Fix auto_pad shape inference bug #2028
Conversation
Merge master
Merge master
| fail_shape_inference("Stride is bigger than Kernel shape"); | ||
| } | ||
| int64_t total_pad = residual == 0 ? kernel_shape[i] - stride : kernel_shape[i] - residual; | ||
| if (total_pad < 0) |
There was a problem hiding this comment.
LGTM.
Just 1 naive question - do we really need this mutation ? Even if total pads is negative, in reality, this means just ignoring a few values in the raw data ("unpad" a few values) right ? Also, I think it should not affect the output dim value when we use the formula - floor((in_dim - kernel_shape + total_pad) / s) + 1. Does this seem reasonable ?
There was a problem hiding this comment.
I checked the corner cases and I think probably there is no impact if you leave total_pad as negative value, but for simplicity and clarity, it's better to keep it as non-negative value.
There was a problem hiding this comment.
@linkerzhang Please review this PR
|
can you also add a test case to cover this bug case please? Thank you! |
* fix auto_pad shape inference bug * add test case for the case stride bigger than kernel shape * Update shape_inference_test.py
* fix auto_pad shape inference bug * add test case for the case stride bigger than kernel shape * Update shape_inference_test.py
No description provided.