зеркало из https://github.com/mozilla/DeepSpeech.git
29 строки
1002 B
Bash
Executable File
29 строки
1002 B
Bash
Executable File
#!/bin/sh
|
|
|
|
set -xe
|
|
|
|
ldc93s1_dir="./data/ldc93s1-tc"
|
|
ldc93s1_csv="${ldc93s1_dir}/ldc93s1.csv"
|
|
|
|
epoch_count=$1
|
|
|
|
if [ ! -f "${ldc93s1_dir}/ldc93s1.csv" ]; then
|
|
echo "Downloading and preprocessing LDC93S1 example data, saving in ${ldc93s1_dir}."
|
|
python -u bin/import_ldc93s1.py ${ldc93s1_dir}
|
|
fi;
|
|
|
|
# Force only one visible device because we have a single-sample dataset
|
|
# and when trying to run on multiple devices (like GPUs), this will break
|
|
export CUDA_VISIBLE_DEVICES=0
|
|
|
|
python -u DeepSpeech.py --noshow_progressbar --noearly_stop \
|
|
--train_files ${ldc93s1_csv} --train_batch_size 1 \
|
|
--feature_cache '/tmp/ldc93s1_cache' \
|
|
--dev_files ${ldc93s1_csv} --dev_batch_size 1 \
|
|
--test_files ${ldc93s1_csv} --test_batch_size 1 \
|
|
--n_hidden 100 --epochs $epoch_count \
|
|
--max_to_keep 1 --checkpoint_dir '/tmp/ckpt' \
|
|
--learning_rate 0.001 --dropout_rate 0.05 --export_dir '/tmp/train' \
|
|
--lm_binary_path 'data/smoke_test/vocab.pruned.lm' \
|
|
--lm_trie_path 'data/smoke_test/vocab.trie'
|