tueimage / crowdskin-bsc-2020 Goto Github PK

"conv_base.trainable = False" should freeze the convolutional base but model.summary returns:
Total params: 33,589,826
Trainable params: 33,589,826
Non-trainable params: 0

This was fixed by replacing "conv_base.trainable = False" with:
for layer in conv_base.layers:
layer.trainable = False

The results returned by the multitask model with frozen weights performs worse then baseline.
auc for frozen and unfrozen weights are shown in the plot below:

Is it still possible to compare the baseline and the multitask model with each other when the baseline uses a frozen conv_base and the multitask model uses a unfrozen conv_base?