fix different reso in batch

library/train_util.py

@@ -678,8 +678,16 @@ class BaseDataset(torch.utils.data.Dataset):
     def cache_latents(self, vae, vae_batch_size=1):
         # ちょっと速くした
         print("caching latents.")
-        infos = []
-        for info in self.image_data.values():
+
+        image_infos= list(self.image_data.values())
+
+        # sort by resolution
+        image_infos.sort(key=lambda info: info.bucket_reso[0] * info.bucket_reso[1])
+
+        # split by resolution
+        batches = []
+        batch = []
+        for info in image_infos:
             subset = self.image_to_subset[info.image_key]
 
             if info.latents_npz is not None:
@@ -690,10 +698,23 @@ class BaseDataset(torch.utils.data.Dataset):
                     info.latents_flipped = torch.FloatTensor(info.latents_flipped)
                 continue
 
-            infos.append(info)
+            # if last member of batch has different resolution, flush the batch
+            if len(batch) > 0 and batch[-1].bucket_reso != info.bucket_reso:
+                batches.append(batch)
+                batch = []
 
-        for i in tqdm(range(0, len(infos), vae_batch_size), smoothing=1, total=len(infos) // vae_batch_size):
-            batch = infos[i : i + vae_batch_size]
+            batch.append(info)
+
+            # if number of data in batch is enough, flush the batch
+            if len(batch) >= vae_batch_size:
+                batches.append(batch)
+                batch = []
+
+        if len(batch) > 0:
+            batches.append(batch)
+
+        # iterate batches   
+        for batch in tqdm(batches, smoothing=1, total=len(batches)):
             images = []
             for info in batch:
                 image = self.load_image(info.absolute_path)