GitLab

Using python2 style '/' will convert int type to float in python3 which will cause the following error when creating FloatTensor:

TypeError: torch.FloatTensor constructor received an invalid combination of arguments - got (float, int, int, int), but expected one of:
 * no arguments
 * (int ...)
      didn't match because some of the arguments have invalid types: (float, int, int, int)

* clipping values, and replacing transposes with permute

Without clipping, any values larger than 255 will be replaced with
int(v mod 256) by byte(), which results in high freq noise in image.

* replaced clip with clamp

This reverts commit 2f579d58.

Without clipping, any values larger than 255 will be replaced with
int(v mod 256) by byte(), which results in high freq noise in image.

In python2, calling `make_grid` won't display the last images if the number of images doesn't divide `nrow`.
Int division...

This should fix that!

* fix for to_tensor when input is np.ndarray of shape [H,W,C]. Issue #48 pytorch/vision

* update cifar datasets to transpose images from CHW -> HWC

* fix flake8 issue on test_transforms.py

    git ls-files | grep '\.py$' | xargs -n1 -P`nproc` autopep8 -i

Previous assertion only allows numbers for `fill`, which rules out the possibility to pad multi-channel colors. For example, now fill=(255,255,255) or fill="white" can be used to achieve white padding for rgb inputs.

YCbCr breaks the code without warning. Changed the code to make `ToTensor()` cope with all PIL image types.

* Add SqueezeNet 1.0 and 1.1 models
* Selectively avoid inplace in SqueezeNet
* Use Glorot uniform initialization in SqueezeNet
* Make all ReLU in SqueezeNet in-place
* Add pretrained SqueezeNet 1.0 and 1.1
* Minor fixes in SqueezeNet models