Batches of size 1 cause error: Expected query.is_contiguous() to be true, but got false

1 files changed, 7 insertions, 20 deletions

diff --git a/dreambooth.py b/dreambooth.py
index 2df6858..0c58ab5 100644
--- a/dreambooth.py
+++ b/dreambooth.py
@@ -433,7 +433,7 @@ class Checkpointer:
             del image_grid
             del stable_latents
 
-        for data, pool in [(train_data, "train"), (val_data, "val")]:
+        for data, pool in [(val_data, "val"), (train_data, "train")]:
             all_samples = []
             filename = f"step_{step}_{pool}.png"
 
@@ -492,12 +492,11 @@ def main():
 
     if args.with_prior_preservation:
         class_images_dir = Path(args.class_data_dir)
-        if not class_images_dir.exists():
-            class_images_dir.mkdir(parents=True)
+        class_images_dir.mkdir(parents=True, exist_ok=True)
         cur_class_images = len(list(class_images_dir.iterdir()))
 
         if cur_class_images < args.num_class_images:
-            torch_dtype = torch.float16 if accelerator.device.type == "cuda" else torch.float32
+            torch_dtype = torch.bfloat16 if accelerator.device.type == "cuda" else torch.float32
             pipeline = StableDiffusionPipeline.from_pretrained(
                 args.pretrained_model_name_or_path, torch_dtype=torch_dtype)
             pipeline.enable_attention_slicing()
@@ -581,7 +580,6 @@ def main():
         eps=args.adam_epsilon,
     )
 
-    # TODO (patil-suraj): laod scheduler using args
     noise_scheduler = DDPMScheduler(
         beta_start=0.00085,
         beta_end=0.012,
@@ -595,7 +593,7 @@ def main():
         pixel_values = [example["instance_images"] for example in examples]
 
         # concat class and instance examples for prior preservation
-        if args.with_prior_preservation:
+        if args.with_prior_preservation and "class_prompt_ids" in examples[0]:
             input_ids += [example["class_prompt_ids"] for example in examples]
             pixel_values += [example["class_images"] for example in examples]
 
@@ -789,6 +787,8 @@ def main():
 
             train_loss /= len(train_dataloader)
 
+            accelerator.wait_for_everyone()
+
             unet.eval()
             val_loss = 0.0
 
@@ -812,18 +812,7 @@ def main():
                     noise_pred, noise = accelerator.gather_for_metrics((noise_pred, noise))
 
                     with accelerator.autocast():
-                        if args.with_prior_preservation:
-                            noise_pred, noise_pred_prior = torch.chunk(noise_pred, 2, dim=0)
-                            noise, noise_prior = torch.chunk(noise, 2, dim=0)
-
-                            loss = F.mse_loss(noise_pred, noise, reduction="none").mean([1, 2, 3]).mean()
-
-                            prior_loss = F.mse_loss(noise_pred_prior, noise_prior,
-                                                    reduction="none").mean([1, 2, 3]).mean()
-
-                            loss = loss + args.prior_loss_weight * prior_loss
-                        else:
-                            loss = F.mse_loss(noise_pred, noise, reduction="none").mean([1, 2, 3]).mean()
+                        loss = F.mse_loss(noise_pred, noise, reduction="none").mean([1, 2, 3]).mean()
 
                     loss = loss.detach().item()
                     val_loss += loss
@@ -851,8 +840,6 @@ def main():
                     global_step,
                     args.resolution, args.resolution, 7.5, 0.0, args.sample_steps)
 
-            accelerator.wait_for_everyone()
-
         # Create the pipeline using using the trained modules and save it.
         if accelerator.is_main_process:
             print("Finished! Saving final checkpoint and resume state.")