V prediction ddim (#1313)

* v diffusion support for ddpm

* quality and style

* variable name consistency

* missing base case

* pass prediction type along in the pipeline

* put prediction type in scheduler config

* style

* try to train on ddim

* changes to ddim

* ddim v prediction works to train butterflies example

* fix bad merge, style and quality

* try to fix broken doc strings

* second pass

* one more

* white space

* Update src/diffusers/schedulers/scheduling_ddim.py

* remove extra lines

* Update src/diffusers/schedulers/scheduling_ddim.py

Co-authored-by: Ben Glickenhaus <[email protected]>
Co-authored-by: Nathan Lambert <[email protected]>

Author: 3 people

examples/v_prediction/train_butterflies.py

@@ -0,0 +1,227 @@
+import glob
+import os
+from dataclasses import dataclass
+
+import torch
+import torch.nn.functional as F
+
+from accelerate import Accelerator
+from datasets import load_dataset
+from diffusers import DDIMPipeline, DDIMScheduler, DDPMPipeline, DDPMScheduler, UNet2DModel
+from diffusers.hub_utils import init_git_repo, push_to_hub
+from diffusers.optimization import get_cosine_schedule_with_warmup
+from PIL import Image
+from torchvision import transforms
+from tqdm.auto import tqdm
+
+
+@dataclass
+class TrainingConfig:
+    image_size = 128  # the generated image resolution
+    train_batch_size = 16
+    eval_batch_size = 16  # how many images to sample during evaluation
+    num_epochs = 50
+    gradient_accumulation_steps = 1
+    learning_rate = 5e-5
+    lr_warmup_steps = 500
+    save_image_epochs = 10
+    save_model_epochs = 30
+    mixed_precision = "fp16"  # `no` for float32, `fp16` for automatic mixed precision
+    output_dir = "ddim-butterflies-128-v-diffusion"  # the model namy locally and on the HF Hub
+
+    push_to_hub = False  # whether to upload the saved model to the HF Hub
+    hub_private_repo = False
+    overwrite_output_dir = True  # overwrite the old model when re-running the notebook
+    seed = 0
+
+
+config = TrainingConfig()
+
+
+config.dataset_name = "huggan/smithsonian_butterflies_subset"
+dataset = load_dataset(config.dataset_name, split="train")
+
+
+preprocess = transforms.Compose(
+    [
+        transforms.Resize((config.image_size, config.image_size)),
+        transforms.RandomHorizontalFlip(),
+        transforms.ToTensor(),
+        transforms.Normalize([0.5], [0.5]),
+    ]
+)
+
+
+def transform(examples):
+    images = [preprocess(image.convert("RGB")) for image in examples["image"]]
+    return {"images": images}
+
+
+dataset.set_transform(transform)
+
+
+train_dataloader = torch.utils.data.DataLoader(dataset, batch_size=config.train_batch_size, shuffle=True)
+
+
+model = UNet2DModel(
+    sample_size=config.image_size,  # the target image resolution
+    in_channels=3,  # the number of input channels, 3 for RGB images
+    out_channels=3,  # the number of output channels
+    layers_per_block=2,  # how many ResNet layers to use per UNet block
+    block_out_channels=(128, 128, 256, 256, 512, 512),  # the number of output channes for each UNet block
+    down_block_types=(
+        "DownBlock2D",  # a regular ResNet downsampling block
+        "DownBlock2D",
+        "DownBlock2D",
+        "DownBlock2D",
+        "AttnDownBlock2D",  # a ResNet downsampling block with spatial self-attention
+        "DownBlock2D",
+    ),
+    up_block_types=(
+        "UpBlock2D",  # a regular ResNet upsampling block
+        "AttnUpBlock2D",  # a ResNet upsampling block with spatial self-attention
+        "UpBlock2D",
+        "UpBlock2D",
+        "UpBlock2D",
+        "UpBlock2D",
+    ),
+)
+
+
+if config.output_dir.startswith("ddpm"):
+    noise_scheduler = DDPMScheduler(
+        num_train_timesteps=1000,
+        beta_schedule="squaredcos_cap_v2",
+        variance_type="v_diffusion",
+        prediction_type="v",
+    )
+else:
+    noise_scheduler = DDIMScheduler(
+        num_train_timesteps=1000,
+        beta_schedule="squaredcos_cap_v2",
+        variance_type="v_diffusion",
+        prediction_type="v",
+    )
+
+
+optimizer = torch.optim.AdamW(model.parameters(), lr=config.learning_rate)
+
+
+lr_scheduler = get_cosine_schedule_with_warmup(
+    optimizer=optimizer,
+    num_warmup_steps=config.lr_warmup_steps,
+    num_training_steps=(len(train_dataloader) * config.num_epochs),
+)
+
+
+def make_grid(images, rows, cols):
+    w, h = images[0].size
+    grid = Image.new("RGB", size=(cols * w, rows * h))
+    for i, image in enumerate(images):
+        grid.paste(image, box=(i % cols * w, i // cols * h))
+    return grid
+
+
+def evaluate(config, epoch, pipeline):
+    # Sample some images from random noise (this is the backward diffusion process).
+    # The default pipeline output type is `List[PIL.Image]`
+    images = pipeline(
+        batch_size=config.eval_batch_size,
+        generator=torch.manual_seed(config.seed),
+    ).images
+
+    # Make a grid out of the images
+    image_grid = make_grid(images, rows=4, cols=4)
+
+    # Save the images
+    test_dir = os.path.join(config.output_dir, "samples")
+    os.makedirs(test_dir, exist_ok=True)
+    image_grid.save(f"{test_dir}/{epoch:04d}.png")
+
+
+def train_loop(config, model, noise_scheduler, optimizer, train_dataloader, lr_scheduler):
+    # Initialize accelerator and tensorboard logging
+    accelerator = Accelerator(
+        mixed_precision=config.mixed_precision,
+        gradient_accumulation_steps=config.gradient_accumulation_steps,
+        log_with="tensorboard",
+        logging_dir=os.path.join(config.output_dir, "logs"),
+    )
+    if accelerator.is_main_process:
+        if config.push_to_hub:
+            repo = init_git_repo(config, at_init=True)
+        accelerator.init_trackers("train_example")
+
+    # Prepare everything
+    # There is no specific order to remember, you just need to unpack the
+    # objects in the same order you gave them to the prepare method.
+    model, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
+        model, optimizer, train_dataloader, lr_scheduler
+    )
+
+    global_step = 0
+
+    if config.output_dir.startswith("ddpm"):
+        pipeline = DDPMPipeline(unet=accelerator.unwrap_model(model), scheduler=noise_scheduler)
+    else:
+        pipeline = DDIMPipeline(unet=accelerator.unwrap_model(model), scheduler=noise_scheduler)
+
+    evaluate(config, 0, pipeline)
+
+    # Now you train the model
+    for epoch in range(config.num_epochs):
+        progress_bar = tqdm(total=len(train_dataloader), disable=not accelerator.is_local_main_process)
+        progress_bar.set_description(f"Epoch {epoch}")
+
+        for step, batch in enumerate(train_dataloader):
+            clean_images = batch["images"]
+            # Sample noise to add to the images
+            noise = torch.randn(clean_images.shape).to(clean_images.device)
+            bs = clean_images.shape[0]
+
+            # Sample a random timestep for each image
+            timesteps = torch.randint(0, noise_scheduler.num_train_timesteps, (bs,), device=clean_images.device).long()
+
+            with accelerator.accumulate(model):
+                # Predict the noise residual
+                alpha_t, sigma_t = noise_scheduler.get_alpha_sigma(clean_images, timesteps, accelerator.device)
+                z_t = alpha_t * clean_images + sigma_t * noise
+                noise_pred = model(z_t, timesteps).sample
+                v = alpha_t * noise - sigma_t * clean_images
+                loss = F.mse_loss(noise_pred, v)
+                accelerator.backward(loss)
+
+                accelerator.clip_grad_norm_(model.parameters(), 1.0)
+                optimizer.step()
+                lr_scheduler.step()
+                optimizer.zero_grad()
+
+            progress_bar.update(1)
+            logs = {"loss": loss.detach().item(), "lr": lr_scheduler.get_last_lr()[0], "step": global_step}
+            progress_bar.set_postfix(**logs)
+            accelerator.log(logs, step=global_step)
+            global_step += 1
+
+        # After each epoch you optionally sample some demo images with evaluate() and save the model
+        if accelerator.is_main_process:
+            if config.output_dir.startswith("ddpm"):
+                pipeline = DDPMPipeline(unet=accelerator.unwrap_model(model), scheduler=noise_scheduler)
+            else:
+                pipeline = DDIMPipeline(unet=accelerator.unwrap_model(model), scheduler=noise_scheduler)
+
+            if (epoch + 1) % config.save_image_epochs == 0 or epoch == config.num_epochs - 1:
+                evaluate(config, epoch, pipeline)
+
+            if (epoch + 1) % config.save_model_epochs == 0 or epoch == config.num_epochs - 1:
+                if config.push_to_hub:
+                    push_to_hub(config, pipeline, repo, commit_message=f"Epoch {epoch}", blocking=True)
+                else:
+                    pipeline.save_pretrained(config.output_dir)
+
+
+args = (config, model, noise_scheduler, optimizer, train_dataloader, lr_scheduler)
+
+train_loop(*args)
+
+sample_images = sorted(glob.glob(f"{config.output_dir}/samples/*.png"))
+Image.open(sample_images[-1])