The trainable state of the weights and the trainable state of the layer

If I set the trainable parameter to True in the build method of the custom layer to create the weights, then during training I set custom_layer.trainable to False to train 5 epochs, and then set custom_layer.trainable to true to train 5 epochs, via self.add_weight The resulting weights don’t get updated during the training of 10 epochs, and I’m wondering why.

Hi @wen_chao_zhou. Welcome to the forum.
Can you please share your code with toy data e.g. in a Colab so that people cna have a better idea of your project?
Thank you.