Ketansomewhere/cifar10_conditional_diffusion1

Here is Custom Pipeline for Class conditioned diffusion model. For training script, pipeline, tutorial nb and sampling please check my Github Repo:- https://github.com/KetanMann/Class_Conditioned_Diffusion_Training_Script Here is Class Conditional Diffusion Pipeline and Sampling.

Firstly install Diffusers

!pip install git+https://github.com/huggingface/diffusers 

Then login to your huggingface account.

from huggingface_hub import notebook_login
notebook_login()

Finally for sampling and model testing. Run these lines of code.

from diffusers import UNet2DModel, DDPMScheduler
from diffusers.utils.torch_utils import randn_tensor
from diffusers.pipelines.pipeline_utils import DiffusionPipeline, ImagePipelineOutput
from huggingface_hub import hf_hub_download
import torch
import os
from PIL import Image
import matplotlib.pyplot as plt
from typing import List, Optional, Tuple, Union

class DDPMPipelinenew(DiffusionPipeline):
    def __init__(self, unet, scheduler, num_classes: int):
        super().__init__()
        self.register_modules(unet=unet, scheduler=scheduler)
        self.num_classes = num_classes
        self._device = unet.device  # Ensure the pipeline knows the device

    @torch.no_grad()
    def __call__(
        self,
        batch_size: int = 64,
        class_labels: Optional[torch.Tensor] = None,
        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
        num_inference_steps: int = 1000,
        output_type: Optional[str] = "pil",
        return_dict: bool = True,
    ) -> Union[ImagePipelineOutput, Tuple]:
        
        # Ensure class_labels is on the same device as the model
        class_labels = class_labels.to(self._device)
        if class_labels.ndim == 0:
            class_labels = class_labels.unsqueeze(0).expand(batch_size)
        else:
            class_labels = class_labels.expand(batch_size)

        # Sample gaussian noise to begin loop
        if isinstance(self.unet.config.sample_size, int):
            image_shape = (
                batch_size,
                self.unet.config.in_channels,
                self.unet.config.sample_size,
                self.unet.config.sample_size,
            )
        else:
            image_shape = (batch_size, self.unet.config.in_channels, *self.unet.config.sample_size)

        image = randn_tensor(image_shape, generator=generator, device=self._device)

        # Set step values
        self.scheduler.set_timesteps(num_inference_steps)

        for t in self.progress_bar(self.scheduler.timesteps):
            # Ensure the class labels are correctly broadcast to match the input tensor shape
            model_output = self.unet(image, t, class_labels).sample

            image = self.scheduler.step(model_output, t, image, generator=generator).prev_sample

        image = (image / 2 + 0.5).clamp(0, 1)
        image = image.cpu().permute(0, 2, 3, 1).numpy()
        if output_type == "pil":
            image = self.numpy_to_pil(image)

        if not return_dict:
            return (image,)

        return ImagePipelineOutput(images=image)

    def to(self, device: torch.device):
        self._device = device
        self.unet.to(device)
        return self

def load_pipeline(repo_id, num_classes, device):
    unet = UNet2DModel.from_pretrained(repo_id, subfolder="unet").to(device)
    scheduler = DDPMScheduler.from_pretrained(repo_id, subfolder="scheduler")
    pipeline = DDPMPipelinenew(unet=unet, scheduler=scheduler, num_classes=num_classes)
    return pipeline.to(device)  # Move the entire pipeline to the device

def save_images_locally(images, save_dir, epoch, class_label):
    os.makedirs(save_dir, exist_ok=True)
    for i, image in enumerate(images):
        image_path = os.path.join(save_dir, f"image_epoch{epoch}_class{class_label}_idx{i}.png")
        image.save(image_path)

def generate_images(pipeline, class_label, batch_size, num_inference_steps, save_dir, epoch):
    generator = torch.Generator(device=pipeline._device).manual_seed(0)
    class_labels = torch.tensor([class_label] * batch_size).to(pipeline._device)
    images = pipeline(
        generator=generator,
        batch_size=batch_size,
        num_inference_steps=num_inference_steps,
        class_labels=class_labels,
        output_type="pil",
    ).images
    save_images_locally(images, save_dir, epoch, class_label)
    return images

def create_image_grid(images, grid_size, save_path):
    assert len(images) == grid_size ** 2, "Number of images must be equal to grid_size squared"
    width, height = images[0].size
    grid_img = Image.new('RGB', (grid_size * width, grid_size * height))
    
    for i, image in enumerate(images):
        x = i % grid_size * width
        y = i // grid_size * height
        grid_img.paste(image, (x, y))
    
    grid_img.save(save_path)
    return grid_img

if __name__ == "__main__":
    repo_id = "Ketansomewhere/cifar10_conditional_diffusion1"
    num_classes = 10  # Adjust to your number of classes
    batch_size = 64
    num_inference_steps = 1000  # Can be as low as 50 for faster generation
    save_dir = "generated_images"
    epoch = 0
    grid_size = 8  # 8x8 grid

    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    pipeline = load_pipeline(repo_id, num_classes, device)
    
    for class_label in range(num_classes):
        images = generate_images(pipeline, class_label, batch_size, num_inference_steps, save_dir, epoch)
        
        # Create and save the grid image
        grid_img_path = os.path.join(save_dir, f"grid_image_class{class_label}.png")
        grid_img = create_image_grid(images, grid_size, grid_img_path)
        
        # Plot the grid image
        plt.figure(figsize=(10, 10))
        plt.imshow(grid_img)
        plt.axis('off')
        plt.title(f'Class {class_label}')
        plt.savefig(os.path.join(save_dir, f"grid_image_class{class_label}.png"))
        plt.show()

Also, check this nb for the above implementation testing.ipynb .