run_infer_golden.sh

#! /usr/bin/env bash

cd ../.. > /dev/null

# download language model(We skip this part since we already have 70GB model)
#cd models/lm > /dev/null
#sh download_lm_ch.sh
#if [ $? -ne 0 ]; then
#    exit 1
#fi
#cd - > /dev/null


# download well-trained model
cd models/aishell > /dev/null
sh download_model.sh
if [ $? -ne 0 ]; then
    exit 1
fi
cd - > /dev/null


# infer
CUDA_VISIBLE_DEVICES=0 \
python -u infer.py \
--num_samples=10 \
--trainer_count=1 \
--beam_size=300 \
--num_proc_bsearch=8 \
--num_conv_layers=2 \
--num_rnn_layers=3 \
--rnn_layer_size=1024 \
--alpha=2.6 \
--beta=5.0 \
--cutoff_prob=0.99 \
--cutoff_top_n=40 \
--use_gru=True \
--use_gpu=True \
--share_rnn_weights=False \
--infer_manifest='data/aishell/manifest.test' \
--mean_std_path='models/aishell/mean_std.npz' \
--vocab_path='models/aishell/vocab.txt' \
--model_path='models/aishell/params.tar.gz' \
--lang_model_path='models/lm/zhidao_giga.klm' \
--decoding_method='ctc_beam_search' \
--error_rate_type='cer' \
--specgram_type='linear'

if [ $? -ne 0 ]; then
    echo "Failed in inference!"
    exit 1
fi


exit 0