrid([coords_h, coords_w]))  # 2, Wh, Ww
+        coords_flatten = torch.flatten(coords, 1)  # 2, Wh*Ww
+        relative_coords = (
+            coords_flatten[:, :, None] - coords_flatten[:, None, :]
+        )  # 2, Wh*Ww, Wh*Ww
+        relative_coords = relative_coords.permute(
+            1, 2, 0
+        ).contiguous()  # Wh*Ww, Wh*Ww, 2
+        relative_coords[:, :, 0] += self.window_size[0] - 1  # shift to start from 0
+        relative_coords[:, :, 1] += self.window_size[1] - 1
+        relative_coords[:, :, 0] *= 2 * self.window_size[1] - 1
+        relative_position_index = relative_coords.sum(-1)  # Wh*Ww, Wh*Ww
+        self.register_buffer("relative_position_index", relative_position_index)
+
+        self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
+        self.attn_drop = nn.Dropout(attn_drop)
+        self.proj = nn.Linear(dim, dim)
+        self.proj_drop = nn.Dropout(proj_drop)
+
+        trunc_normal_(self.relative_position_bias_table, std=0.02)
+        self.softmax = nn.Softmax(dim=-1)
+
+    def forward(self, x, mask=None):
+        """
+        Args:
+            x: input features with shape of (num_windows*B, N, C)
+            mask: (0/-inf) mask with shape of (num_windows, Wh*Ww, Wh*Ww) or None
+        """
+        B_, N, C = x.shape
+        qkv = (
+            self.qkv(x)
+            .reshape(B_, N, 3, self.num_heads, C // self.num_heads)
+            .permute(2, 0, 3, 1, 4)
+        )
+        q, k, v = (
+            qkv[0],
+            qkv[1],
+            qkv[2],
+        )  # make torchscript happy (cannot use tensor as tuple)
+
+        q = q * self.scale
+        attn = q @ k.transpose(-2, -1)
+
+        relative_position_bias = self.relative_position_bias_table[
+            self.relative_position_index.view(-1)
+        ].view(
+            self.window_size[0] * self.window_size[1],
+            self.window_size[0] * self.window_size[1],
+            -1,
+        )  # Wh*Ww,Wh*Ww,nH
+        relative_position_bias = relative_position_bias.permute(
+            2, 0, 1
+        ).contiguous()  # nH, Wh*Ww, Wh*Ww
+        attn = attn + relative_position_bias.unsqueeze(0)
+
+        if mask is not None:
+            nW = mask.shape[0]
+            attn = attn.view(B_ // nW, nW, self.num_heads, N, N) + mask.unsqueeze(
+                1
+            ).unsqueeze(0)
+            attn = attn.view(-1, self.num_heads, N, N)
+            attn = self.softmax(attn)
+        else:
+            attn = self.softmax(attn)
+
+        attn = self.attn_drop(attn)
+
+        x = (attn @ v).transpose(1, 2).reshape(B_, N, C)
+        x = self.proj(x)
+        x = self.proj_drop(x)
+        return x, attn
+
+    def extra_repr(self):
+        return f"dim={self.dim}, window_size={self.window_size}, num_heads={self.num_heads}"
+
+
+# We use the model based on Swintransformer Block, therefore we can use the swin-transformer pretrained model
+class SwinTransformerBlock(nn.Module):
+    r"""Swin Transformer Block.
+    Args:
+        dim (int): Number of input channels.
+        input_resolution (tuple[int]): Input resulotion.
+        num_heads (int): Number of attention heads.
+        window_size (int): Window size.
+        shift_size (int): Shift size for SW-MSA.
+        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
+        qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True
+        qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set.
+        drop (float, optional): Dropout rate. Default: 0.0
+        attn_drop (float, optional): Attention dropout rate. Default: 0.0
+        drop_path (float, optional): Stochastic depth rate. Default: 0.0
+        act_layer (nn.Module, optional): Activation layer. Default: nn.GELU
+        norm_layer (nn.Module, optional): Normalization layer.  Default: nn.LayerNorm
+    """
+
+    def __init__(
+        self,
+        dim,
+        input_resolution,
+        num_heads,
+        window_size=7,
+        shift_size=0,
+        mlp_ratio=4.0,
+        qkv_bias=True,
+        qk_scale=None,
+        drop=0.0,
+        attn_drop=0.0,
+        drop_path=0.0,
+        act_layer=nn.GELU,
+        norm_layer=nn.LayerNorm,
+        norm_before_mlp="ln",
+    ):
+        super().__init__()
+        self.dim = dim
+        self.input_resolution = input_resolution
+        self.num_heads = num_heads
+        self.window_size = window_size
+        self.shift_size = shift_size
+        self.mlp_ratio = mlp_ratio
+        self.norm_before_mlp = norm_before_mlp
+        if min(self.input_resolution) <= self.window_size:
+            # if window size is larger than input resolution, we don't partition windows
+            self.shift_size = 0
+            self.window_size = min(self.input_resolution)
+        assert (
+            0 <= self.shift_size < self.window_size
+        ), "shift_size must in 0-window_size"
+
+        self.norm1 = norm_layer(dim)
+        self.attn = WindowAttention(
+            dim,
+            window_size=to_2tuple(self.window_size),
+            num_heads=num_heads,
+            qkv_bias=qkv_bias,
+            qk_scale=qk_scale,
+            attn_drop=attn_drop,
+            proj_drop=drop,
+        )
+
+        self.drop_path = DropPath(drop_path) if drop_path > 0.0 else nn.Identity()
+        if self.norm_before_mlp == "ln":
+            self.norm2 = nn.LayerNorm(dim)
+        elif self.norm_before_mlp == "bn":
+            self.norm2 = lambda x: nn.BatchNorm1d(dim)(x.transpose(1, 2)).transpose(
+                1, 2
+            )
+        else:
+            raise NotImplementedError
+        mlp_hidden_dim = int(dim * mlp_ratio)
+        self.mlp = Mlp(
+            in_features=dim,
+            hidden_features=mlp_hidden_dim,
+            act_layer=act_layer,
+            drop=drop,
+        )
+
+        if self.shift_size > 0:
+            # calculate attention mask for SW-MSA
+            H, W = self.input_resolution
+            img_mask = torch.zeros((1, H, W, 1))  # 1 H W 1
+            h_slices = (
+                slice(0, -self.window_size),
+                slice(-self.window_size, -self.shift_size),
+                slice(-self.shift_size, None),
+            )
+            w_slices = (
+                slice(0, -self.window_size),
+                slice(-self.window_size, -self.shift_size),
+                slice(-self.shift_size, None),
+            )
+            cnt = 0
+            for h in h_slices:
+                for w in w_slices:
+                    img_mask[:, h, w, :] = cnt
+                    cnt += 1
+
+            mask_windows = window_partition(
+                img_mask, self.window_size
+            )  # nW, window_size, window_size, 1
+            mask_windows = mask_windows.view(-1, self.window_size * self.window_size)
+            attn_mask = mask_windows.unsqueeze(1) - mask_windows.unsqueeze(2)
+            attn_mask = attn_mask.masked_fill(
+                attn_mask != 0, float(-100.0)
+            ).masked_fill(attn_mask == 0, float(0.0))
+        else:
+            attn_mask = None
+
+        self.register_buffer("attn_mask", attn_mask)
+
+    def forward(self, x):
+        # pdb.set_trace()
+        H, W = self.input_resolution
+        # print("H: ", H)
+        # print("W: ", W)
+        # pdb.set_trace()
+        B, L, C = x.shape
+        # assert L == H * W, "input feature has wrong size"
+
+        shortcut = x
+        x = self.norm1(x)
+        x = x.view(B, H, W, C)
+
+        # cyclic shift
+        if self.shift_size > 0:
+            shifted_x = torch.roll(
+                x, shifts=(-self.shift_size, -self.shift_size), dims=(1, 2)
+            )
+        else:
+            shifted_x = x
+
+        # partition windows
+        x_windows = window_partition(
+            shifted_x, self.window_size
+        )  # nW*B, window_size, window_size, C
+        x_windows = x_windows.view(
+            -1, self.window_size * self.window_size, C
+        )  # nW*B, window_size*window_size, C
+
+        # W-MSA/SW-MSA
+        attn_windows, attn = self.attn(
+            x_windows, mask=self.attn_mask
+        )  # nW*B, window_size*window_size, C
+
+        # merge windows
+        attn_windows = attn_windows.view(-1, self.window_size, self.window_size, C)
+        shifted_x = window_reverse(attn_windows, self.window_size, H, W)  # B H' W' C
+
+        # reverse cyclic shift
+        if self.shift_size > 0:
+            x = torch.roll(
+                shifted_x, shifts=(self.shift_size, self.shift_size), dims=(1, 2)
+            )
+        else:
+            x = shifted_x
+        x = x.view(B, H * W, C)
+
+        # FFN
+        x = shortcut + self.drop_path(x)
+        x = x + self.drop_path(self.mlp(self.norm2(x)))
+
+        return x, attn
+
+    def extra_repr(self):
+        return (
+            f"dim={self.dim}, input_resolution={self.input_resolution}, num_heads={self.num_heads}, "
+            f"window_size={self.window_size}, shift_size={self.shift_size}, mlp_ratio={self.mlp_ratio}"
+        )
+
+
+class PatchMerging(nn.Module):
+    r"""Patch Merging Layer.
+    Args:
+        input_resolution (tuple[int]): Resolution of input feature.
+        dim (int): Number of input channels.
+        norm_layer (nn.Module, optional): Normalization layer.  Default: nn.LayerNorm
+    """
+
+    def __init__(self, input_resolution, dim, norm_layer=nn.LayerNorm):
+        super().__init__()
+        self.input_resolution = input_resolution
+        self.dim = dim
+        self.reduction = nn.Linear(4 * dim, 2 * dim, bias=False)
+        self.norm = norm_layer(4 * dim)
+
+    def forward(self, x):
+        """
+        x: B, H*W, C
+        """
+        H, W = self.input_resolution
+        B, L, C = x.shape
+        assert L == H * W, "input feature has wrong size"
+        assert H % 2 == 0 and W % 2 == 0, f"x size ({H}*{W}) are not even."
+
+        x = x.view(B, H, W, C)
+
+        x0 = x[:, 0::2, 0::2, :]  # B H/2 W/2 C
+        x1 = x[:, 1::2, 0::2, :]  # B H/2 W/2 C
+        x2 = x[:, 0::2, 1::2, :]  # B H/2 W/2 C
+        x3 = x[:, 1::2, 1::2, :]  # B H/2 W/2 C
+        x = torch.cat([x0, x1, x2, x3], -1)  # B H/2 W/2 4*C
+        x = x.view(B, -1, 4 * C)  # B H/2*W/2 4*C
+
+        x = self.norm(x)
+        x = self.reduction(x)
+
+        return x
+
+    def extra_repr(self):
+        return f"input_resolution={self.input_resolution}, dim={self.dim}"
+
+
+class BasicLayer(nn.Module):
+    """A basic Swin Transformer layer for one stage.
+    Args:
+        dim (int): Number of input channels.
+        input_resolution (tuple[int]): Input resolution.
+        depth (int): Number of blocks.
+        num_heads (int): Number of attention heads.
+        window_size (int): Local window size.
+        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
+        qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True
+        qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set.
+        drop (float, optional): Dropout rate. Default: 0.0
+        attn_drop (float, optional): Attention dropout rate. Default: 0.0
+        drop_path (float | tuple[float], optional): Stochastic depth rate. Default: 0.0
+        norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm
+        downsample (nn.Module | None, optional): Downsample layer at the end of the layer. Default: None
+        use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False.
+    """
+
+    def __init__(
+        self,
+        dim,
+        input_resolution,
+        depth,
+        num_heads,
+        window_size,
+        mlp_ratio=4.0,
+        qkv_bias=True,
+        qk_scale=None,
+        drop=0.0,
+        attn_drop=0.0,
+        drop_path=0.0,
+        norm_layer=nn.LayerNorm,
+        downsample=None,
+        use_checkpoint=False,
+        norm_before_mlp="ln",
+    ):
+        super().__init__()
+        self.dim = dim
+        self.input_resolution = input_resolution
+        self.depth = depth
+        self.use_checkpoint = use_checkpoint
+
+        # build blocks
+        self.blocks = nn.ModuleList(
+            [
+                SwinTransformerBlock(
+                    dim=dim,
+                    input_resolution=input_resolution,
+                    num_heads=num_heads,
+                    window_size=window_size,
+                    shift_size=0 if (i % 2 == 0) else window_size // 2,
+                    mlp_ratio=mlp_ratio,
+                    qkv_bias=qkv_bias,
+                    qk_scale=qk_scale,
+                    drop=drop,
+                    attn_drop=attn_drop,
+                    drop_path=drop_path[i]
+                    if isinstance(drop_path, list)
+                    else drop_path,
+                    norm_layer=norm_layer,
+                    norm_before_mlp=norm_before_mlp,
+                )
+                for i in range(depth)
+            ]
+        )
+
+        # patch merging layer
+        if downsample is not None:
+            self.downsample = downsample(
+                input_resolution, dim=dim, norm_layer=norm_layer
+            )
+        else:
+            self.downsample = None
+
+    def forward(self, x):
+        attns = []
+        for blk in self.blocks:
+            if self.use_checkpoint:
+                x = checkpoint.checkpoint(blk, x)
+            else:
+                x, attn = blk(x)
+                if not self.training:
+                    attns.append(attn.unsqueeze(0))
+        if self.downsample is not None:
+            x = self.downsample(x)
+        if not self.training:
+            attn = torch.cat(attns, dim=0)
+            attn = torch.mean(attn, dim=0)
+        return x, attn
+
+    def extra_repr(self):
+        return f"dim={self.dim}, input_resolution={self.input_resolution}, depth={self.depth}"
+
+
+# The Core of HTSAT
+class HTSAT_Swin_Transformer(nn.Module):
+    r"""HTSAT based on the Swin Transformer
+    Args:
+        spec_size (int | tuple(int)): Input Spectrogram size. Default 256
+        patch_size (int | tuple(int)): Patch size. Default: 4
+        path_stride (iot | tuple(int)): Patch Stride for Frequency and Time Axis. Default: 4
+        in_chans (int): Number of input image channels. Default: 1 (mono)
+        num_classes (int): Number of classes for classification head. Default: 527
+        embed_dim (int): Patch embedding dimension. Default: 96
+        depths (tuple(int)): Depth of each HTSAT-Swin Transformer layer.
+        num_heads (tuple(int)): Number of attention heads in different layers.
+        window_size (int): Window size. Default: 8
+        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. Default: 4
+        qkv_bias (bool): If True, add a learnable bias to query, key, value. Default: True
+        qk_scale (float): Override default qk scale of head_dim ** -0.5 if set. Default: None
+        drop_rate (float): Dropout rate. Default: 0
+        attn_drop_rate (float): Attention dropout rate. Default: 0
+        drop_path_rate (float): Stochastic depth rate. Default: 0.1
+        norm_layer (nn.Module): Normalization layer. Default: nn.LayerNorm.
+        ape (bool): If True, add absolute position embedding to the patch embedding. Default: False
+        patch_norm (bool): If True, add normalization after patch embedding. Default: True
+        use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False
+        config (module): The configuration Module from config.py
+    """
+
+    def __init__(
+        self,
+        spec_size=256,
+        patch_size=4,
+        patch_stride=(4, 4),
+        in_chans=1,
+        num_classes=527,
+        embed_dim=96,
+        depths=[2, 2, 6, 2],
+        num_heads=[4, 8, 16, 32],
+        window_size=8,
+        mlp_ratio=4.0,
+        qkv_bias=True,
+        qk_scale=None,
+        drop_rate=0.0,
+        attn_drop_rate=0.0,
+        drop_path_rate=0.1,
+        norm_layer=nn.LayerNorm,
+        ape=False,
+        patch_norm=True,
+        use_checkpoint=False,
+        norm_before_mlp="ln",
+        config=None,
+        enable_fusion=False,
+        fusion_type="None",
+        **kwargs,
+    ):
+        super(HTSAT_Swin_Transformer, self).__init__()
+
+        self.config = config
+        self.spec_size = spec_size
+        self.patch_stride = patch_stride
+        self.patch_size = patch_size
+        self.window_size = window_size
+        self.embed_dim = embed_dim
+        self.depths = depths
+        self.ape = ape
+        self.in_chans = in_chans
+        self.num_classes = num_classes
+        self.num_heads = num_heads
+        self.num_layers = len(self.depths)
+        self.num_features = int(self.embed_dim * 2 ** (self.num_layers - 1))
+
+        self.drop_rate = drop_rate
+        self.attn_drop_rate = attn_drop_rate
+        self.drop_path_rate = drop_path_rate
+
+        self.qkv_bias = qkv_bias
+        self.qk_scale = None
+
+        self.patch_norm = patch_norm
+        self.norm_layer = norm_layer if self.patch_norm else None
+        self.norm_before_mlp = norm_before_mlp
+        self.mlp_ratio = mlp_ratio
+
+        self.use_checkpoint = use_checkpoint
+
+        self.enable_fusion = enable_fusion
+        self.fusion_type = fusion_type
+
+        #  process mel-spec ; used only once
+        self.freq_ratio = self.spec_size // self.config.mel_bins
+        window = "hann"
+        center = True
+        pad_mode = "reflect"
+        ref = 1.0
+        amin = 1e-10
+        top_db = None
+        self.interpolate_ratio = 32  # Downsampled ratio
+        # Spectrogram extractor
+        self.spectrogram_extractor = Spectrogram(
+            n_fft=config.window_size,
+            hop_length=config.hop_size,
+            win_length=config.window_size,
+            window=window,
+            center=center,
+            pad_mode=pad_mode,
+            freeze_parameters=True,
+        )
+        # Logmel feature extractor
+        self.logmel_extractor = LogmelFilterBank(
+            sr=config.sample_rate,
+            n_fft=config.window_size,
+            n_mels=config.mel_bins,
+            fmin=config.fmin,
+            fmax=config.fmax,
+            ref=ref,
+            amin=amin,
+            top_db=top_db,
+            freeze_parameters=True,
+        )
+        # Spec augmenter
+        self.spec_augmenter = SpecAugmentation(
+            time_drop_width=64,
+            time_stripes_num=2,
+            freq_drop_width=8,
+            freq_stripes_num=2,
+        )  # 2 2
+        self.bn0 = nn.BatchNorm2d(self.config.mel_bins)
+
+        # split spctrogram into non-overlapping patches
+        self.patch_embed = PatchEmbed(
+            img_size=self.spec_size,
+            patch_size=self.patch_size,
+            in_chans=self.in_chans,
+            embed_dim=self.embed_dim,
+            norm_layer=self.norm_layer,
+            patch_stride=patch_stride,
+            enable_fusion=self.enable_fusion,
+            fusion_type=self.fusion_type,
+        )
+
+        num_patches = self.patch_embed.num_patches
+        patches_resolution = self.patch_embed.grid_size
+        self.patches_resolution = patches_resolution
+
+        # absolute position embedding
+        if self.ape:
+            self.absolute_pos_embed = nn.Parameter(
+                torch.zeros(1, num_patches, self.embed_dim)
+            )
+            trunc_normal_(self.absolute_pos_embed, std=0.02)
+
+        self.pos_drop = nn.Dropout(p=self.drop_rate)
+
+        # stochastic depth
+        dpr = [
+            x.item() for x in torch.linspace(0, self.drop_path_rate, sum(self.depths))
+        ]  # stochastic depth decay rule
+
+        # build layers
+        self.layers = nn.ModuleList()
+        for i_layer in range(self.num_layers):
+            layer = BasicLayer(
+                dim=int(self.embed_dim * 2**i_layer),
+                input_resolution=(
+                    patches_resolution[0] // (2**i_layer),
+                    patches_resolution[1] // (2**i_layer),
+                ),
+                depth=self.depths[i_layer],
+                num_heads=self.num_heads[i_layer],
+                window_size=self.window_size,
+                mlp_ratio=self.mlp_ratio,
+                qkv_bias=self.qkv_bias,
+                qk_scale=self.qk_scale,
+                drop=self.drop_rate,
+                attn_drop=self.attn_drop_rate,
+                drop_path=dpr[
+                    sum(self.depths[:i_layer]) : sum(self.depths[: i_layer + 1])
+                ],
+                norm_layer=self.norm_layer,
+                downsample=PatchMerging if (i_layer < self.num_layers - 1) else None,
+                use_checkpoint=use_checkpoint,
+                norm_before_mlp=self.norm_before_mlp,
+            )
+            self.layers.append(layer)
+
+        self.norm = self.norm_layer(self.num_features)
+        self.avgpool = nn.AdaptiveAvgPool1d(1)
+        self.maxpool = nn.AdaptiveMaxPool1d(1)
+
+        SF = (
+            self.spec_size
+            // (2 ** (len(self.depths) - 1))
+            // self.patch_stride[0]
+            // self.freq_ratio
+        )
+        self.tscam_conv = nn.Conv2d(
+            in_channels=self.num_features,
+            out_channels=self.num_classes,
+            kernel_size=(SF, 3),
+            padding=(0, 1),
+        )
+        self.head = nn.Linear(num_classes, num_classes)
+
+        if (self.enable_fusion) and (
+            self.fusion_type in ["daf_1d", "aff_1d", "iaff_1d"]
+        ):
+            self.mel_conv1d = nn.Sequential(
+                nn.Conv1d(64, 64, kernel_size=5, stride=3, padding=2),
+                nn.BatchNorm1d(64),
+            )
+            if self.fusion_type == "daf_1d":
+                self.fusion_model = DAF()
+            elif self.fusion_type == "aff_1d":
+                self.fusion_model = AFF(channels=64, type="1D")
+            elif self.fusion_type == "iaff_1d":
+                self.fusion_model = iAFF(channels=64, type="1D")
+
+        self.apply(self._init_weights)
+
+    def _init_weights(self, m):
+        if isinstance(m, nn.Linear):
+            trunc_normal_(m.weight, std=0.02)
+            if isinstance(m, nn.Linear) and m.bias is not None:
+                nn.init.constant_(m.bias, 0)
+        elif isinstance(m, nn.LayerNorm):
+            nn.init.constant_(m.bias, 0)
+            nn.init.constant_(m.weight, 1.0)
+
+    @torch.jit.ignore
+    def no_weight_decay(self):
+        return {"absolute_pos_embed"}
+
+    @torch.jit.ignore
+    def no_weight_decay_keywords(self):
+        return {"relative_position_bias_table"}
+
+    def forward_features(self, x, longer_idx=None):
+        # A deprecated optimization for using a hierarchical output from different blocks
+
+        frames_num = x.shape[2]
+        x = self.patch_embed(x, longer_idx=longer_idx)
+        if self.ape:
+            x = x + self.absolute_pos_embed
+        x = self.pos_drop(x)
+        for i, layer in enumerate(self.layers):
+            x, attn = layer(x)
+        # for x
+        x = self.norm(x)
+        B, N, C = x.shape
+        SF = frames_num // (2 ** (len(self.depths) - 1)) // self.patch_stride[0]
+        ST = frames_num // (2 ** (len(self.depths) - 1)) // self.patch_stride[1]
+        x = x.permute(0, 2, 1).contiguous().reshape(B, C, SF, ST)
+        B, C, F, T = x.shape
+        # group 2D CNN
+        c_freq_bin = F // self.freq_ratio
+        x = x.reshape(B, C, F // c_freq_bin, c_freq_bin, T)
+        x = x.permute(0, 1, 3, 2, 4).contiguous().reshape(B, C, c_freq_bin, -1)
+        # get latent_output
+        fine_grained_latent_output = torch.mean(x, dim=2)
+        fine_grained_latent_output = interpolate(
+            fine_grained_latent_output.permute(0, 2, 1).contiguous(),
+            8 * self.patch_stride[1],
+        )
+
+        latent_output = self.avgpool(torch.flatten(x, 2))
+        latent_output = torch.flatten(latent_output, 1)
+
+        # display the attention map, if needed
+
+        x = self.tscam_conv(x)
+        x = torch.flatten(x, 2)  # B, C, T
+
+        fpx = interpolate(
+            torch.sigmoid(x).permute(0, 2, 1).contiguous(), 8 * self.patch_stride[1]
+        )
+
+        x = self.avgpool(x)
+        x = torch.flatten(x, 1)
+
+        output_dict = {
+            "framewise_output": fpx,  # already sigmoided
+            "clipwise_output": torch.sigmoid(x),
+            "fine_grained_embedding": fine_grained_latent_output,
+            "embedding": latent_output,
+        }
+
+        return output_dict
+
+    def crop_wav(self, x, crop_size, spe_pos=None):
+        time_steps = x.shape[2]
+        tx = torch.zeros(x.shape[0], x.shape[1], crop_size, x.shape[3]).to(x.device)
+        for i in range(len(x)):
+            if spe_pos is None:
+                crop_pos = random.randint(0, time_steps - crop_size - 1)
+            else:
+                crop_pos = spe_pos
+            tx[i][0] = x[i, 0, crop_pos : crop_pos + crop_size, :]
+        return tx
+
+    # Reshape the wavform to a img size, if you want to use the pretrained swin transformer model
+    def reshape_wav2img(self, x):
+        B, C, T, F = x.shape
+        target_T = int(self.spec_size * self.freq_ratio)
+        target_F = self.spec_size // self.freq_ratio
+        assert (
+            T <= target_T and F <= target_F
+        ), "the wav size should less than or equal to the swin input size"
+        # to avoid bicubic zero error
+        if T < target_T:
+            x = nn.functional.interpolate(
+                x, (target_T, x.shape[3]), mode="bicubic", align_corners=True
+            )
+        if F < target_F:
+            x = nn.functional.interpolate(
+                x, (x.shape[2], target_F), mode="bicubic", align_corners=True
+            )
+        x = x.permute(0, 1, 3, 2).contiguous()
+        x = x.reshape(
+            x.shape[0],
+            x.shape[1],
+            x.shape[2],
+            self.freq_ratio,
+            x.shape[3] // self.freq_ratio,
+        )
+        # print(x.shape)
+        x = x.permute(0, 1, 3, 2, 4).contiguous()
+        x = x.reshape(x.shape[0], x.shape[1], x.shape[2] * x.shape[3], x.shape[4])
+        return x
+
+    # Repeat the wavform to a img size, if you want to use the pretrained swin transformer model
+    def repeat_wat2img(self, x, cur_pos):
+        B, C, T, F = x.shape
+        target_T = int(self.spec_size * self.freq_ratio)
+        target_F = self.spec_size // self.freq_ratio
+        assert (
+            T <= target_T and F <= target_F
+        ), "the wav size should less than or equal to the swin input size"
+        # to avoid bicubic zero error
+        if T < target_T:
+            x = nn.functional.interpolate(
+                x, (target_T, x.shape[3]), mode="bicubic", align_corners=True
+            )
+        if F < target_F:
+            x = nn.functional.interpolate(
+                x, (x.shape[2], target_F), mode="bicubic", align_corners=True
+            )
+        x = x.permute(0, 1, 3, 2).contiguous()  # B C F T
+        x = x[:, :, :, cur_pos : cur_pos + self.spec_size]
+        x = x.repeat(repeats=(1, 1, 4, 1))
+        return x
+
+    def forward(
+        self, x: torch.Tensor, mixup_lambda=None, infer_mode=False, device=None
+    ):  # out_feat_keys: List[str] = None):
+        if self.enable_fusion and x["longer"].sum() == 0:
+            # if no audio is longer than 10s, then randomly select one audio to be longer
+            x["longer"][torch.randint(0, x["longer"].shape[0], (1,))] = True
+
+        if not self.enable_fusion:
+            x = x["waveform"].to(device=device, non_blocking=True)
+            x = self.spectrogram_extractor(x)  # (batch_size, 1, time_steps, freq_bins)
+            x = self.logmel_extractor(x)  # (batch_size, 1, time_steps, mel_bins)
+            x = x.transpose(1, 3)
+            x = self.bn0(x)
+            x = x.transpose(1, 3)
+            if self.training:
+                x = self.spec_augmenter(x)
+
+            if self.training and mixup_lambda is not None:
+                x = do_mixup(x, mixup_lambda)
+
+            x = self.reshape_wav2img(x)
+            output_dict = self.forward_features(x)
+        else:
+            longer_list = x["longer"].to(device=device, non_blocking=True)
+            x = x["mel_fusion"].to(device=device, non_blocking=True)
+            x = x.transpose(1, 3)
+            x = self.bn0(x)
+            x = x.transpose(1, 3)
+            longer_list_idx = torch.where(longer_list)[0]
+            if self.fusion_type in ["daf_1d", "aff_1d", "iaff_1d"]:
+                new_x = x[:, 0:1, :, :].clone().contiguous()
+                if len(longer_list_idx) > 0:
+                    # local processing
+                    fusion_x_local = x[longer_list_idx, 1:, :, :].clone().contiguous()
+                    FB, FC, FT, FF = fusion_x_local.size()
+                    fusion_x_local = fusion_x_local.view(FB * FC, FT, FF)
+                    fusion_x_local = torch.permute(
+                        fusion_x_local, (0, 2, 1)
+                    ).contiguous()
+                    fusion_x_local = self.mel_conv1d(fusion_x_local)
+                    fusion_x_local = fusion_x_local.view(
+                        FB, FC, FF, fusion_x_local.size(-1)
+                    )
+                    fusion_x_local = (
+                        torch.permute(fusion_x_local, (0, 2, 1, 3))
+                        .contiguous()
+                        .flatten(2)
+                    )
+                    if fusion_x_local.size(-1) < FT:
+                        fusion_x_local = torch.cat(
+                            [
+                                fusion_x_local,
+                                torch.zeros(
+                                    (FB, FF, FT - fusion_x_local.size(-1)),
+                                    device=device,
+                                ),
+                            ],
+                            dim=-1,
+                        )
+                    else:
+                        fusion_x_local = fusion_x_local[:, :, :FT]
+                    # 1D fusion
+                    new_x = new_x.squeeze(1).permute((0, 2, 1)).contiguous()
+                    new_x[longer_list_idx] = self.fusion_model(
+                        new_x[longer_list_idx], fusion_x_local
+                    )
+                    x = new_x.permute((0, 2, 1)).contiguous()[:, None, :, :]
+                else:
+                    x = new_x
+
+            elif self.fusion_type in ["daf_2d", "aff_2d", "iaff_2d", "channel_map"]:
+                x = x  # no change
+
+            if self.training:
+                x = self.spec_augmenter(x)
+            if self.training and mixup_lambda is not None:
+                x = do_mixup(x, mixup_lambda)
+
+            x = self.reshape_wav2img(x)
+            output_dict = self.forward_features(x, longer_idx=longer_list_idx)
+
+        # if infer_mode:
+        #     # in infer mode. we need to handle different length audio input
+        #     frame_num = x.shape[2]
+        #     target_T = int(self.spec_size * self.freq_ratio)
+        #     repeat_ratio = math.floor(target_T / frame_num)
+        #     x = x.repeat(repeats=(1,1,repeat_ratio,1))
+        #     x = self.reshape_wav2img(x)
+        #     output_dict = self.forward_features(x)
+        # else:
+        #     if x.shape[2] > self.freq_ratio * self.spec_size:
+        #         if self.training:
+        #             x = self.crop_wav(x, crop_size=self.freq_ratio * self.spec_size)
+        #             x = self.reshape_wav2img(x)
+        #             output_dict = self.forward_features(x)
+        #         else:
+        #             # Change: Hard code here
+        #             overlap_size = (x.shape[2] - 1) // 4
+        #             output_dicts = []
+        #             crop_size = (x.shape[2] - 1) // 2
+        #             for cur_pos in range(0, x.shape[2] - crop_size - 1, overlap_size):
+        #                 tx = self.crop_wav(x, crop_size = crop_size, spe_pos = cur_pos)
+        #                 tx = self.reshape_wav2img(tx)
+        #                 output_dicts.append(self.forward_features(tx))
+        #             clipwise_output = torch.zeros_like(output_dicts[0]["clipwise_output"]).float().to(x.device)
+        #             framewise_output = torch.zeros_like(output_dicts[0]["framewise_output"]).float().to(x.device)
+        #             for d in output_dicts:
+        #                 clipwise_output += d["clipwise_output"]
+        #                 framewise_output += d["framewise_output"]
+        #             clipwise_output  = clipwise_output / len(output_dicts)
+        #             framewise_output = framewise_output / len(output_dicts)
+        #             output_dict = {
+        #                 'framewise_output': framewise_output,
+        #                 'clipwise_output': clipwise_output
+        #             }
+        #     else: # this part is typically used, and most easy one
+        #         x = self.reshape_wav2img(x)
+        #         output_dict = self.forward_features(x)
+        # x = self.head(x)
+
+        # We process the data in the dataloader part, in that here we only consider the input_T < fixed_T
+
+        return output_dict
+
+
+def create_htsat_model(audio_cfg, enable_fusion=False, fusion_type="None"):
+    try:
+        assert audio_cfg.model_name in [
+            "tiny",
+            "base",
+            "large",
+        ], "model name for HTS-AT is wrong!"
+        if audio_cfg.model_name == "tiny":
+            model = HTSAT_Swin_Transformer(
+                spec_size=256,
+                patch_size=4,
+                patch_stride=(4, 4),
+                num_classes=audio_cfg.class_num,
+                embed_dim=96,
+                depths=[2, 2, 6, 2],
+                num_heads=[4, 8, 16, 32],
+                window_size=8,
+                config=audio_cfg,
+                enable_fusion=enable_fusion,
+                fusion_type=fusion_type,
+            )
+        elif audio_cfg.model_name == "base":
+            model = HTSAT_Swin_Transformer(
+                spec_size=256,
+                patch_size=4,
+                patch_stride=(4, 4),
+                num_classes=audio_cfg.class_num,
+                embed_dim=128,
+                depths=[2, 2, 12, 2],
+                num_heads=[4, 8, 16, 32],
+                window_size=8,
+                config=audio_cfg,
+                enable_fusion=enable_fusion,
+                fusion_type=fusion_type,
+            )
+        elif audio_cfg.model_name == "large":
+            model = HTSAT_Swin_Transformer(
+                spec_size=256,
+                patch_size=4,
+                patch_stride=(4, 4),
+                num_classes=audio_cfg.class_num,
+                embed_dim=256,
+                depths=[2, 2, 12, 2],
+                num_heads=[4, 8, 16, 32],
+                window_size=8,
+                config=audio_cfg,
+                enable_fusion=enable_fusion,
+                fusion_type=fusion_type,
+            )
+
+        return model
+    except:
+        raise RuntimeError(
+            f"Import Model for {audio_cfg.model_name} not found, or the audio cfg parameters are not enough."
+        )

audioldm2/clap/open_clip/loss.py

@@ -0,0 +1,397 @@
+import torch
+import torch.distributed.nn
+from torch import distributed as dist, nn as nn
+from torch.nn import functional as F
+import numpy as np
+from sklearn.metrics import average_precision_score, roc_auc_score, accuracy_score
+
+try:
+    import horovod.torch as hvd
+except ImportError:
+    hvd = None
+
+
+def gather_features(
+    audio_features,
+    text_features,
+    audio_features_mlp=None,
+    text_features_mlp=None,
+    local_loss=False,
+    gather_with_grad=False,
+    rank=0,
+    world_size=1,
+    use_horovod=False,
+    mlp_loss=False,
+):
+    if use_horovod:
+        assert hvd is not None, "Please install horovod"
+        if gather_with_grad:
+            all_audio_features = hvd.allgather(audio_features)
+            all_text_features = hvd.allgather(text_features)
+            if mlp_loss:
+                all_audio_features_mlp = hvd.allgather(audio_features_mlp)
+                all_text_features_mlp = hvd.allgather(text_features_mlp)
+        else:
+            with torch.no_grad():
+                all_audio_features = hvd.allgather(audio_features)
+                all_text_features = hvd.allgather(text_features)
+                if mlp_loss:
+                    all_audio_features_mlp = hvd.allgather(audio_features_mlp)
+                    all_text_features_mlp = hvd.allgather(text_features_mlp)
+            if not local_loss:
+                # ensure grads for local rank when all_* features don't have a gradient
+                gathered_audio_features = list(
+                    all_audio_features.chunk(world_size, dim=0)
+                )
+                gathered_text_features = list(
+                    all_text_features.chunk(world_size, dim=0)
+                )
+                gathered_audio_features[rank] = audio_features
+                gathered_text_features[rank] = text_features
+                all_audio_features = torch.cat(gathered_audio_features, dim=0)
+                all_text_features = torch.cat(gathered_text_features, dim=0)
+                if mlp_loss:
+                    gathered_audio_features_mlp = list(
+                        all_audio_features_mlp.chunk(world_size, dim=0)
+                    )
+                    gathered_text_features_mlp = list(
+                        all_text_features_mlp.chunk(world_size, dim=0)
+                    )
+                    gathered_audio_features_mlp[rank] = audio_features_mlp
+                    gathered_text_features_mlp[rank] = text_features_mlp
+                    all_audio_features_mlp = torch.cat(
+                        gathered_audio_features_mlp, dim=0
+                    )
+                    all_text_features_mlp = torch.cat(gathered_text_features_mlp, dim=0)
+    else:
+        # We gather tensors from all gpus
+        if gather_with_grad:
+            all_audio_features = torch.cat(
+                torch.distributed.nn.all_gather(audio_features), dim=0
+            )
+            all_text_features = torch.cat(
+                torch.distributed.nn.all_gather(text_features), dim=0
+            )
+            if mlp_loss:
+                all_audio_features_mlp = torch.cat(
+                    torch.distributed.nn.all_gather(audio_features_mlp), dim=0
+                )
+                all_text_features_mlp = torch.cat(
+                    torch.distributed.nn.all_gather(text_features_mlp), dim=0
+                )
+        else:
+            gathered_audio_features = [
+                torch.zeros_like(audio_features) for _ in range(world_size)
+            ]
+            gathered_text_features = [
+                torch.zeros_like(text_features) for _ in range(world_size)
+            ]
+            dist.all_gather(gathered_audio_features, audio_features)
+            dist.all_gather(gathered_text_features, text_features)
+            if mlp_loss:
+                gathered_audio_features_mlp = [
+                    torch.zeros_like(audio_features_mlp) for _ in range(world_size)
+                ]
+                gathered_text_features_mlp = [
+                    torch.zeros_like(text_features_mlp) for _ in range(world_size)
+                ]
+                dist.all_gather(gathered_audio_features_mlp, audio_features_mlp)
+                dist.all_gather(gathered_text_features_mlp, text_features_mlp)
+            if not local_loss:
+                # ensure grads for local rank when all_* features don't have a gradient
+                gathered_audio_features[rank] = audio_features
+                gathered_text_features[rank] = text_features
+                if mlp_loss:
+                    gathered_audio_features_mlp[rank] = audio_features_mlp
+                    gathered_text_features_mlp[rank] = text_features_mlp
+
+            all_audio_features = torch.cat(gathered_audio_features, dim=0)
+            all_text_features = torch.cat(gathered_text_features, dim=0)
+            if mlp_loss:
+                all_audio_features_mlp = torch.cat(gathered_audio_features_mlp, dim=0)
+                all_text_features_mlp = torch.cat(gathered_text_features_mlp, dim=0)
+    if mlp_loss:
+        return (
+            all_audio_features,
+            all_text_features,
+            all_audio_features_mlp,
+            all_text_features_mlp,
+        )
+    else:
+        return all_audio_features, all_text_features
+
+
+class ClipLoss(nn.Module):
+    def __init__(
+        self,
+        local_loss=False,
+        gather_with_grad=False,
+        cache_labels=False,
+        rank=0,
+        world_size=1,
+        use_horovod=False,
+        mlp_loss=False,
+        weight_loss_kappa=0,
+    ):
+        super().__init__()
+        self.local_loss = local_loss
+        self.gather_with_grad = gather_with_grad
+        self.cache_labels = cache_labels
+        self.rank = rank
+        self.world_size = world_size
+        self.use_horovod = use_horovod
+        self.mlp_loss = mlp_loss
+        self.weighted_loss = bool(weight_loss_kappa != 0)
+        self.weight_loss_kappa = weight_loss_kappa
+        # cache state
+        self.prev_num_logits = 0
+        self.labels = {}
+
+    def forward(
+        self,
+        audio_features,
+        text_features,
+        logit_scale_a,
+        logit_scale_t=None,
+        audio_features_mlp=None,
+        text_features_mlp=None,
+    ):
+        device = audio_features.device
+        if self.mlp_loss:
+            if self.world_size > 1:
+                (
+                    all_audio_features,
+                    all_text_features,
+                    all_audio_features_mlp,
+                    all_text_features_mlp,
+                ) = gather_features(
+                    audio_features=audio_features,
+                    text_features=text_features,
+                    audio_features_mlp=audio_features_mlp,
+                    text_features_mlp=text_features_mlp,
+                    local_loss=self.local_loss,
+                    gather_with_grad=self.gather_with_grad,
+                    rank=self.rank,
+                    world_size=self.world_size,
+                    use_horovod=self.use_horovod,
+                    mlp_loss=self.mlp_loss,
+                )
+                if self.local_loss:
+                    a_logits_per_audio = (
+                        logit_scale_a * audio_features @ all_text_features_mlp.T
+                    )
+                    a_logits_per_text = (
+                        logit_scale_a * text_features_mlp @ all_audio_features.T
+                    )
+                    t_logits_per_audio = (
+                        logit_scale_t * audio_features_mlp @ all_text_features.T
+                    )
+                    t_logits_per_text = (
+                        logit_scale_t * text_features @ all_audio_features_mlp.T
+                    )
+                else:
+                    a_logits_per_audio = (
+                        logit_scale_a * all_audio_features @ all_text_features_mlp.T
+                    )
+                    a_logits_per_text = a_logits_per_audio.T
+                    t_logits_per_audio = (
+                        logit_scale_t * all_audio_features_mlp @ all_text_features.T
+                    )
+                    t_logits_per_text = t_logits_per_audio.T
+            else:
+                a_logits_per_audio = (
+                    logit_scale_a * audio_features @ text_features_mlp.T
+                )
+                a_logits_per_text = logit_scale_a * text_features_mlp @ audio_features.T
+                t_logits_per_audio = (
+                    logit_scale_t * audio_features_mlp @ text_features.T
+                )
+                t_logits_per_text = logit_scale_t * text_features @ audio_features_mlp.T
+
+            # calculated ground-truth and cache if enabled
+            num_logits = a_logits_per_audio.shape[0]
+            if self.prev_num_logits != num_logits or device not in self.labels:
+                labels = torch.arange(num_logits, device=device, dtype=torch.long)
+                if self.world_size > 1 and self.local_loss:
+                    labels = labels + num_logits * self.rank
+                if self.cache_labels:
+                    self.labels[device] = labels
+                    self.prev_num_logits = num_logits
+            else:
+                labels = self.labels[device]
+
+            if not self.weighted_loss:
+                total_loss = (
+                    F.cross_entropy(a_logits_per_audio, labels)
+                    + F.cross_entropy(a_logits_per_text, labels)
+                    + F.cross_entropy(t_logits_per_audio, labels)
+                    + F.cross_entropy(t_logits_per_text, labels)
+                ) / 4
+            else:
+                audio_weight = (audio_features @ audio_features.T).detach()
+                audio_weight = (
+                    torch.exp(
+                        torch.sum(audio_weight, axis=1)
+                        / (self.weight_loss_kappa * len(audio_weight))
+                    )
+                ).detach()
+                text_weight = (text_features @ text_features.T).detach()
+                text_weight = (
+                    torch.exp(
+                        torch.sum(text_weight, axis=1)
+                        / (self.weight_loss_kappa * len(text_features))
+                    )
+                ).detach()
+                total_loss = (
+                    F.cross_entropy(a_logits_per_audio, labels, weight=audio_weight)
+                    + F.cross_entropy(a_logits_per_text, labels, weight=audio_weight)
+                    + F.cross_entropy(t_logits_per_audio, labels, weight=text_weight)
+                    + F.cross_entropy(t_logits_per_text, labels, weight=text_weight)
+                ) / 4
+        else:
+            if self.world_size > 1:
+                all_audio_features, all_text_features = gather_features(
+                    audio_features=audio_features,
+                    text_features=text_features,
+                    local_loss=self.local_loss,
+                    gather_with_grad=self.gather_with_grad,
+                    rank=self.rank,
+                    world_size=self.world_size,
+                    use_horovod=self.use_horovod,
+                    mlp_loss=self.mlp_loss,
+                )
+
+                if self.local_loss:
+                    logits_per_audio = (
+                        logit_scale_a * audio_features @ all_text_features.T
+                    )
+                    logits_per_text = (
+                        logit_scale_a * text_features @ all_audio_features.T
+                    )
+                else:
+                    logits_per_audio = (
+                        logit_scale_a * all_audio_features @ all_text_features.T
+                    )
+                    logits_per_text = logits_per_audio.T
+            else:
+                logits_per_audio = logit_scale_a * audio_features @ text_features.T
+                logits_per_text = logit_scale_a * text_features @ audio_features.T
+
+            # calculated ground-truth and cache if enabled
+            num_logits = logits_per_audio.shape[0]
+            if self.prev_num_logits != num_logits or device not in self.labels:
+                labels = torch.arange(num_logits, device=device, dtype=torch.long)
+                if self.world_size > 1 and self.local_loss:
+                    labels = labels + num_logits * self.rank
+                if self.cache_labels:
+                    self.labels[device] = labels
+                    self.prev_num_logits = num_logits
+            else:
+                labels = self.labels[device]
+            if not self.weighted_loss:
+                total_loss = (
+                    F.cross_entropy(logits_per_audio, labels)
+                    + F.cross_entropy(logits_per_text, labels)
+                ) / 2
+            else:
+                audio_weight = (all_audio_features @ all_audio_features.T).detach()
+                audio_weight = (
+                    torch.exp(
+                        torch.sum(audio_weight, axis=1)
+                        / (self.weight_loss_kappa * len(all_audio_features))
+                    )
+                ).detach()
+                text_weight = (all_text_features @ all_text_features.T).detach()
+                text_weight = (
+                    torch.exp(
+                        torch.sum(text_weight, axis=1)
+                        / (self.weight_loss_kappa * len(all_text_features))
+                    )
+                ).detach()
+                total_loss = (
+                    F.cross_entropy(logits_per_audio, labels, weight=text_weight)
+                    + F.cross_entropy(logits_per_text, labels, weight=audio_weight)
+                ) / 2
+        return total_loss
+
+
+def lp_gather_features(pred, target, world_size=1, use_horovod=False):
+    if use_horovod:
+        assert hvd is not None, "Please install horovod"
+        with torch.no_grad():
+            all_preds = hvd.allgather(pred)
+            all_targets = hvd.allgath(target)
+    else:
+        gathered_preds = [torch.zeros_like(pred) for _ in range(world_size)]
+        gathered_targets = [torch.zeros_like(target) for _ in range(world_size)]
+
+        dist.all_gather(gathered_preds, pred)
+        dist.all_gather(gathered_targets, target)
+        all_preds = torch.cat(gathered_preds, dim=0)
+        all_targets = torch.cat(gathered_targets, dim=0)
+
+    return all_preds, all_targets
+
+
+def get_map(pred, target):
+    pred = torch.sigmoid(pred).numpy()
+    target = target.numpy()
+    return np.mean(average_precision_score(target, pred, average=None))
+
+
+def get_acc(pred, target):
+    pred = torch.argmax(pred, 1).numpy()
+    target = torch.argmax(target, 1).numpy()
+    return accuracy_score(target, pred)
+
+
+def get_mauc(pred, target):
+    pred = torch.sigmoid(pred).numpy()
+    target = target.numpy()
+    return np.mean(roc_auc_score(target, pred, average=None))
+
+
+class LPMetrics(object):
+    def __init__(self, metric_names=["map", "acc", "mauc"]):
+        self.metrics = []
+        for name in metric_names:
+            self.metrics.append(self.get_metric(name))
+        self.metric_names = metric_names
+
+    def get_metric(self, name):
+        if name == "map":
+            return get_map
+        elif name == "acc":
+            return get_acc
+        elif name == "mauc":
+            return get_mauc
+        else:
+            raise ValueError(f"the metric should be at least one of [map, acc, mauc]")
+
+    def evaluate_mertics(self, pred, target):
+        metric_dict = {}
+        for i in range(len(self.metric_names)):
+            metric_dict[self.metric_names[i]] = self.metrics[i](pred, target)
+        return metric_dict
+
+
+def calc_celoss(pred, target):
+    target = torch.argmax(target, 1).long()
+    return nn.CrossEntropyLoss()(pred, target)
+
+
+class LPLoss(nn.Module):
+    def __init__(self, loss_name):
+        super().__init__()
+        if loss_name == "bce":
+            self.loss_func = nn.BCEWithLogitsLoss()
+        elif loss_name == "ce":
+            self.loss_func = calc_celoss
+        elif loss_name == "mse":
+            self.loss_func = nn.MSELoss()
+        else:
+            raise ValueError(f"the loss func should be at least one of [bce, ce, mse]")
+
+    def forward(self, pred, target):
+        loss = self.loss_func(pred, target)
+        return loss

audioldm2/clap/open_clip/model.py

@@ -0,0 +1,931 @@
+""" CLAP Model
+
+Adapted from CLIP: https://github.com/openai/CLIP. Originally MIT License, Copyright (c) 2021 OpenAI.
+Adapted to the Audio Task.
+"""
+
+from collections import OrderedDict
+from dataclasses import dataclass
+from typing import Tuple, Union, Callable, Optional
+
+import numpy as np
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+import logging
+from .utils import freeze_batch_norm_2d
+
+from .pann_model import create_pann_model
+from .htsat import create_htsat_model
+from transformers import BertModel, RobertaModel, BartModel, RobertaConfig
+
+
+class MLPLayers(nn.Module):
+    def __init__(self, units=[512, 512, 512], nonlin=nn.ReLU(), dropout=0.1):
+        super(MLPLayers, self).__init__()
+        self.nonlin = nonlin
+        self.dropout = dropout
+
+        sequence = []
+        for u0, u1 in zip(units[:-1], units[1:]):
+            sequence.append(nn.Linear(u0, u1))
+            sequence.append(self.nonlin)
+            sequence.append(nn.Dropout(self.dropout))
+        sequence = sequence[:-2]
+
+        self.sequential = nn.Sequential(*sequence)
+
+    def forward(self, X):
+        X = self.sequential(X)
+        return X
+
+
+class Bottleneck(nn.Module):
+    expansion = 4
+
+    def __init__(self, inplanes, planes, stride=1):
+        super().__init__()
+
+        # all conv layers have stride 1. an avgpool is performed after the second convolution when stride > 1
+        self.conv1 = nn.Conv2d(inplanes, planes, 1, bias=False)
+        self.bn1 = nn.BatchNorm2d(planes)
+
+        self.conv2 = nn.Conv2d(planes, planes, 3, padding=1, bias=False)
+        self.bn2 = nn.BatchNorm2d(planes)
+
+        self.avgpool = nn.AvgPool2d(stride) if stride > 1 else nn.Identity()
+
+        self.conv3 = nn.Conv2d(planes, planes * self.expansion, 1, bias=False)
+        self.bn3 = nn.BatchNorm2d(planes * self.expansion)
+
+        self.relu = nn.ReLU(inplace=True)
+        self.downsample = None
+        self.stride = stride
+
+        if stride > 1 or inplanes != planes * Bottleneck.expansion:
+            # downsampling layer is prepended with an avgpool, and the subsequent convolution has stride 1
+            self.downsample = nn.Sequential(
+                OrderedDict(
+                    [
+                        ("-1", nn.AvgPool2d(stride)),
+                        (
+                            "0",
+                            nn.Conv2d(
+                                inplanes,
+                                planes * self.expansion,
+                                1,
+                                stride=1,
+                                bias=False,
+                            ),
+                        ),
+                        ("1", nn.BatchNorm2d(planes * self.expansion)),
+                    ]
+                )
+            )
+
+    def forward(self, x: torch.Tensor):
+        identity = x
+
+        out = self.relu(self.bn1(self.conv1(x)))
+        out = self.relu(self.bn2(self.conv2(out)))
+        out = self.avgpool(out)
+        out = self.bn3(self.conv3(out))
+
+        if self.downsample is not None:
+            identity = self.downsample(x)
+
+        out += identity
+        out = self.relu(out)
+        return out
+
+
+class AttentionPool2d(nn.Module):
+    def __init__(
+        self, spacial_dim: int, embed_dim: int, num_heads: int, output_dim: int = None
+    ):
+        super().__init__()
+        self.positional_embedding = nn.Parameter(
+            torch.randn(spacial_dim**2 + 1, embed_dim) / embed_dim**0.5
+        )
+        self.k_proj = nn.Linear(embed_dim, embed_dim)
+        self.q_proj = nn.Linear(embed_dim, embed_dim)
+        self.v_proj = nn.Linear(embed_dim, embed_dim)
+        self.c_proj = nn.Linear(embed_dim, output_dim or embed_dim)
+        self.num_heads = num_heads
+
+    def forward(self, x):
+        x = x.reshape(x.shape[0], x.shape[1], x.shape[2] * x.shape[3]).permute(
+            2, 0, 1
+        )  # NCHW -> (HW)NC
+        x = torch.cat([x.mean(dim=0, keepdim=True), x], dim=0)  # (HW+1)NC
+        x = x + self.positional_embedding[:, None, :].to(x.dtype)  # (HW+1)NC
+        x, _ = F.multi_head_attention_forward(
+            query=x,
+            key=x,
+            value=x,
+            embed_dim_to_check=x.shape[-1],
+            num_heads=self.num_heads,
+            q_proj_weight=self.q_proj.weight,
+            k_proj_weight=self.k_proj.weight,
+            v_proj_weight=self.v_proj.weight,
+            in_proj_weight=None,
+            in_proj_bias=torch.cat(
+                [self.q_proj.bias, self.k_proj.bias, self.v_proj.bias]
+            ),
+            bias_k=None,
+            bias_v=None,
+            add_zero_attn=False,
+            dropout_p=0,
+            out_proj_weight=self.c_proj.weight,
+            out_proj_bias=self.c_proj.bias,
+            use_separate_proj_weight=True,
+            training=self.training,
+            need_weights=False,
+        )
+
+        return x[0]
+
+
+class ModifiedResNet(nn.Module):
+    """
+    A ResNet class that is similar to torchvision's but contains the following changes:
+    - There are now 3 "stem" convolutions as opposed to 1, with an average pool instead of a max pool.
+    - Performs anti-aliasing strided convolutions, where an avgpool is prepended to convolutions with stride > 1
+    - The final pooling layer is a QKV attention instead of an average pool
+    """
+
+    def __init__(self, layers, output_dim, heads, image_size=224, width=64):
+        super().__init__()
+        self.output_dim = output_dim
+        self.image_size = image_size
+
+        # the 3-layer stem
+        self.conv1 = nn.Conv2d(
+            3, width // 2, kernel_size=3, stride=2, padding=1, bias=False
+        )
+        self.bn1 = nn.BatchNorm2d(width // 2)
+        self.conv2 = nn.Conv2d(
+            width // 2, width // 2, kernel_size=3, padding=1, bias=False
+        )
+        self.bn2 = nn.BatchNorm2d(width // 2)
+        self.conv3 = nn.Conv2d(width // 2, width, kernel_size=3, padding=1, bias=False)
+        self.bn3 = nn.BatchNorm2d(width)
+        self.avgpool = nn.AvgPool2d(2)
+        self.relu = nn.ReLU(inplace=True)
+
+        # residual layers
+        self._inplanes = width  # this is a *mutable* variable used during construction
+        self.layer1 = self._make_layer(width, layers[0])
+        self.layer2 = self._make_layer(width * 2, layers[1], stride=2)
+        self.layer3 = self._make_layer(width * 4, layers[2], stride=2)
+        self.layer4 = self._make_layer(width * 8, layers[3], stride=2)
+
+        embed_dim = width * 32  # the ResNet feature dimension
+        self.attnpool = AttentionPool2d(image_size // 32, embed_dim, heads, output_dim)
+
+        self.init_parameters()
+
+    def _make_layer(self, planes, blocks, stride=1):
+        layers = [Bottleneck(self._inplanes, planes, stride)]
+
+        self._inplanes = planes * Bottleneck.expansion
+        for _ in range(1, blocks):
+            layers.append(Bottleneck(self._inplanes, planes))
+
+        return nn.Sequential(*layers)
+
+    def init_parameters(self):
+        if self.attnpool is not None:
+            std = self.attnpool.c_proj.in_features**-0.5
+            nn.init.normal_(self.attnpool.q_proj.weight, std=std)
+            nn.init.normal_(self.attnpool.k_proj.weight, std=std)
+            nn.init.normal_(self.attnpool.v_proj.weight, std=std)
+            nn.init.normal_(self.attnpool.c_proj.weight, std=std)
+
+        for resnet_block in [self.layer1, self.layer2, self.layer3, self.layer4]:
+            for name, param in resnet_block.named_parameters():
+                if name.endswith("bn3.weight"):
+                    nn.init.zeros_(param)
+
+    def lock(self, unlocked_groups=0, freeze_bn_stats=False):
+        assert (
+            unlocked_groups == 0
+        ), "partial locking not currently supported for this model"
+        for param in self.parameters():
+            param.requires_grad = False
+        if freeze_bn_stats:
+            freeze_batch_norm_2d(self)
+
+    def stem(self, x):
+        for conv, bn in [
+            (self.conv1, self.bn1),
+            (self.conv2, self.bn2),
+            (self.conv3, self.bn3),
+        ]:
+            x = self.relu(bn(conv(x)))
+        x = self.avgpool(x)
+        return x
+
+    def forward(self, x):
+        x = self.stem(x)
+        x = self.layer1(x)
+        x = self.layer2(x)
+        x = self.layer3(x)
+        x = self.layer4(x)
+        x = self.attnpool(x)
+
+        return x
+
+
+class LayerNorm(nn.LayerNorm):
+    """Subclass torch's LayerNorm to handle fp16."""
+
+    def forward(self, x: torch.Tensor):
+        orig_type = x.dtype
+        x = F.layer_norm(x, self.normalized_shape, self.weight, self.bias, self.eps)
+        return x.to(orig_type)
+
+
+class QuickGELU(nn.Module):
+    # NOTE This is slower than nn.GELU or nn.SiLU and uses more GPU memory
+    def forward(self, x: torch.Tensor):
+        return x * torch.sigmoid(1.702 * x)
+
+
+class ResidualAttentionBlock(nn.Module):
+    def __init__(self, d_model: int, n_head: int, act_layer: Callable = nn.GELU):
+        super().__init__()
+
+        self.attn = nn.MultiheadAttention(d_model, n_head)
+        self.ln_1 = LayerNorm(d_model)
+        self.mlp = nn.Sequential(
+            OrderedDict(
+                [
+                    ("c_fc", nn.Linear(d_model, d_model * 4)),
+                    ("gelu", act_layer()),
+                    ("c_proj", nn.Linear(d_model * 4, d_model)),
+                ]
+            )
+        )
+        self.ln_2 = LayerNorm(d_model)
+
+    def attention(self, x: torch.Tensor, attn_mask: Optional[torch.Tensor] = None):
+        return self.attn(x, x, x, need_weights=False, attn_mask=attn_mask)[0]
+
+    def forward(self, x: torch.Tensor, attn_mask: Optional[torch.Tensor] = None):
+        x = x + self.attention(self.ln_1(x), attn_mask=attn_mask)
+        x = x + self.mlp(self.ln_2(x))
+        return x
+
+
+class Transformer(nn.Module):
+    def __init__(
+        self, width: int, layers: int, heads: int, act_layer: Callable = nn.GELU
+    ):
+        super().__init__()
+        self.width = width
+        self.layers = layers
+        self.resblocks = nn.ModuleList(
+            [
+                ResidualAttentionBlock(width, heads, act_layer=act_layer)
+                for _ in range(layers)
+            ]
+        )
+
+    def forward(self, x: torch.Tensor, attn_mask: Optional[torch.Tensor] = None):
+        for r in self.resblocks:
+            x = r(x, attn_mask=attn_mask)
+        return x
+
+
+class VisualTransformer(nn.Module):
+    def __init__(
+        self,
+        image_size: int,
+        patch_size: int,
+        width: int,
+        layers: int,
+        heads: int,
+        output_dim: int,
+        act_layer: Callable = nn.GELU,
+    ):
+        super().__init__()
+        self.image_size = image_size
+        self.output_dim = output_dim
+        self.conv1 = nn.Conv2d(
+            in_channels=3,
+            out_channels=width,
+            kernel_size=patch_size,
+            stride=patch_size,
+            bias=False,
+        )
+
+        scale = width**-0.5
+        self.class_embedding = nn.Parameter(scale * torch.randn(width))
+        self.positional_embedding = nn.Parameter(
+            scale * torch.randn((image_size // patch_size) ** 2 + 1, width)
+        )
+        self.ln_pre = LayerNorm(width)
+
+        self.text_branch = Transformer(width, layers, heads, act_layer=act_layer)
+
+        self.ln_post = LayerNorm(width)
+        self.proj = nn.Parameter(scale * torch.randn(width, output_dim))
+
+    def lock(self, unlocked_groups=0, freeze_bn_stats=False):
+        assert (
+            unlocked_groups == 0
+        ), "partial locking not currently supported for this model"
+        for param in self.parameters():
+            param.requires_grad = False
+
+    def forward(self, x: torch.Tensor):
+        x = self.conv1(x)  # shape = [*, width, grid, grid]
+        x = x.reshape(x.shape[0], x.shape[1], -1)  # shape = [*, width, grid ** 2]
+        x = x.permute(0, 2, 1)  # shape = [*, grid ** 2, width]
+        x = torch.cat(
+            [
+                self.class_embedding.to(x.dtype)
+                + torch.zeros(
+                    x.shape[0], 1, x.shape[-1], dtype=x.dtype, device=x.device
+                ),
+                x,
+            ],
+            dim=1,
+        )  # shape = [*, grid ** 2 + 1, width]
+        x = x + self.positional_embedding.to(x.dtype)
+        x = self.ln_pre(x)
+
+        x = x.permute(1, 0, 2)  # NLD -> LND
+        x = self.text_branch(x)
+        x = x.permute(1, 0, 2)  # LND -> NLD
+
+        x = self.ln_post(x[:, 0, :])
+
+        if self.proj is not None:
+            x = x @ self.proj
+
+        return x
+
+
+@dataclass
+class CLAPVisionCfg:
+    layers: Union[Tuple[int, int, int, int], int] = 12
+    width: int = 768
+    patch_size: int = 16
+    image_size: Union[Tuple[int, int], int] = 224
+    timm_model_name: str = (
+        None  # a valid model name overrides layers, width, patch_size
+    )
+    timm_model_pretrained: bool = (
+        False  # use (imagenet) pretrained weights for named model
+    )
+    timm_pool: str = (
+        "avg"  # feature pooling for timm model ('abs_attn', 'rot_attn', 'avg', '')
+    )
+    timm_proj: str = (
+        "linear"  # linear projection for timm model output ('linear', 'mlp', '')
+    )
+
+
+# Audio Config Class
+@dataclass
+class CLAPAudioCfp:
+    model_type: str = "PANN"
+    model_name: str = "Cnn14"
+    sample_rate: int = 48000
+    # Param
+    audio_length: int = 1024
+    window_size: int = 1024
+    hop_size: int = 1024
+    fmin: int = 50
+    fmax: int = 14000
+    class_num: int = 527
+    mel_bins: int = 64
+    clip_samples: int = 480000
+
+
+@dataclass
+class CLAPTextCfg:
+    context_length: int
+    vocab_size: int
+    width: int
+    heads: int
+    layers: int
+    model_type: str
+
+
+class CLAP(nn.Module):
+    def __init__(
+        self,
+        embed_dim: int,
+        audio_cfg: CLAPAudioCfp,
+        text_cfg: CLAPTextCfg,
+        quick_gelu: bool = False,
+        enable_fusion: bool = False,
+        fusion_type: str = "None",
+        joint_embed_shape: int = 512,
+        mlp_act: str = "relu",
+    ):
+        super().__init__()
+        if isinstance(audio_cfg, dict):
+            audio_cfg = CLAPAudioCfp(**audio_cfg)
+        if isinstance(text_cfg, dict):
+            text_cfg = CLAPTextCfg(**text_cfg)
+
+        self.audio_cfg = audio_cfg
+        self.text_cfg = text_cfg
+        self.enable_fusion = enable_fusion
+        self.fusion_type = fusion_type
+        self.joint_embed_shape = joint_embed_shape
+        self.mlp_act = mlp_act
+
+        self.context_length = text_cfg.context_length
+
+        # OpenAI models are pretrained w/ QuickGELU but native nn.GELU is both faster and more
+        # memory efficient in recent PyTorch releases (>= 1.10).
+        # NOTE: timm models always use native GELU regardless of quick_gelu flag.
+        act_layer = QuickGELU if quick_gelu else nn.GELU
+
+        if mlp_act == "relu":
+            mlp_act_layer = nn.ReLU()
+        elif mlp_act == "gelu":
+            mlp_act_layer = nn.GELU()
+        else:
+            raise NotImplementedError
+
+        # audio branch
+        # audio branch parameters
+        if audio_cfg.model_type == "PANN":
+            self.audio_branch = create_pann_model(audio_cfg, enable_fusion, fusion_type)
+        elif audio_cfg.model_type == "HTSAT":
+            self.audio_branch = create_htsat_model(
+                audio_cfg, enable_fusion, fusion_type
+            )
+        else:
+            logging.error(f"Model config for {audio_cfg.model_type} not found")
+            raise RuntimeError(f"Model config for {audio_cfg.model_type} not found.")
+
+        # text branch
+        # text branch parameters
+        if text_cfg.model_type == "transformer":
+            self.text_branch = Transformer(
+                width=text_cfg.width,
+                layers=text_cfg.layers,
+                heads=text_cfg.heads,
+                act_layer=act_layer,
+            )
+            self.vocab_size = text_cfg.vocab_size
+            self.token_embedding = nn.Embedding(text_cfg.vocab_size, text_cfg.width)
+            self.positional_embedding = nn.Parameter(
+                torch.empty(self.context_length, text_cfg.width)
+            )
+            self.ln_final = LayerNorm(text_cfg.width)
+            self.text_transform = MLPLayers(
+                units=[
+                    self.joint_embed_shape,
+                    self.joint_embed_shape,
+                    self.joint_embed_shape,
+                ],
+                dropout=0.1,
+            )
+            self.text_projection = nn.Sequential(
+                nn.Linear(text_cfg.width, self.joint_embed_shape),
+                mlp_act_layer,
+                nn.Linear(self.joint_embed_shape, self.joint_embed_shape),
+            )
+        elif text_cfg.model_type == "bert":
+            self.text_branch = BertModel.from_pretrained("bert-base-uncased")
+            self.text_transform = MLPLayers(
+                units=[
+                    self.joint_embed_shape,
+                    self.joint_embed_shape,
+                    self.joint_embed_shape,
+                ],
+                dropout=0.1,
+            )
+            self.text_projection = nn.Sequential(
+                nn.Linear(768, self.joint_embed_shape),
+                mlp_act_layer,
+                nn.Linear(self.joint_embed_shape, self.joint_embed_shape),
+            )
+        elif text_cfg.model_type == "roberta":
+            self.text_branch = RobertaModel(
+                RobertaConfig.from_pretrained("roberta-base")
+            )
+            self.text_transform = MLPLayers(
+                units=[
+                    self.joint_embed_shape,
+                    self.joint_embed_shape,
+                    self.joint_embed_shape,
+                ],
+                dropout=0.1,
+            )
+            self.text_projection = nn.Sequential(
+                nn.Linear(768, self.joint_embed_shape),
+                mlp_act_layer,
+                nn.Linear(self.joint_embed_shape, self.joint_embed_shape),
+            )
+        elif text_cfg.model_type == "bart":
+            self.text_branch = BartModel.from_pretrained("facebook/bart-base")
+            self.text_transform = MLPLayers(
+                units=[
+                    self.joint_embed_shape,
+                    self.joint_embed_shape,
+                    self.joint_embed_shape,
+                ],
+                dropout=0.1,
+            )
+            self.text_projection = nn.Sequential(
+                nn.Linear(768, self.joint_embed_shape),
+                mlp_act_layer,
+                nn.Linear(self.joint_embed_shape, self.joint_embed_shape),
+            )
+        else:
+            logging.error(f"Model config for {text_cfg.model_type} not found")
+            raise RuntimeError(f"Model config for {text_cfg.model_type} not found.")
+        self.text_branch_type = text_cfg.model_type
+        # text branch parameters
+
+        # audio branch parameters
+        self.audio_transform = MLPLayers(
+            units=[
+                self.joint_embed_shape,
+                self.joint_embed_shape,
+                self.joint_embed_shape,
+            ],
+            dropout=0.1,
+        )
+
+        # below here is text branch parameters
+
+        # ============================================================================================================
+        self.audio_projection = nn.Sequential(
+            nn.Linear(embed_dim, self.joint_embed_shape),
+            mlp_act_layer,
+            nn.Linear(self.joint_embed_shape, self.joint_embed_shape),
+        )
+
+        self.logit_scale_a = nn.Parameter(torch.ones([]) * np.log(1 / 0.07))
+        self.logit_scale_t = nn.Parameter(torch.ones([]) * np.log(1 / 0.07))
+        self.register_buffer("attn_mask", self.build_attention_mask(), persistent=False)
+
+        self.init_text_branch_parameters()
+
+    def init_text_branch_parameters(self):
+        if self.text_branch_type == "transformer":
+            nn.init.normal_(self.token_embedding.weight, std=0.02)
+            nn.init.normal_(self.positional_embedding, std=0.01)
+            proj_std = (self.text_branch.width**-0.5) * (
+                (2 * self.text_branch.layers) ** -0.5
+            )
+            attn_std = self.text_branch.width**-0.5
+            fc_std = (2 * self.text_branch.width) ** -0.5
+            for block in self.text_branch.resblocks:
+                nn.init.normal_(block.attn.in_proj_weight, std=attn_std)
+                nn.init.normal_(block.attn.out_proj.weight, std=proj_std)
+                nn.init.normal_(block.mlp.c_fc.weight, std=fc_std)
+                nn.init.normal_(block.mlp.c_proj.weight, std=proj_std)
+        if self.text_branch_type == "bert" or self.text_branch_type == "roberta":
+            self.text_branch.embeddings.word_embeddings.weight.shape[-1]
+        elif self.text_branch_type == "bart":
+            self.text_branch.shared.weight.shape[-1]
+        else:
+            self.text_branch.width
+        nn.init.constant_(self.logit_scale_a, np.log(1 / 0.07))
+        nn.init.constant_(self.logit_scale_t, np.log(1 / 0.07))
+
+        # deprecated
+        # if hasattr(self.visual, 'init_parameters'):
+        # self.visual.init_parameters()
+
+        # if self.text_projection is not None:
+        #     nn.init.normal_(self.text_projection, std=width**-0.5)
+
+    def build_attention_mask(self):
+        # lazily create causal attention mask, with full attention between the vision tokens
+        # pytorch uses additive attention mask; fill with -inf
+        mask = torch.empty(self.context_length, self.context_length)
+        mask.fill_(float("-inf"))
+        mask.triu_(1)  # zero out the lower diagonal
+        return mask
+
+    def encode_audio(self, audio, device):
+        return self.audio_branch(
+            audio, mixup_lambda=None, device=device
+        )  # mix lambda needs to add
+
+    # def list_of_dict_of_tensor2dict_of_tensor(self, x, device):
+    #     tmp = {}
+    #     for k in x[0].keys():
+    #         tmp[k] = []
+    #         for i in range(len(x)):
+    #             tmp[k].append(x[i][k][:77])
+    #     for k in x[0].keys():
+    #         tmp[k] = torch.tensor(tmp[k]).to(device=device, non_blocking=True)
+    #     return tmp
+
+    def encode_text(self, text, device):
+        if self.text_branch_type == "transformer":
+            text = text.to(device=device, non_blocking=True)
+            x = self.token_embedding(text)  # [batch_size, n_ctx, d_model]
+
+            x = x + self.positional_embedding
+            x = x.permute(1, 0, 2)  # NLD -> LND
+            x = self.text_branch(x, attn_mask=self.attn_mask)
+            x = x.permute(1, 0, 2)  # LND -> NLD
+            x = self.ln_final(x)
+
+            # x.shape = [batch_size, n_ctx, transformer.width]
+            # take features from the eot embedding (eot_token is the highest number in each sequence)
+            x = self.text_projection(x[torch.arange(x.shape[0]), text.argmax(dim=-1)])
+        elif self.text_branch_type == "bert":
+            # text = self.list_of_dict_of_tensor2dict_of_tensor(text, device)
+            # text = BatchEncoding(text)
+            x = self.text_branch(
+                input_ids=text["input_ids"].to(device=device, non_blocking=True),
+                attention_mask=text["attention_mask"].to(
+                    device=device, non_blocking=True
+                ),
+                token_type_ids=text["token_type_ids"].to(
+                    device=device, non_blocking=True
+                ),
+            )["pooler_output"]
+            x = self.text_projection(x)
+        elif self.text_branch_type == "roberta":
+            x = self.text_branch(
+                input_ids=text["input_ids"].to(device=device, non_blocking=True),
+                attention_mask=text["attention_mask"].to(
+                    device=device, non_blocking=True
+                ),
+            )["pooler_output"]
+            x = self.text_projection(x)
+        elif self.text_branch_type == "bart":
+            x = torch.mean(
+                self.text_branch(
+                    input_ids=text["input_ids"].to(device=device, non_blocking=True),
+                    attention_mask=text["attention_mask"].to(
+                        device=device, non_blocking=True
+                    ),
+                )["encoder_last_hidden_state"],
+                axis=1,
+            )
+            x = self.text_projection(x)
+        else:
+            logging.error(f"Model type {self.text_branch_type} not found")
+            raise RuntimeError(f"Model type {self.text_branch_type} not found.")
+        return x
+
+    def forward(self, audio, text, device=None):
+        """Forward audio and text into the CLAP
+
+        Parameters
+        ----------
+        audio: torch.Tensor (batch_size, audio_length)
+            the time-domain audio input / the batch of mel_spec and longer list.
+        text: torch.Tensor () // need to add
+            the text token input
+        """
+        if device is None:
+            if audio is not None:
+                device = audio.device
+            elif text is not None:
+                device = text.device
+        if audio is None and text is None:
+            # a hack to get the logit scale
+            return self.logit_scale_a.exp(), self.logit_scale_t.exp()
+        elif audio is None:
+            return self.encode_text(text, device=device)
+        elif text is None:
+            return self.audio_projection(
+                self.encode_audio(audio, device=device)["embedding"]
+            )
+        audio_features = self.audio_projection(
+            self.encode_audio(audio, device=device)["embedding"]
+        )
+        audio_features = F.normalize(audio_features, dim=-1)
+
+        text_features = self.encode_text(text, device=device)
+        # print("text_features", text_features)
+        # print("text_features.shape", text_features.shape)
+        # print("text_features.type", type(text_features))
+        text_features = F.normalize(text_features, dim=-1)
+
+        audio_features_mlp = self.audio_transform(audio_features)
+        text_features_mlp = self.text_transform(text_features)
+        # Four outputs: audio features (basic & MLP), text features (basic & MLP)
+        return (
+            audio_features,
+            text_features,
+            audio_features_mlp,
+            text_features_mlp,
+            self.logit_scale_a.exp(),
+            self.logit_scale_t.exp(),
+        )
+
+    def get_logit_scale(self):
+        return self.logit_scale_a.exp(), self.logit_scale_t.exp()
+
+    def get_text_embedding(self, data):
+        """Get the text embedding from the model
+
+        Parameters
+        ----------
+        data: torch.Tensor
+            a tensor of text embedding
+
+        Returns
+        ----------
+        text_embed: torch.Tensor
+            a tensor of text_embeds (N, D)
+
+        """
+        device = next(self.parameters()).device
+        for k in data:
+            data[k] = data[k].to(device)
+        text_embeds = self.encode_text(data, device=device)
+        text_embeds = F.normalize(text_embeds, dim=-1)
+
+        return text_embeds
+
+    def get_audio_embedding(self, data):
+        """Get the audio embedding from the model
+
+        Parameters
+        ----------
+        data: a list of dict
+            the audio input dict list from 'get_audio_feature' method
+
+        Returns
+        ----------
+        audio_embed: torch.Tensor
+            a tensor of audio_embeds (N, D)
+
+        """
+        device = next(self.parameters()).device
+        # input_dict = {}
+        # keys = data[0].keys()
+        # for k in keys:
+        #     input_dict[k] = torch.cat([d[k].unsqueeze(0) for d in data], dim=0).to(
+        #         device
+        #     )
+        audio_embeds = self.audio_projection(
+            self.encode_audio(data, device=device)["embedding"]
+        )
+        audio_embeds = F.normalize(audio_embeds, dim=-1)
+
+        return audio_embeds
+
+    def audio_infer(self, audio, hopsize=None, device=None):
+        """Forward one audio and produce the audio embedding
+
+        Parameters
+        ----------
+        audio:  (audio_length)
+            the time-domain audio input, notice that it must be only one input
+        hopsize: int
+            the overlap hopsize as the sliding window
+
+        Returns
+        ----------
+        output_dict: {
+            key: [n, (embedding_shape)] if "HTS-AT"
+            or
+            key: [(embedding_shape)] if "PANN"
+        }
+            the list of key values of the audio branch
+
+        """
+
+        assert not self.training, "the inference mode must be run at eval stage"
+        output_dict = {}
+        # PANN
+        if self.audio_cfg.model_type == "PANN":
+            audio_input = audio.unsqueeze(dim=0)
+            output_dict[key] = self.encode_audio(audio_input, device=device)[
+                key
+            ].squeeze(dim=0)
+        elif self.audio_cfg.model_type == "HTSAT":
+            # repeat
+            audio_len = len(audio)
+            k = self.audio_cfg.clip_samples // audio_len
+            if k > 1:
+                audio = audio.repeat(k)
+                audio_len = len(audio)
+
+            if hopsize is None:
+                hopsize = min(hopsize, audio_len)
+
+            if audio_len > self.audio_cfg.clip_samples:
+                audio_input = [
+                    audio[pos : pos + self.audio_cfg.clip_samples].clone()
+                    for pos in range(
+                        0, audio_len - self.audio_cfg.clip_samples, hopsize
+                    )
+                ]
+                audio_input.append(audio[-self.audio_cfg.clip_samples :].clone())
+                audio_input = torch.stack(audio_input)
+                output_dict[key] = self.encode_audio(audio_input, device=device)[key]
+            else:
+                audio_input = audio.unsqueeze(dim=0)
+                output_dict[key] = self.encode_audio(audio_input, device=device)[
+                    key
+                ].squeeze(dim=0)
+
+        return output_dict
+
+
+def convert_weights_to_fp16(model: nn.Module):
+    """Convert applicable model parameters to fp16"""
+
+    def _convert_weights_to_fp16(l):
+        if isinstance(l, (nn.Conv1d, nn.Conv2d, nn.Linear)):
+            l.weight.data = l.weight.data.half()
+            if l.bias is not None:
+                l.bias.data = l.bias.data.half()
+
+        if isinstance(l, nn.MultiheadAttention):
+            for attr in [
+                *[f"{s}_proj_weight" for s in ["in", "q", "k", "v"]],
+                "in_proj_bias",
+                "bias_k",
+                "bias_v",
+            ]:
+                tensor = getattr(l, attr)
+                if tensor is not None:
+                    tensor.data = tensor.data.half()
+
+        for name in ["text_projection", "proj"]:
+            if hasattr(l, name):
+                attr = getattr(l, name)
+                if attr is not None:
+                    attr.data = attr.data.half()
+
+    model.apply(_convert_weights_to_fp16)
+
+
+# Ignore the state dict of the vision part
+def build_model_from_openai_state_dict(
+    state_dict: dict, model_cfg, enable_fusion: bool = False, fusion_type: str = "None"
+):
+    embed_dim = model_cfg["embed_dim"]
+    audio_cfg = model_cfg["audio_cfg"]
+    text_cfg = model_cfg["text_cfg"]
+    state_dict["positional_embedding"].shape[0]
+    state_dict["token_embedding.weight"].shape[0]
+    transformer_width = state_dict["ln_final.weight"].shape[0]
+    transformer_width // 64
+    transformer_layers = len(
+        set(
+            k.split(".")[2]
+            for k in state_dict
+            if k.startswith(f"transformer.resblocks")
+        )
+    )
+
+    audio_cfg = CLAPAudioCfp(**audio_cfg)
+    text_cfg = CLAPTextCfg(**text_cfg)
+
+    model = CLAP(
+        embed_dim,
+        audio_cfg=audio_cfg,
+        text_cfg=text_cfg,
+        quick_gelu=True,  # OpenAI models were trained with QuickGELU
+        enable_fusion=enable_fusion,
+        fusion_type=fusion_type,
+    )
+    state_dict["logit_scale_a"] = state_dict["logit_scale"]
+    state_dict["logit_scale_t"] = state_dict["logit_scale"]
+    pop_keys = list(state_dict.keys())[::]
+    # pop the visual branch saved weights
+    for key in pop_keys:
+        if key.startswith("visual."):
+            state_dict.pop(key, None)
+
+    for key in ["logit_scale", "input_resolution", "context_length", "vocab_size"]:
+        state_dict.pop(key, None)
+
+    # not use fp16
+    # convert_weights_to_fp16(model)
+    model.load_state_dict(state_dict, strict=False)
+    return model.eval()
+
+
+def trace_model(model, batch_size=256, device=torch.device("cpu")):
+    model.eval()
+    audio_length = model.audio_cfg.audio_length
+    example_audio = torch.ones((batch_size, audio_length), device=device)
+    example_text = torch.zeros(
+        (batch_size, model.context_length), dtype=torch.int, device=device
+    )
+    model = torch.jit.trace_module(
+        model,
+        inputs=dict(
+            forward=(example_audio, example_text),
+            encode_text=(example_text,),
+            encode_image=(example_audio,),
+        ),
+    )
+    model.audio_cfg.audio_length = audio_length  # Question: what does this do?
+    return model

audioldm2/clap/open_clip/model_configs/HTSAT-base.json

@@ -0,0 +1,23 @@
+{
+    "embed_dim": 1024,
+    "audio_cfg": {
+        "audio_length": 1024,
+        "clip_samples": 480000,
+        "mel_bins": 64,
+        "sample_rate": 48000,
+        "window_size": 1024,
+        "hop_size": 480,
+        "fmin": 50,
+        "fmax": 14000,
+        "class_num": 527,
+        "model_type": "HTSAT",
+        "model_name": "base"
+    },
+    "text_cfg": {
+        "context_length": 77,
+        "vocab_size": 49408,
+        "width": 512,
+        "heads": 8,
+        "layers": 12
+    }
+}

audioldm2/clap/open_clip/model_configs/HTSAT-large.json

@@ -0,0 +1,23 @@
+{
+    "embed_dim": 2048,
+    "audio_cfg": {
+        "audio_length": 1024,
+        "clip_samples": 480000,
+        "mel_bins": 64,
+        "sample_rate": 48000,
+        "window_size": 1024,
+        "hop_size": 480,
+        "fmin": 50,
+        "fmax": 14000,
+        "class_num": 527,
+        "model_type": "HTSAT",
+        "model_name": "large"
+    },
+    "text_cfg": {
+        "context_length": 77,
+        "vocab_size": 49408,
+        "width": 512,
+        "heads": 8,
+        "layers": 12
+    }
+}

audioldm2/clap/open_clip/model_configs/HTSAT-tiny-win-1536.json

@@ -0,0 +1,23 @@
+{
+    "embed_dim": 768,
+    "audio_cfg": {
+        "audio_length": 1024,
+        "clip_samples": 480000,
+        "mel_bins": 64,
+        "sample_rate": 48000,
+        "window_size": 1536,
+        "hop_size": 480,
+        "fmin": 50,
+        "fmax": 14000,
+        "class_num": 527,
+        "model_type": "HTSAT",
+        "model_name": "tiny"
+    },
+    "text_cfg": {
+        "context_length": 77,
+        "vocab_size": 49408,
+        "width": 512,
+        "heads": 8,
+        "layers": 12
+    }
+}

audioldm2/clap/open_clip/model_configs/HTSAT-tiny.json

@@ -0,0 +1,23 @@
+{
+    "embed_dim": 768,
+    "audio_cfg": {
+        "audio_length": 1024,
+        "clip_samples": 480000,
+        "mel_bins": 64,
+        "sample_rate": 48000,
+        "window_size": 1024,
+        "hop_size": 480,
+        "fmin": 50,
+        "fmax": 14000,
+        "class_num": 527,
+        "model_type": "HTSAT",
+        "model_name": "tiny"
+    },
+    "text_cfg": {
+        "context_length": 77,
+        "vocab_size": 49408,
+        "width": 512,
+        "heads": 8,
+        "layers": 12
+    }
+}

audioldm2/clap/open_clip/model_configs/PANN-10.json

@@ -0,0 +1,23 @@
+{
+    "embed_dim": 1024,
+    "audio_cfg": {
+        "audio_length": 1024,
+        "clip_samples": 480000,
+        "mel_bins": 64,
+        "sample_rate": 48000,
+        "window_size": 1024,
+        "hop_size": 480,
+        "fmin": 50,
+        "fmax": 14000,
+        "class_num": 527,
+        "model_type": "PANN",
+        "model_name": "Cnn10"
+    },
+    "text_cfg": {
+        "context_length": 77,
+        "vocab_size": 49408,
+        "width": 512,
+        "heads": 8,
+        "layers": 12
+    }
+}

audioldm2/clap/open_clip/model_configs/PANN-14-fmax-18k.json

@@ -0,0 +1,23 @@
+{
+    "embed_dim": 2048,
+    "audio_cfg": {
+        "audio_length": 1024,
+        "clip_samples": 480000,
+        "mel_bins": 64,
+        "sample_rate": 48000,
+        "window_size": 1024,
+        "hop_size": 480,
+        "fmin": 50,
+        "fmax": 18000,
+        "class_num": 527,
+        "model_type": "PANN",
+        "model_name": "Cnn14"
+    },
+    "text_cfg": {
+        "context_length": 77,
+        "vocab_size": 49408,
+        "width": 512,
+        "heads": 8,
+        "layers": 12
+    }
+}

audioldm2/clap/open_clip/model_configs/PANN-14-fmax-8k-20s.json

@@ -0,0 +1,23 @@
+{
+    "embed_dim": 2048,
+    "audio_cfg": {
+        "audio_length": 1024,
+        "clip_samples": 960000,
+        "mel_bins": 64,
+        "sample_rate": 48000,
+        "window_size": 1024,
+        "hop_size": 360,
+        "fmin": 50,
+        "fmax": 8000,
+        "class_num": 527,
+        "model_type": "PANN",
+        "model_name": "Cnn14"
+    },
+    "text_cfg": {
+        "context_length": 77,
+        "vocab_size": 49408,
+        "width": 512,
+        "heads": 8,
+        "layers": 12
+    }
+}

audioldm2/clap/open_clip/model_configs/PANN-14-tiny-transformer.json

@@ -0,0 +1,23 @@
+{
+    "embed_dim": 2048,
+    "audio_cfg": {
+        "audio_length": 1024,
+        "clip_samples": 480000,
+        "mel_bins": 64,
+        "sample_rate": 48000,
+        "window_size": 1024,
+        "hop_size": 480,
+        "fmin": 50,
+        "fmax": 14000,
+        "class_num": 527,
+        "model_type": "PANN",
+        "model_name": "Cnn14"
+    },
+    "text_cfg": {
+        "context_length": 77,
+        "vocab_size": 49408,
+        "width": 512,
+        "heads": 8,
+        "layers": 4
+    }
+}

audioldm2/clap/open_clip/model_configs/PANN-14-win-1536.json

@@ -0,0 +1,23 @@
+{
+    "embed_dim": 2048,
+    "audio_cfg": {
+        "audio_length": 1024,
+        "clip_samples": 480000,
+        "mel_bins": 64,
+        "sample_rate": 48000,
+        "window_size": 1536,
+        "hop_size": 480,
+        "fmin": 50,
+        "fmax": 14000,
+        "class_num": 527,
+        "model_type": "PANN",
+        "model_name": "Cnn14"
+    },
+    "text_cfg": {
+        "context_length": 77,
+        "vocab_size": 49408,
+        "width": 512,
+        "heads": 8,
+        "layers": 12
+    }
+}

audioldm2/clap/open_clip/model_configs/PANN-14.json

@@ -0,0 +1,23 @@
+{
+    "embed_dim": 2048,
+    "audio_cfg": {
+        "audio_length": 1024,
+        "clip_samples": 480000,
+        "mel_bins": 64,
+        "sample_rate": 48000,
+        "window_size": 1024,
+        "hop_size": 480,
+        "fmin": 50,
+        "fmax": 14000,
+        "class_num": 527,
+        "model_type": "PANN",
+        "model_name": "Cnn14"
+    },
+    "text_cfg": {
+        "context_length": 77,
+        "vocab_size": 49408,
+        "width": 512,
+        "heads": 8,
+        "layers": 12
+    }
+}

audioldm2/clap/open_clip/model_configs/PANN-6.json

@@ -0,0 +1,23 @@
+{
+    "embed_dim": 512,
+    "audio_cfg": {
+        "audio_length": 1024,
+        "clip_samples": 480000,
+        "mel_bins": 64,
+        "sample_rate": 48000,
+        "window_size": 1024,
+        "hop_size": 480,
+        "fmin": 50,
+        "fmax": 14000,
+        "class_num": 527,
+        "model_type": "PANN",
+        "model_name": "Cnn6"
+    },
+    "text_cfg": {
+        "context_length": 77,
+        "vocab_size": 49408,
+        "width": 512,
+        "heads": 8,
+        "layers": 12
+    }
+}

audioldm2/clap/open_clip/model_configs/RN101-quickgelu.json

@@ -0,0 +1,22 @@
+{
+    "embed_dim": 512,
+    "quick_gelu": true,
+    "vision_cfg": {
+        "image_size": 224,
+        "layers": [
+            3,
+            4,
+            23,
+            3
+        ],
+        "width": 64,
+        "patch_size": null
+    },
+    "text_cfg": {
+        "context_length": 77,
+        "vocab_size": 49408,
+        "width": 512,
+        "heads": 8,
+        "layers": 12
+    }
+}