About conv operation in HW4

[DeliMm]

Maybe there is a mistake in the conv's gradient function with the following place

CMU10-714/homework/hw4/python/needle/ops.py

Line 613 in 99180c4

X_grad = conv(out_grad, W_permute, padding=K-1-self.padding)

I have a test case with :

• A's shape: (1, 16, 16, 1)
• B's shape: (7, 7, 1, 1)
• stride=3 & padding=2

When I test the above case, there are some errors with

tests/hw4/test_conv.py::test_op_conv[backward-needle.backend_ndarray.ndarray_backend_cuda-Z_shape16-W_shape16-1-0]  
- ValueError: operands could not be broadcast together with shapes (1,16,16,1) (1,17,17,1)

This function may not have considered the situation of a large convolution kernel during implementation, especially when $(H - 2p + k) \pmod s \ne 0$.

May I ask if it‘s possible to check such case and carry out improvement/perfection? Thanks for your reply.

[Skylight-Lark]

Maybe there is a mistake in the conv's gradient function with the following place

CMU10-714/homework/hw4/python/needle/ops.py

Line 613 in 99180c4

X_grad = conv(out_grad, W_permute, padding=K-1-self.padding)

I have a test case with :

• A's shape: (1, 16, 16, 1)
• B's shape: (7, 7, 1, 1)
• stride=3 & padding=2

When I test the above case, there are some errors with

tests/hw4/test_conv.py::test_op_conv[backward-needle.backend_ndarray.ndarray_backend_cuda-Z_shape16-W_shape16-1-0]  
- ValueError: operands could not be broadcast together with shapes (1,16,16,1) (1,17,17,1)

This function may not have considered the situation of a large convolution kernel during implementation, especially when (H−2p+k)(mods)≠0.

May I ask if it‘s possible to check such case and carry out improvement/perfection? Thanks for your reply.

@PKUFlyingPig @DeliMm
Hello! I think the caculation of OFM's size is wrong:

CMU10-714/homework/hw4/python/needle/ops.py

Line 595 in 99180c4

out_H, out_W = (H-K+1)//self.stride, (W-K+1)//self.stride

As the pytorch doc says:
https://pytorch.org/docs/stable/generated/torch.nn.Conv2d.html#torch.nn.Conv2d
The right size is:
(H + 2 * P - K) // self.stride + 1
So the caculation of conv gradient has also some problems, like X.grad conv padding