def build_optimizer(self, named_params):
no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
optimizer_grouped_parameters = [{'params': [p for n, p in named_params if not any(nd in n for nd in no_decay)], 'weight_decay': self.opt.weight_decay},
{'params': [p for n, p in named_params if any(nd in n for nd in no_decay)], 'weight_decay': 0.0}]
fused_adam = FusedAdam(optimizer_grouped_parameters,
lr=self.opt.learning_rate,
bias_correction=False,
max_grad_norm=self.opt.clip)
#params = [p for n, p in named_params if p.requires_grad]
#fused_adam = BertAdam(params, lr=self.opt.learning_rate, max_grad_norm=self.opt.clip, weight_decay=self.opt.weight_decay)
return FP16_Optimizer(fused_adam, dynamic_loss_scale=True)
def train(output_directory, log_directory, checkpoint_path, warm_start, n_gpus,
rank, group_name, hparams):
"""Training and validation logging results to tensorboard and stdout
Params
------
output_directory (string): directory to save checkpoints
log_directory (string) directory to save tensorboard logs
checkpoint_path(string): checkpoint path
n_gpus (int): number of gpus
rank (int): rank of current gpu
hparams (object): comma separated list of "name=value" pairs.
"""
if hparams.distributed_run:
init_distributed(hparams, n_gpus, rank, group_name)
torch.manual_seed(hparams.seed)
torch.cuda.manual_seed(hparams.seed)
torch.nn.functional.sigmoid
model = load_model(hparams)
learning_rate = hparams.learning_rate
#lr = args.lr * (0.1 ** (epoch // 30))
optimizer = torch.optim.Adam(model.parameters(),
lr=learning_rate,
weight_decay=hparams.weight_decay)
# optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate, momentum=0.9, dampening=0, weight_decay=hparams.weight_decay)
if hparams.fp16_run:
optimizer = FP16_Optimizer(
optimizer, dynamic_loss_scale=hparams.dynamic_loss_scaling)
if hparams.distributed_run:
model = apply_gradient_allreduce(model)
criterion = Tacotron2Loss(hparams)
logger = prepare_directories_and_logger(output_directory, log_directory,
rank)
train_loader, valset, collate_fn = prepare_dataloaders(hparams)
# Load checkpoint if one exists
iteration = 0
epoch_offset = 0
if checkpoint_path is not None:
if warm_start:
model = warm_start_model(checkpoint_path, model)
else:
model, optimizer, _learning_rate, iteration = load_checkpoint(
checkpoint_path, model, optimizer)
if hparams.use_saved_learning_rate:
learning_rate = _learning_rate
iteration += 1 # next iteration is iteration + 1
epoch_offset = max(0, int(iteration / len(train_loader)))
model.train()
# ================ MAIN TRAINNIG LOOP! ===================
step = 0
for epoch in range(epoch_offset, hparams.epochs):
print("Epoch: {}".format(epoch))
for i, batch in enumerate(train_loader):
start = time.perf_counter()
for param_group in optimizer.param_groups:
param_group['lr'] = learning_rate
model.zero_grad()
x, y = model.parse_batch(batch)
y_pred = model(x)
loss, recon_loss, S_kl_loss, R_kl_loss, speaker_loss, augment_loss, alignment_loss = criterion(
y_pred, y, iteration)
if hparams.distributed_run:
reduced_loss = reduce_tensor(loss.data, n_gpus).item()
else:
reduced_loss = loss.item()
if hparams.fp16_run:
optimizer.backward(loss)
grad_norm = optimizer.clip_fp32_grads(hparams.grad_clip_thresh)
else:
loss.backward()
grad_norm = torch.nn.utils.clip_grad_norm_(
model.parameters(), hparams.grad_clip_thresh)
optimizer.step()
overflow = optimizer.overflow if hparams.fp16_run else False
if not overflow and not math.isnan(reduced_loss) and rank == 0:
duration = time.perf_counter() - start
print(
"Train loss {} {:.6f} Grad Norm {:.6f} {:.2f}s/it".format(
iteration, reduced_loss, grad_norm, duration))
logger.log_training(reduced_loss, grad_norm, learning_rate, duration, recon_loss, S_kl_loss, R_kl_loss, \
speaker_loss, augment_loss, alignment_loss, iteration)
if not overflow and (iteration % hparams.iters_per_checkpoint
== 0):
validate(model, criterion, valset, iteration,
hparams.batch_size, n_gpus, collate_fn, logger,
hparams.distributed_run, rank)
if rank == 0:
checkpoint_path = os.path.join(
output_directory, "checkpoint_{}".format(iteration))
save_checkpoint(model, optimizer, learning_rate, iteration,
checkpoint_path)
iteration += 1
def train(output_directory, log_directory, checkpoint_path, warm_start, n_gpus,
rank, group_name, hparams):
"""Training and validation logging results to tensorboard and stdout
Params
------
output_directory (string): directory to save checkpoints
log_directory (string) directory to save tensorboard logs
checkpoint_path(string): checkpoint path
n_gpus (int): number of gpus
rank (int): rank of current gpu
hparams (object): comma separated list of "name=value" pairs.
"""
if hparams.distributed_run:
init_distributed(hparams, n_gpus, rank, group_name)
torch.manual_seed(hparams.seed)
torch.cuda.manual_seed(hparams.seed)
print("Loading models...")
model = load_model(hparams)
print("Initializing optimizer...")
learning_rate = hparams.learning_rate
optimizer = torch.optim.Adam(model.parameters(),
lr=learning_rate,
weight_decay=hparams.weight_decay)
if hparams.fp16_run:
optimizer = FP16_Optimizer(
optimizer, dynamic_loss_scale=hparams.dynamic_loss_scaling)
criterion = Tacotron2Loss()
print("Initializing logger...")
logger = prepare_directories_and_logger(output_directory, log_directory,
rank)
print("Initializing dataloader...")
train_loader, valset, collate_fn = prepare_dataloaders(hparams)
print("Loading checkpoints...")
# Load checkpoint if one exists
iteration = 0
epoch_offset = 0
if checkpoint_path is not None:
if warm_start:
model = warm_start_model(checkpoint_path, model)
else:
model, optimizer, _learning_rate, iteration = load_checkpoint(
checkpoint_path, model, optimizer)
if hparams.use_saved_learning_rate:
learning_rate = _learning_rate
iteration += 1 # next iteration is iteration + 1
epoch_offset = max(0, int(iteration / len(train_loader)))
model.train()
if hparams.distributed_run or torch.cuda.device_count() > 1:
batch_parser = model.module.parse_batch
else:
batch_parser = model.parse_batch
# ================ MAIN TRAINNIG LOOP! ===================
for epoch in range(epoch_offset, hparams.epochs):
print("Epoch: {}".format(epoch))
for i, batch in enumerate(train_loader):
start = time.perf_counter()
for param_group in optimizer.param_groups:
param_group['lr'] = learning_rate
model.zero_grad()
x, y = batch_parser(batch)
y_pred = model(x)
loss = criterion(y_pred, y)
reduced_loss = reduce_tensor(loss.data, n_gpus)[0] \
if hparams.distributed_run else loss.data[0]
if hparams.fp16_run:
optimizer.backward(loss)
grad_norm = optimizer.clip_fp32_grads(hparams.grad_clip_thresh)
else:
loss.backward()
grad_norm = torch.nn.utils.clip_grad_norm(
model.parameters(), hparams.grad_clip_thresh)
optimizer.step()
overflow = optimizer.overflow if hparams.fp16_run else False
if not overflow and not math.isnan(reduced_loss) and rank == 0:
duration = time.perf_counter() - start
print(
"Train loss {} {:.6f} Grad Norm {:.6f} {:.2f}s/it".format(
iteration, reduced_loss, grad_norm, duration))
logger.log_training(reduced_loss, grad_norm, learning_rate,
duration, iteration)
if not overflow and (iteration % hparams.iters_per_checkpoint
== 0):
reduced_val_loss = validate(model, criterion, valset,
iteration, hparams.batch_size,
n_gpus, collate_fn, logger,
hparams.distributed_run, rank)
if rank == 0:
print("Validation loss {}: {:9f} ".format(
iteration, reduced_val_loss))
checkpoint_path = os.path.join(
output_directory, "checkpoint_{}".format(iteration))
save_checkpoint(model, optimizer, learning_rate, iteration,
checkpoint_path)
logger.log_validation(reduced_val_loss, model, x, y,
y_pred, iteration, hparams)
iteration += 1
def train(cfg, local_rank, distributed, random_number_generator=None):
if (torch._C, '_jit_set_profiling_executor') :
torch._C._jit_set_profiling_executor(False)
if (torch._C, '_jit_set_profiling_mode') :
torch._C._jit_set_profiling_mode(False)
# Model logging
log_event(key=constants.GLOBAL_BATCH_SIZE, value=cfg.SOLVER.IMS_PER_BATCH)
log_event(key=constants.NUM_IMAGE_CANDIDATES, value=cfg.MODEL.RPN.FPN_POST_NMS_TOP_N_TRAIN)
model = build_detection_model(cfg)
device = torch.device(cfg.MODEL.DEVICE)
model.to(device)
# Initialize mixed-precision training
is_fp16 = (cfg.DTYPE == "float16")
if is_fp16:
# convert model to FP16
model.half()
optimizer = make_optimizer(cfg, model)
# Optimizer logging
log_event(key=constants.OPT_NAME, value="sgd_with_momentum")
log_event(key=constants.OPT_BASE_LR, value=cfg.SOLVER.BASE_LR)
log_event(key=constants.OPT_LR_WARMUP_STEPS, value=cfg.SOLVER.WARMUP_ITERS)
log_event(key=constants.OPT_LR_WARMUP_FACTOR, value=cfg.SOLVER.WARMUP_FACTOR)
log_event(key=constants.OPT_LR_DECAY_FACTOR, value=cfg.SOLVER.GAMMA)
log_event(key=constants.OPT_LR_DECAY_STEPS, value=cfg.SOLVER.STEPS)
log_event(key=constants.MIN_IMAGE_SIZE, value=cfg.INPUT.MIN_SIZE_TRAIN[0])
log_event(key=constants.MAX_IMAGE_SIZE, value=cfg.INPUT.MAX_SIZE_TRAIN)
scheduler = make_lr_scheduler(cfg, optimizer)
# disable the garbage collection
gc.disable()
if distributed:
model = DDP(model, delay_allreduce=True)
arguments = {}
arguments["iteration"] = 0
arguments["nhwc"] = cfg.NHWC
output_dir = cfg.OUTPUT_DIR
save_to_disk = get_rank() == 0
checkpointer = DetectronCheckpointer(
cfg, model, optimizer, scheduler, output_dir, save_to_disk
)
arguments["save_checkpoints"] = cfg.SAVE_CHECKPOINTS
extra_checkpoint_data = checkpointer.load(cfg.MODEL.WEIGHT, cfg.NHWC)
arguments.update(extra_checkpoint_data)
if is_fp16:
optimizer = FP16_Optimizer(optimizer, dynamic_loss_scale=True)
log_end(key=constants.INIT_STOP)
barrier()
log_start(key=constants.RUN_START)
barrier()
data_loader, iters_per_epoch = make_data_loader(
cfg,
is_train=True,
is_distributed=distributed,
start_iter=arguments["iteration"],
random_number_generator=random_number_generator,
)
log_event(key=constants.TRAIN_SAMPLES, value=len(data_loader))
checkpoint_period = cfg.SOLVER.CHECKPOINT_PERIOD
# set the callback function to evaluate and potentially
# early exit each epoch
if cfg.PER_EPOCH_EVAL:
per_iter_callback_fn = functools.partial(
mlperf_test_early_exit,
iters_per_epoch=iters_per_epoch,
tester=functools.partial(test, cfg=cfg),
model=model,
distributed=distributed,
min_bbox_map=cfg.MLPERF.MIN_BBOX_MAP,
min_segm_map=cfg.MLPERF.MIN_SEGM_MAP)
else:
per_iter_callback_fn = None
start_train_time = time.time()
success = do_train(
model,
data_loader,
optimizer,
scheduler,
checkpointer,
device,
checkpoint_period,
arguments,
cfg.DISABLE_REDUCED_LOGGING,
per_iter_start_callback_fn=functools.partial(mlperf_log_epoch_start, iters_per_epoch=iters_per_epoch),
per_iter_end_callback_fn=per_iter_callback_fn,
)
end_train_time = time.time()
total_training_time = end_train_time - start_train_time
print(
"&&&& MLPERF METRIC THROUGHPUT={:.4f} iterations / s".format((arguments["iteration"] * cfg.SOLVER.IMS_PER_BATCH) / total_training_time)
)
return model, success
def style(model, style_image, content_image, iterations):
n_iter = [0]
t0 = time.time()
style_image = prep(style_image, False)
content_image = prep(content_image)
targets = compute_targets(model, style_image, content_image)
loss_layers = model.style_layers + model.content_layers
if args.optimizer.lower() == 'adam':
optimizer = optim.Adam([content_image],
lr=args.lr,
eps=args.eps,
betas=(args.beta1, 0.999))
elif args.optimizer.lower() == 'sgd':
optimizer = torch.optim.SGD([content_image],
lr=args.lr,
momentum=0.99999)
else:
optimizer = optim.LBFGS([content_image],
lr=args.lr,
history_size=args.history_size,
tolerance_grad=args.tolerance_grad,
tolerance_change=args.tolerance_change,
max_iter=args.max_iter,
max_eval=args.max_eval)
if args.half:
optimizer = FP16_Optimizer(optimizer,
scale=args.static_loss_scale,
dynamic_scale=args.dynamic_loss_scale)
def closure():
optimizer.zero_grad()
out = model(content_image, loss_layers)
layer_losses = [
weights[a] * loss_fns[a](A.float(), targets[a])
for a, A in enumerate(out)
]
loss = sum(layer_losses)
if args.half:
optimizer.backward(loss)
else:
loss.backward()
# print(content_image.grad.data)
# print(optimizer.fp32_params[0].grad.data)
# quit()
n_iter[0] += 1
if n_iter[0] % args.log_interval == 1:
print('Iteration: %d, loss: %d time : %s' %
(n_iter[0], int(loss.data[0]), time.time() - t0))
# print([loss_layers[li] + ': ' + str(l.data[0]) for li,l in enumerate(layer_losses)]) #loss of each layer
return loss
while n_iter[0] <= iterations:
optimizer.step(closure)
if args.save_interval > 0 and n_iter[
0] % args.save_interval == 0 and n_iter[0] > 0:
postp(content_image.data[0].float().cpu().squeeze()).save(
outfile + '-i' + str(n_iter[0]),
format='JPEG',
subsampling=0,
quality=60)
return postp(content_image.data[0].float().cpu().squeeze())
def train(output_directory, log_directory, checkpoint_path, warm_start, n_gpus,
rank, group_name, hyper_params, train_loader, valset, collate_fn):
"""Training and validation method with logging results to tensorboard and stdout
:param output_directory (string): directory to save checkpoints
:param log_directory (string): directory to save tensorboard logs
:param checkpoint_path (string): checkpoint path
:param n_gpus (int): number of gpus
:param rank (int): rank of current gpu
:param hyper_params (object dictionary): dictionary with all hyper parameters
"""
# Check whether is a distributed running
if hyper_params['distributed_run']:
init_distributed(hyper_params, n_gpus, rank, group_name)
# set the same fixed seed to reproduce same results everytime we train
torch.manual_seed(hyper_params['seed'])
torch.cuda.manual_seed(hyper_params['seed'])
model = load_model(hyper_params)
learning_rate = hyper_params['learning_rate']
optimizer = torch.optim.Adam(model.parameters(),
lr=learning_rate,
weight_decay=hyper_params['weight_decay'])
if hyper_params['fp16_run']:
optimizer = FP16_Optimizer(
optimizer, dynamic_loss_scale=hyper_params['dynamic_loss_scaling'])
# Define the criterion of the loss function. The objective.
criterion = Tacotron2Loss()
logger = prepare_directories_and_logger(output_directory, log_directory,
rank)
# logger = ''
iteration = 0
epoch_offset = 0
if checkpoint_path is not None:
if warm_start:
# Re-start the model from the last checkpoint if we save the parameters and don't want to start from 0
model = warm_start_model(checkpoint_path, model)
else:
# CHECK THIS OUT!!!
model, optimizer, _learning_rate, iteration = load_checkpoint(
checkpoint_path, model, optimizer)
if hyper_params['use_saved_learning_rate']:
learning_rate = _learning_rate
iteration += 1 # next iteration is iteration + 1
epoch_offset = max(0, int(iteration / len(train_loader)))
# Set this to make all modules and regularization aware this is the training stage:
model.train()
# MAIN LOOP
for epoch in range(epoch_offset, hyper_params['epochs']):
print("Epoch: {}".format(epoch))
for i, batch in enumerate(train_loader):
start = time.perf_counter()
# CHECK THIS OUT!!!
for param_group in optimizer.param_groups:
param_group['lr'] = learning_rate
model.zero_grad()
input_data, output_target = model.parse_batch(batch)
output_predicted = model(input_data)
loss = criterion(output_predicted, output_target)
if hyper_params['distributed_run']:
reduced_loss = reduce_tensor(loss.data, n_gpus).item()
else:
reduced_loss = loss.item()
if hyper_params['fp16_run']:
optimizer.backward(
loss) # transformed optimizer into fp16 type
grad_norm = optimizer.clip_fp32_grads(
hyper_params['grad_clip_thresh'])
else:
loss.backward()
grad_norm = torch.nn.utils.clip_grad_norm_(
model.parameters(), hyper_params['grad_clip_thresh'])
# Performs a single optimization step (parameter update)
optimizer.step()
# This boolean controls overflow when running in fp16 optimizer
overflow = optimizer.overflow if hyper_params['fp16_run'] else False
# If overflow is True, it will not enter. If isnan is True, it will not enter neither.
if not overflow and not math.isnan(reduced_loss) and rank == 0:
duration = time.perf_counter() - start
print(
"Train loss {} {:.6f} Grand Norm {:.6f} {:.2f}s/it".format(
iteration, reduced_loss, grad_norm, duration))
# logs training information of the current iteration
logger.log_training(reduced_loss, grad_norm, learning_rate,
duration, iteration)
# Every iters_per_checkpoint steps there is a validation of the model and its updated parameters
if not overflow and (iteration %
hyper_params['iters_per_checkpoint'] == 0):
validate(model, criterion, valset, iteration,
hyper_params['batch_size'], n_gpus, collate_fn,
logger, hyper_params['distributed_run'], rank)
if rank == 0:
checkpoint_path = os.path.join(
output_directory, "checkpoint_{}".format(iteration))
save_checkpoint(model, optimizer, learning_rate, iteration,
checkpoint_path)
iteration += 1
