do two forward passes for prior preservation

Author: patil-suraj

examples/dreambooth/train_dreambooth.py

@@ -442,20 +442,25 @@ def collate_fn(examples):
         input_ids = [example["instance_prompt_ids"] for example in examples]
         pixel_values = [example["instance_images"] for example in examples]
 
-        # concat class and instance examples for prior preservation
-        if args.with_prior_preservation:
-            input_ids += [example["class_prompt_ids"] for example in examples]
-            pixel_values += [example["class_images"] for example in examples]
-
-        pixel_values = torch.stack(pixel_values)
-        pixel_values = pixel_values.to(memory_format=torch.contiguous_format).float()
-
+        pixel_values = torch.stack(pixel_values).to(memory_format=torch.contiguous_format).float()
         input_ids = tokenizer.pad({"input_ids": input_ids}, padding=True, return_tensors="pt").input_ids
 
         batch = {
             "input_ids": input_ids,
             "pixel_values": pixel_values,
         }
+
+        if args.with_prior_preservation:
+            class_input_ids = [example["class_prompt_ids"] for example in examples]
+            class_pixel_values = [example["class_images"] for example in examples]
+
+            class_pixel_values = torch.stack(class_pixel_values).to(memory_format=torch.contiguous_format).float()
+            class_input_ids = tokenizer.pad(
+                {"input_ids": class_input_ids}, padding=True, return_tensors="pt"
+            ).input_ids
+            batch["class_input_ids"] = class_input_ids
+            batch["class_pixel_values"] = class_pixel_values
+
         return batch
 
     train_dataloader = torch.utils.data.DataLoader(
@@ -516,33 +521,41 @@ def collate_fn(examples):
         unet.train()
         for step, batch in enumerate(train_dataloader):
             with accelerator.accumulate(unet):
-                # Convert images to latent space
-                with torch.no_grad():
-                    latents = vae.encode(batch["pixel_values"]).latent_dist.sample()
-                    latents = latents * 0.18215
-
-                # Sample noise that we'll add to the latents
-                noise = torch.randn(latents.shape).to(latents.device)
-                bsz = latents.shape[0]
-                # Sample a random timestep for each image
-                timesteps = torch.randint(
-                    0, noise_scheduler.config.num_train_timesteps, (bsz,), device=latents.device
-                ).long()
-
-                # Add noise to the latents according to the noise magnitude at each timestep
-                # (this is the forward diffusion process)
-                noisy_latents = noise_scheduler.add_noise(latents, noise, timesteps)
-
-                # Get the text embedding for conditioning
-                with torch.no_grad():
-                    encoder_hidden_states = text_encoder(batch["input_ids"])[0]
-
-                # Predict the noise residual
-                noise_pred = unet(noisy_latents, timesteps, encoder_hidden_states).sample
-
-                loss = F.mse_loss(noise_pred, noise, reduction="none").mean([1, 2, 3]).mean()
-                accelerator.backward(loss)
 
+                def _forward(input_ids, pixel_values):
+                    # Convert images to latent space
+                    with torch.no_grad():
+                        latents = vae.encode(pixel_values).latent_dist.sample()
+                        latents = latents * 0.18215
+
+                    # Sample noise that we'll add to the latents
+                    noise = torch.randn(latents.shape).to(latents.device)
+                    bsz = latents.shape[0]
+                    # Sample a random timestep for each image
+                    timesteps = torch.randint(
+                        0, noise_scheduler.config.num_train_timesteps, (bsz,), device=latents.device
+                    ).long()
+
+                    # Add noise to the latents according to the noise magnitude at each timestep
+                    # (this is the forward diffusion process)
+                    noisy_latents = noise_scheduler.add_noise(latents, noise, timesteps)
+
+                    # Get the text embedding for conditioning
+                    with torch.no_grad():
+                        encoder_hidden_states = text_encoder(input_ids)[0]
+
+                    # Predict the noise residual
+                    noise_pred = unet(noisy_latents, timesteps, encoder_hidden_states).sample
+                    loss = F.mse_loss(noise_pred, noise, reduction="none").mean([1, 2, 3]).mean()
+                    return loss
+
+                loss = _forward(batch["input_ids"], batch["pixel_values"])
+
+                if args.with_prior_preservation:
+                    prior_loss = _forward(batch["class_input_ids"], batch["class_pixel_values"])
+                    loss = loss + prior_loss
+
+                accelerator.backward(loss)
                 optimizer.step()
                 lr_scheduler.step()
                 optimizer.zero_grad()