[common]
# method can be one of the followings - train,predict,load
mode = train
#ex: gpu0,gpu1,gpu2,gpu3
context = gpu0
# checkpoint prefix, check point will be saved under checkpoints folder with prefix
prefix = test_fc
# when mode is load or predict, model will be loaded from the file name with model_file under checkpoints
model_file = test_fc-0001
batch_size = 2
# log will be saved by the log_filename
log_filename = test.log
# checkpoint set n to save checkpoints after n epoch
save_checkpoint_every_n_epoch = 1
save_checkpoint_every_n_batch = 1000
is_bi_graphemes = False
tensorboard_log_dir = tblog/libri_sample
# if random_seed is -1 then it gets random seed from timestamp
mx_random_seed = -1
random_seed = -1

[data]
train_json = ./Libri_sample.json
test_json = ./Libri_sample.json
val_json = ./Libri_sample.json

language = en
width = 161
height = 1
channel = 1
stride = 1

[arch]
channel_num = 32
conv_layer1_filter_dim = [11, 41]
conv_layer1_stride = [2, 2]
conv_layer2_filter_dim = [11, 21]
conv_layer2_stride = [1, 2]

num_rnn_layer = 3
num_hidden_rnn_list = [1760, 1760, 1760]
num_hidden_proj = 0

num_rear_fc_layers = 0
num_hidden_rear_fc_list = []
act_type_rear_fc_list = []

#network: lstm, bilstm, gru, bigru
rnn_type = bigru
#vanilla_lstm or fc_lstm (no effect when network_type is gru, bigru)
lstm_type = fc_lstm
is_batchnorm = True

[train]
num_epoch = 70

learning_rate = 0.005
# constant learning rate annealing by factor
learning_rate_annealing = 1.1
# supports only sgd and adam
optimizer = adam
# for sgd
momentum = 0.9
# set to 0 to disable gradient clipping
clip_gradient = 0

initializer = Xavier
init_scale = 2
factor_type = in
weight_decay = 0.00001
# show progress every nth batches
show_every = 1
save_optimizer_states = True
normalize_target_k = 2
# overwrite meta files(feats_mean,feats_std,unicode_en_baidu_bi_graphemes.csv)
overwrite_meta_files = True
enable_logging_train_metric = True
enable_logging_validation_metric = True

[load]
load_optimizer_states = True
is_start_from_batch = False