split_targets = []
offset = 0
for size in target_sizes:
split_targets.append(targets[offset:offset + size])
offset += size
if args.cuda:
inputs = inputs.cuda()
out = model(inputs)
out = out.transpose(0, 1) # TxNxH
seq_length = out.size(0)
sizes = Variable(input_percentages.mul_(int(seq_length)).int())
decoded_output = decoder.decode(out.data, sizes)
target_strings = decoder.process_strings(
decoder.convert_to_strings(split_targets))
wer, cer = 0, 0
for x in range(len(target_strings)):
wer += decoder.wer(decoded_output[x], target_strings[x]) / float(
len(target_strings[x].split()))
cer += decoder.cer(decoded_output[x], target_strings[x]) / float(
len(target_strings[x]))
total_cer += cer
total_wer += wer
wer = total_wer / len(test_loader.dataset)
cer = total_cer / len(test_loader.dataset)
print('Validation Summary \t'
'Average WER {wer:.3f}\t'
'Average CER {cer:.3f}\t'.format(wer=wer * 100, cer=cer * 100))
def main():
args = parser.parse_args()
save_folder = args.save_folder
loss_results, cer_results, wer_results = None, None, None
if args.visdom:
from visdom import Visdom
viz = Visdom(server=args.visdom_server)
opts = [
dict(title='Loss', ylabel='Loss', xlabel='Epoch'),
dict(title='WER', ylabel='WER', xlabel='Epoch'),
dict(title='CER', ylabel='CER', xlabel='Epoch')
]
viz_windows = [None, None, None]
loss_results, cer_results, wer_results = torch.Tensor(
args.epochs), torch.Tensor(args.epochs), torch.Tensor(args.epochs)
epochs = torch.arange(1, args.epochs + 1)
if args.tensorboard:
from logger import Logger
try:
os.makedirs(args.log_dir)
except OSError as e:
if e.errno == errno.EEXIST:
print('Directory already exists.')
for file in os.listdir(args.log_dir):
file_path = os.path.join(args.log_dir, file)
try:
if os.path.isfile(file_path):
os.unlink(file_path)
except Exception as e:
raise
else:
raise
logger = TensorBoardLogger(args.log_dir)
try:
os.makedirs(save_folder)
except OSError as e:
if e.errno == errno.EEXIST:
print('Directory already exists.')
else:
raise
criterion = CTCLoss()
with open(args.labels_path) as label_file:
labels = str(''.join(json.load(label_file)))
audio_conf = dict(sample_rate=args.sample_rate,
window_size=args.window_size,
window_stride=args.window_stride,
window=args.window,
noise_dir=args.noise_dir,
noise_prob=args.noise_prob,
noise_levels=(args.noise_min, args.noise_max))
train_dataset = SpectrogramDataset(audio_conf=audio_conf,
manifest_filepath=args.train_manifest,
labels=labels,
normalize=True,
augment=args.augment)
test_dataset = SpectrogramDataset(audio_conf=audio_conf,
manifest_filepath=args.val_manifest,
labels=labels,
normalize=True,
augment=False)
train_loader = AudioDataLoader(train_dataset,
batch_size=args.batch_size,
num_workers=args.num_workers)
test_loader = AudioDataLoader(test_dataset,
batch_size=args.batch_size,
num_workers=args.num_workers)
rnn_type = args.rnn_type.lower()
assert rnn_type in supported_rnns, "rnn_type should be either lstm, rnn or gru"
model = DeepSpeech(rnn_hidden_size=args.hidden_size,
nb_layers=args.hidden_layers,
labels=labels,
rnn_type=supported_rnns[rnn_type],
audio_conf=audio_conf,
bidirectional=True)
parameters = model.parameters()
optimizer = torch.optim.SGD(parameters,
lr=args.lr,
momentum=args.momentum,
nesterov=True)
decoder = ArgMaxDecoder(labels)
if args.continue_from:
print("Loading checkpoint model %s" % args.continue_from)
package = torch.load(args.continue_from)
model.load_state_dict(package['state_dict'])
optimizer.load_state_dict(package['optim_dict'])
start_epoch = int(package.get(
'epoch', 1)) - 1 # Python index start at 0 for training
start_iter = package.get('iteration', None)
if start_iter is None:
start_epoch += 1 # Assume that we saved a model after an epoch finished, so start at the next epoch.
start_iter = 0
else:
start_iter += 1
avg_loss = int(package.get('avg_loss', 0))
loss_results, cer_results, wer_results = package[
'loss_results'], package['cer_results'], package['wer_results']
if args.visdom and \
package['loss_results'] is not None and start_epoch > 0: # Add previous scores to visdom graph
x_axis = epochs[0:start_epoch]
y_axis = [
loss_results[0:start_epoch], wer_results[0:start_epoch],
cer_results[0:start_epoch]
]
for x in range(len(viz_windows)):
viz_windows[x] = viz.line(
X=x_axis,
Y=y_axis[x],
opts=opts[x],
)
if args.tensorboard and \
package['loss_results'] is not None and start_epoch > 0: # Previous scores to tensorboard logs
for i in range(start_epoch):
info = {
'Avg Train Loss': loss_results[i],
'Avg WER': wer_results[i],
'Avg CER': cer_results[i]
}
for tag, val in info.items():
logger.scalar_summary(tag, val, i + 1)
else:
avg_loss = 0
start_epoch = 0
start_iter = 0
if args.cuda:
model = torch.nn.DataParallel(model).cuda()
print(model)
print("Number of parameters: %d" % DeepSpeech.get_param_size(model))
batch_time = AverageMeter()
data_time = AverageMeter()
losses = AverageMeter()
inf = float("inf")
for epoch in range(start_epoch, args.epochs):
model.train()
end = time.time()
for i, (data) in enumerate(train_loader, start=start_iter):
if i == len(train_loader):
break
inputs, targets, input_percentages, target_sizes = data
# measure data loading time
data_time.update(time.time() - end)
inputs = Variable(inputs, requires_grad=True)
target_sizes = Variable(target_sizes, requires_grad=False)
targets = Variable(targets, requires_grad=False)
if args.cuda:
inputs = inputs.cuda()
out = model(inputs)
out = out.transpose(0, 1) # TxNxH
seq_length = out.size(0)
sizes = Variable(input_percentages.mul_(int(seq_length)).int(),
requires_grad=False)
loss = criterion(out, targets, sizes, target_sizes)
loss = loss / inputs.size(0) # average the loss by minibatch
loss_sum = loss.data.sum()
if loss_sum == inf or loss_sum == -inf:
print("WARNING: received an inf loss, setting loss value to 0")
loss_value = 0
else:
loss_value = loss.data[0]
avg_loss += loss_value
losses.update(loss_value, inputs.size(0))
# compute gradient
optimizer.zero_grad()
loss.backward()
torch.nn.utils.clip_grad_norm(model.parameters(), args.max_norm)
# SGD step
optimizer.step()
if args.cuda:
torch.cuda.synchronize()
# measure elapsed time
batch_time.update(time.time() - end)
end = time.time()
if not args.silent:
print('Epoch: [{0}][{1}/{2}]\t'
'Time {batch_time.val:.3f} ({batch_time.avg:.3f})\t'
'Data {data_time.val:.3f} ({data_time.avg:.3f})\t'
'Loss {loss.val:.4f} ({loss.avg:.4f})\t'.format(
(epoch + 1), (i + 1),
len(train_loader),
batch_time=batch_time,
data_time=data_time,
loss=losses))
if args.checkpoint_per_batch > 0 and i > 0 and (
i + 1) % args.checkpoint_per_batch == 0:
file_path = '%s/deepspeech_checkpoint_epoch_%d_iter_%d.pth.tar' % (
save_folder, epoch + 1, i + 1)
print("Saving checkpoint model to %s" % file_path)
torch.save(
DeepSpeech.serialize(model,
optimizer=optimizer,
epoch=epoch,
iteration=i,
loss_results=loss_results,
wer_results=wer_results,
cer_results=cer_results,
avg_loss=avg_loss), file_path)
del loss
del out
avg_loss /= len(train_loader)
print('Training Summary Epoch: [{0}]\t'
'Average Loss {loss:.3f}\t'.format(epoch + 1, loss=avg_loss))
start_iter = 0 # Reset start iteration for next epoch
total_cer, total_wer = 0, 0
model.eval()
for i, (data) in enumerate(test_loader): # test
inputs, targets, input_percentages, target_sizes = data
inputs = Variable(inputs, volatile=True, requires_grad=False)
# unflatten targets
split_targets = []
offset = 0
for size in target_sizes:
split_targets.append(targets[offset:offset + size])
offset += size
if args.cuda:
inputs = inputs.cuda()
out = model(inputs)
out = out.transpose(0, 1) # TxNxH
seq_length = out.size(0)
sizes = Variable(input_percentages.mul_(int(seq_length)).int(),
volatile=True,
requires_grad=False)
decoded_output = decoder.decode(out.data, sizes)
target_strings = decoder.process_strings(
decoder.convert_to_strings(split_targets))
wer, cer = 0, 0
for x in range(len(target_strings)):
wer += decoder.wer(decoded_output[x],
target_strings[x]) / float(
len(target_strings[x].split()))
cer += decoder.cer(decoded_output[x],
target_strings[x]) / float(
len(target_strings[x]))
total_cer += cer
total_wer += wer
if args.cuda:
torch.cuda.synchronize()
del out
wer = total_wer / len(test_loader.dataset)
cer = total_cer / len(test_loader.dataset)
wer *= 100
cer *= 100
loss_results[epoch] = avg_loss
wer_results[epoch] = wer
cer_results[epoch] = cer
print('Validation Summary Epoch: [{0}]\t'
'Average WER {wer:.3f}\t'
'Average CER {cer:.3f}\t'.format(epoch + 1, wer=wer, cer=cer))
if args.visdom:
# epoch += 1
x_axis = epochs[0:epoch + 1]
y_axis = [
loss_results[0:epoch + 1], wer_results[0:epoch + 1],
cer_results[0:epoch + 1]
]
for x in range(len(viz_windows)):
if viz_windows[x] is None:
viz_windows[x] = viz.line(
X=x_axis,
Y=y_axis[x],
opts=opts[x],
)
else:
viz.line(
X=x_axis,
Y=y_axis[x],
win=viz_windows[x],
update='replace',
)
if args.tensorboard:
info = {'Avg Train Loss': avg_loss, 'Avg WER': wer, 'Avg CER': cer}
for tag, val in info.items():
logger.scalar_summary(tag, val, epoch + 1)
if args.log_params:
for tag, value in model.named_parameters():
tag = tag.replace('.', '/')
logger.histo_summary(tag, to_np(value), epoch + 1)
logger.histo_summary(tag + '/grad', to_np(value.grad),
epoch + 1)
if args.checkpoint:
file_path = '%s/deepspeech_%d.pth.tar' % (save_folder, epoch + 1)
torch.save(
DeepSpeech.serialize(model,
optimizer=optimizer,
epoch=epoch,
loss_results=loss_results,
wer_results=wer_results,
cer_results=cer_results), file_path)
# anneal lr
optim_state = optimizer.state_dict()
optim_state['param_groups'][0][
'lr'] = optim_state['param_groups'][0]['lr'] / args.learning_anneal
optimizer.load_state_dict(optim_state)
print('Learning rate annealed to: {lr:.6f}'.format(
lr=optim_state['param_groups'][0]['lr']))
avg_loss = 0
if not args.no_bucketing and epoch == 0:
print("Switching to bucketing sampler for following epochs")
train_dataset = SpectrogramDatasetWithLength(
audio_conf=audio_conf,
manifest_filepath=args.train_manifest,
labels=labels,
normalize=True,
augment=args.augment)
sampler = BucketingSampler(train_dataset)
train_loader.sampler = sampler
torch.save(DeepSpeech.serialize(model, optimizer=optimizer),
args.final_model_path)
def main():
args = parser.parse_args()
save_folder = args.save_folder
if args.visdom:
from visdom import Visdom
viz = Visdom()
opts = [
dict(title='Loss', ylabel='Loss', xlabel='Epoch'),
dict(title='WER', ylabel='WER', xlabel='Epoch'),
dict(title='CER', ylabel='CER', xlabel='Epoch')
]
viz_windows = [None, None, None]
loss_results, cer_results, wer_results = torch.Tensor(
args.epochs), torch.Tensor(args.epochs), torch.Tensor(args.epochs)
epochs = torch.range(1, args.epochs)
try:
os.makedirs(save_folder)
except OSError as e:
if e.errno == errno.EEXIST:
print('Directory already exists.')
else:
raise
criterion = CTCLoss()
with open(args.labels_path) as label_file:
labels = str(''.join(json.load(label_file)))
audio_conf = dict(sample_rate=args.sample_rate,
window_size=args.window_size,
window_stride=args.window_stride,
window=args.window)
train_dataset = SpectrogramDataset(audio_conf=audio_conf,
manifest_filepath=args.train_manifest,
labels=labels,
normalize=True)
test_dataset = SpectrogramDataset(audio_conf=audio_conf,
manifest_filepath=args.val_manifest,
labels=labels,
normalize=True)
train_loader = AudioDataLoader(train_dataset,
batch_size=args.batch_size,
num_workers=args.num_workers)
test_loader = AudioDataLoader(test_dataset,
batch_size=args.batch_size,
num_workers=args.num_workers)
model = DeepSpeech(rnn_hidden_size=args.hidden_size,
nb_layers=args.hidden_layers,
num_classes=len(labels))
decoder = ArgMaxDecoder(labels)
if args.cuda:
model = torch.nn.DataParallel(model).cuda()
print(model)
parameters = model.parameters()
optimizer = torch.optim.SGD(parameters,
lr=args.lr,
momentum=args.momentum,
nesterov=True)
batch_time = AverageMeter()
data_time = AverageMeter()
losses = AverageMeter()
for epoch in range(args.epochs):
model.train()
end = time.time()
avg_loss = 0
for i, (data) in enumerate(train_loader):
inputs, targets, input_percentages, target_sizes = data
# measure data loading time
data_time.update(time.time() - end)
inputs = Variable(inputs)
target_sizes = Variable(target_sizes)
targets = Variable(targets)
if args.cuda:
inputs = inputs.cuda()
out = model(inputs)
out = out.transpose(0, 1) # TxNxH
seq_length = out.size(0)
sizes = Variable(input_percentages.mul_(int(seq_length)).int())
loss = criterion(out, targets, sizes, target_sizes)
loss = loss / inputs.size(0) # average the loss by minibatch
loss_sum = loss.data.sum()
inf = float("inf")
if loss_sum == inf or loss_sum == -inf:
print("WARNING: received an inf loss, setting loss value to 0")
loss_value = 0
else:
loss_value = loss.data[0]
avg_loss += loss_value
losses.update(loss_value, inputs.size(0))
# compute gradient
optimizer.zero_grad()
loss.backward()
torch.nn.utils.clip_grad_norm(model.parameters(), args.max_norm)
# SGD step
optimizer.step()
# measure elapsed time
batch_time.update(time.time() - end)
end = time.time()
if not args.silent:
print('Epoch: [{0}][{1}/{2}]\t'
'Time {batch_time.val:.3f} ({batch_time.avg:.3f})\t'
'Data {data_time.val:.3f} ({data_time.avg:.3f})\t'
'Loss {loss.val:.4f} ({loss.avg:.4f})\t'.format(
(epoch + 1), (i + 1),
len(train_loader),
batch_time=batch_time,
data_time=data_time,
loss=losses))
avg_loss /= len(train_loader)
print('Training Summary Epoch: [{0}]\t'
'Average Loss {loss:.3f}\t'.format(epoch + 1, loss=avg_loss))
total_cer, total_wer = 0, 0
for i, (data) in enumerate(test_loader): # test
inputs, targets, input_percentages, target_sizes = data
inputs = Variable(inputs)
# unflatten targets
split_targets = []
offset = 0
for size in target_sizes:
split_targets.append(targets[offset:offset + size])
offset += size
if args.cuda:
inputs = inputs.cuda()
out = model(inputs)
out = out.transpose(0, 1) # TxNxH
seq_length = out.size(0)
sizes = Variable(input_percentages.mul_(int(seq_length)).int())
decoded_output = decoder.decode(out.data, sizes)
target_strings = decoder.process_strings(
decoder.convert_to_strings(split_targets))
wer, cer = 0, 0
for x in range(len(target_strings)):
wer += decoder.wer(decoded_output[x],
target_strings[x]) / float(
len(target_strings[x].split()))
cer += decoder.cer(decoded_output[x],
target_strings[x]) / float(
len(target_strings[x]))
total_cer += cer
total_wer += wer
wer = total_wer / len(test_loader.dataset)
cer = total_cer / len(test_loader.dataset)
wer *= 100
cer *= 100
print('Validation Summary Epoch: [{0}]\t'
'Average WER {wer:.0f}\t'
'Average CER {cer:.0f}\t'.format(epoch + 1, wer=wer, cer=cer))
if args.visdom:
loss_results[epoch] = avg_loss
wer_results[epoch] = wer
cer_results[epoch] = cer
epoch += 1
x_axis = epochs[0:epoch]
y_axis = [
loss_results[0:epoch], wer_results[0:epoch],
cer_results[0:epoch]
]
for x in range(len(viz_windows)):
if viz_windows[x] is None:
viz_windows[x] = viz.line(
X=x_axis,
Y=y_axis[x],
opts=opts[x],
)
else:
viz.line(
X=x_axis,
Y=y_axis[x],
win=viz_windows[x],
update='replace',
)
if args.epoch_save:
file_path = '%s/deepspeech_%d.pth.tar' % (save_folder, epoch)
torch.save(checkpoint(model, args, len(labels), epoch), file_path)
torch.save(checkpoint(model, args, len(labels)), args.final_model_path)