Continuing training from a frozen graph - Unable to use the trained new model to predict new data

(kmrambo) #1

I have continued the training from frozen graph and the training completed successfully. Below is the command I used for training purpose:

python --n_hidden 2048 --initialize_from_frozen_model /Users/kmrambo/Deepspeech/models/output_graph.pb --checkpoint_dir fine_tuning_checkpoints --epoch 3 --train_files /Users/kmrambo/DeepSpeech-master/sampledata/cv-valid-train.csv --dev_files /Users/kmrambo/DeepSpeech-master/sampledata/cv-valid-dev.csv --test_files /Users/kmrambo/DeepSpeech-master/sampledata/cv-valid-test.csv --learning_rate 0.0001 —decoder_library_path /Users/kmrambo/Documents/GitHub/DeepSpeech/

Now I see that the Checkpoint directory has been updated with the checkpoint files after training. Below is a snapshot of the outputs in my checkpoint directory:

I now want to use this re-trained model to predict new data. I see that the following is the command to predict new data:

./deepspeech models/output_graph.pb audio_input.wav models/alphabet.txt models/lm.binary models/trie

But in the above command, I have to supply the output.pb , trie files of the re-trained model. Where do I find those files or should I follow any additional steps for those files to get generated after re-training from a frozen model?

  • OS Platform and Distribution (e.g., Linux Ubuntu 16.04): MacOS High Sierra
  • TensorFlow version (use command below): 1.4.0
  • Python version: 2.7.14
  • GPU model and memory: No GPU, Only CPU

(Reuben Morais) #2

Rerun the training command and pass --export_dir /path/to/some/folder. output_model.pb will be exported into the folder you pass.

(kmrambo) #3

Thank You @reuben. It worked!!!