Hi guys,
Could you please help me to understand why the model is giving a lower accuracy?
I fine-tuned the model 0.9.3 with a small dataset. And by evaluation test, before training, the accuracy was 19% (WER). However, after trained, it changed to 28%. Why?
I used the same parameters as mentioned in https://github.com/mozilla/DeepSpeech/releases
I just didn’t use augmentation and cache_for_epochs. The other parameters are the same, as follow:
!python DeepSpeech.py --n_hidden 2048 \
--checkpoint_dir fine_tuning_checkpoints \
--train_files /data/train.csv \
--dev_files /data/dev.csv \
--test_files /data/test.csv \
--train_batch_size 128 \
--dev_batch_size 128 \
--test_batch_size 128 \
--n_hidden 2048 \
--learning_rate 0.0001 \
--dropout_rate 0.40 \
--epochs 200 \
--export_dir output_models/ \
--train_cudnn
Thank you