app.py

import jax
import jax.numpy as jnp
from flax.training import train_state
import optax
from diffusers import FlaxStableDiffusionPipeline
from datasets import load_dataset
from tqdm.auto import tqdm
import os
import pickle
from PIL import Image
import numpy as np
import gc


from diffusers.schedulers import PNDMScheduler

class CustomPNDMScheduler(PNDMScheduler):
    def add_noise(self, state, original_samples, noise, timesteps):
        # Explicitly cast timesteps to int32
        timesteps = timesteps.astype(jnp.int32)
        return super().add_noise(state, original_samples, noise, timesteps)

        
# Force JAX to use CPU
jax.config.update('jax_platform_name', 'cpu')

print("Using CPU for computations")

# Set up cache directories
cache_dir = "/tmp/huggingface_cache"
model_cache_dir = os.path.join(cache_dir, "stable_diffusion_model")
os.makedirs(model_cache_dir, exist_ok=True)

print(f"Cache directory: {cache_dir}")
print(f"Model cache directory: {model_cache_dir}")

# Function to load or download the model
def get_model(model_id, revision):
    model_cache_file = os.path.join(model_cache_dir, f"{model_id.replace('/', '_')}_{revision}.pkl")
    print(f"Model cache file: {model_cache_file}")
    if os.path.exists(model_cache_file):
        print("Loading model from cache...")
        with open(model_cache_file, 'rb') as f:
            return pickle.load(f)
    else:
        print("Downloading model...")
        pipeline, params = FlaxStableDiffusionPipeline.from_pretrained(
            model_id, 
            revision=revision,
            dtype=jnp.float32,  # Use float32 for CPU
        )
        with open(model_cache_file, 'wb') as f:
            pickle.dump((pipeline, params), f)
        return pipeline, params

# Load the pre-trained model
model_id = "CompVis/stable-diffusion-v1-4"
pipeline, params = get_model(model_id, "flax")

# Extract UNet from pipeline
unet = pipeline.unet



# After loading the pipeline
custom_scheduler = CustomPNDMScheduler.from_config(pipeline.scheduler.config)
pipeline.scheduler = custom_scheduler


# Load and preprocess your dataset
def preprocess_images(examples):
    def process_image(image):
        if isinstance(image, str):
            image = Image.open(image)
        if not isinstance(image, Image.Image):
            raise ValueError(f"Unexpected image type: {type(image)}")
        image = image.convert("RGB").resize((512, 512))
        image = np.array(image).astype(np.float32) / 255.0
        return image.transpose(2, 0, 1)

    return {"pixel_values": [process_image(img) for img in examples["image"]]}

# Load dataset from Hugging Face
dataset_name = "uruguayai/montevideo"
dataset_cache_file = os.path.join(cache_dir, "montevideo_dataset.pkl")

print(f"Dataset name: {dataset_name}")
print(f"Dataset cache file: {dataset_cache_file}")

try:
    if os.path.exists(dataset_cache_file):
        print("Loading dataset from cache...")
        with open(dataset_cache_file, 'rb') as f:
            processed_dataset = pickle.load(f)
    else:
        print("Loading dataset from Hugging Face...")
        dataset = load_dataset(dataset_name, split="train[:500]")  # Load only first 500 samples
        print("Processing dataset...")
        processed_dataset = dataset.map(preprocess_images, batched=True, remove_columns=dataset.column_names)
        with open(dataset_cache_file, 'wb') as f:
            pickle.dump(processed_dataset, f)

    print(f"Processed dataset size: {len(processed_dataset)}")

except Exception as e:
    print(f"Error loading or processing dataset: {str(e)}")
    raise ValueError("Unable to load or process the dataset.")

# Function to clear JIT cache
def clear_jit_cache():
    jax.clear_caches()
    gc.collect()

# Training function
def train_step(state, batch, rng):
    def compute_loss(params, pixel_values, rng):
        print("pixel_values dtype:", pixel_values.dtype)
        print("params dtypes:", jax.tree_map(lambda x: x.dtype, params))
        print("rng dtype:", rng.dtype)
        
        # Ensure pixel_values are float32
        pixel_values = jnp.array(pixel_values, dtype=jnp.float32)
        
        # Encode images to latent space
        latents = pipeline.vae.apply(
            {"params": params["vae"]},
            pixel_values,
            method=pipeline.vae.encode
        ).latent_dist.sample(rng)
        latents = latents * jnp.float32(0.18215)

        # Generate random noise
        noise = jax.random.normal(rng, latents.shape, dtype=jnp.float32)
        
        # Sample random timesteps
        timesteps = jax.random.randint(
            rng, (latents.shape[0],), 0, pipeline.scheduler.config.num_train_timesteps
        )
        
        print("timesteps dtype:", timesteps.dtype)
        print("latents dtype:", latents.dtype)
        print("noise dtype:", noise.dtype)
        
        # Add noise to latents
        noisy_latents = pipeline.scheduler.add_noise(
            pipeline.scheduler.create_state(),
            original_samples=latents,
            noise=noise,
            timesteps=timesteps
        )
        
        # Generate random encoder hidden states (simulating text embeddings)
        encoder_hidden_states = jax.random.normal(
            rng, 
            (latents.shape[0], pipeline.text_encoder.config.hidden_size),
            dtype=jnp.float32
        )
        
        # Predict noise
        model_output = state.apply_fn(
            {'params': params["unet"]},
            noisy_latents,
            timesteps,
            encoder_hidden_states=encoder_hidden_states,
            train=True,
        )
        
        # Compute loss
        return jnp.mean((model_output - noise) ** 2)

    grad_fn = jax.grad(compute_loss, argnums=0, allow_int=True)
    rng, step_rng = jax.random.split(rng)
    
    grads = grad_fn(state.params, batch["pixel_values"], step_rng)
    loss = compute_loss(state.params, batch["pixel_values"], step_rng)
    state = state.apply_gradients(grads=grads)
    return state, loss

# Initialize training state
learning_rate = 1e-5
optimizer = optax.adam(learning_rate)
float32_params = jax.tree_map(lambda x: x.astype(jnp.float32) if x.dtype != jnp.int32 else x, params)
state = train_state.TrainState.create(
    apply_fn=unet.__call__,
    params=float32_params,
    tx=optimizer,
)

# Modify the train_step function
def train_step(state, batch, rng):
    def compute_loss(params, pixel_values, rng):
        # Ensure pixel_values are float32
        pixel_values = jnp.array(pixel_values, dtype=jnp.float32)
        
        # Encode images to latent space
        latents = pipeline.vae.apply(
            {"params": params["vae"]},
            pixel_values,
            method=pipeline.vae.encode
        ).latent_dist.sample(rng)
        latents = latents * jnp.float32(0.18215)

        # Generate random noise
        noise = jax.random.normal(rng, latents.shape, dtype=jnp.float32)
        
        # Sample random timesteps
        timesteps = jax.random.randint(
            rng, (latents.shape[0],), 0, pipeline.scheduler.config.num_train_timesteps
        )
        timesteps = jnp.array(timesteps, dtype=jnp.float32)
        
        # Add noise to latents
        noisy_latents = pipeline.scheduler.add_noise(
            pipeline.scheduler.create_state(),
            original_samples=latents,
            noise=noise,
            timesteps=timesteps
        )
        
        # Generate random encoder hidden states (simulating text embeddings)
        encoder_hidden_states = jax.random.normal(
            rng, 
            (latents.shape[0], pipeline.text_encoder.config.hidden_size),
            dtype=jnp.float32
        )
        
        # Predict noise
        model_output = state.apply_fn(
            {'params': params["unet"]},
            noisy_latents,
            timesteps,
            encoder_hidden_states=encoder_hidden_states,
            train=True,
        )
        
        # Compute loss
        return jnp.mean((model_output - noise) ** 2)

    grad_fn = jax.grad(compute_loss, argnums=0, allow_int=True)
    rng, step_rng = jax.random.split(rng)
    
    grads = grad_fn(state.params, batch["pixel_values"], step_rng)
    loss = compute_loss(state.params, batch["pixel_values"], step_rng)
    state = state.apply_gradients(grads=grads)
    return state, loss



# Training loop (remains the same)
num_epochs = 3
batch_size = 1
rng = jax.random.PRNGKey(0)

for epoch in range(num_epochs):
    epoch_loss = 0
    num_batches = 0
    for batch in tqdm(processed_dataset.batch(batch_size)):
        batch['pixel_values'] = jnp.array(batch['pixel_values'], dtype=jnp.float32)
        rng, step_rng = jax.random.split(rng)
        state, loss = train_step(state, batch, step_rng)
        epoch_loss += loss
        num_batches += 1
        
        if num_batches % 10 == 0:
            clear_jit_cache()
    
    avg_loss = epoch_loss / num_batches
    print(f"Epoch {epoch+1}/{num_epochs}, Average Loss: {avg_loss}")
    clear_jit_cache()

    
# Save the fine-tuned model
output_dir = "/tmp/montevideo_fine_tuned_model"
os.makedirs(output_dir, exist_ok=True)
unet.save_pretrained(output_dir, params=state.params["unet"])

print(f"Model saved to {output_dir}")