You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
I just found that dropout_op often outputs tensor with all 0s even with a big size. I test the dropout_op with GPU by simple codes (the dropout_prob is 0.1 and the input data is 1s with shape [64, 32, 512]), and print the max and min of the output tensor 10 times as following:
Bugs may exist in the GPU kernel of dropout_op.
I just found that dropout_op often outputs tensor with all 0s even with a big size. I test the dropout_op with GPU by simple codes (the
dropout_prob
is 0.1 and the input data is 1s with shape[64, 32, 512]
), and print the max and min of the output tensor 10 times as following:The test codes are as following:
Running the code with CPU gets the proper outputs.
The text was updated successfully, but these errors were encountered: