aroraaman's picture
Add all of `fourm`
3424266
raw
history blame contribute delete
No virus
48.5 kB
# Copyright 2024 EPFL and Apple Inc.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from typing import List, Tuple, Dict, Optional, Union, Any
from contextlib import nullcontext
import copy
import torch
import torch.nn as nn
import torch.nn.functional as F
from einops import rearrange, repeat
from diffusers.schedulers.scheduling_utils import SchedulerMixin
from diffusers import StableDiffusionPipeline
from huggingface_hub import PyTorchModelHubMixin
from fourm.vq.quantizers import VectorQuantizerLucid, Memcodes
import fourm.vq.models.vit_models as vit_models
import fourm.vq.models.unet.unet as unet
import fourm.vq.models.uvit as uvit
import fourm.vq.models.controlnet as controlnet
from fourm.vq.models.mlp_models import build_mlp
from fourm.vq.scheduling import DDPMScheduler, DDIMScheduler, PNDMScheduler, PipelineCond
from fourm.utils import denormalize
# If freeze_enc is True, the following modules will be frozen
FREEZE_MODULES = ['encoder', 'quant_proj', 'quantize', 'cls_emb']
class VQ(nn.Module, PyTorchModelHubMixin):
"""Base class for VQVAE and DiVAE models. Implements the encoder and quantizer, and can be used as such without a decoder
after training.
Args:
image_size: Input and target image size.
image_size_enc: Input image size for the encoder. Defaults to image_size. Change this when loading weights
from a tokenizer trained on a different image size.
n_channels: Number of input channels.
n_labels: Number of classes for semantic segmentation.
enc_type: String identifier specifying the encoder architecture. See vq/vit_models.py and vq/mlp_models.py
for available architectures.
patch_proj: Whether or not to use a ViT-style patch-wise linear projection in the encoder.
post_mlp: Whether or not to add a small point-wise MLP before the quantizer.
patch_size: Patch size for the encoder.
quant_type: String identifier specifying the quantizer implementation. Can be 'lucid', or 'memcodes'.
codebook_size: Number of codebook entries.
num_codebooks: Number of "parallel" codebooks to use. Only relevant for 'lucid' and 'memcodes' quantizers.
When using this, the tokens will be of shape B N_C H_Q W_Q, where N_C is the number of codebooks.
latent_dim: Dimensionality of the latent code. Can be small when using norm_codes=True,
see ViT-VQGAN (https://arxiv.org/abs/2110.04627) paper for details.
norm_codes: Whether or not to normalize the codebook entries to the unit sphere.
See ViT-VQGAN (https://arxiv.org/abs/2110.04627) paper for details.
norm_latents: Whether or not to normalize the latent codes to the unit sphere for computing commitment loss.
sync_codebook: Enable this when training on multiple GPUs, and disable for single GPUs, e.g. at inference.
ema_decay: Decay rate for the exponential moving average of the codebook entries.
threshold_ema_dead_code: Threshold for replacing stale codes that are used less than the
indicated exponential moving average of the codebook entries.
code_replacement_policy: Policy for replacing stale codes. Can be 'batch_random' or 'linde_buzo_gray'.
commitment_weight: Weight for the quantizer commitment loss.
kmeans_init: Whether or not to initialize the codebook entries with k-means clustering.
ckpt_path: Path to a checkpoint to load the model weights from.
ignore_keys: List of keys to ignore when loading the state_dict from the above checkpoint.
freeze_enc: Whether or not to freeze the encoder weights. See FREEZE_MODULES for the list of modules.
undo_std: Whether or not to undo any ImageNet standardization and transform the images to [-1,1]
before feeding the input to the encoder.
config: Dictionary containing the model configuration. Only used when loading
from Huggingface Hub. Ignore otherwise.
"""
def __init__(self,
image_size: int = 224,
image_size_enc: Optional[int] = None,
n_channels: str = 3,
n_labels: Optional[int] = None,
enc_type: str = 'vit_b_enc',
patch_proj: bool = True,
post_mlp: bool = False,
patch_size: int = 16,
quant_type: str = 'lucid',
codebook_size: Union[int, str] = 16384,
num_codebooks: int = 1,
latent_dim: int = 32,
norm_codes: bool = True,
norm_latents: bool = False,
sync_codebook: bool = True,
ema_decay: float = 0.99,
threshold_ema_dead_code: float = 0.25,
code_replacement_policy: str = 'batch_random',
commitment_weight: float = 1.0,
kmeans_init: bool = False,
ckpt_path: Optional[str] = None,
ignore_keys: List[str] = [
'decoder', 'loss',
'post_quant_conv', 'post_quant_proj',
'encoder.pos_emb',
],
freeze_enc: bool = False,
undo_std: bool = False,
config: Optional[Dict[str, Any]] = None,
**kwargs):
if config is not None:
config = copy.deepcopy(config)
self.__init__(**config)
return
super().__init__()
self.image_size = image_size
self.n_channels = n_channels
self.n_labels = n_labels
self.enc_type = enc_type
self.patch_proj = patch_proj
self.post_mlp = post_mlp
self.patch_size = patch_size
self.quant_type = quant_type
self.codebook_size = codebook_size
self.num_codebooks = num_codebooks
self.latent_dim = latent_dim
self.norm_codes = norm_codes
self.norm_latents = norm_latents
self.sync_codebook = sync_codebook
self.ema_decay = ema_decay
self.threshold_ema_dead_code = threshold_ema_dead_code
self.code_replacement_policy = code_replacement_policy
self.commitment_weight = commitment_weight
self.kmeans_init = kmeans_init
self.ckpt_path = ckpt_path
self.ignore_keys = ignore_keys
self.freeze_enc = freeze_enc
self.undo_std = undo_std
# For semantic segmentation
if n_labels is not None:
self.cls_emb = nn.Embedding(num_embeddings=n_labels, embedding_dim=n_channels)
self.colorize = torch.randn(3, n_labels, 1, 1)
else:
self.cls_emb = None
# Init encoder
image_size_enc = image_size_enc or image_size
if 'vit' in enc_type:
self.encoder = getattr(vit_models, enc_type)(
in_channels=n_channels, patch_size=patch_size,
resolution=image_size_enc, patch_proj=patch_proj, post_mlp=post_mlp
)
self.enc_dim = self.encoder.dim_tokens
elif 'MLP' in enc_type:
self.encoder = build_mlp(model_id=enc_type, dim_in=n_channels, dim_out=None)
self.enc_dim = self.encoder.dim_out
else:
raise NotImplementedError(f'{enc_type} not implemented.')
# Encoder -> quantizer projection
self.quant_proj = torch.nn.Conv2d(self.enc_dim, self.latent_dim, 1)
# Init quantizer
if quant_type == 'lucid':
self.quantize = VectorQuantizerLucid(
dim=latent_dim,
codebook_size=codebook_size,
codebook_dim=latent_dim,
heads=num_codebooks,
use_cosine_sim = norm_codes,
threshold_ema_dead_code = threshold_ema_dead_code,
code_replacement_policy=code_replacement_policy,
sync_codebook = sync_codebook,
decay = ema_decay,
commitment_weight=self.commitment_weight,
norm_latents = norm_latents,
kmeans_init=kmeans_init,
)
elif quant_type == 'memcodes':
self.quantize = Memcodes(
dim=latent_dim, codebook_size=codebook_size,
heads=num_codebooks, temperature=1.,
)
else:
raise ValueError(f'{quant_type} not a valid quant_type.')
# Load checkpoint
if ckpt_path is not None:
self.init_from_ckpt(ckpt_path, ignore_keys=ignore_keys)
# Freeze encoder
if freeze_enc:
for module_name, module in self.named_children():
if module_name not in FREEZE_MODULES:
continue
for param in module.parameters():
param.requires_grad = False
module.eval()
def train(self, mode: bool = True) -> 'VQ':
"""Override the default train() to set the training mode to all modules
except the encoder if freeze_enc is True.
Args:
mode: Whether to set the model to training mode (True) or evaluation mode (False).
"""
if not isinstance(mode, bool):
raise ValueError("training mode is expected to be boolean")
self.training = mode
for module_name, module in self.named_children():
if self.freeze_enc and module_name in FREEZE_MODULES:
continue
module.train(mode)
return self
def init_from_ckpt(self, path: str, ignore_keys: List[str] = list()) -> 'VQ':
"""Loads the state_dict from a checkpoint file and initializes the model with it.
Renames the keys in the state_dict if necessary (e.g. when loading VQ-GAN weights).
Args:
path: Path to the checkpoint file.
ignore_keys: List of keys to ignore when loading the state_dict.
Returns:
self
"""
ckpt = torch.load(path, map_location="cpu")
sd = ckpt['model'] if 'model' in ckpt else ckpt['state_dict']
# Compatibility with ViT-VQGAN weights
if 'quant_conv.0.weight' in sd and 'quant_conv.0.bias' in sd:
print("Renaming quant_conv.0 to quant_proj")
sd['quant_proj.weight'] = sd['quant_conv.0.weight']
sd['quant_proj.bias'] = sd['quant_conv.0.bias']
del sd['quant_conv.0.weight']
del sd['quant_conv.0.bias']
elif 'quant_conv.weight' in sd and 'quant_conv.bias' in sd:
print("Renaming quant_conv to quant_proj")
sd['quant_proj.weight'] = sd['quant_conv.weight']
sd['quant_proj.bias'] = sd['quant_conv.bias']
del sd['quant_conv.weight']
del sd['quant_conv.bias']
if 'post_quant_conv.0.weight' in sd and 'post_quant_conv.0.bias' in sd:
print("Renaming post_quant_conv.0 to post_quant_proj")
sd['post_quant_proj.weight'] = sd['post_quant_conv.0.weight']
sd['post_quant_proj.bias'] = sd['post_quant_conv.0.bias']
del sd['post_quant_conv.0.weight']
del sd['post_quant_conv.0.bias']
elif 'post_quant_conv.weight' in sd and 'post_quant_conv.bias' in sd:
print("Renaming post_quant_conv to post_quant_proj")
sd['post_quant_proj.weight'] = sd['post_quant_conv.weight']
sd['post_quant_proj.bias'] = sd['post_quant_conv.bias']
del sd['post_quant_conv.weight']
del sd['post_quant_conv.bias']
keys = list(sd.keys())
for k in keys:
for ik in ignore_keys:
if k.startswith(ik):
print("Deleting key {} from state_dict.".format(k))
del sd[k]
msg = self.load_state_dict(sd, strict=False)
print(msg)
print(f"Restored from {path}")
return self
def prepare_input(self, x: torch.Tensor) -> torch.Tensor:
"""Preprocesses the input image tensor before feeding it to the encoder.
If self.undo_std, the input is first denormalized from the ImageNet
standardization to [-1, 1]. If semantic segmentation is performed, the
class indices are embedded.
Args:
x: Input image tensor of shape B C H W
or B H W in case of semantic segmentation
Returns:
Preprocessed input tensor of shape B C H W
"""
if self.undo_std:
x = 2.0 * denormalize(x) - 1.0
if self.cls_emb is not None:
x = rearrange(self.cls_emb(x), 'b h w c -> b c h w')
return x
def to_rgb(self, x: torch.Tensor) -> torch.Tensor:
"""When semantic segmentation is performed, this function converts the
class embeddings to RGB.
Args:
x: Input tensor of shape B C H W
Returns:
RGB tensor of shape B C H W
"""
x = F.conv2d(x, weight=self.colorize)
x = (x-x.min())/(x.max()-x.min())
return x
def encode(self, x: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor, torch.LongTensor]:
"""Encodes an input image tensor and quantizes the latent code.
Args:
x: Input image tensor of shape B C H W
or B H W in case of semantic segmentation
Returns:
quant: Quantized latent code of shape B D_Q H_Q W_Q
code_loss: Codebook loss
tokens: Quantized indices of shape B H_Q W_Q
"""
x = self.prepare_input(x)
h = self.encoder(x)
h = self.quant_proj(h)
quant, code_loss, tokens = self.quantize(h)
return quant, code_loss, tokens
def tokenize(self, x: torch.Tensor) -> torch.LongTensor:
"""Tokenizes an input image tensor.
Args:
x: Input image tensor of shape B C H W
or B H W in case of semantic segmentation
Returns:
Quantized indices of shape B H_Q W_Q
"""
_, _, tokens = self.encode(x)
return tokens
def autoencode(self, x: torch.Tensor, **kwargs) -> torch.Tensor:
"""Autoencodes an input image tensor by encoding it, quantizing the latent code,
and decoding it back to an image.
Args:
x: Input image tensor of shape B C H W
or B H W in case of semantic segmentation
Returns:
Reconstructed image tensor of shape B C H W
"""
pass
def decode_quant(self, quant: torch.Tensor, **kwargs) -> torch.Tensor:
"""Decodes quantized latent codes back to an image.
Args:
quant: Quantized latent code of shape B D_Q H_Q W_Q
Returns:
Decoded image tensor of shape B C H W
"""
pass
def tokens_to_embedding(self, tokens: torch.LongTensor) -> torch.Tensor:
"""Look up the codebook entries corresponding the discrete tokens.
Args:
tokens: Quantized indices of shape B H_Q W_Q
Returns:
Quantized latent code of shape B D_Q H_Q W_Q
"""
return self.quantize.indices_to_embedding(tokens)
def decode_tokens(self, tokens: torch.LongTensor, **kwargs) -> torch.Tensor:
"""Decodes discrete tokens back to an image.
Args:
tokens: Quantized indices of shape B H_Q W_Q
Returns:
Decoded image tensor of shape B C H W
"""
quant = self.tokens_to_embedding(tokens)
dec = self.decode_quant(quant, **kwargs)
return dec
def forward(self, x: torch.Tensor, **kwargs) -> Tuple[torch.Tensor, torch.Tensor]:
"""Forward pass of the encoder and quantizer.
Args:
x: Input image tensor of shape B C H W
or B H W in case of semantic segmentation
Returns:
quant: Quantized latent code of shape B D_Q H_Q W_Q
code_loss: Codebook loss
"""
quant, code_loss, _ = self.encode(x)
return quant, code_loss
class VQVAE(VQ):
"""VQ-VAE model = simple encoder + decoder with a discrete bottleneck and
basic reconstruction loss (optionall with perceptual loss), i.e. no diffusion,
nor GAN discriminator.
Args:
dec_type: String identifier specifying the decoder architecture.
See vq/vit_models.py and vq/mlp_models.py for available architectures.
out_conv: Whether or not to add final conv layers to the ViT decoder.
image_size_dec: Image size for the decoder. Defaults to self.image_size.
Change this when loading weights from a tokenizer decoder trained on a
different image size.
patch_size_dec: Patch size for the decoder. Defaults to self.patch_size.
config: Dictionary containing the model configuration. Only used when loading
from Huggingface Hub. Ignore otherwise.
"""
def __init__(self,
dec_type: str = 'vit_b_dec',
out_conv: bool = False,
image_size_dec: int = None,
patch_size_dec: int = None,
config: Optional[Dict[str, Any]] = None,
*args,
**kwargs):
if config is not None:
config = copy.deepcopy(config)
self.__init__(**config)
return
# Don't want to load the weights just yet
self.original_ckpt_path = kwargs.get('ckpt_path', None)
kwargs['ckpt_path'] = None
super().__init__(*args, **kwargs)
self.ckpt_path = self.original_ckpt_path
# Init decoder
out_channels = self.n_channels if self.n_labels is None else self.n_labels
image_size_dec = image_size_dec or self.image_size
patch_size = patch_size_dec or self.patch_size
if 'vit' in dec_type:
self.decoder = getattr(vit_models, dec_type)(
out_channels=out_channels, patch_size=patch_size,
resolution=image_size_dec, out_conv=out_conv, post_mlp=self.post_mlp,
patch_proj=self.patch_proj
)
self.dec_dim = self.decoder.dim_tokens
elif 'MLP' in dec_type:
self.decoder = build_mlp(model_id=dec_type, dim_in=None, dim_out=out_channels)
self.dec_dim = self.decoder.dim_in
else:
raise NotImplementedError(f'{dec_type} not implemented.')
# Quantizer -> decoder projection
self.post_quant_proj = torch.nn.Conv2d(self.latent_dim, self.dec_dim, 1)
# Load checkpoint
if self.ckpt_path is not None:
self.init_from_ckpt(self.ckpt_path, ignore_keys=self.ignore_keys)
def decode_quant(self, quant: torch.Tensor, **kwargs) -> torch.Tensor:
"""Decodes quantized latent codes back to an image.
Args:
quant: Quantized latent code of shape B D_Q H_Q W_Q
Returns:
Decoded image tensor of shape B C H W
"""
quant = self.post_quant_proj(quant)
dec = self.decoder(quant)
return dec
def forward(self, x: torch.Tensor, **kwargs) -> Tuple[torch.Tensor, torch.Tensor]:
"""Forward pass of the encoder, quantizer, and decoder.
Args:
x: Input image tensor of shape B C H W
or B H W in case of semantic segmentation
Returns:
dec: Decoded image tensor of shape B C H W
code_loss: Codebook loss
"""
with torch.no_grad() if self.freeze_enc else nullcontext():
quant, code_loss, _ = self.encode(x)
dec = self.decode_quant(quant)
return dec, code_loss
def autoencode(self, x: torch.Tensor, **kwargs) -> torch.Tensor:
"""Autoencodes an input image tensor by encoding it, quantizing the
latent code, and decoding it back to an image.
Args:
x: Input image tensor of shape B C H W
or B H W in case of semantic segmentation
Returns:
Reconstructed image tensor of shape B C H W
"""
dec, _ = self.forward(x)
return dec
class DiVAE(VQ):
"""DiVAE ("Diffusion VQ-VAE") model = simple encoder + diffusion decoder with
a discrete bottleneck, inspired by https://arxiv.org/abs/2206.00386.
Args:
dec_type: String identifier specifying the decoder architecture.
See vq/models/unet/unet.py and vq/models/uvit.py for available architectures.
num_train_timesteps: Number of diffusion timesteps to use for training.
cls_free_guidance_dropout: Dropout probability for classifier-free guidance.
masked_cfg: Whether or not to randomly mask out conditioning tokens.
cls_free_guidance_dropout must be > 0.0 for this to have any effect, and
decides how often masking is performed. E.g. with 0.5, half of the time
the conditioning tokens will be randomly masked, and half the time they
will be kept as is.
masked_cfg_low: Lower bound of number of tokens to mask out.
masked_cfg_high: Upper bound of number of tokens to mask out (inclusive).
Defaults to total number of tokens (H_Q * W_Q) if it is set to None.
scheduler: String identifier specifying the diffusion scheduler to use.
Can be 'ddpm' or 'ddim'.
beta_schedule: String identifier specifying the beta schedule to use for
the diffusion process. Can be 'linear', 'squaredcos_cap_v2' (cosine),
'shifted_cosine:{shift_amount}'; see vq/scheduling for details.
prediction_type: String identifier specifying the type of prediction to use.
Can be 'sample', 'epsilon', or 'v_prediction'; see vq/scheduling for details.
clip_sample: Whether or not to clip the samples to [-1, 1], at inference only.
thresholding: Whether or not to use dynamic thresholding (introduced by Imagen,
https://arxiv.org/abs/2205.11487) for the diffusion process, at inference only.
conditioning: String identifier specifying the way to condition the diffusion
decoder. Can be 'concat' or 'xattn'. See models for details (only relevant to UViT).
dec_transformer_dropout: Dropout rate for the transformer layers in the
diffusion decoder (only relevant to UViT models).
zero_terminal_snr: Whether or not to enforce zero terminal SNR, i.e. the SNR
at the last timestep is set to zero. This is useful for preventing the model
from "cheating" by using information in the last timestep to reconstruct the image.
See https://arxiv.org/abs/2305.08891.
image_size_dec: Image size for the decoder. Defaults to image_size.
Change this when loading weights from a tokenizer decoder trained on a
different image size.
config: Dictionary containing the model configuration. Only used when loading
from Huggingface Hub. Ignore otherwise.
"""
def __init__(self,
dec_type: str = 'unet_patched',
num_train_timesteps: int = 1000,
cls_free_guidance_dropout: float = 0.0,
masked_cfg: bool = False,
masked_cfg_low: int = 0,
masked_cfg_high: Optional[int] = None,
scheduler: str = 'ddpm',
beta_schedule: str = 'squaredcos_cap_v2',
prediction_type: str = 'v_prediction',
clip_sample: bool = False,
thresholding: bool = True,
conditioning: str = 'concat',
dec_transformer_dropout: float = 0.2,
zero_terminal_snr: bool = True,
image_size_dec: Optional[int] = None,
config: Optional[Dict[str, Any]] = None,
*args, **kwargs):
if config is not None:
config = copy.deepcopy(config)
self.__init__(**config)
return
# Don't want to load the weights just yet
self.original_ckpt_path = kwargs.get('ckpt_path', None)
kwargs['ckpt_path'] = None
super().__init__(*args, **kwargs)
self.ckpt_path = self.original_ckpt_path
self.num_train_timesteps = num_train_timesteps
self.beta_schedule = beta_schedule
self.prediction_type = prediction_type
self.clip_sample = clip_sample
self.thresholding = thresholding
self.zero_terminal_snr = zero_terminal_snr
if cls_free_guidance_dropout > 0.0:
self.cfg_dist = torch.distributions.Bernoulli(probs=cls_free_guidance_dropout)
else:
self.cfg_dist = None
self.masked_cfg = masked_cfg
self.masked_cfg_low = masked_cfg_low
self.masked_cfg_high = masked_cfg_high
# Init diffusion decoder
image_size_dec = image_size_dec or self.image_size
if 'unet_' in dec_type:
self.decoder = getattr(unet, dec_type)(
in_channels=self.n_channels,
out_channels=self.n_channels,
cond_channels=self.latent_dim,
image_size=image_size_dec,
)
elif 'uvit_' in dec_type:
self.decoder = getattr(uvit, dec_type)(
sample_size=image_size_dec,
in_channels=self.n_channels,
out_channels=self.n_channels,
cond_dim=self.latent_dim,
cond_type=conditioning,
mid_drop_rate=dec_transformer_dropout,
)
else:
raise NotImplementedError(f'dec_type {dec_type} not implemented.')
# Init training diffusion scheduler / default pipeline for generation
scheduler_cls = DDPMScheduler if scheduler == 'ddpm' else DDIMScheduler
self.noise_scheduler = scheduler_cls(
num_train_timesteps=num_train_timesteps,
thresholding=thresholding,
clip_sample=clip_sample,
beta_schedule=beta_schedule,
prediction_type=prediction_type,
zero_terminal_snr=zero_terminal_snr,
)
self.pipeline = PipelineCond(model=self.decoder, scheduler=self.noise_scheduler)
# Load checkpoint
if self.ckpt_path is not None:
self.init_from_ckpt(self.ckpt_path, ignore_keys=self.ignore_keys)
def sample_mask(self, quant: torch.Tensor, low: int = 0, high: Optional[int] = None) -> torch.BoolTensor:
"""Returns a mask of shape B H_Q W_Q, where True = masked-out, False = keep.
Args:
quant: Dequantized latent tensor of shape B D_Q H_Q W_Q
low: Lower bound of number of tokens to mask out
high: Upper bound of number of tokens to mask out (inclusive).
Defaults to total number of tokens (H_Q * W_Q) if it is set to None.
Returns:
Boolean mask of shape B H_Q W_Q
"""
B, _, H_Q, W_Q = quant.shape
num_tokens = H_Q * W_Q
high = high if high is not None else num_tokens
zero_idxs = torch.randint(low=low, high=high+1, size=(B,), device=quant.device)
noise = torch.rand(B, num_tokens, device=quant.device)
ids_arange_shuffle = torch.argsort(noise, dim=1) # ascend: small is keep, large is remove
mask = torch.where(ids_arange_shuffle < zero_idxs.unsqueeze(1), 0, 1)
mask = rearrange(mask, 'b (h w) -> b h w', h=H_Q, w=W_Q).bool()
return mask
def _get_pipeline(self, scheduler: Optional[SchedulerMixin] = None) -> PipelineCond:
"""Creates a conditional diffusion pipeline with the given scheduler.
Args:
scheduler: Scheduler to use for the diffusion pipeline.
If None, the default scheduler will be used.
Returns:
Conditional diffusion pipeline.
"""
return PipelineCond(model=self.decoder, scheduler=scheduler) if scheduler is not None else self.pipeline
def decode_quant(self,
quant: torch.Tensor,
timesteps: Optional[int] = None,
scheduler: Optional[SchedulerMixin] = None,
generator: Optional[torch.Generator] = None,
image_size: Optional[Union[Tuple[int, int], int]] = None,
verbose: bool = False,
scheduler_timesteps_mode: str = 'trailing',
orig_res: Optional[Union[torch.LongTensor, Tuple[int, int]]] = None) -> torch.Tensor:
"""Decodes quantized latent codes back to an image.
Args:
quant: Quantized latent code of shape B D_Q H_Q W_Q
timesteps: Number of diffusion timesteps to use. Defaults to self.num_train_timesteps.
scheduler: Scheduler to use for the diffusion pipeline. Defaults to the training scheduler.
generator: Random number generator to use for sampling. By default generations are stochastic.
image_size: Image size to use for the diffusion pipeline. Defaults to decoder image size.
verbose: Whether or not to print progress bar.
scheduler_timesteps_mode: The mode to use for DDIMScheduler. One of `trailing`, `linspace`,
`leading`. See https://arxiv.org/abs/2305.08891 for more details.
orig_res: The original resolution of the image to condition the diffusion on. Ignored if None.
See SDXL https://arxiv.org/abs/2307.01952 for more details.
Returns:
Decoded image tensor of shape B C H W
"""
pipeline = self._get_pipeline(scheduler)
dec = pipeline(
quant, timesteps=timesteps, generator=generator, image_size=image_size,
verbose=verbose, scheduler_timesteps_mode=scheduler_timesteps_mode, orig_res=orig_res
)
return dec
def decode_tokens(self, tokens: torch.LongTensor, **kwargs) -> torch.Tensor:
"""See `decode_quant` for details on the optional args."""
return super().decode_tokens(tokens, **kwargs)
def autoencode(self,
input_clean: torch.Tensor,
timesteps: Optional[int] = None,
scheduler: Optional[SchedulerMixin] = None,
generator: Optional[torch.Generator] = None,
verbose: bool = True,
scheduler_timesteps_mode: str = 'trailing',
orig_res: Optional[Union[torch.LongTensor, Tuple[int, int]]] = None,
**kwargs) -> torch.Tensor:
"""Autoencodes an input image tensor by encoding it, quantizing the latent code,
and decoding it back to an image.
Args:
input_clean: Input image tensor of shape B C H W
or B H W in case of semantic segmentation
timesteps: Number of diffusion timesteps to use. Defaults to self.num_train_timesteps.
scheduler: Scheduler to use for the diffusion pipeline. Defaults to the training scheduler.
generator: Random number generator to use for sampling. By default generations are stochastic.
verbose: Whether or not to print progress bar.
scheduler_timesteps_mode: The mode to use for DDIMScheduler. One of `trailing`, `linspace`,
`leading`. See https://arxiv.org/abs/2305.08891 for more details.
orig_res: The original resolution of the image to condition the diffusion on. Ignored if None.
See SDXL https://arxiv.org/abs/2307.01952 for more details.
Returns:
Reconstructed image tensor of shape B C H W
"""
pipeline = self._get_pipeline(scheduler)
quant, _, _ = self.encode(input_clean)
image_size = input_clean.shape[-1]
dec = pipeline(
quant, timesteps=timesteps, generator=generator, image_size=image_size,
verbose=verbose, scheduler_timesteps_mode=scheduler_timesteps_mode, orig_res=orig_res
)
return dec
def forward(self,
input_clean: torch.Tensor,
input_noised: torch.Tensor,
timesteps: Union[torch.Tensor, float, int],
cond_mask: Optional[torch.Tensor] = None,
orig_res: Optional[Union[torch.LongTensor, Tuple[int, int]]] = None) -> Tuple[torch.Tensor, torch.Tensor]:
"""Forward pass of the encoder, quantizer, and decoder.
Args:
input_clean: Clean input image tensor of shape B C H W
or B H W in case of semantic segmentation. Used for encoding.
input_noised: Noised input image tensor of shape B C H W. Used as
input to the diffusion decoder.
timesteps: Timesteps for conditioning the diffusion decoder on.
cond_mask: Optional mask for the diffusion conditioning.
True = masked-out, False = keep.
orig_res: The original resolution of the image to condition the diffusion on. Ignored if None.
See SDXL https://arxiv.org/abs/2307.01952 for more details.
Returns:
dec: Decoded image tensor of shape B C H W
code_loss: Codebook loss
"""
with torch.no_grad() if self.freeze_enc else nullcontext():
quant, code_loss, _ = self.encode(input_clean)
if cond_mask is None and self.cfg_dist is not None and self.training:
# Create a random mask for each batch element. True = masked-out, False = keep
B, _, H_Q, W_Q = quant.shape
cond_mask = self.cfg_dist.sample((B,)).to(quant.device, dtype=torch.bool)
cond_mask = repeat(cond_mask, 'b -> b h w', h=H_Q, w=W_Q)
if self.masked_cfg:
mask = self.sample_mask(quant, low=self.masked_cfg_low, high=self.masked_cfg_high)
cond_mask = (mask * cond_mask)
dec = self.decoder(input_noised, timesteps, quant, cond_mask=cond_mask, orig_res=orig_res)
return dec, code_loss
class VQControlNet(VQ):
"""VQControlNet model = simple pertrained encoder + a ControlNet decoder conditioned on tokens.
Args:
sd_path: Path to the Stable Diffusion weights for training the ControlNet.
image_size_sd: Stable diffusion input image size. Defaults to image_size.
Change this to the image size that Stable Diffusion is trained on.
pretrained_cn: Whether to use pretrained Stable Diffusion weights for the control model.
cls_free_guidance_dropout: Dropout probability for classifier-free guidance.
masked_cfg: Whether or not to randomly mask out conditioning tokens.
cls_free_guidance_dropout must be > 0.0 for this to have any effect, and
decides how often masking is performed. E.g. with 0.5, half of the time
the conditioning tokens will be randomly masked, and half the time they
will be kept as is.
masked_cfg_low: Lower bound of number of tokens to mask out.
masked_cfg_high: Upper bound of number of tokens to mask out (inclusive).
Defaults to total number of tokens (H_Q * W_Q) if it is set to None.
enable_xformer: Enables xFormers.
adapter: Path to the adapter model weights. The adapter model is initialy trained to map
the tokens to a VAE latent-like representation. Then the output of the adapter model
is passed as the condition to train the ControlNet. By default there is no adapter usage.
config: Dictionary containing the model configuration. Only used when loading
from Huggingface Hub. Ignore otherwise.
"""
def __init__(self,
sd_path: str = "runwayml/stable-diffusion-v1-5",
image_size_sd: Optional[int] = None,
pretrained_cn: bool = False,
cls_free_guidance_dropout: float = 0.0,
masked_cfg: bool = False,
masked_cfg_low: int = 0,
masked_cfg_high: Optional[int] = None,
enable_xformer: bool = False,
adapter: Optional[str] = None,
config: Optional[Dict[str, Any]] = None,
*args, **kwargs):
if config is not None:
config = copy.deepcopy(config)
self.__init__(**config)
return
# Don't want to load the weights just yet
self.original_ckpt_path = kwargs.get('ckpt_path', None)
kwargs['ckpt_path'] = None
super().__init__(*args, **kwargs)
self.ckpt_path = self.original_ckpt_path
if cls_free_guidance_dropout > 0.0:
self.cfg_dist = torch.distributions.Bernoulli(probs=cls_free_guidance_dropout)
else:
self.cfg_dist = None
self.masked_cfg = masked_cfg
self.masked_cfg_low = masked_cfg_low
self.masked_cfg_high = masked_cfg_high
self.image_size_sd = self.image_size if image_size_sd is None else image_size_sd
sd_pipeline = StableDiffusionPipeline.from_pretrained(sd_path)
try:
import xformers
XFORMERS_AVAILABLE = True
except ImportError:
print("xFormers not available")
XFORMERS_AVAILABLE = False
enable_xformer = enable_xformer and XFORMERS_AVAILABLE
if enable_xformer:
print('Enabling xFormer for Stable Diffusion')
sd_pipeline.enable_xformers_memory_efficient_attention()
self.decoder = getattr(controlnet, 'controlnet')(
in_channels=4,
cond_channels=self.latent_dim,
sd_pipeline=sd_pipeline,
image_size=self.image_size_sd,
pretrained_cn=pretrained_cn,
enable_xformer=enable_xformer,
adapter=adapter,
)
# Use the defualt controlnet pipeline both for training and generation
self.noise_scheduler = PNDMScheduler(**sd_pipeline.scheduler.config)
self.vae = sd_pipeline.vae
self._freeze_vae()
self.pipeline = PipelineCond(model=self.decoder, scheduler=self.noise_scheduler)
# Load checkpoint
if self.ckpt_path is not None:
self.init_from_ckpt(self.ckpt_path, ignore_keys=self.ignore_keys)
def sample_mask(self, quant: torch.Tensor, low: int = 0, high: Optional[int] = None) -> torch.BoolTensor:
"""Returns a mask of shape B H_Q W_Q, where True = masked-out, False = keep.
Args:
quant: Dequantized latent tensor of shape B D_Q H_Q W_Q
low: Lower bound of number of tokens to mask out
high: Upper bound of number of tokens to mask out (inclusive).
Defaults to total number of tokens (H_Q * W_Q) if it is set to None.
Returns:
Boolean mask of shape B H_Q W_Q
"""
B, _, H_Q, W_Q = quant.shape
num_tokens = H_Q * W_Q
high = high if high is not None else num_tokens
zero_idxs = torch.randint(low=low, high=high+1, size=(B,), device=quant.device)
noise = torch.rand(B, num_tokens, device=quant.device)
ids_arange_shuffle = torch.argsort(noise, dim=1) # ascend: small is keep, large is remove
mask = torch.where(ids_arange_shuffle < zero_idxs.unsqueeze(1), 0, 1)
mask = rearrange(mask, 'b (h w) -> b h w', h=H_Q, w=W_Q).bool()
return mask
def decode_quant(self,
quant: torch.Tensor,
timesteps: Optional[int] = None,
generator: Optional[torch.Generator] = None,
image_size: Optional[Union[Tuple[int, int], int]] = None,
verbose: bool = False,
vae_decode: bool = False,
scheduler_timesteps_mode: str = 'leading',
prompt: Optional[Union[List[str], str]]= None,
orig_res: Optional[Union[torch.LongTensor, Tuple[int, int]]] = None,
guidance_scale: int = 0.0,
cond_scale: int = 1.0) -> torch.Tensor:
"""Decodes quantized latent codes back to an image.
Args:
quant: Quantized latent code of shape B D_Q H_Q W_Q
timesteps: Number of diffusion timesteps to use. Defaults to self.num_train_timesteps.
generator: Random number generator to use for sampling. By default generations are stochastic.
image_size: Image size to use for the diffusion pipeline. Defaults to decoder image size.
verbose: Whether or not to print progress bar.
vae_decode: If set to True decodes the latent output of stable diffusion
scheduler_timesteps_mode: The mode to use for DDIMScheduler. One of `trailing`, `linspace`,
`leading`. See https://arxiv.org/abs/2305.08891 for more details.
prompt: the input prompts for controlnet.
orig_res: The original resolution of the image to condition the diffusion on. Ignored if None.
See SDXL https://arxiv.org/abs/2307.01952 for more details.
guidance_scale: Classifier free guidance scale.
cond_scale: Scale that is multiplied by the output of control model before being added
to stable diffusion layers in controlnet.
Returns:
Decoded tensor of shape B C H W
"""
dec = self.pipeline(
quant, timesteps=timesteps, generator=generator, image_size=image_size,
verbose=verbose, scheduler_timesteps_mode=scheduler_timesteps_mode, prompt=prompt,
guidance_scale=guidance_scale, cond_scale=cond_scale,
)
if vae_decode:
return self.vae_decode(dec)
return dec
def decode_tokens(self, tokens: torch.LongTensor, **kwargs) -> torch.Tensor:
"""See `decode_quant` for details on the optional args."""
return super().decode_tokens(tokens, **kwargs)
@torch.no_grad()
def vae_encode(self, x: torch.Tensor):
"""Encodes the input image into vae latent representaiton.
Args:
x: Input images
Returns:
Encoded latent tensor of shape B C H W
"""
z = self.vae.encode(x).latent_dist.sample()
z = z * self.vae.config.scaling_factor
return z
@torch.no_grad()
def vae_decode(self, x: torch.Tensor, clip: bool = True) -> torch.Tensor:
"""Decodes the vae latent representation into vae latent representaiton.
Args:
x: VAE latent representation
clip: If set True clips the decoded image between -1 and 1.
Returns:
Decoded image of shape B C H W
"""
x = self.vae.decode(x / self.vae.config.scaling_factor).sample
if clip:
x = torch.clip(x, min=-1, max=1)
return x
def autoencode(self,
input_clean: torch.Tensor,
timesteps: Optional[int] = None,
generator: Optional[torch.Generator] = None,
image_size: Optional[Union[Tuple[int, int], int]] = None,
verbose: bool = False,
vae_decode: bool = False,
scheduler_timesteps_mode: str = 'leading',
prompt: Optional[Union[List[str], str]]= None,
orig_res: Optional[Union[torch.LongTensor, Tuple[int, int]]] = None,
guidance_scale: int = 0.0,
cond_scale: int = 1.0) -> torch.Tensor:
"""Autoencodes an input image tensor by encoding it, quantizing the latent code,
and decoding it back to an image.
Args:
input_clean: Input image tensor of shape B C H W
or B H W in case of semantic segmentation
timesteps: Number of diffusion timesteps to use. Defaults to self.num_train_timesteps.
scheduler: Scheduler to use for the diffusion pipeline. Defaults to the training scheduler.
generator: Random number generator to use for sampling. By default generations are stochastic.
image_size: Image size to use for the diffusion pipeline. Defaults to decoder image size.
verbose: Whether or not to print progress bar.
vae_decode: If set to True, decodes the latent output of stable diffusion
scheduler_timesteps_mode: The mode to use for DDIMScheduler. One of `trailing`, `linspace`,
`leading`. See https://arxiv.org/abs/2305.08891 for more details.
prompt: the input prompts for controlnet.
orig_res: The original resolution of the image to condition the diffusion on. Ignored if None.
See SDXL https://arxiv.org/abs/2307.01952 for more details.
guidance_scale: Classifier free guidance scale.
cond_scale: Scale that is multiplied by the output of control model before being added
to stable diffusion layers in controlnet.
Returns:
Reconstructed tensor of shape B C H W
"""
quant, _, _ = self.encode(input_clean)
dec = self.pipeline(
quant, timesteps=timesteps, generator=generator,
verbose=verbose, scheduler_timesteps_mode=scheduler_timesteps_mode, prompt=prompt,
guidance_scale=guidance_scale, cond_scale=cond_scale,
)
if vae_decode:
return self.vae_decode(dec)
return dec
def forward(self,
input_clean: torch.Tensor,
input_noised: torch.Tensor,
timesteps: Union[torch.Tensor, float, int],
cond_mask: Optional[torch.Tensor] = None,
prompt: Optional[Union[List[str], str]] = None,
orig_res: Optional[Union[torch.LongTensor, Tuple[int, int]]] = None) -> Tuple[torch.Tensor, torch.Tensor]:
"""Forward pass of the encoder, quantizer, and decoder.
Args:
input_clean: Clean input image tensor of shape B C H W
or B H W in case of semantic segmentation. Used for encoding.
input_noised: Noised input image tensor of shape B C H W. Used as
input to the diffusion decoder.
timesteps: Timesteps for conditioning the diffusion decoder on.
cond_mask: Optional mask for the diffusion conditioning.
True = masked-out, False = keep.
prompt: ControlNet input prompt. Defaults to an empty string.
orig_res: The original resolution of the image to condition the diffusion on. Ignored if None.
See SDXL https://arxiv.org/abs/2307.01952 for more details.
Returns:
dec: Decoded image tensor of shape B C H W
code_loss: Codebook loss
"""
with torch.no_grad() if self.freeze_enc else nullcontext():
quant, code_loss, _ = self.encode(input_clean)
if cond_mask is None and self.cfg_dist is not None and self.training:
# Create a random mask for each batch element. True = masked-out, False = keep
B, _, H_Q, W_Q = quant.shape
cond_mask = self.cfg_dist.sample((B,)).to(quant.device, dtype=torch.bool)
cond_mask = repeat(cond_mask, 'b -> b h w', h=H_Q, w=W_Q)
if self.masked_cfg:
mask = self.sample_mask(quant, low=self.masked_cfg_low, high=self.masked_cfg_high)
cond_mask = (mask * cond_mask)
dec = self.decoder(input_noised, timesteps, quant, cond_mask=cond_mask, orig_res=orig_res, prompt=prompt)
return dec, code_loss
def _freeze_vae(self):
"""Freezes VAE"""
for param in self.vae.parameters():
param.requires_grad = False