Trainable = False for layers/embedding layer

[sravanbabuiitm]

Hi,

While training a deep network with pretrained embeddings, LSTM , one would have a choice to freeze the pretrained embeddings or let the pretrained embeddings get trained too.

Other than not setting a trainer for the embeddings layer in which case it wont get updated, is there a diff way to explicitly mention while creating the layer to not train/update it ?

I need to create a trainer for rest of the network, which I m trying to avoid.

[szha]

you can do net.layer1.collect_params().setattr('grad_req', 'null') to freeze the weights in layer1.