[GitHub] [incubator-mxnet] Zha0q1 commented on issue #20265: [v1.x] [MKLDNN] 2 Conv Tests Failed with MKLDNN + ACL

[GitBox <gi...@apache.org>]

Zha0q1 commented on issue #20265:
URL: https://github.com/apache/incubator-mxnet/issues/20265#issuecomment-843570689


   Here is the source code of the first test case `test_operator.test_convolution_grouping `
   https://github.com/apache/incubator-mxnet/blob/master/tests/python/unittest/test_operator.py#L1674-L1703
   
   ```python
   @with_seed()
   def test_convolution_grouping():
       for dim in [1, 2, 3]:
           num_filter = 4
           for num_group in [1, 2]:
               kernel = (3,) * dim
               shape = (1, 4) + (9,) * dim
   ```
   The test will only fail when `dim=2` and `num_group=2`, i.e. when `kernel=(3, 3)` and `shape=(1, 4, 9, 9)`. 
   
   So the test case will check if the Conv op will produce the same output and gradients (for x, w, b) when `num_group=2` as those of when we do the "grouping" manually. So it looks like for all of output, x gradient, w gradient, and b gradient the first halves of then tensors are the same, but the second halves differ.
   
   ```
   Mismatched elements: 98 / 196 (50%)
   Max absolute difference: 16.347755
   Max relative difference: 23.329699
    x: array([[[[ 1.597207e-01,  1.540166e+00,  6.755731e+00, -5.728347e+00,
              1.296894e+00, -1.326614e+00,  6.375989e+00],
            [ 1.170041e+00,  3.974693e-01,  5.603638e+00, -1.179677e+00,...
    y: array([[[[ 1.597207e-01,  1.540166e+00,  6.755731e+00, -5.728347e+00,
              1.296894e+00, -1.326614e+00,  6.375989e+00],
            [ 1.170040e+00,  3.974691e-01,  5.603638e+00, -1.179677e+00,...
   ```


-- 
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

For queries about this service, please contact Infrastructure at:
users@infra.apache.org



---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscribe@mxnet.apache.org
For additional commands, e-mail: issues-help@mxnet.apache.org