app.py

import gradio as gr
from random import randint
from all_models import models

from externalmod import gr_Interface_load, randomize_seed

import asyncio
import os
from threading import RLock

# Create a lock to ensure thread safety when accessing shared resources
lock = RLock()
# Load Hugging Face token from environment variable, if available
HF_TOKEN = os.environ.get("HF_TOKEN") if os.environ.get("HF_TOKEN") else None # If private or gated models aren't used, ENV setting is unnecessary.

# Function to load all models specified in the 'models' list
def load_fn(models):
    global models_load
    models_load = {}
    
    # Iterate through all models to load them
    for model in models:
        if model not in models_load.keys():
            try:
                # Log model loading attempt
                print(f"Attempting to load model: {model}")
                # Load model interface using externalmod function
                m = gr_Interface_load(f'models/{model}', hf_token=HF_TOKEN)
                print(f"Successfully loaded model: {model}")
            except Exception as error:
                # In case of an error, print it and create a placeholder interface
                print(f"Error loading model {model}: {error}")
                m = gr.Interface(lambda: None, ['text'], ['image'])
            # Update the models_load dictionary with the loaded model
            models_load.update({model: m})

# Load all models defined in the 'models' list
print("Loading models...")
load_fn(models)
print("Models loaded successfully.")

num_models = 6

# Set the default models to use for inference
default_models = models[:num_models]
inference_timeout = 600
MAX_SEED = 3999999999
# Generate a starting seed randomly between 1941 and 2024
starting_seed = randint(1941, 2024)
print(f"Starting seed: {starting_seed}")

# Extend the choices list to ensure it contains 'num_models' elements
def extend_choices(choices):
    print(f"Extending choices: {choices}")
    extended = choices[:num_models] + (num_models - len(choices[:num_models])) * ['NA']
    print(f"Extended choices: {extended}")
    return extended

# Update the image boxes based on selected models
def update_imgbox(choices):
    print(f"Updating image boxes with choices: {choices}")
    choices_plus = extend_choices(choices[:num_models])
    imgboxes = [gr.Image(None, label=m, visible=(m != 'NA')) for m in choices_plus]
    print(f"Updated image boxes: {imgboxes}")
    return imgboxes

# Asynchronous function to perform inference on a given model
async def infer(model_str, prompt, seed=1, timeout=inference_timeout):
    from pathlib import Path
    kwargs = {}
    noise = ""
    kwargs["seed"] = seed
    # Create an asynchronous task to run the model inference
    print(f"Starting inference for model: {model_str} with prompt: '{prompt}' and seed: {seed}")
    task = asyncio.create_task(asyncio.to_thread(models_load[model_str].fn,
                               prompt=f'{prompt} {noise}', **kwargs, token=HF_TOKEN))
    await asyncio.sleep(0)  # Allow other tasks to run
    try:
        # Wait for the task to complete within the specified timeout
        result = await asyncio.wait_for(task, timeout=timeout)
        print(f"Inference completed for model: {model_str}")
    except (Exception, asyncio.TimeoutError) as e:
        # Handle any exceptions or timeout errors
        print(f"Error during inference for model {model_str}: {e}")
        if not task.done():
            task.cancel()
            print(f"Task cancelled for model: {model_str}")
        result = None
    # If the task completed successfully, save the result as an image
    if task.done() and result is not None:
        with lock:
            png_path = "image.png"
            result.save(png_path)
            image = str(Path(png_path).resolve())
            print(f"Result saved as image: {image}")
        return image
    print(f"No result for model: {model_str}")
    return None

# Function to generate an image based on the given model, prompt, and seed
def gen_fnseed(model_str, prompt, seed=1):
    if model_str == 'NA':
        print(f"Model is 'NA', skipping generation.")
        return None
    try:
        # Create a new event loop to run the asynchronous inference function
        print(f"Generating image for model: {model_str} with prompt: '{prompt}' and seed: {seed}")
        loop = asyncio.new_event_loop()
        result = loop.run_until_complete(infer(model_str, prompt, seed, inference_timeout))
    except (Exception, asyncio.CancelledError) as e:
        # Handle any exceptions or cancelled tasks
        print(f"Error during generation for model {model_str}: {e}")
        result = None
    finally:
        # Close the event loop
        loop.close()
        print(f"Event loop closed for model: {model_str}")
    return result

# Create the Gradio Blocks interface with a custom theme
print("Creating Gradio interface...")
with gr.Blocks(theme="Nymbo/Nymbo_Theme") as demo:
    gr.HTML("<center><h1>Compare-6</h1></center>")
    with gr.Tab('Compare-6'):
        # Text input for user prompt
        txt_input = gr.Textbox(label='Your prompt:', lines=4)
        # Button to generate images
        gen_button = gr.Button('Generate up to 6 images in up to 3 minutes total')
        with gr.Row():
            # Slider to select a seed for reproducibility
            seed = gr.Slider(label="Use a seed to replicate the same image later (maximum 3999999999)", minimum=0, maximum=MAX_SEED, step=1, value=starting_seed, scale=3)
            # Button to randomize the seed
            seed_rand = gr.Button("Randomize Seed 🎲", size="sm", variant="secondary", scale=1)    
        # Set up click event to randomize the seed
        seed_rand.click(randomize_seed, None, [seed], queue=False)
        print("Seed randomization button set up.")
        # Button click to start generation
        gen_button.click(lambda s: gr.update(interactive=True), None)
        print("Generation button set up.")

        with gr.Row():
            # Create image output components for each model
            output = [gr.Image(label=m, min_width=480) for m in default_models]
            # Create hidden textboxes to store the current models
            current_models = [gr.Textbox(m, visible=False) for m in default_models]
            
            # Set up generation events for each model and output image
            for m, o in zip(current_models, output):
                print(f"Setting up generation event for model: {m.value}")
                gen_event = gr.on(triggers=[gen_button.click, txt_input.submit], fn=gen_fnseed,
                                  inputs=[m, txt_input, seed], outputs=[o], concurrency_limit=None, queue=False)
                # The commented stop button could be used to cancel the generation event
                #stop_button.click(lambda s: gr.update(interactive=False), None, stop_button, cancels=[gen_event])
        # Accordion to allow model selection
        with gr.Accordion('Model selection'):
            # Checkbox group to select up to 'num_models' different models
            model_choice = gr.CheckboxGroup(models, label=f'Choose up to {int(num_models)} different models from the {len(models)} available!', value=default_models, interactive=True)
            # Update image boxes and current models based on model selection
            model_choice.change(update_imgbox, model_choice, output)
            model_choice.change(extend_choices, model_choice, current_models)
            print("Model selection setup complete.")
        with gr.Row():
            # Placeholder HTML to add additional UI elements if needed
            gr.HTML(
)

# Queue settings for handling multiple concurrent requests
print("Setting up queue...")
demo.queue(default_concurrency_limit=200, max_size=200)
print("Launching Gradio interface...")
demo.launch(show_api=False, max_threads=400)
print("Gradio interface launched successfully.")