def main(args):
'''
There are 5 simple steps to this program
'''
# 1. combine all data into 2 dataframes (train, valid)
print("Getting data from arguments")
train_dataprops, df_train = combine_all_wavs_and_trans_from_csvs(
args.train_files)
valid_dataprops, df_valid = combine_all_wavs_and_trans_from_csvs(
args.valid_files)
# check any special data model requirments e.g. a spectrogram
if (args.model_arch == 1):
model_input_type = "mfcc"
elif (args.model_arch == 2 or args.model_arch == 5):
print("Spectrogram required")
# spectrogram = True
model_input_type = "spectrogram"
else:
model_input_type = "mfcc"
## 2. init data generators
print("Creating data batch generators")
traindata = BatchGenerator(dataframe=df_train,
training=True,
batch_size=args.batchsize,
model_input_type=model_input_type)
validdata = BatchGenerator(dataframe=df_valid,
training=False,
batch_size=args.batchsize,
model_input_type=model_input_type)
inputs, outputs = traindata.get_batch(0)
input_shape = inputs['the_input'].shape[1:]
output_shape = inputs['the_labels'].shape[1:]
output_dir = os.path.join('checkpoints/results', 'model')
if not os.path.isdir(output_dir):
os.makedirs(output_dir)
## 3. Load existing or create new model
if args.loadcheckpointpath:
# load existing
print("Loading model")
cp = args.loadcheckpointpath
assert (os.path.isdir(cp))
model_path = os.path.join(cp, "model")
# assert(os.path.isfile(model_path))
model = load_model_checkpoint(model_path)
print("Model loaded")
else:
# new model recipes here
print('New model DS{}'.format(args.model_arch))
if (args.model_arch == 0):
# DeepSpeech1 with Dropout
model = ds1_dropout(input_dim=26,
fc_size=args.fc_size,
rnn_size=args.rnn_size,
dropout=[0.1, 0.1, 0.1],
output_dim=29)
elif (args.model_arch == 1):
# DeepSpeech1 - no dropout
model = ds1(input_dim=26,
fc_size=args.fc_size,
rnn_size=args.rnn_size,
output_dim=29)
elif (args.model_arch == 2):
# DeepSpeech2 model
model = ds2_gru_model(input_dim=161,
fc_size=args.fc_size,
rnn_size=args.rnn_size,
output_dim=29)
elif (args.model_arch == 3):
# own model
model = ownModel(input_shape,
output_shape,
fc_size=args.fc_size,
rnn_size=args.rnn_size,
dropout=[0.1, 0.1, 0.1],
output_dim=29)
elif (args.model_arch == 4):
# graves model
model = graves(input_dim=26,
rnn_size=args.rnn_size,
output_dim=29,
std=0.5)
elif (args.model_arch == 5):
# cnn city
model = cnn_city(input_dim=161,
fc_size=args.fc_size,
rnn_size=args.rnn_size,
output_dim=29)
elif (args.model_arch == 6):
# constrained model
model = const(input_dim=26,
fc_size=args.fc_size,
rnn_size=args.rnn_size,
output_dim=29)
else:
raise ("model not found")
print(model.summary(line_length=140))
# required to save the JSON
save_model(model, output_dir)
if (args.opt.lower() == 'sgd'):
opt = SGD(lr=args.learning_rate,
decay=1e-6,
momentum=0.9,
nesterov=True,
clipnorm=5)
elif (args.opt.lower() == 'adam'):
opt = Adam(lr=args.learning_rate,
beta_1=0.9,
beta_2=0.999,
epsilon=1e-8,
clipnorm=5)
elif (args.opt.lower() == 'nadam'):
opt = Nadam(lr=args.learning_rate,
beta_1=0.9,
beta_2=0.999,
epsilon=1e-8,
clipnorm=5)
else:
raise Exception("optimiser not recognised")
model.compile(optimizer=opt, loss=ctc)
## 4. train
if args.train_steps == 0:
args.train_steps = len(df_train.index) // args.batchsize
# print(args.train_steps)
# we use 1/xth of the validation data at each epoch end to test val score
if args.valid_steps == 0:
args.valid_steps = (len(df_valid.index) // args.batchsize)
# print(args.valid_steps)
if args.memcheck:
cb_list = [MemoryCallback()]
else:
cb_list = []
if args.tensorboard:
tb_cb = TensorBoard(log_dir='./tensorboard/{}/'.format(args.name),
write_graph=False,
write_images=True)
cb_list.append(tb_cb)
y_pred = model.get_layer('ctc').input[0]
input_data = model.get_layer('the_input').input
report = K.function([input_data, K.learning_phase()], [y_pred])
report_cb = ReportCallback(report, validdata, model, args.name, save=True)
cb_list.append(report_cb)
model.fit_generator(
generator=traindata.next_batch(),
steps_per_epoch=args.train_steps,
epochs=args.epochs,
callbacks=cb_list,
validation_data=validdata.next_batch(),
validation_steps=args.valid_steps,
)
## These are the most important metrics
print("Mean WER :", report_cb.mean_wer_log)
print("Mean LER :", report_cb.mean_ler_log)
print("NormMeanLER:", report_cb.norm_mean_ler_log)
# export to csv?
K.clear_session()
tr_batch_size) # 每个 epoch 中包含的 batch 数
display_batch = int(tr_batch_num / display_num) # 每训练 display_batch 之后输出一次
saver = tf.train.Saver(max_to_keep=10) # 最多保存的模型数量
for epoch in range(max_max_epoch):
_lr = 1e-4
if epoch > max_epoch:
_lr = _lr * ((decay)**(epoch - max_epoch))
print('EPOCH %d， lr=%g' % (epoch + 1, _lr))
start_time = time.time()
_costs = 0.0
_accs = 0.0
show_accs = 0.0
show_costs = 0.0
for batch in range(tr_batch_num):
fetches = [accuracy, cost, train_op]
X_batch, X_tag_batch, y_batch = data_train.next_batch(tr_batch_size)
feed_dict = {
X_inputs: X_batch,
X_tag_inputs: X_tag_batch,
y_inputs: y_batch,
lr: _lr,
batch_size: tr_batch_size,
keep_prob: 0.5
}
_acc, _cost, _ = sess.run(
fetches, feed_dict) # the cost is the mean cost of one batch
_accs += _acc
_costs += _cost
show_accs += _acc
show_costs += _cost
if (batch + 1) % display_batch == 0:
