Bug: torchvision/transforms/functional/to_pil_image always converts 1-channel (gray) FloatTensor images to 8-bit unsigned int

• OS: Ubuntu 16.04.4 LTS x64
• PyTorch version: 0.3.0
• Torchvision version: 0.2.0
• How you installed PyTorch (conda, pip, source): conda
• Python version: 3
• CUDA/cuDNN version: 8.0
• GPU models and configuration: Titan X (Maxwell)

ERROR:
ValueError: Incorrect mode (<class 'float'>) supplied for input type <class 'numpy.dtype'>. Should be L

The torchvision transform ToPILImage(mode=float) will always break for input of type torch.FloatTensor
ToPILImage() uses the internal function to_pil_image found in torchvision/transforms/functional.py

In https://github.com/pytorch/vision/blob/master/torchvision/transforms/functional.py:
Line 104 checks if the input is of type torch.FloatTensor
If so, line 105 scales the input by 255, but then converts it to byte
Lines 113-127 check if the user-specified mode is the expected mode, and throws an error if not.
The expected mode is assigned by npimg.dtype, which return np.uint8 if line 105 is executed

I believe the bug can be fixed by changing line 105 from:
pic = pic.mul(255).byte()
-to-
pic = pic.mul(255)

Test script:
import torch
from torchvision import transforms
a = torch.FloatTensor(1,64,64)
tform = transforms.Compose([transforms.ToPILImage(mode='F')])
b = tform(a)

Please let me know if I am in error.
Thank you.