Andrew Lavin

Winograd F(4x4, 5x5)

1

Here are transforms for Winograd F(4x4, 5x5). That means a 5x5 kernel with a 4x4 output tile. I imagine the code should be a relatively simple adaptation of F(6x6, 3x3),...

Support for cl_khr_image2d_from_buffer

2

It appears that OpenCL 1.2 extension cl_khr_image2d_from_buffer is not completely supported. This allows zero-copy creation of a 2D image from a buffer, provided that the buffer satisfies CL_DEVICE_IMAGE_BASE_ADDRESS_ALIGNMENT and CL_DEVICE_IMAGE_PITCH_ALIGNMENT...

https://github.com/gan3sh500/octaveconv-pytorch/blob/d4d1ee0a4397a3405cc149ca02fc1dff5e654a85/octconv.py#L14 The emerging consensus is that this is probably the wrong way to implement strided OctConv. See https://github.com/terrychenism/OctaveConv/issues/4

Support cuDNN v4

9

Placeholder for cuDNN v4 support development effort. I am only set up to test on Linux, so it would be helpful if somebody else could test on Windows and Mac.

Strided convolutions not implemented

2

It appears that the stride parameter does not have the desired effect in this implementation of OctConv. It is used to compute the padding, but otherwise has no effect on...

Fixes #319.

**Is your feature request related to a problem? Please describe.** torch.nn.Conv2d has a "groups" attribute set to a value greater than 1 for EfficientNet, ResNeXt, ConvNeXt, and other current vision...

Fixes #77.