Support reuse cuDNNLSTM-trained checkpoints by multi-layer LSTM(Block)Cell
* Add tensor stitching/partition in RNNParamSaveable for saving/restoring properly shaped and formatted weights/biases to share w/ LSTM(Block)Cell * Add remapped names for canonical tensors during saving. * Unittests PiperOrigin-RevId: 159634913
Loading
Please sign in to comment