Improved batch padding

1 files changed, 26 insertions, 29 deletions

diff --git a/data/csv.py b/data/csv.py
index 480e9f2..233f5d8 100644
--- a/data/csv.py
+++ b/data/csv.py
@@ -399,23 +399,20 @@ class VlpnDataset(IterableDataset):
             mask[:start] = False
             mask[end:] = False
 
-        while mask.any():
+        while mask.any() or len(batch) != 0:
+            if len(batch) >= batch_size:
+                yield batch
+                batch = []
+
             bucket_mask = mask.logical_and(self.bucket_assignments == bucket)
             bucket_items = self.bucket_items[bucket_mask]
 
-            if len(batch) >= batch_size:
+            if len(bucket_items) == 0 and len(batch) != 0 and not self.fill_batch:
                 yield batch
                 batch = []
+                continue
 
-            if len(bucket_items) == 0:
-                if len(batch) != 0:
-                    if self.fill_batch:
-                        fill_items = self.bucket_items[self.bucket_assignments == bucket]
-                        fill_perm = torch.randint(len(fill_items), (batch_size - len(batch),), generator=self.generator)
-                        batch += fill_items[fill_perm]
-                    yield batch
-                    batch = []
-
+            if len(bucket_items) == 0 and len(batch) == 0:
                 bucket = self.bucket_assignments[mask][0]
                 ratio = self.buckets[bucket]
                 width = int(self.size * ratio) if ratio > 1 else self.size
@@ -430,30 +427,30 @@ class VlpnDataset(IterableDataset):
                         transforms.Normalize([0.5], [0.5]),
                     ]
                 )
+
+                continue
+
+            if len(bucket_items) == 0:
+                bucket_items = self.bucket_items[self.bucket_assignments == bucket]
+                item_index = bucket_items[torch.randint(len(bucket_items), (1,), generator=self.generator)]
             else:
                 item_index = bucket_items[0]
-                item = self.items[item_index]
                 mask[self.bucket_item_range[bucket_mask][0]] = False
 
-                example = {}
+            item = self.items[item_index]
 
-                example["prompt_ids"] = self.get_input_ids(keywords_to_prompt(item.prompt))
-                example["nprompt_ids"] = self.get_input_ids(item.nprompt)
+            example = {}
 
-                example["instance_prompt_ids"] = self.get_input_ids(
-                    keywords_to_prompt(item.prompt, self.dropout, True)
-                )
-                example["instance_images"] = image_transforms(get_image(item.instance_image_path))
+            example["prompt_ids"] = self.get_input_ids(keywords_to_prompt(item.prompt))
+            example["nprompt_ids"] = self.get_input_ids(item.nprompt)
 
-                if self.num_class_images != 0:
-                    example["class_prompt_ids"] = self.get_input_ids(item.cprompt)
-                    example["class_images"] = image_transforms(get_image(item.class_image_path))
+            example["instance_prompt_ids"] = self.get_input_ids(
+                keywords_to_prompt(item.prompt, self.dropout, True)
+            )
+            example["instance_images"] = image_transforms(get_image(item.instance_image_path))
 
-                batch.append(example)
+            if self.num_class_images != 0:
+                example["class_prompt_ids"] = self.get_input_ids(item.cprompt)
+                example["class_images"] = image_transforms(get_image(item.class_image_path))
 
-        if len(batch) != 0:
-            if self.fill_batch:
-                fill_items = self.bucket_items[self.bucket_assignments == bucket]
-                fill_perm = torch.randint(len(fill_items), (batch_size - len(batch),), generator=self.generator)
-                batch += fill_items[fill_perm]
-            yield batch
+            batch.append(example)