FIX: Conv[DepthWise] forward implementation and add tests
Context
Call to Conv(kernel_size=[3,3], stride = [2,2])
was returning a random Tensor.
After careful review, it appears the new Conv implementation !101 (merged), if faster, splits the computation in more small cases that were not all tested. For example, Conv(kernel_size=[3,3], stride = [2,2])
and Conv(kernel_size=[3,3], stride = [2,2])
don't go through the same loop.
This MR adds tests for each of the following cases for Conv
and ConvDepthWise
:
- kernel [3, 3]
- no stride, no dilation
- stride [2,2], no dilation
- stride [2,2], dilation [2,2]
- kernel [1,1]
- no stride, no dilation
- stride [3,3], no dilation
- stride [3,3], dilation [2,2]
- kernel [5,5]
- no stride, no dilation
- stride [2,2], no dilation
- stride [2,2], dilation [2,2]
Some issues were found thanks to these new tests and fixed.
Merge request reports
Activity
added Fix 🔥🔥 StatusReview Ready TopicOperator Unit-test 🧪 labels
assigned to @pineapple
enabled an automatic merge when all merge checks for 0cd78e35 pass
mentioned in commit 7db58d25