Replaced ToTensor with a combination of PILToTensor and ConvertImageDtype

[prabhat00155]

Training logs before these changes:
training_log_no_changes.txt
Training logs after these changes:
training_log_after_changes.txt

cc @datumbox

[datumbox]

The changes LGTM, but there are additional references of ToTensor in other places. I think it's worth removing them from code usages and documentation.

Here are some examples:

vision/torchvision/transforms/transforms.py

Line 38 in c8e3b2a

>>> transforms.ToTensor(),

vision/torchvision/transforms/transforms.py

Line 1573 in c8e3b2a

>>> transforms.ToTensor(),

vision/references/detection/presets.py

Line 9 in c8e3b2a

T.ToTensor(),

vision/references/segmentation/presets.py

Line 14 in c8e3b2a

T.ToTensor(),

Feel free to merge this PR if you want and open another one to address the rest, or do the changes here. Whatever works for you.

[prabhat00155]

The changes LGTM, but there are additional references of ToTensor in other places. I think it's worth removing them from code usages and documentation.

Here are some examples:

vision/torchvision/transforms/transforms.py

Line 38 in c8e3b2a

>>> transforms.ToTensor(),

vision/torchvision/transforms/transforms.py

Line 1573 in c8e3b2a

>>> transforms.ToTensor(),

vision/references/detection/presets.py

Line 9 in c8e3b2a

T.ToTensor(),

vision/references/segmentation/presets.py

Line 14 in c8e3b2a

T.ToTensor(),

Feel free to merge this PR if you want and open another one to address the rest, or do the changes here. Whatever works for you.

Thanks @datumbox! I already replaced T.ToTensor() in detection and segmentation to be a combination of pil_to_tensor() and convert_image_dtype() in #4452. I'll update the example usages under torchvision in a separate PR.

[datumbox]

@prabhat00155 You are right, I missed the fact that we use a custom ToTensor class which internally calls the F.pil_to_tensor() and F.convert_image_dtype() methods.

How would you feel about replacing the custom ToTensor class from the segmentation/detection references with 2 similar wrapper classes called ConvertImageDtype and PILToTensor? This is just to reinforce the fact that ToTensor is not advised. It won't have any real effects on what you actually do. Let me know your thoughts.

[prabhat00155]

@prabhat00155 You are right, I missed the fact that we use a custom ToTensor class which internally calls the F.pil_to_tensor() and F.convert_image_dtype() methods.

How would you feel about replacing the custom ToTensor class from the segmentation/detection references with 2 similar wrapper classes called ConvertImageDtype and PILToTensor? This is just to reinforce the fact that ToTensor is not advised. It won't have any real effects on what you actually do. Let me know your thoughts.

Yeah, sure. I can make that change.