Error on loading 0.5.1 checkpoints with current master DeepSpeech codebase

I downloaded deepspeech checkpoint from github. then i kept the checkpoint in the appropriate directory. I want to retrain the pre-trained model with my own dataset. I run the following and following error. Any suggestion how to resolve the error.

deepspeech-venv) nayan@nayan-virtual-machine:~/Desktop/DeepSpeech$ python3 --n_hidden 2048 --checkpoint_dir /home/nayan/.local/share/deepspeech/checkpoints/deepspeech-0.5.0-checkpoint --epochs 3 --train_files data/david/david.csv --dev_files data/david/david.csv --test_files data/david/david.csv --learning_rate 0.0001

Restoring from checkpoint failed. This is most likely due to a Variable name or other graph key that is missing from the checkpoint. Please ensure that you have not altered the graph expected based on the checkpoint. Original error:

Key cudnn_lstm/rnn/multi_rnn_cell/cell_0/cudnn_compatible_lstm_cell/bias not found in checkpoint
[[node save/RestoreV2 (defined at ]]

1 Like

Yes, @reuben landed changes on the current master branch that leverages CuDNN for faster training process. I don’t think we intend to support loading a non-cudnn checkpoint.

1 Like

We don’t currently have a flag or anything like that in the tree because this situation will only last until our next release model with 0.6.0, but you can convert the 0.5.1 model to the new graph. A converted version which we use for our tests is available here:

1 Like