Hi, I'm trying to migrate this sample to TL2.
But I found in train.py : 116 that optimizer only works for Decoder's variables, and there are codes restoring weights from .npz file to Encoders before training (and these weights will not be changed during training).
In order to migrate to TL2, and without pretrained Encoder's weights at hand, I need to write a model composed of 3 Encoders and 1 Decoders, and then config a tf.keras.optimizers.Adam optimizer to apply gradients to all trainable variables of the model, am i right?
Hi, I'm trying to migrate this sample to TL2. But I found in
train.py : 116that optimizer only works for Decoder's variables, and there are codes restoring weights from .npz file to Encoders before training (and these weights will not be changed during training). In order to migrate to TL2, and without pretrained Encoder's weights at hand, I need to write a model composed of 3 Encoders and 1 Decoders, and then config atf.keras.optimizers.Adamoptimizer to apply gradients to all trainable variables of the model, am i right?