Normalize images

train_model.py

@@ -31,8 +31,9 @@ def save_and_test(pipeline, epoch):
     pipeline.save_pretrained(model_file)
 
 def transform_images(image):
-    res = torch.Tensor(SAMPLE_NUM_CHANNELS, SAMPLE_SIZE, SAMPLE_SIZE)
+    res = torch.Tensor((SAMPLE_NUM_CHANNELS, SAMPLE_SIZE, SAMPLE_SIZE))
     pil_to_tensor = T.PILToTensor()
+    tensor_to_pil = T.ToPILImage()
 
     res_index = 0
     scale_factor = np.minimum(SAMPLE_SIZE / image.width, SAMPLE_SIZE / image.height)
@@ -40,7 +41,13 @@ def transform_images(image):
     
     new_image = PIL.Image.new('RGB', (SAMPLE_SIZE, SAMPLE_SIZE))
     new_image.paste(image, box=(int((SAMPLE_SIZE - image.width)/2), int((SAMPLE_SIZE - image.height)/2)))
+
+    #data = np.array(new_image, dtype=np.float32)
+    #data = (data / 128.0 - 1.0)
+    #res = torch.from_numpy(data)
     res = pil_to_tensor(new_image)
+    res.to(dtype=torch.float32)
+    res = res / torch.Tensor([128.0]) - torch.Tensor([1.0])
     return res
         
 def convert_images(dataset):