0.9.12

Author: divideconcept

torchstudio/datasets/randomgenerator.py

@@ -10,7 +10,7 @@ class RandomGenerator(Dataset):
             Size of the dataset (number of samples)
         tensors:
             A list of tuples defining tensor properties: shape, type, range
-            All properties are optionals. Defaults are null, float, [0,1]
+            All properties are optionals. Defaults are null, torch.float, [0,1]
     """
 
     def __init__(self, size:int=256, tensors=[(3,64,64), (int,[0,9])]):
@@ -29,12 +29,12 @@ def __getitem__(self, idx):
         sample = []
         for properties in self.tensors:
             shape=[]
-            dtype=float
+            dtype=torch.float
             drange=[0,1]
             for property in properties:
                 if type(property)==int:
                     shape.append(property)
-                elif inspect.isclass(property):
+                elif type(property)==type or type(property)==torch.dtype:
                     dtype=property
                 elif type(property) is list:
                     drange=property

torchstudio/modeltrain.py

@@ -237,6 +237,11 @@ def deepcopy_cpu(value):
                 scaler = torch.cuda.amp.GradScaler()
             if mode=='BF16':
                 os.environ["TORCH_CUDNN_V8_API_ENABLED"] = "1" #https://discuss.pytorch.org/t/bfloat16-has-worse-performance-than-float16-for-conv2d/154373
+        train_type=None
+        if mode=='FP16':
+            train_type=torch.float16
+        if mode=='BF16':
+            train_type=torch.bfloat16
         print("Training... epoch "+str(scheduler.last_epoch)+"\n", file=sys.stderr)
 
     if msg_type == 'TrainOneEpoch' and modules_valid:
@@ -252,7 +257,7 @@ def deepcopy_cpu(value):
             targets = [tensors[i].to(device) for i in output_tensors_id]
             optimizer.zero_grad()
 
-            with torch.autocast(device_type='cuda' if 'cuda' in device_id else 'cpu', dtype=torch.bfloat16 if mode=='BF16' else torch.float16, enabled=True if '16' in mode else False):
+            with torch.autocast(device_type='cuda' if 'cuda' in device_id else 'cpu', dtype=train_type, enabled=True if train_type else False):
                 outputs = model(*inputs)
                 outputs = outputs if type(outputs) is not torch.Tensor else [outputs]
                 loss = 0