deepmedic Output node names

Hi, I am trying to freeze the graph before converting it to a TFTRT model. The freeze_graph CLI command in tf requires a output_node_names argument. I looked at the graph in tensorboard and found net/Softmax. But there are so many other nodes. Can you share what name you used for the output_nodes?

Mar 11 '21 13:03 chihyanghsu0805

Hello, thanks for your interest!

Yes, sorry, I have not been naming the nodes in the graph :-/

What you need for prediction is indeed this tensor created by the softmax: https://github.com/deepmedic/deepmedic/blob/7a646c912920fa09c447613b3d4416f2bb8feca6/deepmedic/neuralnet/blocks.py#L212

There is only 1 softmax, so that should do.

Let me know if this works out or not.

Thanks!

Mar 12 '21 11:03 Kamnitsask

Hi, thank you for the response. I found that a TFTRT model is only supported by SaveModel in TF2, frozen graph does not work anymore, or easily converted. Therefore, I have two more questions. First, if I uncommented the TF2 commands (just the saver part), would it save SavedModels? Second, the model conversion also need input node names, can you share it?

Mar 17 '21 18:03 chihyanghsu0805