Hi all,
I am working with a resnet50 that is ending with an fc layer.
I would like to add a softmax layer in the optimization step when the alls script is applied.
Using logits_layer1 = logits_layer(resnet50/fc1, softmax)
results in a layer with FP32 precision
However, I need UINT8 precision in the output layer.
I found that there are additional parameters in the logits_layer function e.g. in the efficientnet from the model zoo : logits_layer1 = logits_layer(efficientnet_lite0/fc1, softmax, -1, cpu)
parameter 3 “-1” and parameter 4 “cpu” are not explained in the DFC3.30 documentation.
What are these parameters ? Do they influence the precision of the created layer?