take care of gradient accumulation automatically for researchers, by passing in a max_batch_size on the decoder or diffusion prior trainer forward

dalle2_pytorch/train.py

@@ -68,16 +68,20 @@ def split(t, split_size = None):
 
 def split_args_and_kwargs(x, *args, split_size = None, **kwargs):
     batch_size = len(x)
-    chunk_size = ceil(batch_size / default(split_size, batch_size))
+    split_size = default(split_size, batch_size)
+    chunk_size = ceil(batch_size / split_size)
 
     dict_len = len(kwargs)
     dict_keys = kwargs.keys()
     all_args = (x, *args, *kwargs.values())
+    len_all_args = len(all_args)
+    split_index = len_all_args - dict_len
+
     split_all_args = [split(arg, split_size = split_size) if exists(arg) and isinstance(arg, (torch.Tensor, Iterable)) else ((arg,) * chunk_size) for arg in all_args]
     chunk_sizes = tuple(map(len, split_all_args[0]))
 
     for (chunk_size, *chunked_all_args) in tuple(zip(chunk_sizes, *split_all_args)):
-        chunked_args, chunked_kwargs_values = chunked_all_args[:-dict_len], chunked_all_args[-dict_len:]
+        chunked_args, chunked_kwargs_values = chunked_all_args[:split_index], chunked_all_args[split_index:]
         chunked_kwargs = dict(tuple(zip(dict_keys, chunked_kwargs_values)))
         yield chunk_size, (chunked_args, chunked_kwargs)
 
@@ -249,22 +253,23 @@ class DiffusionPriorTrainer(nn.Module):
 
     def forward(
         self,
+        x,
         *args,
         max_batch_size = None,
         **kwargs
     ):
+        batch_size = x.shape[0]
         total_samples = 0
         total_loss = 0.
 
-        for chunk_size, (chunked_args, chunked_kwargs) in split_args_and_kwargs(*args, split_size = max_batch_size, **kwargs):
+        for chunk_size, (chunked_args, chunked_kwargs) in split_args_and_kwargs(x, *args, split_size = max_batch_size, **kwargs):
             with autocast(enabled = self.amp):
-                loss = self.diffusion_prior(*args, **kwargs)
+                loss = self.diffusion_prior(*chunked_args, **chunked_kwargs)
 
                 total_loss += loss.item() * chunk_size
                 total_samples += chunk_size
 
-                scaled_loss = self.scaler.scale(loss)
-                scaled_loss.backward()
+                self.scaler.scale(loss * (chunk_size / batch_size)).backward()
 
         return total_loss / total_samples
 
@@ -380,6 +385,7 @@ class DecoderTrainer(nn.Module):
         max_batch_size = None,
         **kwargs
     ):
+        batch_size = x.shape[0]
         total_samples = 0
         total_loss = 0.
 
@@ -390,7 +396,6 @@ class DecoderTrainer(nn.Module):
                 total_loss += loss.item() * chunk_size
                 total_samples += chunk_size
 
-                scaled_loss = self.scale(loss, unet_number = unet_number)
-                scaled_loss.backward()
+                self.scale(loss * (chunk_size / batch_size), unet_number = unet_number).backward()
 
         return total_loss / total_samples

setup.py

@@ -10,7 +10,7 @@ setup(
       'dream = dalle2_pytorch.cli:dream'
     ],
   },
-  version = '0.2.25',
+  version = '0.2.26',
   license='MIT',
   description = 'DALL-E 2',
   author = 'Phil Wang',