dd766c68d9
LeakyReLU returns a tensor with float64 dtype. It is stupid, but this line actually produces a float64 array: ``` 0.5*np.array(0.2, dtype=np.float32) ``` The theano nnet.relu function does something similar like this with the LeakyReLU alpha parameter, which lead to a float64 tensor. The solution is to not cast the alpha to float32. Furthermore I tighten the `test_utils.layer_test`. It is now required that the layer's output dtype is equal to the input dtype. |
||
---|---|---|
.. | ||
integration_tests | ||
keras | ||
test_loss_masking.py | ||
test_loss_weighting.py |