The check is too strong when run with: TF_XLA_FLAGS=--tf_xla_auto_jit=1.
Compiling with xla changes the computation, resulting in failure with the following message:
Traceback (most recent call last):
File "/opt/tensorflow/nvidia-examples/OpenSeq2Seq/open_seq2seq/losses/sequence_loss_test.py", line 60, in test_compute_loss
self.assertEqual(loss1, loss2)
AssertionError: 32.645237 != 32.64524
Only comparing the first 4 decimals makes the test succeed again
okuchaiev
commented
5 years ago
…eLossTest
The check is too strong when run with: TF_XLA_FLAGS=--tf_xla_auto_jit=1. Compiling with xla changes the computation, resulting in failure with the following message:
Traceback (most recent call last): File "/opt/tensorflow/nvidia-examples/OpenSeq2Seq/open_seq2seq/losses/sequence_loss_test.py", line 60, in test_compute_loss self.assertEqual(loss1, loss2) AssertionError: 32.645237 != 32.64524
Only comparing the first 4 decimals makes the test succeed again