def main(libri_dir=c.DATASET_DIR):
print('Looking for fbank features [.npy] files in {}.'.format(libri_dir))
libri = data_catalog(libri_dir)
# filename speaker_id
# 0 audio/LibriSpeechSamples/train-clean-100-npy/1-100-0001.npy 1
# 1 audio/LibriSpeechSamples/train-clean-100-npy/1-100-0002.npy 1
unique_speakers = libri['speaker_id'].unique() # 251 speaker
transform=transforms.Compose([transforms.ToTensor()])
train_dir = stochastic_mini_batch(libri)
train_loader = DataLoader(train_dir, batch_size=c.BATCH_SIZE, shuffle=True)
model = DeepSpeakerModel(embedding_size=c.EMBEDDING_SIZE,num_classes=c.NUM_SPEAKERS)
optimizer = optim.Adam(model.parameters(), lr=0.001, weight_decay=0.0)
epoch = 0
model.cuda()
summary(model, input_size=(1, 160, 64))
for epoch in range(100):
model.train()
for batch_idx, (data_a, data_p, data_n,label_a,label_p,label_n) in tqdm(enumerate(train_loader)):
data_a, data_p, data_n = data_a.type(torch.FloatTensor),data_p.type(torch.FloatTensor),data_n.type(torch.FloatTensor)
data_a, data_p, data_n = data_a.cuda(), data_p.cuda(), data_n.cuda()
data_a, data_p, data_n = Variable(data_a), Variable(data_p), Variable(data_n)
out_a, out_p, out_n = model(data_a), model(data_p), model(data_n)
triplet_loss = TripletMarginLoss(0.2).forward(out_a, out_p, out_n)
loss = triplet_loss
# compute gradient and update weights
optimizer.zero_grad()
loss.backward()
optimizer.step()
print('selected_triplet_loss', triplet_loss.data)
print("epoch:",epoch)
torch.save(model.state_dict(),"checkpoint_{}.pt".format(epoch))
def main(libri_dir=c.DATASET_DIR):
PRE_TRAIN = c.PRE_TRAIN
logging.info(
'Looking for fbank features [.npy] files in {}.'.format(libri_dir))
libri = data_catalog(libri_dir)
if len(libri) == 0:
logging.warning(
'Cannot find npy files, we will load audio, extract features and save it as npy file'
)
logging.warning('Waiting for preprocess...')
preprocess_and_save(c.WAV_DIR, c.DATASET_DIR)
libri = data_catalog(libri_dir)
if len(libri) == 0:
logging.warning(
'Have you converted flac files to wav? If not, run audio/convert_flac_2_wav.sh'
)
exit(1)
unique_speakers = libri['speaker_id'].unique()
spk_utt_dict, unique_speakers = create_dict(libri['filename'].values,
libri['speaker_id'].values,
unique_speakers)
select_batch.create_data_producer(unique_speakers, spk_utt_dict)
batch = stochastic_mini_batch(libri,
batch_size=c.BATCH_SIZE,
unique_speakers=unique_speakers)
batch_size = c.BATCH_SIZE * c.TRIPLET_PER_BATCH
x, y = batch.to_inputs()
b = x[0]
num_frames = b.shape[0]
train_batch_size = batch_size
#batch_shape = [batch_size * num_frames] + list(b.shape[1:]) # A triplet has 3 parts.
input_shape = (num_frames, b.shape[1], b.shape[2])
logging.info('num_frames = {}'.format(num_frames))
logging.info('batch size: {}'.format(batch_size))
logging.info('input shape: {}'.format(input_shape))
logging.info('x.shape : {}'.format(x.shape))
orig_time = time()
model = convolutional_model(input_shape=input_shape,
batch_size=batch_size,
num_frames=num_frames)
logging.info(model.summary())
gru_model = None
if c.COMBINE_MODEL:
gru_model = recurrent_model(input_shape=input_shape,
batch_size=batch_size,
num_frames=num_frames)
logging.info(gru_model.summary())
grad_steps = 0
if PRE_TRAIN:
last_checkpoint = get_last_checkpoint_if_any(c.PRE_CHECKPOINT_FOLDER)
if last_checkpoint is not None:
logging.info(
'Found pre-training checkpoint [{}]. Resume from here...'.
format(last_checkpoint))
x = model.output
x = Dense(len(unique_speakers),
activation='softmax',
name='softmax_layer')(x)
pre_model = Model(model.input, x)
pre_model.load_weights(last_checkpoint)
grad_steps = int(last_checkpoint.split('_')[-2])
logging.info('Successfully loaded pre-training model')
else:
last_checkpoint = get_last_checkpoint_if_any(c.CHECKPOINT_FOLDER)
if last_checkpoint is not None:
logging.info('Found checkpoint [{}]. Resume from here...'.format(
last_checkpoint))
model.load_weights(last_checkpoint)
grad_steps = int(last_checkpoint.split('_')[-2])
logging.info('[DONE]')
if c.COMBINE_MODEL:
last_checkpoint = get_last_checkpoint_if_any(
c.GRU_CHECKPOINT_FOLDER)
if last_checkpoint is not None:
logging.info(
'Found checkpoint [{}]. Resume from here...'.format(
last_checkpoint))
gru_model.load_weights(last_checkpoint)
logging.info('[DONE]')
#adam = Adam(lr=0.01, beta_1=0.9, beta_2=0.999, epsilon=None, decay=0.0, amsgrad=False)
model.compile(optimizer='adam', loss=deep_speaker_loss)
if c.COMBINE_MODEL:
gru_model.compile(optimizer='adam', loss=deep_speaker_loss)
print("model_build_time", time() - orig_time)
logging.info('Starting training...')
lasteer = 10
eer = 1
while True:
orig_time = time()
x, _ = select_batch.best_batch(model, batch_size=c.BATCH_SIZE)
print("select_batch_time:", time() - orig_time)
y = np.random.uniform(size=(x.shape[0], 1))
logging.info('== Presenting step #{0}'.format(grad_steps))
orig_time = time()
loss = model.train_on_batch(x, y)
logging.info(
'== Processed in {0:.2f}s by the network, training loss = {1}.'.
format(time() - orig_time, loss))
if c.COMBINE_MODEL:
loss1 = gru_model.train_on_batch(x, y)
logging.info(
'== Processed in {0:.2f}s by the gru-network, training loss = {1}.'
.format(time() - orig_time, loss1))
with open(c.GRU_CHECKPOINT_FOLDER + '/losses_gru.txt', "a") as f:
f.write("{0},{1}\n".format(grad_steps, loss1))
# record training loss
with open(c.LOSS_LOG, "a") as f:
f.write("{0},{1}\n".format(grad_steps, loss))
if (grad_steps) % 10 == 0:
fm1, tpr1, acc1, eer1, fprr1, tprr1, auc1 = eval_model(
model,
train_batch_size,
test_dir=c.DATASET_DIR,
check_partial=True,
gru_model=gru_model)
logging.info(
'test training data EER = {0:.3f}, F-measure = {1:.3f}, Accuracy = {2:.3f}, AUC = {2:.3f} '
.format(eer1, fm1, acc1, auc1))
with open(c.CHECKPOINT_FOLDER + '/train_acc_eer.txt', "a") as f:
f.write("{0},{1},{2},{3}\n".format(grad_steps, eer1, fm1,
acc1))
if (grad_steps) % c.TEST_PER_EPOCHS == 0:
fm, tpr, acc, eer, fprr, tprr, auc = eval_model(
model,
train_batch_size,
test_dir=c.TEST_DIR,
gru_model=gru_model)
logging.info(
'== Testing model after batch #{0}'.format(grad_steps))
logging.info(
'EER = {0:.3f}, F-measure = {1:.3f}, Accuracy = {2:.3f}, AUC = {2:.3f} '
.format(eer, fm, acc, auc))
with open(c.TEST_LOG, "a") as f:
f.write("{0},{1},{2},{3}\n".format(grad_steps, eer, fm, acc))
# checkpoints are really heavy so let's just keep the last one.
if (grad_steps) % c.SAVE_PER_EPOCHS == 0:
create_dir_and_delete_content(c.CHECKPOINT_FOLDER)
model.save_weights('{0}/model_{1}_{2:.5f}.h5'.format(
c.CHECKPOINT_FOLDER, grad_steps, loss))
if c.COMBINE_MODEL:
gru_model.save_weights('{0}/grumodel_{1}_{2:.5f}.h5'.format(
c.GRU_CHECKPOINT_FOLDER, grad_steps, loss1))
if eer < lasteer:
files = sorted(filter(
lambda f: os.path.isfile(f) and f.endswith(".h5"),
map(lambda f: os.path.join(c.BEST_CHECKPOINT_FOLDER, f),
os.listdir(c.BEST_CHECKPOINT_FOLDER))),
key=lambda file: file.split('/')[-1].split('.')[
-2],
reverse=True)
lasteer = eer
for file in files[:-4]:
logging.info("removing old model: {}".format(file))
os.remove(file)
model.save_weights(
c.BEST_CHECKPOINT_FOLDER +
'/best_model{0}_{1:.5f}.h5'.format(grad_steps, eer))
if c.COMBINE_MODEL:
files = sorted(
filter(
lambda f: os.path.isfile(f) and f.endswith(".h5"),
map(
lambda f: os.path.join(
c.BEST_CHECKPOINT_FOLDER, f),
os.listdir(c.BEST_CHECKPOINT_FOLDER))),
key=lambda file: file.split('/')[-1].split('.')[-2],
reverse=True)
lasteer = eer
for file in files[:-4]:
logging.info("removing old model: {}".format(file))
os.remove(file)
gru_model.save_weights(c.BEST_CHECKPOINT_FOLDER +
'/best_gru_model{0}_{1:.5f}.h5'.
format(grad_steps, eer))
grad_steps += 1
def main(libri_dir=c.DATASET_DIR):
PRE_TRAIN = c.PRE_TRAIN
logging.info('Looking for fbank features [.npy] files in {}.'.format(libri_dir))
libri = data_catalog(libri_dir)
if len(libri) == 0:
logging.warning('Cannot find npy files, we will load audio, extract features and save it as npy file')
logging.warning('Waiting for preprocess...')
preprocess_and_save(c.WAV_DIR, c.DATASET_DIR)
libri = data_catalog(libri_dir)
if len(libri) == 0:
logging.warning('Have you converted flac files to wav? If not, run audio/convert_flac_2_wav.sh')
exit(1)
unique_speakers = libri['speaker_id'].unique()
spk_utt_dict, unique_speakers = create_dict(libri['filename'].values,libri['speaker_id'].values,unique_speakers)
select_batch.create_data_producer(unique_speakers, spk_utt_dict)
batch = stochastic_mini_batch(libri, batch_size=c.BATCH_SIZE, unique_speakers=unique_speakers)
batch_size = c.BATCH_SIZE * c.TRIPLET_PER_BATCH
x, y = batch.to_inputs()
b = x[0]
num_frames = b.shape[0]
train_batch_size = batch_size
#batch_shape = [batch_size * num_frames] + list(b.shape[1:]) # A triplet has 3 parts.
input_shape = (num_frames, b.shape[1], b.shape[2])
logging.info('num_frames = {}'.format(num_frames))
logging.info('batch size: {}'.format(batch_size))
logging.info('input shape: {}'.format(input_shape))
logging.info('x.shape : {}'.format(x.shape))
# 按理x.shape:(batchsize, num_frames, 64, 1)
orig_time = time()
model = convolutional_model(input_shape=input_shape, batch_size=batch_size, num_frames=num_frames)
logging.info(model.summary())
gru_model = None
if c.COMBINE_MODEL:
gru_model = recurrent_model(input_shape=input_shape, batch_size=batch_size, num_frames=num_frames)
logging.info(gru_model.summary())
grad_steps = 0
if PRE_TRAIN:
last_checkpoint = get_last_checkpoint_if_any(c.PRE_CHECKPOINT_FOLDER)
if last_checkpoint is not None:
logging.info('Found pre-training checkpoint [{}]. Resume from here...'.format(last_checkpoint))
x = model.output
x = Dense(len(unique_speakers), activation='softmax', name='softmax_layer')(x)
pre_model = Model(model.input, x)
pre_model.load_weights(last_checkpoint)
grad_steps = int(last_checkpoint.split('_')[-2])
logging.info('Successfully loaded pre-training model')
else:
last_checkpoint = get_last_checkpoint_if_any(c.CHECKPOINT_FOLDER)
if last_checkpoint is not None:
logging.info('Found checkpoint [{}]. Resume from here...'.format(last_checkpoint))
model.load_weights(last_checkpoint)
grad_steps = int(last_checkpoint.split('_')[-2])
logging.info('[DONE]')
if c.COMBINE_MODEL:
last_checkpoint = get_last_checkpoint_if_any(c.GRU_CHECKPOINT_FOLDER)
if last_checkpoint is not None:
logging.info('Found checkpoint [{}]. Resume from here...'.format(last_checkpoint))
gru_model.load_weights(last_checkpoint)
logging.info('[DONE]')
#adam = Adam(lr=0.01, beta_1=0.9, beta_2=0.999, epsilon=None, decay=0.0, amsgrad=False)
model.compile(optimizer='adam', loss=deep_speaker_loss)
if c.COMBINE_MODEL:
gru_model.compile(optimizer='adam', loss=deep_speaker_loss)
print("model_build_time",time()-orig_time)
logging.info('Starting training...')
lasteer = 10
eer = 1
# ======================================================================2020/05/21 10：38
train_times = [] # ===========================================================================2020/05/20 16:30
total_times = 0 # ===========================================================================2020/05/20 16:30
# 迭代10个epoch，每个epoch200个batch
# while True:
os.makedirs(c.BEST_CHECKPOINT_FOLDER, exist_ok=True)
while grad_steps < 2001 :
# ======================================================================结束
orig_time = time()
x, _ = select_batch.best_batch(model, batch_size=c.BATCH_SIZE)
print("select_batch_time:", time() - orig_time)
y = np.random.uniform(size=(x.shape[0], 1))
# If "ValueError: Error when checking target: expected ln to have shape (None, 512) but got array with shape (96, 1)"
# please modify line 121 to following line
# y = np.random.uniform(size=(x.shape[0], 512))
logging.info('== Presenting step #{0}'.format(grad_steps))
orig_time = time()
# ======================================================================2020/05/21 10：38
# 记录训练batch时间
# 记录迭代训练开始时间
begin_time = time() # ===========================================================================2020/05/20 16:30
loss = model.train_on_batch(x, y)
# 记录迭代训练结束时间
train_end_time = time() # ===========================================================================2020/05/20 16:30
# ======================================================================结束
logging.info('== Processed in {0:.2f}s by the network, training loss = {1}.'.format(time() - orig_time, loss))
if c.COMBINE_MODEL:
loss1 = gru_model.train_on_batch(x, y)
logging.info( '== Processed in {0:.2f}s by the gru-network, training loss = {1}.'.format(time() - orig_time, loss1))
with open(c.GRU_CHECKPOINT_FOLDER + '/losses_gru.txt', "a") as f:
f.write("{0},{1}\n".format(grad_steps, loss1))
# record training loss
with open(c.LOSS_LOG, "a") as f:
f.write("{0},{1}\n".format(grad_steps, loss))
if (grad_steps) % 10 == 0:
fm1, tpr1, acc1, eer1 = eval_model(model, train_batch_size, test_dir=c.DATASET_DIR, check_partial=True, gru_model=gru_model)
logging.info('test training data EER = {0:.3f}, F-measure = {1:.3f}, Accuracy = {2:.3f} '.format(eer1, fm1, acc1))
with open(c.CHECKPOINT_FOLDER + '/train_acc_eer.txt', "a") as f:
f.write("{0},{1},{2},{3}\n".format(grad_steps, eer1, fm1, acc1))
if (grad_steps ) % c.TEST_PER_EPOCHS == 0 :
fm, tpr, acc, eer = eval_model(model,train_batch_size, test_dir=c.TEST_DIR,gru_model=gru_model)
logging.info('== Testing model after batch #{0}'.format(grad_steps))
logging.info('EER = {0:.3f}, F-measure = {1:.3f}, Accuracy = {2:.3f} '.format(eer, fm, acc))
with open(c.TEST_LOG, "a") as f:
f.write("{0},{1},{2},{3}\n".format(grad_steps, eer, fm, acc))
# checkpoints are really heavy so let's just keep the last one.
if (grad_steps ) % c.SAVE_PER_EPOCHS == 0:
create_dir_and_delete_content(c.CHECKPOINT_FOLDER)
model.save_weights('{0}/model_{1}_{2:.5f}.h5'.format(c.CHECKPOINT_FOLDER, grad_steps, loss))
if c.COMBINE_MODEL:
gru_model.save_weights('{0}/grumodel_{1}_{2:.5f}.h5'.format(c.GRU_CHECKPOINT_FOLDER, grad_steps, loss1))
if eer < lasteer:
files = sorted(filter(lambda f: os.path.isfile(f) and f.endswith(".h5"),
map(lambda f: os.path.join(c.BEST_CHECKPOINT_FOLDER, f), os.listdir(c.BEST_CHECKPOINT_FOLDER))),
key=lambda file: file.split('/')[-1].split('.')[-2], reverse=True)
lasteer = eer
for file in files[:-4]:
logging.info("removing old model: {}".format(file))
os.remove(file)
model.save_weights(c.BEST_CHECKPOINT_FOLDER+'/best_model{0}_{1:.5f}.h5'.format(grad_steps, eer))
if c.COMBINE_MODEL:
files = sorted(filter(lambda f: os.path.isfile(f) and f.endswith(".h5"),
map(lambda f: os.path.join(c.BEST_CHECKPOINT_FOLDER, f),
os.listdir(c.BEST_CHECKPOINT_FOLDER))),
key=lambda file: file.split('/')[-1].split('.')[-2], reverse=True)
lasteer = eer
for file in files[:-4]:
logging.info("removing old model: {}".format(file))
os.remove(file)
gru_model.save_weights(c.BEST_CHECKPOINT_FOLDER+'/best_gru_model{0}_{1:.5f}.h5'.format(grad_steps, eer))
grad_steps += 1
end_time = time() # ===========================================================================2020/05/20 16:30
total_times += train_end_time - begin_time # ===========================================================================2020/05/20 16:30
train_times.append(str(begin_time) + '_' + str(train_end_time) + '_' + str(end_time) + '_' + str(train_end_time - begin_time)) # ===========================================================================2020/05/20 16:30
print("步数:{},耗时:{}s".format(grad_steps, str(train_end_time - begin_time))) # ===========================================================================2020/05/20 16:30
# ===========================================================================2020/05/20 16:30
# 将时间写入文件
with open('DeepSpeaker_epoch10_spk{}_suttperspk{}_迭代耗时.txt'.format(str(c.Spk_num), str(c.UttPerSpk)), mode='w', encoding='utf-8') as wf:
wf.write("步数{}_平均每次训练耗时:{}\n".format(grad_steps, total_times / grad_steps))
wf.write("开始训练时间_结束训练时间_结束步数训练时间(包括验证读写文件等)_耗时(结束训练时间-开始训练时间)\n")
for line in train_times:
wf.write(line + '\n')