PierrickPochelu
PierrickPochelu
Why use keras if you can use more dedicated inference frameworks such as tensorrt, openvino, onnx-runtime, TF ...
If you ask the "`crossentropy`" or "`categorical_crossentropy`" metrics you will have matching results. If you ask "`binary_crossentropy`" it will not match. I am curious to know if someone has more...
I dive deeper into BinaryCrossentropy, here is 2 important pieces of information: 1) There are 2 potential cross entropy distance: Penalize when label=1 and pred=0 --> `loss=label * log(pred)`. The...
The models output have different meaning according the final activation function (sigmoid or argmax). Therefore, the loss (or the person developing the loss) has different input expectations. Let's take an...
Agree, I removed the arg and the conditional in my last commit.
I have seen some logical mistakes in the unit tests. This wrong input have been used in multiple unit tests : [0.5, 0.89, 0.6] in combination of from_logits=False . Assuming...
The 8 wrong test units have been corrected.
Any other edge case to handle ?