def setup_training(args):
assert (torch.cuda.is_available())
if args.local_rank == -1:
device = torch.device("cuda")
args.n_gpu = torch.cuda.device_count()
args.allreduce_post_accumulation = False
args.allreduce_post_accumulation_fp16 = False
else:
torch.cuda.set_device(args.local_rank)
device = torch.device("cuda", args.local_rank)
# Initializes the distributed backend which will take care of sychronizing nodes/GPUs
torch.distributed.init_process_group(backend='smddp',
init_method='env://')
args.n_gpu = 1
if args.gradient_accumulation_steps == 1:
args.allreduce_post_accumulation = False
args.allreduce_post_accumulation_fp16 = False
if is_main_process():
dllogger.init(backends=[
dllogger.JSONStreamBackend(verbosity=dllogger.Verbosity.VERBOSE,
filename=args.json_summary),
dllogger.StdOutBackend(verbosity=dllogger.Verbosity.VERBOSE,
step_format=format_step)
])
else:
dllogger.init(backends=[])
print(
"device: {} n_gpu: {}, distributed training: {}, 16-bits training: {}".
format(device, args.n_gpu, bool(args.local_rank != -1), args.fp16))
if args.gradient_accumulation_steps < 1:
raise ValueError(
"Invalid gradient_accumulation_steps parameter: {}, should be >= 1"
.format(args.gradient_accumulation_steps))
if args.train_batch_size % args.gradient_accumulation_steps != 0:
raise ValueError(
"Invalid gradient_accumulation_steps parameter: {}, batch size {} should be divisible"
.format(args.gradient_accumulation_steps, args.train_batch_size))
args.train_batch_size = args.train_batch_size // args.gradient_accumulation_steps
if not args.do_train:
raise ValueError(" `do_train` must be True.")
if not args.resume_from_checkpoint and os.path.exists(
args.output_dir) and (os.listdir(args.output_dir) and any(
[i.startswith('ckpt') for i in os.listdir(args.output_dir)])):
raise ValueError(
"Output directory ({}) already exists and is not empty.".format(
args.output_dir))
if (not args.resume_from_checkpoint
or not os.path.exists(args.output_dir)) and is_main_process():
os.makedirs(args.output_dir, exist_ok=True)
return device, args
def setup_training(args):
assert (torch.cuda.is_available())
global ort_supplement
import ort_supplement.ort_supplement as ort_supplement
device = ort_supplement.setup_onnxruntime_with_mpi(args)
if is_main_process():
dllogger.init(backends=[
dllogger.JSONStreamBackend(verbosity=dllogger.Verbosity.VERBOSE,
filename=args.json_summary),
dllogger.StdOutBackend(verbosity=dllogger.Verbosity.VERBOSE,
step_format=format_step)
])
else:
dllogger.init(backends=[])
print(
"device: {} n_gpu: {}, distributed training: {}, 16-bits training: {}".
format(device, args.n_gpu, bool(args.local_rank != -1), args.fp16))
if args.gradient_accumulation_steps < 1:
raise ValueError(
"Invalid gradient_accumulation_steps parameter: {}, should be >= 1"
.format(args.gradient_accumulation_steps))
if args.train_batch_size % args.gradient_accumulation_steps != 0:
raise ValueError(
"Invalid gradient_accumulation_steps parameter: {}, batch size {} should be divisible"
.format(args.gradient_accumulation_steps, args.train_batch_size))
args.train_batch_size = args.train_batch_size // args.gradient_accumulation_steps
if not args.do_train:
raise ValueError(" `do_train` must be True.")
if not args.resume_from_checkpoint and os.path.exists(
args.output_dir) and (os.listdir(args.output_dir) and any(
[i.startswith('ckpt') for i in os.listdir(args.output_dir)])):
raise ValueError(
"Output directory ({}) already exists and is not empty.".format(
args.output_dir))
if (not args.resume_from_checkpoint
or not os.path.exists(args.output_dir)) and is_main_process():
os.makedirs(args.output_dir, exist_ok=True)
return device, args
def setup_training(args):
assert torch.cuda.is_available()
if args.smp > 0:
# Initialize SMP. The configuration is obtained from the parameters passed to
# the Sagemaker PyTorch estimator.
smp.init()
# SMP: Set the device to the GPU ID used by the current process.
# Input tensors should be transferred to this device.
torch.cuda.set_device(smp.local_rank())
device = torch.device("cuda", smp.local_rank())
args.n_gpu = 1
# if args.local_rank == -1:
# device = torch.device("cuda")
# args.n_gpu = torch.cuda.device_count()
# args.allreduce_post_accumulation = False
# args.allreduce_post_accumulation_fp16 = False
# else:
# torch.cuda.set_device(args.local_rank)
# device = torch.device("cuda", args.local_rank)
# # Initializes the distributed backend which will take care of sychronizing nodes/GPUs
# torch.distributed.init_process_group(backend='nccl', init_method='env://')
# args.n_gpu = 1
if args.gradient_accumulation_steps == 1:
args.allreduce_post_accumulation = False
args.allreduce_post_accumulation_fp16 = False
print(
"device: {} n_gpu: {}, mp_rank: {}, rank: {}, distributed training: {}, 16-bits training: {}"
.format(device, args.n_gpu, smp.mp_rank(), smp.rank(),
bool(args.local_rank != -1), args.fp16))
if args.gradient_accumulation_steps < 1:
raise ValueError(
"Invalid gradient_accumulation_steps parameter: {}, should be >= 1"
.format(args.gradient_accumulation_steps))
if args.train_batch_size % args.gradient_accumulation_steps != 0:
raise ValueError(
"Invalid gradient_accumulation_steps parameter: {}, batch size {} should be divisible"
.format(args.gradient_accumulation_steps, args.train_batch_size))
args.train_batch_size = args.train_batch_size // args.gradient_accumulation_steps
if (not args.resume_from_checkpoint and os.path.exists(args.output_dir) and
(os.listdir(args.output_dir)
and any([i.startswith("ckpt")
for i in os.listdir(args.output_dir)]))):
raise ValueError(
"Output directory ({}) already exists and is not empty.".format(
args.output_dir))
if (not args.resume_from_checkpoint
or not os.path.exists(args.output_dir)) and is_main_process():
os.makedirs(args.output_dir, exist_ok=True)
return device, args
def train_one_epoch(model, optimizer, data_loader, device, epoch, print_freq):
model.train()
metric_logger = utils.MetricLogger(delimiter=" ")
metric_logger.add_meter(
'lr', utils.SmoothedValue(window_size=1, fmt='{value:.6f}'))
header = 'Epoch: [{}]'.format(epoch)
lr_scheduler = None
if epoch == 0:
warmup_factor = 1. / 1000
warmup_iters = min(1000, len(data_loader) - 1)
lr_scheduler = utils.warmup_lr_scheduler(optimizer, warmup_iters,
warmup_factor)
for images, targets in metric_logger.log_every(data_loader, print_freq,
header):
images = list(image.to(device) for image in images)
targets = [{k: v.to(device) for k, v in t.items()} for t in targets]
loss_dict = model(images, targets)
# applying logging only in the main process
# ### OUR CODE ###
if utils.is_main_process():
# let's track the losses here by adding scalars
tensorboard.logger.add_scalar_dict(
# passing the dictionary of losses (pairs - loss_key: loss_value)
loss_dict,
# passing the global step (number of iterations)
global_step=tensorboard.global_iter,
# adding the tag to combine plots in a subgroup
tag="loss")
# incrementing the global step (number of iterations)
tensorboard.global_iter += 1
# ### END OF OUR CODE ###
losses = sum(loss for loss in loss_dict.values())
# reduce losses over all GPUs for logging purposes
loss_dict_reduced = utils.reduce_dict(loss_dict)
losses_reduced = sum(loss for loss in loss_dict_reduced.values())
loss_value = losses_reduced.item()
if not math.isfinite(loss_value):
print("Loss is {}, stopping training".format(loss_value))
print(loss_dict_reduced)
sys.exit(1)
optimizer.zero_grad()
losses.backward()
optimizer.step()
if lr_scheduler is not None:
lr_scheduler.step()
metric_logger.update(loss=losses_reduced, **loss_dict_reduced)
metric_logger.update(lr=optimizer.param_groups[0]["lr"])
return metric_logger
def take_optimizer_step(args, optimizer, model, overflow_buf, global_step):
global skipped_steps
if args.allreduce_post_accumulation:
# manually allreduce gradients after all accumulation steps
# check for Inf/NaN
# 1. allocate an uninitialized buffer for flattened gradient
loss_scale = _amp_state.loss_scalers[0].loss_scale() if args.fp16 else 1
master_grads = [p.grad for p in amp.master_params(optimizer) if p.grad is not None]
flat_grad_size = sum(p.numel() for p in master_grads)
allreduce_dtype = torch.float16 if args.allreduce_post_accumulation_fp16 else torch.float32
flat_raw = torch.empty(flat_grad_size, device='cuda', dtype=allreduce_dtype)
# 2. combine unflattening and predivision of unscaled 'raw' gradient
allreduced_views = apex_C.unflatten(flat_raw, master_grads)
overflow_buf.zero_()
amp_C.multi_tensor_scale(65536,
overflow_buf,
[master_grads, allreduced_views],
loss_scale / (get_world_size() * args.gradient_accumulation_steps))
# 3. sum gradient across ranks. Because of the predivision, this averages the gradient
torch.distributed.all_reduce(flat_raw)
# 4. combine unscaling and unflattening of allreduced gradient
overflow_buf.zero_()
amp_C.multi_tensor_scale(65536,
overflow_buf,
[allreduced_views, master_grads],
1./loss_scale)
# 5. update loss scale
if args.fp16:
scaler = _amp_state.loss_scalers[0]
old_overflow_buf = scaler._overflow_buf
scaler._overflow_buf = overflow_buf
had_overflow = scaler.update_scale()
scaler._overfloat_buf = old_overflow_buf
else:
had_overflow = 0
# 6. call optimizer step function
if had_overflow == 0:
optimizer.step()
global_step += 1
else:
# Overflow detected, print message and clear gradients
skipped_steps += 1
if is_main_process():
scaler = _amp_state.loss_scalers[0]
dllogger.log(step="PARAMETER", data={"loss_scale": scaler.loss_scale()})
if _amp_state.opt_properties.master_weights:
for param in optimizer._amp_stash.all_fp32_from_fp16_params:
param.grad = None
for param in model.parameters():
param.grad = None
else:
if args.apply_optimizer > 0:
optimizer.step()
# optimizer.zero_grad()
for param in model.parameters():
param.grad = None
global_step += 1
return global_step
def checkpoint_step(args, epoch, global_step, model, optimizer, grad_scaler, last3_checkpoint_paths) :
torch.cuda.synchronize()
if is_main_process() and not args.skip_checkpoint:
# Save a trained model
dllogger.log(step="PARAMETER", data={"checkpoint_step": global_step})
model_to_save = model.module if hasattr(model,
'module') else model # Only save the model it-self
if args.resume_step < 0 or not args.phase2:
output_save_file = os.path.join(args.output_dir, "ckpt_{}.pt".format(global_step))
else:
output_save_file = os.path.join(args.output_dir, "ckpt_{}.pt".format(global_step + args.phase1_end_step))
if args.do_train:
torch.save({'model': model_to_save.state_dict(),
'optimizer': optimizer.state_dict(),
'grad_scaler': grad_scaler.state_dict(),
'epoch': epoch}, output_save_file)
# The new checkpoint could have a name already in
# last3_checkpoint_paths. In this case, torch.save will overwrite
# the old file; thus, we need to take the name out of
# last3_checkpoint_paths and append it to the last.
if output_save_file in last3_checkpoint_paths:
last3_checkpoint_paths.remove(output_save_file)
last3_checkpoint_paths.append(output_save_file)
if len(last3_checkpoint_paths) > 3:
ckpt_to_be_removed = last3_checkpoint_paths.pop(0)
os.remove(ckpt_to_be_removed)
def get_train_features(data_dir, bert_model, max_seq_length, do_lower_case,
local_rank, train_batch_size,
gradient_accumulation_steps, num_train_epochs,
tokenizer, processor):
cached_train_features_file = os.path.join(
data_dir,
'{0}_{1}_{2}'.format(
list(filter(None, bert_model.split('/'))).pop(),
str(max_seq_length),
str(do_lower_case),
),
)
train_features = None
try:
with open(cached_train_features_file, "rb") as reader:
train_features = pickle.load(reader)
logger.info("Loaded pre-processed features from {}".format(
cached_train_features_file))
except:
logger.info("Did not find pre-processed features from {}".format(
cached_train_features_file))
train_examples = processor.get_train_examples(data_dir)
train_features, _ = convert_examples_to_features(
train_examples,
processor.get_labels(),
max_seq_length,
tokenizer,
)
if is_main_process():
logger.info(" Saving train features into cached file %s",
cached_train_features_file)
with open(cached_train_features_file, "wb") as writer:
pickle.dump(train_features, writer)
return train_features
def voc_evaluate(model, data_loader, device):
n_threads = torch.get_num_threads()
torch.set_num_threads(1)
cpu_device = torch.device("cpu")
model.eval()
metric_logger = utils.MetricLogger(delimiter=" ")
header = 'Test:'
all_boxes = [[] for i in range(21)]
image_index = []
for image, targets in metric_logger.log_every(data_loader, 100, header):
image = list(img.to(device) for img in image)
targets = [{k: v.to(device) for k, v in t.items()} for t in targets]
torch.cuda.synchronize()
model_time = time.time()
outputs = model(image)
name = ''.join([chr(i) for i in targets[0]['name'].tolist()])
image_index.append(name)
outputs = [{k: v.to(cpu_device) for k, v in t.items()} for t in outputs]
image_boxes = [[] for i in range(21)]
for o in outputs:
for i in range(o['boxes'].shape[0]):
image_boxes[o['labels'][i]].extend([
torch.cat([o['boxes'][i],o['scores'][i].unsqueeze(0)], dim=0)
])
#makes sure that the all_boxes is filled with empty array when
#there are no boxes in image_boxes
for i in range(21):
if image_boxes[i] != []:
all_boxes[i].append([torch.stack(image_boxes[i])])
else:
all_boxes[i].append([])
model_time = time.time() - model_time
metric_logger.synchronize_between_processes()
all_boxes_gathered = utils.all_gather(all_boxes)
image_index_gathered = utils.all_gather(image_index)
# results from all processes are gathered here
if utils.is_main_process():
all_boxes = [[] for i in range(21)]
for abgs in all_boxes_gathered:
for ab,abg in zip(all_boxes,abgs):
ab += abg
image_index = []
for iig in image_index_gathered:
image_index+=iig
_write_voc_results_file(all_boxes,image_index, data_loader.dataset.root,
data_loader.dataset._transforms.transforms[0].CLASSES)
_do_python_eval(data_loader)
torch.set_num_threads(n_threads)
def _should_plot(self, epoch, iteration, total_iterations):
if not utils.is_main_process() or not self.plot_freq:
return False
if epoch % self.plot_freq == 0 or epoch == self.start_epoch + self.epochs - 1:
period = max(total_iterations // 10, 1)
if iteration % period == 0:
return True
return False
def maybe_load(self):
self.global_step = None
self.f_id = None
self.files = None
checkpoint = None
if chio.exists(self.args.output_dir):
model_names = [f for f in chio.list(
self.args.output_dir)
if f.endswith(".pt.{}".format(self.team))]
if len(model_names) != 0:
self.args.resume_step = max(
[int(x.split(
'.pt.{}'.format(self.team))[0].split('_')[1].strip())
for x in model_names])
self.global_step = self.args.resume_step
if self.global_step is not None:
print("Load from {}".format(os.path.join(self.args.output_dir,
"ckpt_{}.pt.{}".format(
self.global_step,
self.team))))
with chio.open(os.path.join(self.args.output_dir,
"ckpt_{}.pt.{}".format(
self.global_step, self.team)),
"rb") as f:
checkpoint = torch.load(f, map_location="cpu")
self.model.load_state_dict(checkpoint['model'],
strict=False)
self.another_model.load_state_dict(
checkpoint['another_model'], strict=False)
if self.args.phase2:
self.global_step -= self.args.phase1_end_step
if is_main_process():
print("resume step from ", self.args.resume_step)
if self.args.phase2:
keys = list(checkpoint['optimizer']['state'].keys())
# Override hyperparameters from Phase 1
for key in keys:
checkpoint['optimizer']['state'][key]['step'] = \
self.global_step
for iter, item in enumerate(
checkpoint['optimizer']['param_groups']):
checkpoint['optimizer']['param_groups'][iter]['t_total'] =\
self.args.max_steps
checkpoint['optimizer']['param_groups'][iter]['warmup'] = \
self.args.warmup_proportion
checkpoint['optimizer']['param_groups'][iter]['lr'] = \
self.args.learning_rate
self.optimizer.load_state_dict(checkpoint['optimizer'])
# Restore AMP master parameters
self.f_id = checkpoint['files'][0]
self.files = checkpoint['files'][1:]
def write_scalars(self, scalars, global_step, name=''):
if not utils.is_main_process():
return
if name:
scalars = {
f'{tag}/{name}': value
for tag, value in scalars.items()
}
for tag, value in scalars.items():
self.writer.add_scalar(tag, value, global_step)
def train_one_epoch(model, optimizer, data_loader, device, epoch, print_freq, writer):
model.train()
metric_logger = utils.MetricLogger(delimiter=" ")
metric_logger.add_meter("lr", utils.SmoothedValue(window_size=1, fmt="{value:.6f}"))
header = "Epoch: [{}]".format(epoch)
lr_scheduler = None
if epoch == 0:
warmup_factor = 1.0 / 1000
warmup_iters = min(1000, len(data_loader) - 1)
lr_scheduler = utils.warmup_lr_scheduler(optimizer, warmup_iters, warmup_factor)
step = 0
for images, targets in metric_logger.log_every(data_loader, print_freq, header):
images = list(image.to(device) for image in images)
targets = [
{k: v.to(device) for k, v in t.items() if k in ["boxes", "labels", "area", "iscrowd"]}
for t in targets
]
loss_dict = model(images, targets)
losses = sum(loss for loss in loss_dict.values())
# reduce losses over all GPUs for logging purposes
loss_dict_reduced = utils.reduce_dict(loss_dict)
losses_reduced = sum(loss for loss in loss_dict_reduced.values())
loss_value = losses_reduced.item()
if not math.isfinite(loss_value):
print("Loss is {}, stopping training".format(loss_value))
print(loss_dict_reduced)
sys.exit(1)
optimizer.zero_grad()
losses.backward()
optimizer.step()
if lr_scheduler is not None:
lr_scheduler.step()
metric_logger.update(loss=losses_reduced, **loss_dict_reduced)
metric_logger.update(lr=optimizer.param_groups[0]["lr"])
if is_main_process():
writer.add_scalar(
"lr", optimizer.param_groups[0]["lr"], epoch * len(data_loader) + step
)
step += 1
return metric_logger
def prepare_model(args, device):
# Prepare model
config = modeling.BertConfig.from_json_file(args.config_file)
# Padding for divisibility by 8
if config.vocab_size % 8 != 0:
config.vocab_size += 8 - (config.vocab_size % 8)
model = modeling.BertForPreTraining(config)
criterion = BertPretrainingCriterion(config.vocab_size,
args.train_batch_size,
args.max_seq_length)
model.enable_apex(False)
model = bert_model_with_loss(model, criterion)
model = ort_supplement.create_ort_trainer(args, device, model)
checkpoint = None
if not args.resume_from_checkpoint:
global_step = 0
else:
if args.resume_step == -1 and not args.init_checkpoint:
model_names = [
f for f in os.listdir(args.output_dir) if f.endswith(".pt")
]
args.resume_step = max([
int(x.split('.pt')[0].split('_')[1].strip())
for x in model_names
])
global_step = args.resume_step if not args.init_checkpoint else 0
if not args.init_checkpoint:
checkpoint = torch.load(os.path.join(
args.output_dir, "ckpt_{}.pt".format(global_step)),
map_location="cpu")
else:
checkpoint = torch.load(args.init_checkpoint, map_location="cpu")
model.load_state_dict(checkpoint['model'], strict=False)
if args.phase2 and not args.init_checkpoint:
global_step -= args.phase1_end_step
if is_main_process():
print("resume step from ", args.resume_step)
return model, checkpoint, global_step
def _setup_output(self, output_dir, out_file, overwrite=False):
self.output_dir = Path(output_dir)
self.output_dir.mkdir(parents=True, exist_ok=True)
self.out_file = None
if utils.is_main_process():
if overwrite:
for filename in self.output_dir.iterdir():
if filename.is_dir():
shutil.rmtree(filename)
elif filename.is_file():
filename.unlink()
if out_file:
self.out_file = (self.output_dir / 'train.txt')
self.writer = SummaryWriter(output_dir)
def __new__(cls, config=None) -> Any:
# this is a SUPER TRICKY HACK to work with singleton in multiprocessing (enable num_workers > 0 on windows)
# it exploits the IMPL DETAIL - this class will not be constructed in the spawn threads
# (the multiprocessing is used only in pytorch dataloader which do not create such object)
# if you know the less terrible and devastating way to write this, please share
if config is None and not is_main_process() and is_os_windows():
return object.__new__(
cls) # we do not need this object (from IMPL DETAIL)
# this is a singleton - once initialized (in the __init__ of experiment) it is available from everywhere
if cls._instance is None:
assert config is not None, "maybe you used relative module import of this file? don't do this"
cls._instance = object.__new__(cls)
else:
assert config is None
return cls._instance
def create_checkpoint(self, model, optimizer, epoch, lr_scheduler):
if not utils.is_main_process():
return
if isinstance(model,
(nn.DataParallel, nn.parallel.DistributedDataParallel)):
model_state_dict = model.module.state_dict()
else:
model_state_dict = model.state_dict()
torch.save(
{
'epoch': epoch,
'model_state_dict': model_state_dict,
'optimizer_state_dict': optimizer.state_dict(),
'lr_scheduler': lr_scheduler.state_dict()
}, self.output_dir / f'checkpoint{epoch:03}.tar')
def train_one_epoch(model, criterion, optimizer, data_loader, lr_scheduler,
device, epoch, print_freq):
model.train()
metric_logger = utils.MetricLogger(delimiter=" ")
metric_logger.add_meter('lr',
utils.SmoothedValue(window_size=1, fmt='{value}'))
header = 'Epoch: [{}]'.format(epoch)
for data, target in metric_logger.log_every(data_loader, print_freq,
header):
global n_iter
n_iter = n_iter + 1
optimizer.zero_grad()
target = target.to(device)
output = model(data)
loss = criterion(output, target)
loss = loss.mean()
#visualization
segmap = torch.argmax(output['out'], dim=1)
loss.backward()
optimizer.step()
lr_scheduler.step()
metric_logger.update(loss=loss.item(),
lr=optimizer.param_groups[0]["lr"])
if n_iter % args.print_freq == 0:
if args.tensorboard and utils.is_main_process():
args.writer.add_scalar('SupLoss', loss.item(), n_iter)
if n_iter % (args.print_freq * 100) == 0:
grid = torchvision.utils.make_grid(data[:1])
grid = (grid - grid.min()) / (grid.max() - grid.min() +
1e-5)
args.writer.add_image('sup images', grid, n_iter)
segmap = args.colormap[segmap[0].detach().cpu().numpy()]
segmap = segmap / 255.
args.writer.add_image('sup segmaps',
segmap.transpose((2, 0, 1)), n_iter)
def __init__(self,
lr=0.02,
momentum=0.9,
weight_decay=1e-4,
lr_steps=None,
lr_gamma=0.1,
data_path='.',
output_dir='.',
out_file=False,
batch_size=32,
device='cpu',
epochs=1,
num_workers=4,
dist_url='env://',
print_freq=100,
plot_freq=None,
data_parallel=False,
overwrite=False):
self._setup_output(output_dir, out_file, overwrite)
self.lr = lr
self.momentum = momentum
self.weight_decay = weight_decay
self.lr_steps = lr_steps
self.lr_gamma = lr_gamma
self.plot_freq = plot_freq if utils.is_main_process() else None
self.print_freq = print_freq
self.epochs = epochs
self.data_path = data_path
self.batch_size = batch_size
self.num_workers = num_workers
self.start_epoch = 0
self.dist_url = dist_url
device_index = self._init_distributed_mode()
self.device = torch.device(f'{device}:{device_index}')
self.data_parallel = data_parallel
assert not (self.data_parallel and self.distributed
), 'use either DataParallel or DistributedDataParallel'
self.checkpoint = infer_checkpoint(self.output_dir)
def accumulate_predictions_from_multiple_gpus(predictions_per_gpu):
all_predictions = all_gather(predictions_per_gpu)
if not is_main_process():
return
# merge the list of dicts
predictions = {}
for p in all_predictions:
predictions.update(p)
# convert a dict where the key is the index in a list
image_ids = list(sorted(predictions.keys()))
if len(image_ids) != image_ids[-1] + 1:
logger = logging.getLogger("RetinaNet.inference")
logger.warning(
"Number of images that were gathered from multiple processes is not "
"a contiguous set. Some images might be missing from the evaluation"
)
# convert to a list
predictions = [predictions[i] for i in image_ids]
return predictions
def record_hparams(self, metrics=None):
if not metrics:
metrics = {}
if not utils.is_main_process():
return
hparams_dict = {
'optimizer': 'SGD',
'lr': self.lr,
'momentum': self.momentum,
'weight_decay': self.weight_decay,
'gamma': self.lr_gamma,
'bsize': self.batch_size * self.world_size,
}
if self.lr_steps:
hparams_dict['lr_steps'] = ', '.join(str(s) for s in self.lr_steps)
metrics_dict = {
f'hparam/{name}': value
for name, value in metrics.items()
}
self.writer.add_hparams(hparams_dict, metrics_dict)
def inference(model, data_loader, dataset_name, device='cuda', output_folder=None,
expected_results=(), expected_results_sigma_tol=4):
device = torch.device(device)
num_devices = get_world_size()
logger = logging.getLogger("RetinaNet.inference")
dataset = data_loader.dataset
logger.info("Start evaluation on {} dataset({} images).".format(dataset_name, len(dataset)))
total_timer = Timer()
inference_timer = Timer()
total_timer.tic()
predictions = compute_on_dataset(model, data_loader, device, inference_timer)
# wait for all processes to complete before measuring the time
synchronize()
total_time = total_timer.toc()
total_time_str = get_time_str(total_time)
logger.info(
"Total run time: {} ({} s / img per device, on {} devices)".format(
total_time_str, total_time * num_devices / len(dataset), num_devices
)
)
predictions = accumulate_predictions_from_multiple_gpus(predictions)
if not is_main_process():
return
if output_folder:
torch.save(predictions, os.path.join(output_folder, "predictions.pth"))
extra_args = dict(
expected_results=expected_results,
expected_results_sigma_tol=expected_results_sigma_tol,
)
return evaluate(dataset=dataset,
predictions=predictions,
output_folder=output_folder,
**extra_args)
def main():
# args = parse_arguments()
# del args.local_rank
# print(args)
# args_to_yaml(args, 'config_finetune_train_glue_mrpc.yaml')
# exit(0)
config_yaml, local_rank = parse_my_arguments()
args = args_from_yaml(config_yaml)
args.local_rank = local_rank
""" Experiment Setup """
if args.server_ip and args.server_port:
# Distant debugging - see https://code.visualstudio.com/docs/python/debugging#_attach-to-a-local-script
import ptvsd
print("Waiting for debugger attach")
ptvsd.enable_attach(address=(args.server_ip, args.server_port),
redirect_output=True)
ptvsd.wait_for_attach()
if args.local_rank == -1 or args.no_cuda:
device = torch.device("cuda" if torch.cuda.is_available()
and not args.no_cuda else "cpu")
n_gpu = torch.cuda.device_count()
else:
torch.cuda.set_device(args.local_rank)
device = torch.device("cuda", args.local_rank)
n_gpu = 1
# Initializes the distributed backend which will take care of sychronizing nodes/GPUs
torch.distributed.init_process_group(backend='nccl')
logger.info(
"device: {} n_gpu: {}, distributed training: {}, 16-bits training: {}".
format(device, n_gpu, bool(args.local_rank != -1), args.fp16))
if args.gradient_accumulation_steps < 1:
raise ValueError(
"Invalid gradient_accumulation_steps parameter: {}, should be >= 1"
.format(args.gradient_accumulation_steps))
args.train_batch_size = args.train_batch_size // args.gradient_accumulation_steps
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
if n_gpu > 0:
torch.cuda.manual_seed_all(args.seed)
if not args.do_train and not args.do_eval:
raise ValueError(
"At least one of `do_train` or `do_eval` must be True.")
if os.path.exists(args.output_dir) and os.listdir(
args.output_dir) and args.do_train:
print(
"WARNING: Output directory ({}) already exists and is not empty.".
format(args.output_dir))
if not os.path.exists(args.output_dir):
os.makedirs(args.output_dir)
processors = {
"cola": ColaProcessor,
"mnli": MnliProcessor,
"mrpc": MrpcProcessor,
}
num_labels_task = {
"cola": 2,
"mnli": 3,
"mrpc": 2,
}
task_name = args.task_name.lower()
if task_name not in processors:
raise ValueError("Task not found: %s" % task_name)
processor = processors[task_name]()
num_labels = num_labels_task[task_name]
label_list = processor.get_labels()
tokenizer = BertTokenizer.from_pretrained(args.bert_model,
do_lower_case=args.do_lower_case)
train_examples = None
num_train_optimization_steps = None
if args.do_train:
train_examples = processor.get_train_examples(args.data_dir)
num_train_optimization_steps = int(
len(train_examples) / args.train_batch_size /
args.gradient_accumulation_steps) * args.num_train_epochs
if args.local_rank != -1:
num_train_optimization_steps = num_train_optimization_steps // torch.distributed.get_world_size(
)
""" Prepare Model """
# Prepare model
cache_dir = args.cache_dir if args.cache_dir else os.path.join(
PYTORCH_PRETRAINED_BERT_CACHE, 'distributed_{}'.format(
args.local_rank))
model = BertForSequenceClassification.from_pretrained(
args.bert_model, cache_dir=cache_dir, num_labels=num_labels)
state_dict = torch.load(args.init_checkpoint, map_location='cpu')
state_dict = state_dict.get(
'model', state_dict
) # in a full checkpoint weights are saved in state_dict['model']
model.load_state_dict(state_dict, strict=False)
if args.fp16:
model.half()
model.to(device)
if args.local_rank != -1:
try:
from apex.parallel import DistributedDataParallel as DDP
except ImportError:
raise ImportError(
"Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training."
)
model = DDP(model)
elif n_gpu > 1:
model = torch.nn.DataParallel(model)
plain_model = getattr(model, 'module', model)
with open(args.sparsity_config, 'r') as f:
raw_dict = yaml.load(f, Loader=yaml.SafeLoader)
masks = dict.fromkeys(raw_dict['prune_ratios'].keys())
for param_name in list(masks.keys()):
if get_parameter_by_name(plain_model, param_name) is None:
print(f'[WARNING] Cannot find {param_name}')
del masks[param_name]
for param_name in masks:
param = get_parameter_by_name(plain_model, param_name)
non_zero_mask = torch.ne(param, 0).to(param.dtype)
masks[param_name] = non_zero_mask
""" Prepare Optimizer"""
# Prepare optimizer
param_optimizer = list(model.named_parameters())
no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
optimizer_grouped_parameters = [{
'params':
[p for n, p in param_optimizer if not any(nd in n for nd in no_decay)],
'weight_decay':
0.01
}, {
'params':
[p for n, p in param_optimizer if any(nd in n for nd in no_decay)],
'weight_decay':
0.0
}]
if args.fp16:
try:
from apex.fp16_utils.fp16_optimizer import FP16_Optimizer
from apex.optimizers import FusedAdam
except ImportError:
raise ImportError(
"Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training."
)
optimizer = FusedAdam(optimizer_grouped_parameters,
lr=args.learning_rate,
bias_correction=False,
max_grad_norm=1.0)
if args.loss_scale == 0:
optimizer = FP16_Optimizer(optimizer, dynamic_loss_scale=True)
else:
optimizer = FP16_Optimizer(optimizer,
static_loss_scale=args.loss_scale)
else:
optimizer = BertAdam(optimizer_grouped_parameters,
lr=args.learning_rate,
warmup=args.warmup_proportion,
t_total=num_train_optimization_steps)
global_step = 0
nb_tr_steps = 0
tr_loss = 0
if args.do_train:
""" Prepare Dataset """
train_features = convert_examples_to_features(train_examples,
label_list,
args.max_seq_length,
tokenizer)
logger.info("***** Running training *****")
logger.info(" Num examples = %d", len(train_examples))
logger.info(" Batch size = %d", args.train_batch_size)
logger.info(" Num steps = %d", num_train_optimization_steps)
all_input_ids = torch.tensor([f.input_ids for f in train_features],
dtype=torch.long)
all_input_mask = torch.tensor([f.input_mask for f in train_features],
dtype=torch.long)
all_segment_ids = torch.tensor([f.segment_ids for f in train_features],
dtype=torch.long)
all_label_ids = torch.tensor([f.label_id for f in train_features],
dtype=torch.long)
train_data = TensorDataset(all_input_ids, all_input_mask,
all_segment_ids, all_label_ids)
if args.local_rank == -1:
train_sampler = RandomSampler(train_data)
else:
train_sampler = DistributedSampler(train_data)
train_dataloader = DataLoader(train_data,
sampler=train_sampler,
batch_size=args.train_batch_size)
""" Training Loop """
model.train()
for _ in trange(int(args.num_train_epochs), desc="Epoch"):
tr_loss = 0
nb_tr_examples, nb_tr_steps = 0, 0
for step, batch in enumerate(
tqdm(train_dataloader, desc="Iteration")):
if args.max_steps > 0 and global_step > args.max_steps:
break
batch = tuple(t.to(device) for t in batch)
input_ids, input_mask, segment_ids, label_ids = batch
loss = model(input_ids, segment_ids, input_mask, label_ids)
if n_gpu > 1:
loss = loss.mean() # mean() to average on multi-gpu.
if args.gradient_accumulation_steps > 1:
loss = loss / args.gradient_accumulation_steps
if args.fp16:
optimizer.backward(loss)
else:
loss.backward()
tr_loss += loss.item()
nb_tr_examples += input_ids.size(0)
nb_tr_steps += 1
if (step + 1) % args.gradient_accumulation_steps == 0:
if args.fp16:
# modify learning rate with special warm up BERT uses
# if args.fp16 is False, BertAdam is used that handles this automatically
lr_this_step = args.learning_rate * warmup_linear(
global_step / num_train_optimization_steps,
args.warmup_proportion)
for param_group in optimizer.param_groups:
param_group['lr'] = lr_this_step
optimizer.step()
optimizer.zero_grad()
global_step += 1
plain_model = getattr(model, 'module', model)
for param_name, mask in masks.items():
get_parameter_by_name(plain_model,
param_name).data *= mask
""" Load Model for Evaluation """
if args.do_train:
# Save a trained model and the associated configuration
output_model_file = os.path.join(args.output_dir, WEIGHTS_NAME)
output_config_file = os.path.join(args.output_dir, CONFIG_NAME)
if is_main_process(
): # only the main process should save the trained model
model_to_save = model.module if hasattr(
model, 'module') else model # Only save the model it-self
torch.save(model_to_save.state_dict(), output_model_file)
with open(output_config_file, 'w') as f:
f.write(model_to_save.config.to_json_string())
# Load a trained model and config that you have fine-tuned
config = BertConfig(output_config_file)
model = BertForSequenceClassification(config, num_labels=num_labels)
model.load_state_dict(torch.load(output_model_file))
else:
model = BertForSequenceClassification.from_pretrained(
args.bert_model, num_labels=num_labels)
state_dict = torch.load(args.init_checkpoint, map_location='cpu')
state_dict = state_dict.get('model', state_dict)
model.load_state_dict(state_dict, strict=False)
model.to(device)
""" Run Evaluation """
if args.do_eval and (args.local_rank == -1
or torch.distributed.get_rank() == 0):
eval_examples = processor.get_dev_examples(args.data_dir)
eval_features = convert_examples_to_features(eval_examples, label_list,
args.max_seq_length,
tokenizer)
logger.info("***** Running evaluation *****")
logger.info(" Num examples = %d", len(eval_examples))
logger.info(" Batch size = %d", args.eval_batch_size)
all_input_ids = torch.tensor([f.input_ids for f in eval_features],
dtype=torch.long)
all_input_mask = torch.tensor([f.input_mask for f in eval_features],
dtype=torch.long)
all_segment_ids = torch.tensor([f.segment_ids for f in eval_features],
dtype=torch.long)
all_label_ids = torch.tensor([f.label_id for f in eval_features],
dtype=torch.long)
eval_data = TensorDataset(all_input_ids, all_input_mask,
all_segment_ids, all_label_ids)
# Run prediction for full data
eval_sampler = SequentialSampler(eval_data)
eval_dataloader = DataLoader(eval_data,
sampler=eval_sampler,
batch_size=args.eval_batch_size)
model.eval()
eval_loss, eval_accuracy = 0, 0
nb_eval_steps, nb_eval_examples = 0, 0
for input_ids, input_mask, segment_ids, label_ids in tqdm(
eval_dataloader, desc="Evaluating"):
input_ids = input_ids.to(device)
input_mask = input_mask.to(device)
segment_ids = segment_ids.to(device)
label_ids = label_ids.to(device)
with torch.no_grad():
tmp_eval_loss = model(input_ids, segment_ids, input_mask,
label_ids)
logits = model(input_ids, segment_ids, input_mask)
logits = logits.detach().cpu().numpy()
label_ids = label_ids.to('cpu').numpy()
tmp_eval_accuracy = accuracy(logits, label_ids)
eval_loss += tmp_eval_loss.mean().item()
eval_accuracy += tmp_eval_accuracy
nb_eval_examples += input_ids.size(0)
nb_eval_steps += 1
eval_loss = eval_loss / nb_eval_steps
eval_accuracy = eval_accuracy / nb_eval_examples
loss = tr_loss / nb_tr_steps if args.do_train else None
result = {
'eval_loss': eval_loss,
'eval_accuracy': eval_accuracy,
'global_step': global_step,
'loss': loss
}
output_eval_file = os.path.join(args.output_dir, "eval_results.txt")
with open(output_eval_file, "w") as writer:
logger.info("***** Eval results *****")
for key in sorted(result.keys()):
logger.info(" %s = %s", key, str(result[key]))
writer.write("%s = %s\n" % (key, str(result[key])))
def main(args):
utils.init_distributed_mode(args)
print(args)
if args.distillation_type != 'none' and args.finetune and not args.eval:
raise NotImplementedError(
"Finetuning with distillation not yet supported")
device = torch.device(args.device)
# fix the seed for reproducibility
seed = args.seed + utils.get_rank()
torch.manual_seed(seed)
np.random.seed(seed)
# random.seed(seed)
cudnn.benchmark = True
dataset_train, args.nb_classes = build_dataset(is_train=True, args=args)
dataset_val, _ = build_dataset(is_train=False, args=args)
if True: # args.distributed:
num_tasks = utils.get_world_size()
global_rank = utils.get_rank()
if args.repeated_aug:
sampler_train = RASampler(dataset_train,
num_replicas=num_tasks,
rank=global_rank,
shuffle=True)
else:
sampler_train = torch.utils.data.DistributedSampler(
dataset_train,
num_replicas=num_tasks,
rank=global_rank,
shuffle=True)
if args.dist_eval:
if len(dataset_val) % num_tasks != 0:
print(
'Warning: Enabling distributed evaluation with an eval dataset not divisible by process number. '
'This will slightly alter validation results as extra duplicate entries are added to achieve '
'equal num of samples per-process.')
sampler_val = torch.utils.data.DistributedSampler(
dataset_val,
num_replicas=num_tasks,
rank=global_rank,
shuffle=False)
else:
sampler_val = torch.utils.data.SequentialSampler(dataset_val)
else:
sampler_train = torch.utils.data.RandomSampler(dataset_train)
sampler_val = torch.utils.data.SequentialSampler(dataset_val)
data_loader_train = torch.utils.data.DataLoader(
dataset_train,
sampler=sampler_train,
batch_size=args.batch_size,
num_workers=args.num_workers,
pin_memory=args.pin_mem,
drop_last=True,
)
data_loader_val = torch.utils.data.DataLoader(dataset_val,
sampler=sampler_val,
batch_size=int(
1.5 * args.batch_size),
num_workers=args.num_workers,
pin_memory=args.pin_mem,
drop_last=False)
mixup_fn = None
mixup_active = args.mixup > 0 or args.cutmix > 0. or args.cutmix_minmax is not None
if mixup_active:
mixup_fn = Mixup(mixup_alpha=args.mixup,
cutmix_alpha=args.cutmix,
cutmix_minmax=args.cutmix_minmax,
prob=args.mixup_prob,
switch_prob=args.mixup_switch_prob,
mode=args.mixup_mode,
label_smoothing=args.smoothing,
num_classes=args.nb_classes)
print(f"Creating model: {args.model}")
model = create_model(
args.model,
pretrained=False,
num_classes=args.nb_classes,
drop_rate=args.drop,
drop_path_rate=args.drop_path,
drop_block_rate=None,
)
if args.finetune:
if args.finetune.startswith('https'):
checkpoint = torch.hub.load_state_dict_from_url(args.finetune,
map_location='cpu',
check_hash=True)
else:
checkpoint = torch.load(args.finetune, map_location='cpu')
checkpoint_model = checkpoint['model']
state_dict = model.state_dict()
for k in [
'head.weight', 'head.bias', 'head_dist.weight',
'head_dist.bias'
]:
if k in checkpoint_model and checkpoint_model[
k].shape != state_dict[k].shape:
print(f"Removing key {k} from pretrained checkpoint")
del checkpoint_model[k]
# interpolate position embedding
pos_embed_checkpoint = checkpoint_model['pos_embed']
embedding_size = pos_embed_checkpoint.shape[-1]
num_patches = model.patch_embed.num_patches
num_extra_tokens = model.pos_embed.shape[-2] - num_patches
# height (== width) for the checkpoint position embedding
orig_size = int(
(pos_embed_checkpoint.shape[-2] - num_extra_tokens)**0.5)
# height (== width) for the new position embedding
new_size = int(num_patches**0.5)
# class_token and dist_token are kept unchanged
extra_tokens = pos_embed_checkpoint[:, :num_extra_tokens]
# only the position tokens are interpolated
pos_tokens = pos_embed_checkpoint[:, num_extra_tokens:]
pos_tokens = pos_tokens.reshape(-1, orig_size, orig_size,
embedding_size).permute(0, 3, 1, 2)
pos_tokens = torch.nn.functional.interpolate(pos_tokens,
size=(new_size, new_size),
mode='bicubic',
align_corners=False)
pos_tokens = pos_tokens.permute(0, 2, 3, 1).flatten(1, 2)
new_pos_embed = torch.cat((extra_tokens, pos_tokens), dim=1)
checkpoint_model['pos_embed'] = new_pos_embed
model.load_state_dict(checkpoint_model, strict=False)
model.to(device)
model_ema = None
if args.model_ema:
# Important to create EMA model after cuda(), DP wrapper, and AMP but before SyncBN and DDP wrapper
model_ema = ModelEma(model,
decay=args.model_ema_decay,
device='cpu' if args.model_ema_force_cpu else '',
resume='')
model_without_ddp = model
if args.distributed:
model = torch.nn.parallel.DistributedDataParallel(
model, device_ids=[args.gpu])
model_without_ddp = model.module
n_parameters = sum(p.numel() for p in model.parameters()
if p.requires_grad)
print('number of params:', n_parameters)
linear_scaled_lr = args.lr * args.batch_size * utils.get_world_size(
) / 512.0
args.lr = linear_scaled_lr
optimizer = create_optimizer(args, model_without_ddp)
loss_scaler = NativeScaler()
lr_scheduler, _ = create_scheduler(args, optimizer)
criterion = LabelSmoothingCrossEntropy()
if args.mixup > 0.:
# smoothing is handled with mixup label transform
criterion = SoftTargetCrossEntropy()
elif args.smoothing:
criterion = LabelSmoothingCrossEntropy(smoothing=args.smoothing)
else:
criterion = torch.nn.CrossEntropyLoss()
teacher_model = None
if args.distillation_type != 'none':
assert args.teacher_path, 'need to specify teacher-path when using distillation'
print(f"Creating teacher model: {args.teacher_model}")
teacher_model = create_model(
args.teacher_model,
pretrained=False,
num_classes=args.nb_classes,
global_pool='avg',
)
if args.teacher_path.startswith('https'):
checkpoint = torch.hub.load_state_dict_from_url(args.teacher_path,
map_location='cpu',
check_hash=True)
else:
checkpoint = torch.load(args.teacher_path, map_location='cpu')
teacher_model.load_state_dict(checkpoint['model'])
teacher_model.to(device)
teacher_model.eval()
# wrap the criterion in our custom DistillationLoss, which
# just dispatches to the original criterion if args.distillation_type is 'none'
criterion = DistillationLoss(criterion, teacher_model,
args.distillation_type,
args.distillation_alpha,
args.distillation_tau)
output_dir = Path(args.output_dir)
if args.resume:
if args.resume.startswith('https'):
checkpoint = torch.hub.load_state_dict_from_url(args.resume,
map_location='cpu',
check_hash=True)
else:
checkpoint = torch.load(args.resume, map_location='cpu')
model_without_ddp.load_state_dict(checkpoint['model'])
if not args.eval and 'optimizer' in checkpoint and 'lr_scheduler' in checkpoint and 'epoch' in checkpoint:
optimizer.load_state_dict(checkpoint['optimizer'])
lr_scheduler.load_state_dict(checkpoint['lr_scheduler'])
args.start_epoch = checkpoint['epoch'] + 1
if args.model_ema:
utils._load_checkpoint_for_ema(model_ema,
checkpoint['model_ema'])
if 'scaler' in checkpoint:
loss_scaler.load_state_dict(checkpoint['scaler'])
if args.eval:
test_stats = evaluate(data_loader_val, model, device)
print(
f"Accuracy of the network on the {len(dataset_val)} test images: {test_stats['acc1']:.1f}%"
)
return
print(f"Start training for {args.epochs} epochs")
start_time = time.time()
max_accuracy = 0.0
for epoch in range(args.start_epoch, args.epochs):
if args.distributed:
data_loader_train.sampler.set_epoch(epoch)
train_stats = train_one_epoch(
model,
criterion,
data_loader_train,
optimizer,
device,
epoch,
loss_scaler,
args.clip_grad,
model_ema,
mixup_fn,
set_training_mode=args.finetune ==
'' # keep in eval mode during finetuning
)
lr_scheduler.step(epoch)
if args.output_dir:
checkpoint_paths = [output_dir / ('checkpoint_%04d.pth' % (epoch))]
for checkpoint_path in checkpoint_paths:
utils.save_on_master(
{
'model': model_without_ddp.state_dict(),
'optimizer': optimizer.state_dict(),
'lr_scheduler': lr_scheduler.state_dict(),
'epoch': epoch,
'model_ema': get_state_dict(model_ema),
'scaler': loss_scaler.state_dict(),
'args': args,
}, checkpoint_path)
if not args.train_without_eval:
test_stats = evaluate(data_loader_val, model, device)
print(
f"Accuracy of the network on the {len(dataset_val)} test images: {test_stats['acc1']:.1f}%"
)
max_accuracy = max(max_accuracy, test_stats["acc1"])
print(f'Max accuracy: {max_accuracy:.2f}%')
log_stats = {
**{f'train_{k}': v
for k, v in train_stats.items()},
**{f'test_{k}': v
for k, v in test_stats.items()}, 'epoch': epoch,
'n_parameters': n_parameters
}
else:
log_stats = {
**{f'train_{k}': v
for k, v in train_stats.items()}, 'epoch': epoch,
'n_parameters': n_parameters
}
if args.output_dir and utils.is_main_process():
with (output_dir / "log.txt").open("a") as f:
f.write(json.dumps(log_stats) + "\n")
total_time = time.time() - start_time
total_time_str = str(datetime.timedelta(seconds=int(total_time)))
print('Training time {}'.format(total_time_str))
input_transform = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize([.485, .456, .406], [.229, .224, .225]),
])
data_kwargs = {
'base_size': args.base_size,
'crop_size': args.crop_size,
'transform': input_transform
}
val_dataset = get_segmentation_dataset(args.dataset,
split=args.split,
mode=args.mode,
**data_kwargs)
sampler = make_data_sampler(val_dataset, False, distributed)
batch_sampler = data.BatchSampler(sampler=sampler,
batch_size=args.batch_size,
drop_last=False)
val_data = data.DataLoader(val_dataset,
shuffle=False,
batch_sampler=batch_sampler,
num_workers=args.num_workers)
metric = SegmentationMetric(val_dataset.num_class)
metric = validate(model, val_data, metric, device)
ptutil.synchronize()
pixAcc, mIoU = ptutil.accumulate_metric(metric)
if ptutil.is_main_process():
print('pixAcc: %.4f, mIoU: %.4f' % (pixAcc, mIoU))
def eval_linear(args):
utils.init_distributed_mode(args)
print("git:\n {}\n".format(utils.get_sha()))
print("\n".join("%s: %s" % (k, str(v)) for k, v in sorted(dict(vars(args)).items())))
cudnn.benchmark = True
# ============ preparing data ... ============
train_transform = pth_transforms.Compose([
pth_transforms.RandomResizedCrop(224),
pth_transforms.RandomHorizontalFlip(),
pth_transforms.ToTensor(),
pth_transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225)),
])
val_transform = pth_transforms.Compose([
pth_transforms.Resize(256, interpolation=3),
pth_transforms.CenterCrop(224),
pth_transforms.ToTensor(),
pth_transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225)),
])
dataset_train = datasets.ImageFolder(os.path.join(args.data_path, "train"), transform=train_transform)
dataset_val = datasets.ImageFolder(os.path.join(args.data_path, "val"), transform=val_transform)
sampler = torch.utils.data.distributed.DistributedSampler(dataset_train)
train_loader = torch.utils.data.DataLoader(
dataset_train,
sampler=sampler,
batch_size=args.batch_size_per_gpu,
num_workers=args.num_workers,
pin_memory=True,
)
val_loader = torch.utils.data.DataLoader(
dataset_val,
batch_size=args.batch_size_per_gpu,
num_workers=args.num_workers,
pin_memory=True,
)
print(f"Data loaded with {len(dataset_train)} train and {len(dataset_val)} val imgs.")
# ============ building network ... ============
model = vits.__dict__[args.arch](patch_size=args.patch_size, num_classes=0)
model.cuda()
model.eval()
print(f"Model {args.arch} {args.patch_size}x{args.patch_size} built.")
# load weights to evaluate
utils.load_pretrained_weights(model, args.pretrained_weights, args.checkpoint_key, args.arch, args.patch_size)
linear_classifier = LinearClassifier(model.embed_dim * (args.n_last_blocks + int(args.avgpool_patchtokens)), num_labels=args.num_labels)
linear_classifier = linear_classifier.cuda()
linear_classifier = nn.parallel.DistributedDataParallel(linear_classifier, device_ids=[args.gpu])
# set optimizer
optimizer = torch.optim.SGD(
linear_classifier.parameters(),
args.lr * (args.batch_size_per_gpu * utils.get_world_size()) / 256., # linear scaling rule
momentum=0.9,
weight_decay=0, # we do not apply weight decay
)
scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(optimizer, args.epochs, eta_min=0)
# Optionally resume from a checkpoint
to_restore = {"epoch": 0, "best_acc": 0.}
utils.restart_from_checkpoint(
os.path.join(args.output_dir, "checkpoint.pth.tar"),
run_variables=to_restore,
state_dict=linear_classifier,
optimizer=optimizer,
scheduler=scheduler,
)
start_epoch = to_restore["epoch"]
best_acc = to_restore["best_acc"]
for epoch in range(start_epoch, args.epochs):
train_loader.sampler.set_epoch(epoch)
train_stats = train(model, linear_classifier, optimizer, train_loader, epoch, args.n_last_blocks, args.avgpool_patchtokens)
scheduler.step()
log_stats = {**{f'train_{k}': v for k, v in train_stats.items()},
'epoch': epoch}
if epoch % args.val_freq == 0 or epoch == args.epochs - 1:
test_stats = validate_network(val_loader, model, linear_classifier, args.n_last_blocks, args.avgpool_patchtokens)
print(f"Accuracy at epoch {epoch} of the network on the {len(dataset_val)} test images: {test_stats['acc1']:.1f}%")
best_acc = max(best_acc, test_stats["acc1"])
print(f'Max accuracy so far: {best_acc:.2f}%')
log_stats = {**{k: v for k, v in log_stats.items()},
**{f'test_{k}': v for k, v in test_stats.items()}}
if utils.is_main_process():
with (Path(args.output_dir) / "log.txt").open("a") as f:
f.write(json.dumps(log_stats) + "\n")
save_dict = {
"epoch": epoch + 1,
"state_dict": linear_classifier.state_dict(),
"optimizer": optimizer.state_dict(),
"scheduler": scheduler.state_dict(),
"best_acc": best_acc,
}
torch.save(save_dict, os.path.join(args.output_dir, "checkpoint.pth.tar"))
print("Training of the supervised linear classifier on frozen features completed.\n"
"Top-1 test accuracy: {acc:.1f}".format(acc=best_acc))
def main():
args = parse_args()
hvd.init()
set_affinity(hvd.local_rank())
if is_main_process():
log("Running total processes: {}".format(get_world_size()))
log("Starting process: {}".format(get_rank()))
if is_main_process():
dllogger.init(backends=[dllogger.JSONStreamBackend(verbosity=dllogger.Verbosity.VERBOSE,
filename=args.json_summary),
dllogger.StdOutBackend(verbosity=dllogger.Verbosity.VERBOSE, step_format=format_step)])
else:
dllogger.init(backends=[])
tf.random.set_seed(args.seed)
dllogger.log(step="PARAMETER", data={"SEED": args.seed})
# script parameters
BATCH_SIZE = args.train_batch_size
EVAL_BATCH_SIZE = args.predict_batch_size
USE_XLA = args.xla
USE_AMP = args.amp
EPOCHS = args.num_train_epochs
if not args.do_train:
EPOCHS = args.num_train_epochs = 1
log("Since running inference only, setting args.num_train_epochs to 1")
if not os.path.exists(args.output_dir) and is_main_process():
os.makedirs(args.output_dir)
# TensorFlow configuration
gpus = tf.config.experimental.list_physical_devices('GPU')
if gpus:
for gpu in gpus:
tf.config.experimental.set_memory_growth(gpu, True)
tf.config.experimental.set_visible_devices(gpus[hvd.local_rank()], 'GPU')
tf.config.optimizer.set_jit(USE_XLA)
#tf.config.optimizer.set_experimental_options({"auto_mixed_precision": USE_AMP})
if args.amp:
policy = tf.keras.mixed_precision.experimental.Policy("mixed_float16", loss_scale="dynamic")
tf.keras.mixed_precision.experimental.set_policy(policy)
print('Compute dtype: %s' % policy.compute_dtype) # Compute dtype: float16
print('Variable dtype: %s' % policy.variable_dtype) # Variable dtype: float32
if is_main_process():
log("***** Loading tokenizer and model *****")
# Load tokenizer and model from pretrained model/vocabulary. Specify the number of labels to classify (2+: classification, 1: regression)
electra_model = args.electra_model
config = ElectraConfig.from_pretrained(electra_model, cache_dir=args.cache_dir)
config.update({"amp": args.amp})
if args.vocab_file is None:
tokenizer = ElectraTokenizer.from_pretrained(electra_model, cache_dir=args.cache_dir)
else:
tokenizer = ElectraTokenizer(
vocab_file=args.vocab_file,
do_lower_case=args.do_lower_case)
model = TFElectraForQuestionAnswering.from_pretrained(electra_model, config=config, cache_dir=args.cache_dir, args=args)
if is_main_process():
log("***** Loading dataset *****")
# Load data
processor = SquadV2Processor() if args.version_2_with_negative else SquadV1Processor()
train_examples = processor.get_train_examples(args.data_dir) if args.do_train else None
dev_examples = processor.get_dev_examples(args.data_dir) if args.do_predict else None
if is_main_process():
log("***** Loading features *****")
# Load cached features
squad_version = '2.0' if args.version_2_with_negative else '1.1'
if args.cache_dir is None:
args.cache_dir = args.data_dir
cached_train_features_file = args.cache_dir.rstrip('/') + '/' + 'TF2_train-v{4}.json_{1}_{2}_{3}'.format(
electra_model.split("/")[1], str(args.max_seq_length), str(args.doc_stride),
str(args.max_query_length), squad_version)
cached_dev_features_file = args.cache_dir.rstrip('/') + '/' + 'TF2_dev-v{4}.json_{1}_{2}_{3}'.format(
electra_model.split("/")[1], str(args.max_seq_length), str(args.doc_stride),
str(args.max_query_length), squad_version)
try:
with open(cached_train_features_file, "rb") as reader:
train_features = pickle.load(reader) if args.do_train else []
with open(cached_dev_features_file, "rb") as reader:
dev_features = pickle.load(reader) if args.do_predict else []
except:
train_features = ( # TODO: (yy) do on rank 0?
squad_convert_examples_to_features(
examples=train_examples,
tokenizer=tokenizer,
max_seq_length=args.max_seq_length,
doc_stride=args.doc_stride,
max_query_length=args.max_query_length,
is_training=True,
return_dataset="",
)
if args.do_train
else []
)
dev_features = (
squad_convert_examples_to_features(
examples=dev_examples,
tokenizer=tokenizer,
max_seq_length=args.max_seq_length,
doc_stride=args.doc_stride,
max_query_length=args.max_query_length,
is_training=False,
return_dataset="",
)
if args.do_predict
else []
)
# Dump Cached features
if not args.skip_cache and is_main_process():
if args.do_train:
log("***** Building Cache Files: {} *****".format(cached_train_features_file))
with open(cached_train_features_file, "wb") as writer:
pickle.dump(train_features, writer)
if args.do_predict:
log("***** Building Cache Files: {} *****".format(cached_dev_features_file))
with open(cached_dev_features_file, "wb") as writer:
pickle.dump(dev_features, writer)
len_train_features = len(train_features)
total_train_steps = int((len_train_features * EPOCHS / BATCH_SIZE) / get_world_size()) + 1
train_steps_per_epoch = int((len_train_features / BATCH_SIZE) / get_world_size()) + 1
len_dev_features = len(dev_features)
total_dev_steps = int((len_dev_features / EVAL_BATCH_SIZE)) + 1
train_dataset = get_dataset_from_features(train_features, BATCH_SIZE,
v2=args.version_2_with_negative) if args.do_train else []
dev_dataset = get_dataset_from_features(dev_features, EVAL_BATCH_SIZE, drop_remainder=False, ngpu=1, mode="dev",
v2=args.version_2_with_negative) if args.do_predict else []
opt = create_optimizer(init_lr=args.learning_rate, num_train_steps=total_train_steps,
num_warmup_steps=int(args.warmup_proportion * total_train_steps),
weight_decay_rate=args.weight_decay_rate,
layerwise_lr_decay=args.layerwise_lr_decay,
n_transformer_layers=model.num_hidden_layers)
if USE_AMP:
# loss scaling is currently required when using mixed precision
opt = tf.keras.mixed_precision.experimental.LossScaleOptimizer(opt, "dynamic")
# Define loss function
loss = tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True)
loss_class = tf.keras.losses.BinaryCrossentropy(
from_logits=True,
name='binary_crossentropy'
)
metric = tf.keras.metrics.SparseCategoricalAccuracy("accuracy")
model.compile(optimizer=opt, loss=loss, metrics=[metric])
train_loss_results = []
if args.do_train and is_main_process():
log("***** Running training *****")
log(" Num examples = ", len_train_features)
log(" Num Epochs = ", args.num_train_epochs)
log(" Instantaneous batch size per GPU = ", args.train_batch_size)
log(
" Total train batch size (w. parallel, distributed & accumulation) = ",
args.train_batch_size
* get_world_size(),
)
log(" Total optimization steps =", total_train_steps)
total_train_time = 0
latency = []
for epoch in range(EPOCHS):
if args.do_train:
epoch_loss_avg = tf.keras.metrics.Mean()
epoch_perf_avg = tf.keras.metrics.Mean()
epoch_start = time.time()
epoch_iterator = tqdm(train_dataset, total=train_steps_per_epoch, desc="Iteration", mininterval=5,
disable=not is_main_process())
for iter, inputs in enumerate(epoch_iterator):
# breaking criterion if max_steps if > 1
if args.max_steps > 0 and (epoch * train_steps_per_epoch + iter) > args.max_steps:
break
iter_start = time.time()
# Optimize the model
loss_value = train_step(model, inputs, loss, USE_AMP, opt, (iter == 0 and epoch == 0),
v2=args.version_2_with_negative, loss_class=loss_class, fp16=USE_AMP)
epoch_perf_avg.update_state(1. * BATCH_SIZE / (time.time() - iter_start))
if iter % args.log_freq == 0:
if is_main_process():
log("\nEpoch: {:03d}, Step:{:6d}, Loss:{:12.8f}, Perf:{:5.0f}, loss_scale:{}, opt_step:{}".format(epoch, iter, loss_value,
epoch_perf_avg.result() * get_world_size(), opt.loss_scale if config.amp else 1,
int(opt.iterations)))
dllogger.log(step=(epoch, iter,), data={"step_loss": float(loss_value.numpy()),
"train_perf": float( epoch_perf_avg.result().numpy() * get_world_size())})
# Track progress
epoch_loss_avg.update_state(loss_value) # Add current batch loss
# End epoch
train_loss_results.append(epoch_loss_avg.result())
total_train_time += float(time.time() - epoch_start)
# Summarize and save checkpoint at the end of each epoch
if is_main_process():
dllogger.log(step=tuple(), data={"e2e_train_time": total_train_time,
"training_sequences_per_second": float(
epoch_perf_avg.result().numpy() * get_world_size()),
"final_loss": float(epoch_loss_avg.result().numpy())})
if not args.skip_checkpoint:
if args.ci:
checkpoint_name = "{}/electra_base_qa_v2_{}_epoch_{}_ckpt".format(args.output_dir, args.version_2_with_negative, epoch + 1)
else:
checkpoint_name = "checkpoints/electra_base_qa_v2_{}_epoch_{}_ckpt".format(args.version_2_with_negative, epoch + 1)
if is_main_process():
model.save_weights(checkpoint_name)
if args.do_predict and (args.evaluate_during_training or epoch == args.num_train_epochs - 1):
if not args.do_train:
log("***** Loading checkpoint: {} *****".format(args.init_checkpoint))
model.load_weights(args.init_checkpoint).expect_partial()
current_feature_id = 0
all_results = []
if is_main_process():
log("***** Running evaluation *****")
log(" Num Batches = ", total_dev_steps)
log(" Batch size = ", args.predict_batch_size)
raw_infer_start = time.time()
if is_main_process():
infer_perf_avg = tf.keras.metrics.Mean()
dev_iterator = tqdm(dev_dataset, total=total_dev_steps, desc="Iteration", mininterval=5,
disable=not is_main_process())
for input_ids, input_mask, segment_ids, start_positions, end_positions, cls_index, p_mask, is_impossible in dev_iterator:
# training=False is needed only if there are layers with different
# behavior during training versus inference (e.g. Dropout).
iter_start = time.time()
if not args.joint_head:
batch_start_logits, batch_end_logits = infer_step(model, input_ids,
attention_mask=input_mask,
token_type_ids=segment_ids,
)[:2]
#Synchronize with GPU to compute time
_ = batch_start_logits.numpy()
else:
outputs = infer_step(model, input_ids,
attention_mask=input_mask,
token_type_ids=segment_ids,
cls_index=cls_index,
p_mask=p_mask,
)
#Synchronize with GPU to compute time
_ = outputs[0].numpy()
infer_time = (time.time() - iter_start)
infer_perf_avg.update_state(1. * EVAL_BATCH_SIZE / infer_time)
latency.append(infer_time)
for iter_ in range(input_ids.shape[0]):
if not args.joint_head:
start_logits = batch_start_logits[iter_].numpy().tolist()
end_logits = batch_end_logits[iter_].numpy().tolist()
dev_feature = dev_features[current_feature_id]
current_feature_id += 1
unique_id = int(dev_feature.unique_id)
all_results.append(RawResult(unique_id=unique_id,
start_logits=start_logits,
end_logits=end_logits))
else:
dev_feature = dev_features[current_feature_id]
current_feature_id += 1
unique_id = int(dev_feature.unique_id)
output = [output[iter_].numpy().tolist() for output in outputs]
start_logits = output[0]
start_top_index = output[1]
end_logits = output[2]
end_top_index = output[3]
cls_logits = output[4]
result = SquadResult(
unique_id,
start_logits,
end_logits,
start_top_index=start_top_index,
end_top_index=end_top_index,
cls_logits=cls_logits,
)
all_results.append(result)
# Compute and save predictions
answers, nbest_answers = get_answers(dev_examples, dev_features, all_results, args)
output_prediction_file = os.path.join(args.output_dir, "predictions.json")
output_nbest_file = os.path.join(args.output_dir, "nbest_predictions.json")
e2e_infer_time = time.time() - raw_infer_start
# if args.version_2_with_negative:
# output_null_log_odds_file = os.path.join(args.output_dir, "null_odds.json")
# else:
# output_null_log_odds_file = None
with open(output_prediction_file, "w") as f:
f.write(json.dumps(answers, indent=4) + "\n")
with open(output_nbest_file, "w") as f:
f.write(json.dumps(nbest_answers, indent=4) + "\n")
if args.do_eval:
if args.version_2_with_negative:
dev_file = "dev-v2.0.json"
else:
dev_file = "dev-v1.1.json"
eval_out = subprocess.check_output([sys.executable, args.eval_script,
args.data_dir + "/" + dev_file, output_prediction_file])
log(eval_out.decode('UTF-8'))
scores = str(eval_out).strip()
exact_match = float(scores.split(":")[1].split(",")[0])
if args.version_2_with_negative:
f1 = float(scores.split(":")[2].split(",")[0])
else:
f1 = float(scores.split(":")[2].split("}")[0])
log("Epoch: {:03d} Results: {}".format(epoch, eval_out.decode('UTF-8')))
log("**EVAL SUMMARY** - Epoch: {:03d}, EM: {:6.3f}, F1: {:6.3f}, Infer_Perf: {:4.0f} seq/s"
.format(epoch, exact_match, f1, infer_perf_avg.result()))
latency_all = sorted(latency)[:-2]
log(
"**LATENCY SUMMARY** - Epoch: {:03d}, Ave: {:6.3f} ms, 90%: {:6.3f} ms, 95%: {:6.3f} ms, 99%: {:6.3f} ms"
.format(epoch, sum(latency_all) / len(latency_all) * 1000,
sum(latency_all[:int(len(latency_all) * 0.9)]) / int(len(latency_all) * 0.9) * 1000,
sum(latency_all[:int(len(latency_all) * 0.95)]) / int(len(latency_all) * 0.95) * 1000,
sum(latency_all[:int(len(latency_all) * 0.99)]) / int(len(latency_all) * 0.99) * 1000,
))
dllogger.log(step=tuple(),
data={"inference_sequences_per_second": float(infer_perf_avg.result().numpy()),
"e2e_inference_time": e2e_infer_time})
if is_main_process() and args.do_train and args.do_eval:
log(
"**RESULTS SUMMARY** - EM: {:6.3f}, F1: {:6.3f}, Train_Time: {:4.0f} s, Train_Perf: {:4.0f} seq/s, Infer_Perf: {:4.0f} seq/s"
.format(exact_match, f1, total_train_time, epoch_perf_avg.result() * get_world_size(),
infer_perf_avg.result()))
dllogger.log(step=tuple(), data={"exact_match": exact_match, "F1": f1})
def main(args):
utils.init_distributed_mode(args)
update_config_from_file(args.cfg)
print(args)
args_text = yaml.safe_dump(args.__dict__, default_flow_style=False)
device = torch.device(args.device)
# fix the seed for reproducibility
seed = args.seed + utils.get_rank()
torch.manual_seed(seed)
np.random.seed(seed)
# random.seed(seed)
cudnn.benchmark = True
dataset_train, args.nb_classes = build_dataset(is_train=True, args=args)
dataset_val, _ = build_dataset(is_train=False, args=args)
if args.distributed:
num_tasks = utils.get_world_size()
global_rank = utils.get_rank()
if args.repeated_aug:
sampler_train = RASampler(dataset_train,
num_replicas=num_tasks,
rank=global_rank,
shuffle=True)
else:
sampler_train = torch.utils.data.DistributedSampler(
dataset_train,
num_replicas=num_tasks,
rank=global_rank,
shuffle=True)
if args.dist_eval:
if len(dataset_val) % num_tasks != 0:
print(
'Warning: Enabling distributed evaluation with an eval dataset not divisible by process number. '
'This will slightly alter validation results as extra duplicate entries are added to achieve '
'equal num of samples per-process.')
sampler_val = torch.utils.data.DistributedSampler(
dataset_val,
num_replicas=num_tasks,
rank=global_rank,
shuffle=False)
else:
sampler_val = torch.utils.data.SequentialSampler(dataset_val)
else:
sampler_val = torch.utils.data.SequentialSampler(dataset_val)
sampler_train = torch.utils.data.RandomSampler(dataset_train)
data_loader_train = torch.utils.data.DataLoader(
dataset_train,
sampler=sampler_train,
batch_size=args.batch_size,
num_workers=args.num_workers,
pin_memory=args.pin_mem,
drop_last=True,
)
data_loader_val = torch.utils.data.DataLoader(dataset_val,
batch_size=int(
2 * args.batch_size),
sampler=sampler_val,
num_workers=args.num_workers,
pin_memory=args.pin_mem,
drop_last=False)
mixup_fn = None
mixup_active = args.mixup > 0 or args.cutmix > 0. or args.cutmix_minmax is not None
if mixup_active:
mixup_fn = Mixup(mixup_alpha=args.mixup,
cutmix_alpha=args.cutmix,
cutmix_minmax=args.cutmix_minmax,
prob=args.mixup_prob,
switch_prob=args.mixup_switch_prob,
mode=args.mixup_mode,
label_smoothing=args.smoothing,
num_classes=args.nb_classes)
print(f"Creating SuperVisionTransformer")
print(cfg)
model = Vision_TransformerSuper(
img_size=args.input_size,
patch_size=args.patch_size,
embed_dim=cfg.SUPERNET.EMBED_DIM,
depth=cfg.SUPERNET.DEPTH,
num_heads=cfg.SUPERNET.NUM_HEADS,
mlp_ratio=cfg.SUPERNET.MLP_RATIO,
qkv_bias=True,
drop_rate=args.drop,
drop_path_rate=args.drop_path,
gp=args.gp,
num_classes=args.nb_classes,
max_relative_position=args.max_relative_position,
relative_position=args.relative_position,
change_qkv=args.change_qkv,
abs_pos=not args.no_abs_pos)
choices = {
'num_heads': cfg.SEARCH_SPACE.NUM_HEADS,
'mlp_ratio': cfg.SEARCH_SPACE.MLP_RATIO,
'embed_dim': cfg.SEARCH_SPACE.EMBED_DIM,
'depth': cfg.SEARCH_SPACE.DEPTH
}
model.to(device)
if args.teacher_model:
teacher_model = create_model(
args.teacher_model,
pretrained=True,
num_classes=args.nb_classes,
)
teacher_model.to(device)
teacher_loss = LabelSmoothingCrossEntropy(smoothing=args.smoothing)
else:
teacher_model = None
teacher_loss = None
model_ema = None
model_without_ddp = model
if args.distributed:
model = torch.nn.parallel.DistributedDataParallel(
model, device_ids=[args.gpu], find_unused_parameters=True)
model_without_ddp = model.module
n_parameters = sum(p.numel() for p in model.parameters()
if p.requires_grad)
print('number of params:', n_parameters)
linear_scaled_lr = args.lr * args.batch_size * utils.get_world_size(
) / 512.0
args.lr = linear_scaled_lr
optimizer = create_optimizer(args, model_without_ddp)
loss_scaler = NativeScaler()
lr_scheduler, _ = create_scheduler(args, optimizer)
# criterion = LabelSmoothingCrossEntropy()
if args.mixup > 0.:
# smoothing is handled with mixup label transform
criterion = SoftTargetCrossEntropy()
elif args.smoothing:
criterion = LabelSmoothingCrossEntropy(smoothing=args.smoothing)
else:
criterion = torch.nn.CrossEntropyLoss()
output_dir = Path(args.output_dir)
if not output_dir.exists():
output_dir.mkdir(parents=True)
# save config for later experiments
with open(output_dir / "config.yaml", 'w') as f:
f.write(args_text)
if args.resume:
if args.resume.startswith('https'):
checkpoint = torch.hub.load_state_dict_from_url(args.resume,
map_location='cpu',
check_hash=True)
else:
checkpoint = torch.load(args.resume, map_location='cpu')
model_without_ddp.load_state_dict(checkpoint['model'])
if not args.eval and 'optimizer' in checkpoint and 'lr_scheduler' in checkpoint and 'epoch' in checkpoint:
optimizer.load_state_dict(checkpoint['optimizer'])
lr_scheduler.load_state_dict(checkpoint['lr_scheduler'])
args.start_epoch = checkpoint['epoch'] + 1
if 'scaler' in checkpoint:
loss_scaler.load_state_dict(checkpoint['scaler'])
if args.model_ema:
utils._load_checkpoint_for_ema(model_ema,
checkpoint['model_ema'])
retrain_config = None
if args.mode == 'retrain' and "RETRAIN" in cfg:
retrain_config = {
'layer_num': cfg.RETRAIN.DEPTH,
'embed_dim': [cfg.RETRAIN.EMBED_DIM] * cfg.RETRAIN.DEPTH,
'num_heads': cfg.RETRAIN.NUM_HEADS,
'mlp_ratio': cfg.RETRAIN.MLP_RATIO
}
if args.eval:
print(retrain_config)
test_stats = evaluate(data_loader_val,
model,
device,
mode=args.mode,
retrain_config=retrain_config)
print(
f"Accuracy of the network on the {len(dataset_val)} test images: {test_stats['acc1']:.1f}%"
)
return
print("Start training")
start_time = time.time()
max_accuracy = 0.0
for epoch in range(args.start_epoch, args.epochs):
if args.distributed:
data_loader_train.sampler.set_epoch(epoch)
train_stats = train_one_epoch(
model,
criterion,
data_loader_train,
optimizer,
device,
epoch,
loss_scaler,
args.clip_grad,
model_ema,
mixup_fn,
amp=args.amp,
teacher_model=teacher_model,
teach_loss=teacher_loss,
choices=choices,
mode=args.mode,
retrain_config=retrain_config,
)
lr_scheduler.step(epoch)
if args.output_dir:
checkpoint_paths = [output_dir / 'checkpoint.pth']
for checkpoint_path in checkpoint_paths:
utils.save_on_master(
{
'model': model_without_ddp.state_dict(),
'optimizer': optimizer.state_dict(),
'lr_scheduler': lr_scheduler.state_dict(),
'epoch': epoch,
# 'model_ema': get_state_dict(model_ema),
'scaler': loss_scaler.state_dict(),
'args': args,
},
checkpoint_path)
test_stats = evaluate(data_loader_val,
model,
device,
amp=args.amp,
choices=choices,
mode=args.mode,
retrain_config=retrain_config)
print(
f"Accuracy of the network on the {len(dataset_val)} test images: {test_stats['acc1']:.1f}%"
)
max_accuracy = max(max_accuracy, test_stats["acc1"])
print(f'Max accuracy: {max_accuracy:.2f}%')
log_stats = {
**{f'train_{k}': v
for k, v in train_stats.items()},
**{f'test_{k}': v
for k, v in test_stats.items()}, 'epoch': epoch,
'n_parameters': n_parameters
}
if args.output_dir and utils.is_main_process():
with (output_dir / "log.txt").open("a") as f:
f.write(json.dumps(log_stats) + "\n")
total_time = time.time() - start_time
total_time_str = str(datetime.timedelta(seconds=int(total_time)))
print('Training time {}'.format(total_time_str))
def main():
global timeout_sent
args = parse_arguments()
random.seed(args.seed + args.local_rank)
np.random.seed(args.seed + args.local_rank)
torch.manual_seed(args.seed + args.local_rank)
torch.cuda.manual_seed(args.seed + args.local_rank)
worker_init = WorkerInitObj(args.seed + args.local_rank)
device, args = setup_training(args)
dllogger.log(step="PARAMETER", data={"Config": [str(args)]})
# Prepare optimizer
model, optimizer, lr_scheduler, checkpoint, global_step, criterion = prepare_model_and_optimizer(args, device)
if args.disable_weight_tying:
# Sanity Check that new param is in optimizer
print ("SANITY CHECK OPTIMIZER: ", id(model.module.cls.predictions.decoder.weight) in [id(g) for g in optimizer.param_groups[0]['params']])
assert id(model.module.cls.predictions.decoder.weight) in [id(g) for g in optimizer.param_groups[0]['params']]
print (f"SAVING EVERY {args.num_steps_per_checkpoint} STEPS!")
if is_main_process():
dllogger.log(step="PARAMETER", data={"SEED": args.seed})
raw_train_start = None
if args.do_train:
if is_main_process():
dllogger.log(step="PARAMETER", data={"train_start": True})
dllogger.log(step="PARAMETER", data={"batch_size_per_gpu": args.train_batch_size})
dllogger.log(step="PARAMETER", data={"learning_rate": args.learning_rate})
model.train()
most_recent_ckpts_paths = []
average_loss = 0.0 # averaged loss every args.log_freq steps
epoch = 0
training_steps = 0
pool = ProcessPoolExecutor(1)
# Note: We loop infinitely over epochs, termination is handled via iteration count
while True:
thread = None
restored_data_loader = None
if not args.resume_from_checkpoint or epoch > 0 or (args.phase2 and global_step < 1) or args.init_checkpoint:
files = [os.path.join(args.input_dir, f) for f in os.listdir(args.input_dir) if
os.path.isfile(os.path.join(args.input_dir, f)) and ('training' in f or 'train' in f)]
files.sort()
num_files = len(files)
random.Random(args.seed + epoch).shuffle(files)
f_start_id = 0
else:
f_start_id = checkpoint['files'][0]
files = checkpoint['files'][1:]
args.resume_from_checkpoint = False
num_files = len(files)
# may not exist in all checkpoints
epoch = checkpoint.get('epoch', 0)
restored_data_loader = checkpoint.get('data_loader', None)
shared_file_list = {}
if torch.distributed.is_initialized() and get_world_size() > num_files:
remainder = get_world_size() % num_files
data_file = files[(f_start_id*get_world_size()+get_rank() + remainder*f_start_id)%num_files]
else:
data_file = files[(f_start_id*get_world_size()+get_rank())%num_files]
previous_file = data_file
if restored_data_loader is None:
train_data = pretraining_dataset(data_file, args.max_predictions_per_seq)
train_sampler = RandomSampler(train_data)
train_dataloader = DataLoader(train_data, sampler=train_sampler,
batch_size=args.train_batch_size * args.n_gpu,
num_workers=4, worker_init_fn=worker_init,
pin_memory=True)
# shared_file_list["0"] = (train_dataloader, data_file)
else:
train_dataloader = restored_data_loader
restored_data_loader = None
overflow_buf = None
if args.allreduce_post_accumulation:
overflow_buf = torch.cuda.IntTensor([0])
for f_id in range(f_start_id + 1 , len(files)):
if get_world_size() > num_files:
data_file = files[(f_id*get_world_size()+get_rank() + remainder*f_id)%num_files]
else:
data_file = files[(f_id*get_world_size()+get_rank())%num_files]
previous_file = data_file
dataset_future = pool.submit(create_pretraining_dataset, data_file, args.max_predictions_per_seq, shared_file_list, args, worker_init)
train_iter = tqdm(train_dataloader, desc="Iteration", disable=args.disable_progress_bar) if is_main_process() else train_dataloader
if raw_train_start is None:
raw_train_start = time.time()
for step, batch in enumerate(train_iter):
training_steps += 1
batch = [t.to(device) for t in batch]
input_ids, segment_ids, input_mask, masked_lm_labels, next_sentence_labels = batch
prediction_scores, seq_relationship_score = model(input_ids=input_ids, token_type_ids=segment_ids, attention_mask=input_mask)
loss, mlm_loss, ns_loss = criterion(prediction_scores, seq_relationship_score, masked_lm_labels, next_sentence_labels)
if args.n_gpu > 1:
loss = loss.mean() # mean() to average on multi-gpu.
mlm_loss = mlm_loss.detach().mean()
ns_loss = ns_loss.detach().mean()
divisor = args.gradient_accumulation_steps
if args.gradient_accumulation_steps > 1:
if not args.allreduce_post_accumulation:
# this division was merged into predivision
loss = loss / args.gradient_accumulation_steps
mlm_loss = mlm_loss.detach() / args.gradient_accumulation_steps
ns_loss = ns_loss.detach() / args.gradient_accumulation_steps
divisor = 1.0
if args.fp16:
with amp.scale_loss(loss, optimizer, delay_overflow_check=args.allreduce_post_accumulation) as scaled_loss:
scaled_loss.backward()
else:
loss.backward()
average_loss += loss.item()
if training_steps % args.gradient_accumulation_steps == 0:
lr_scheduler.step() # learning rate warmup
global_step = take_optimizer_step(args, optimizer, model, overflow_buf, global_step)
if global_step >= args.steps_this_run or timeout_sent:
train_time_raw = time.time() - raw_train_start
last_num_steps = int(training_steps / args.gradient_accumulation_steps) % args.log_freq
last_num_steps = args.log_freq if last_num_steps == 0 else last_num_steps
average_loss = torch.tensor(average_loss, dtype=torch.float32).cuda()
average_loss = average_loss / (last_num_steps * divisor)
if (torch.distributed.is_initialized()):
average_loss /= get_world_size()
torch.distributed.all_reduce(average_loss)
final_loss = average_loss.item()
if is_main_process():
dllogger.log(step=(epoch, global_step, ), data={"final_loss": final_loss})
elif training_steps % (args.log_freq * args.gradient_accumulation_steps) == 0:
if is_main_process():
dllogger.log(step=(epoch, global_step, ), data={"average_loss": average_loss / (args.log_freq * divisor),
"step_loss": loss.item() * args.gradient_accumulation_steps / divisor,
"learning_rate": optimizer.param_groups[0]['lr'],
"mlm_loss" : mlm_loss.item(),
"ns_loss" : ns_loss.item()})
average_loss = 0
if global_step >= args.steps_this_run or training_steps % (
args.num_steps_per_checkpoint * args.gradient_accumulation_steps) == 0 or timeout_sent:
if is_main_process() and not args.skip_checkpoint:
# Save a trained model
dllogger.log(step="PARAMETER", data={"checkpoint_step": global_step})
model_to_save = model.module if hasattr(model,
'module') else model # Only save the model it-self
if args.resume_step < 0 or not args.phase2:
output_save_file = os.path.join(args.output_dir, "ckpt_{}.pt".format(global_step))
else:
output_save_file = os.path.join(args.output_dir, "ckpt_{}.pt".format(global_step + args.phase1_end_step))
if args.do_train:
torch.save({'model': model_to_save.state_dict(),
'optimizer': optimizer.state_dict(),
'master params': list(amp.master_params(optimizer)),
'files': [f_id] + files,
'epoch': epoch,
'data_loader': None if global_step >= args.max_steps else train_dataloader}, output_save_file)
most_recent_ckpts_paths.append(output_save_file)
# Exiting the training due to hitting max steps, or being sent a
# timeout from the cluster scheduler
if global_step >= args.steps_this_run or timeout_sent:
del train_dataloader
# thread.join()
return args, final_loss, train_time_raw, global_step
del train_dataloader
# thread.join()
# Make sure pool has finished and switch train_dataloader
# NOTE: Will block until complete
train_dataloader, data_file = dataset_future.result(timeout=None)
epoch += 1
def prepare_model_and_optimizer(args, device):
# Prepare model
config = modeling.BertConfig.from_json_file(args.config_file)
# Padding for divisibility by 8
if config.vocab_size % 8 != 0:
config.vocab_size += 8 - (config.vocab_size % 8)
modeling.ACT2FN["bias_gelu"] = modeling.bias_gelu_training
model = modeling.BertForPreTraining(config)
if args.disable_weight_tying:
import torch.nn as nn
print ("WARNING!!!!!!! Disabling weight tying for this run")
print ("BEFORE ", model.cls.predictions.decoder.weight is model.bert.embeddings.word_embeddings.weight)
model.cls.predictions.decoder.weight = torch.nn.Parameter(model.cls.predictions.decoder.weight.clone().detach())
print ("AFTER ", model.cls.predictions.decoder.weight is model.bert.embeddings.word_embeddings.weight)
assert (model.cls.predictions.decoder.weight is model.bert.embeddings.word_embeddings.weight) == False
checkpoint = None
if not args.resume_from_checkpoint:
global_step = 0
else:
if args.resume_step == -1 and not args.init_checkpoint:
model_names = [f for f in os.listdir(args.output_dir) if f.endswith(".pt")]
args.resume_step = max([int(x.split('.pt')[0].split('_')[1].strip()) for x in model_names])
global_step = args.resume_step if not args.init_checkpoint else 0
if not args.init_checkpoint:
checkpoint = torch.load(os.path.join(args.output_dir, "ckpt_{}.pt".format(global_step)), map_location="cpu")
else:
checkpoint = torch.load(args.init_checkpoint, map_location="cpu")
model.load_state_dict(checkpoint['model'], strict=False)
if args.phase2 and not args.init_checkpoint:
global_step -= args.phase1_end_step
if is_main_process():
print("resume step from ", args.resume_step)
model.to(device)
param_optimizer = list(model.named_parameters())
no_decay = ['bias', 'gamma', 'beta', 'LayerNorm']
optimizer_grouped_parameters = [
{'params': [p for n, p in param_optimizer if not any(nd in n for nd in no_decay)], 'weight_decay': 0.01},
{'params': [p for n, p in param_optimizer if any(nd in n for nd in no_decay)], 'weight_decay': 0.0}]
optimizer = FusedAdam(optimizer_grouped_parameters,
lr=args.learning_rate)
lr_scheduler = PolyWarmUpScheduler(optimizer,
warmup=args.warmup_proportion,
total_steps=args.max_steps,
degree=1)
if args.fp16:
if args.loss_scale == 0:
model, optimizer = amp.initialize(model, optimizer, opt_level="O2", loss_scale="dynamic", cast_model_outputs=torch.float16)
else:
model, optimizer = amp.initialize(model, optimizer, opt_level="O2", loss_scale=args.loss_scale, cast_model_outputs=torch.float16)
amp._amp_state.loss_scalers[0]._loss_scale = args.init_loss_scale
model.checkpoint_activations(args.checkpoint_activations)
if args.resume_from_checkpoint:
if args.phase2 or args.init_checkpoint:
keys = list(checkpoint['optimizer']['state'].keys())
#Override hyperparameters from previous checkpoint
for key in keys:
checkpoint['optimizer']['state'][key]['step'] = global_step
for iter, item in enumerate(checkpoint['optimizer']['param_groups']):
checkpoint['optimizer']['param_groups'][iter]['step'] = global_step
checkpoint['optimizer']['param_groups'][iter]['t_total'] = args.max_steps
checkpoint['optimizer']['param_groups'][iter]['warmup'] = args.warmup_proportion
checkpoint['optimizer']['param_groups'][iter]['lr'] = args.learning_rate
optimizer.load_state_dict(checkpoint['optimizer']) # , strict=False)
# Restore AMP master parameters
if args.fp16:
optimizer._lazy_init_maybe_master_weights()
optimizer._amp_stash.lazy_init_called = True
optimizer.load_state_dict(checkpoint['optimizer'])
for param, saved_param in zip(amp.master_params(optimizer), checkpoint['master params']):
param.data.copy_(saved_param.data)
if args.local_rank != -1:
if not args.allreduce_post_accumulation:
model = DDP(model, message_size=250000000, gradient_predivide_factor=get_world_size())
else:
flat_dist_call([param.data for param in model.parameters()], torch.distributed.broadcast, (0,) )
elif args.n_gpu > 1:
model = torch.nn.DataParallel(model)
criterion = BertPretrainingCriterion(config.vocab_size)
if args.disable_weight_tying:
# Sanity Check that new param is in optimizer
print ("SANITY CHECK OPTIMIZER: ", id(model.module.cls.predictions.decoder.weight) in [id(g) for g in optimizer.param_groups[0]['params']])
assert id(model.module.cls.predictions.decoder.weight) in [id(g) for g in optimizer.param_groups[0]['params']]
return model, optimizer, lr_scheduler, checkpoint, global_step, criterion
del train_dataloader
# thread.join()
return args, final_loss, train_time_raw, global_step
del train_dataloader
# thread.join()
# Make sure pool has finished and switch train_dataloader
# NOTE: Will block until complete
train_dataloader, data_file = dataset_future.result(timeout=None)
epoch += 1
if __name__ == "__main__":
now = time.time()
args, final_loss, train_time_raw, global_step = main()
gpu_count = args.n_gpu
global_step += args.phase1_end_step if (args.phase2 and args.resume_step > 0) else 0
if args.resume_step == -1:
args.resume_step = 0
if torch.distributed.is_initialized():
gpu_count = get_world_size()
if is_main_process():
e2e_time = time.time() - now
training_perf = args.train_batch_size * args.gradient_accumulation_steps * gpu_count\
* (global_step - args.resume_step + skipped_steps) / train_time_raw
dllogger.log(step=tuple(), data={"e2e_train_time": e2e_time, "training_sequences_per_second": training_perf,
"final_loss": final_loss, "raw_train_time": train_time_raw })
dllogger.flush()