Pre-trained model is not loaded

I am trying to do some fine tune of the pre-trained model with common voice en. The version of the DeepSpeech is v0.6.1.

This is my parameters:

python3 -u DeepSpeech.py \
  --train_files /home/ubuntu/clips/train.csv \
  --dev_files /home/ubuntu/clips/dev.csv \
  --test_files /home/ubuntu/clips/test.csv \
  --train_batch_size 64 \
  --dev_batch_size 64 \
  --test_batch_size 64 \
  --epochs 3 \
  --report_count 20 \
  --learning_rate 0.000005\
  --n_hidden 2048 \
  --checkpoint_dir  /home/ubuntu/commonvoice/checkoutpoint \
  --alphabet_config_path /home/ubuntu/DeepSpeech/data/alphabet.txt \
  --lm_binary_path /home/ubuntu/commonvoice/lm.binary \
  --lm_trie_path /home/ubuntu/commonvoice/trie \
  --export_dir /home/ubuntu/commonvoice/export \
  "$@"

The lm_binary, trie and pre-trained model output_graph.pb(put in checkoutpoint directory) was downloaded with curl -LO https://github.com/mozilla/DeepSpeech/releases/download/v0.6.1/audio-0.6.1.tar.gz.

But it seems that the pre-trained model is not loaded bucause the loss starts at 120 and stays at 90.

Can you be more explicit? Common Voice is already included as much as I remember, so you would be overfitting.

There is not “it seems”: training logs would say explicitely that variables are restored.

I don’t see any CUDNN flags, in your command line.

Is this a typo ? Those are audio samples, not checkpoint.

Also, you need to extract the tarball and point to the directory containing the checkpoints.

Sorry, I posted the wrong link, it’s from https://github.com/mozilla/DeepSpeech/releases/download/v0.6.1/deepspeech-0.6.1-models.tar.gz.

I have used deepspeech-0.6.1-models/output_graph.pb for the pre-tained model.

Also this is the command line:

(deepspeech-train-venv) ubuntu@ip-172-31-36-41:~/DeepSpeech$ sh ../commonvoice/train.sh
+ [ ! -f DeepSpeech.py ]
+ python3 -u DeepSpeech.py --train_files /home/ubuntu/clips/train.csv --dev_files /home/ubuntu/clips/dev.csv --test_files /home/ubuntu/clips/test.csv --train_batch_size 128 --dev_batch_size 128 --test_batch_size 6128 --epochs 1 --report_count 20 --learning_rate 0.000005 --n_hidden 2048 --checkpoint_dir /home/ubuntu/commonvoice/checkoutpoint --alphabet_config_path /home/ubuntu/DeepSpeech/data/alphabet.txt --lm_binary_path /home/ubuntu/commonvoice/lm.binary --lm_trie_path /home/ubuntu/commonvoice/trie --export_dir /home/ubuntu/commonvoice/export
I Initializing variables...
I STARTING Optimization
Epoch 0 |   Training | Elapsed Time: 0:05:25 | Steps: 94 | Loss: 93.352414

output_graph.pb is in /home/ubuntu/commonvoice/checkoutpoint

So please read carefully the documentation, this is not a checkpoint, this is the final model. Please use proper link: https://github.com/mozilla/DeepSpeech/releases/download/v0.6.1/deepspeech-0.6.1-checkpoint.tar.gz