It has been suggested to run these numerical tests prone to cause numerical issues (flaky tests) for a high number of iterations before commit / PR is done. The MXNET_TEST_COUNT variable can be used before running the test with nose as in:
MXNET_TEST_COUNT=100000 nosetests -s -v tes_xxx.py:test_name On Tue, Jul 3, 2018 at 7:17 PM Alex Zai <[email protected]> wrote: > We are preparing a stable release of MKL-DNN integration in 1.3.0 > (experimental since 1.2.0), which supports acceleration of operations such > as Convolution, Deconvolution, FullyConnected, Pooling, Batch > Normalization, Activation, LRN, Softmax. Currently the RNN operator is not > supported as the MKL-DNN API is still experimental; however, they hope to > release a more stable version RNN API this or next week in MKL-DNN 0.15. > > We will have CPP unit test support on these operators and I am planning to > write python unit tests to compare a RNN network's results from the MKLDNN > backend with that of the GPU to test accuracy. Is there any additional > coverage that you think we should cover in the next two weeks? > > Alex >
