зеркало из https://github.com/mozilla/DeepSpeech.git
32 строки
1.0 KiB
Bash
Executable File
32 строки
1.0 KiB
Bash
Executable File
#!/bin/sh
|
|
|
|
set -xe
|
|
|
|
ldc93s1_dir="./data/smoke_test"
|
|
ldc93s1_csv="${ldc93s1_dir}/ldc93s1.csv"
|
|
|
|
if [ ! -f "${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 \
|
|
--dev_files ${ldc93s1_csv} --dev_batch_size 1 \
|
|
--test_files ${ldc93s1_csv} --test_batch_size 1 \
|
|
--n_hidden 100 --epochs 1 \
|
|
--max_to_keep 1 --checkpoint_dir '/tmp/ckpt_bytes' --bytes_output_mode \
|
|
--learning_rate 0.001 --dropout_rate 0.05 \
|
|
--scorer_path 'data/smoke_test/pruned_lm.bytes.scorer' | tee /tmp/resume.log
|
|
|
|
if ! grep "Loading best validating checkpoint from" /tmp/resume.log; then
|
|
echo "Did not resume training from checkpoint"
|
|
exit 1
|
|
else
|
|
exit 0
|
|
fi
|