Continuing training from a frozen graph


(Andrzej) #1

Is there some indicator in logs that the model was loaded (https://github.com/mozilla/DeepSpeech#continuing-training-from-a-frozen-graph)? I used pre-trained model to boost learning but there is only a slight difference in results, so I wonder if I’ve done everything right.

python -u DeepSpeech.py --n_hidden 2048 --initialize_from_frozen_model output_graph.pb --train_files train.csv --dev_files valid.csv --test_files test.csv --train_batch_size 64 --dev_batch_size 64 --test_batch_size 64 --epoch 1 --validation_step 1 --dropout_rate 0.30 --default_stddev 0.046875 --learning_rate 0.0001 --checkpoint_dir checkpoint_transfer/ --summary_secs 600 --summary_dir tensorboard_transfer/


(Reuben Morais) #2

Yes. At the beginning of training we log "Initializing from frozen model: " followed by the path of the frozen model.


(Vincent Foucault) #3

Yes, it seems, but how do you call your new model, for inferences ??

you called the existing model to adapt : --initialize_from_frozen_model
did you save the new, modified one ? : --export_dir

test your new model.

I personnaly tested with my homemade limited model, adapting my model with my 8 years old son (first time).
Inferences, not perfect were good.
Without adaptation, impossible to read a correct inference about my child voice.
So, frozen_model adaptation successed !
hope to help