def main():
opt = 'tangent'
url, filename = (
"https://raw.githubusercontent.com/pytorch/hub/master/imagenet_classes.txt",
"imagenet_classes.txt")
urllib.request.urlretrieve(url, filename)
with open("imagenet_classes.txt", "r") as f:
categories = [s.strip() for s in f.readlines()]
model = timm.create_model('vit_base_patch32_384', pretrained=True)
model.eval()
config = resolve_data_config({}, model=model)
transform = create_transform(**config)
# TODO: FOR PARA PERCORRER IMAGENS DE TESTE
# TODO: SACAR METRICAS COM O GROUND TRUTH
img_name = 'class_256.jpg'
topk_prob, topk_catid = test(img_name,
model,
config,
transform,
opt,
topk=10)
for i in range(topk_prob.size(0)):
print(categories[topk_catid[i]], topk_prob[i].item())
def main():
with open("/home/paulo/projects/pytorch-image-models/imagenet_classes.txt","r") as f:
categories = [s.strip() for s in f.readlines()]
categories = np.array(categories, dtype=object)
parser = argparse.ArgumentParser(description='ViT')
parser.add_argument('--data_dir', default='data/s-mnist')
parser.add_argument('--dataset', default='smnist')
parser.add_argument('--resume', default='s2r-mnist-sgd-normalmodel_last.pth')
parser.add_argument('--mode', default='normal')
parser.add_argument('--image_size', default=60)
parser.add_argument('--patch_size', default=10)
parser.add_argument('--num_classes', default=10)
parser.add_argument('--batch', default=1024)
parser.add_argument('--cuda', default=True)
parser.add_argument('--output_csv')
args = parser.parse_args()
model = timm.create_model('vit_base_patch32_384', pretrained=True)
cuda = args.cuda
batch = args.batch
dataset = {'smnist': SMNIST, 'dvsc': DVSC, 'simagenet': SIMAGENET}
model.eval()
if cuda:
model = model.cuda()
P=np.array([],dtype=int)
config = resolve_data_config({}, model=model)
transform = create_transform(**config)
test_data = dataset[args.dataset](args.data_dir, 'test',transform, 384, 384, None)
test_loader = DataLoader(dataset=test_data, batch_size=batch, shuffle=False)
print("number of files: {}".format(len(test_data.image_ids)))
print("Test Start")
for i, data in enumerate(tqdm(test_loader)):
img, target = data
if cuda:
img = img.cuda()
with torch.no_grad():
preds = model(img)
probabilities = torch.nn.functional.softmax(preds, dim=1)
preds = torch.argmax(probabilities, dim =1)
P = np.concatenate([P,preds.cpu().numpy()])
pred_class = categories[P]
result = pd.DataFrame({'pred_class': pred_class})
result.to_csv(args.output_csv)
def __init__(self,
model_name,
detail=False,
device='cuda',
torchscript=False,
precision='float32',
fuser='',
num_warm_iter=10,
num_bench_iter=50,
use_train_size=False,
**kwargs):
self.model_name = model_name
self.detail = detail
self.device = device
self.use_amp, self.model_dtype, self.data_dtype = resolve_precision(
precision)
self.channels_last = kwargs.pop('channels_last', False)
self.amp_autocast = torch.cuda.amp.autocast if self.use_amp else suppress
if fuser:
set_jit_fuser(fuser)
self.model = create_model(
model_name,
num_classes=kwargs.pop('num_classes', None),
in_chans=3,
global_pool=kwargs.pop('gp', 'fast'),
scriptable=torchscript,
drop_rate=kwargs.pop('drop', 0.),
drop_path_rate=kwargs.pop('drop_path', None),
drop_block_rate=kwargs.pop('drop_block', None),
)
self.model.to(
device=self.device,
dtype=self.model_dtype,
memory_format=torch.channels_last if self.channels_last else None)
self.num_classes = self.model.num_classes
self.param_count = count_params(self.model)
_logger.info('Model %s created, param count: %d' %
(model_name, self.param_count))
self.scripted = False
if torchscript:
self.model = torch.jit.script(self.model)
self.scripted = True
data_config = resolve_data_config(kwargs,
model=self.model,
use_test_size=not use_train_size)
self.input_size = data_config['input_size']
self.batch_size = kwargs.pop('batch_size', 256)
self.example_inputs = None
self.num_warm_iter = num_warm_iter
self.num_bench_iter = num_bench_iter
self.log_freq = num_bench_iter // 5
if 'cuda' in self.device:
self.time_fn = partial(cuda_timestamp, device=self.device)
else:
self.time_fn = timestamp
def __init__(self, args):
super(PiT_Image_Encoder, self).__init__()
self.encoder = timm.create_model('pit_b_224',
pretrained=args.image_enc_pre_trained,
num_classes=args.label_num)
self.config = resolve_data_config({}, model=self.encoder)
self.transform = create_transform(**self.config)
for para in self.encoder.parameters():
assert para.requires_grad is True
def __init__(self, args):
super(SwinT_Image_Encoder, self).__init__()
self.encoder = timm.create_model('swin_base_patch4_window7_224',
pretrained=args.image_enc_pre_trained,
num_classes=3)
self.config = resolve_data_config({}, model=self.encoder)
self.transform = create_transform(**self.config)
for para in self.encoder.parameters():
assert para.requires_grad is True
def data_creator(config):
# torch.manual_seed(args.seed + torch.distributed.get_rank())
args = config["args"]
train_dir = join(args.data, "train")
val_dir = join(args.data, "val")
if args.mock_data:
util.mock_data(train_dir, val_dir)
# todo: verbose should depend on rank
data_config = resolve_data_config(vars(args), verbose=True)
dataset_train = Dataset(join(args.data, "train"))
dataset_eval = Dataset(join(args.data, "val"))
collate_fn = None
if args.prefetcher and args.mixup > 0:
# collate conflict (need to support deinterleaving in collate mixup)
assert args.num_aug_splits == 0
collate_fn = FastCollateMixup(args.mixup, args.smoothing,
args.num_classes)
common_params = dict(
input_size=data_config["input_size"],
use_prefetcher=args.prefetcher,
mean=data_config["mean"],
std=data_config["std"],
num_workers=1,
distributed=args.distributed,
pin_memory=args.pin_mem)
train_loader = create_loader(
dataset_train,
is_training=True,
batch_size=config[BATCH_SIZE],
re_prob=args.reprob,
re_mode=args.remode,
re_count=args.recount,
re_split=args.resplit,
collate_fn=collate_fn,
color_jitter=args.color_jitter,
auto_augment=args.aa,
interpolation=args.train_interpolation,
num_aug_splits=args.num_aug_splits, # always 0 right now
**common_params)
eval_loader = create_loader(
dataset_eval,
is_training=False,
batch_size=args.validation_batch_size_multiplier * config[BATCH_SIZE],
interpolation=data_config["interpolation"],
crop_pct=data_config["crop_pct"],
**common_params)
return train_loader, eval_loader
def main():
global args
setup_default_logging()
args, args_text = _parse_args()
global alpha
alpha = args.alpha
assert alpha == 0.67749, 'Make sure alpha = 0.67749'
args.device = 'cuda:0'
model = create_model(args.model,
pretrained=args.pretrained,
num_classes=args.num_classes,
drop_rate=args.drop,
global_pool=args.gp,
bn_tf=False,
bn_momentum=None,
bn_eps=None,
checkpoint_path=args.initial_checkpoint)
assert args.binarizable == 'T', 'Make sure --binarizable T'
if args.binarizable:
Model_binary_patch(model)
model.cuda()
use_amp = False
if has_apex and args.amp:
print('Using amp with --opt-level {}.'.format(args.opt_level))
model = amp.initialize(model, opt_level=args.opt_level)
use_amp = True
else:
print('Do NOT use amp.')
data_config = resolve_data_config(vars(args),
model=model,
verbose=args.local_rank == 0)
dataset_eval = torchvision.datasets.CIFAR100(root='~/Downloads/CIFAR100',
train=False,
download=True)
loader_eval = create_loader_CIFAR100(
dataset_eval,
input_size=data_config['input_size'],
batch_size=4 * args.batch_size,
is_training=False,
use_prefetcher=False,
interpolation=None,
mean=data_config['mean'],
std=data_config['std'],
num_workers=args.workers,
distributed=False,
)
validate_loss_fn = nn.CrossEntropyLoss().cuda()
eval_metrics = validate(model, loader_eval, validate_loss_fn, args)
def validate(args):
rng = jax.random.PRNGKey(0)
model, variables = create_model(args.model, pretrained=True, rng=rng)
print(f'Created {args.model} model. Validating...')
if args.no_jit:
eval_step = lambda images, labels: eval_forward(
model, variables, images, labels)
else:
eval_step = jax.jit(lambda images, labels: eval_forward(
model, variables, images, labels))
if os.path.splitext(args.data)[1] == '.tar' and os.path.isfile(args.data):
dataset = DatasetTar(args.data)
else:
dataset = Dataset(args.data)
data_config = resolve_data_config(vars(args), model=model)
loader = create_loader(dataset,
input_size=data_config['input_size'],
batch_size=args.batch_size,
use_prefetcher=False,
interpolation=data_config['interpolation'],
mean=data_config['mean'],
std=data_config['std'],
num_workers=8,
crop_pct=data_config['crop_pct'])
batch_time = AverageMeter()
correct_top1, correct_top5 = 0, 0
total_examples = 0
start_time = prev_time = time.time()
for batch_index, (images, labels) in enumerate(loader):
images = images.numpy().transpose(0, 2, 3, 1)
labels = labels.numpy()
top1_count, top5_count = eval_step(images, labels)
correct_top1 += top1_count
correct_top5 += top5_count
total_examples += images.shape[0]
batch_time.update(time.time() - prev_time)
if batch_index % 20 == 0 and batch_index > 0:
print(
f'Test: [{batch_index:>4d}/{len(loader)}] '
f'Rate: {images.shape[0] / batch_time.val:>5.2f}/s ({images.shape[0] / batch_time.avg:>5.2f}/s) '
f'Acc@1: {100 * correct_top1 / total_examples:>7.3f} '
f'Acc@5: {100 * correct_top5 / total_examples:>7.3f}')
prev_time = time.time()
acc_1 = 100 * correct_top1 / total_examples
acc_5 = 100 * correct_top5 / total_examples
print(
f'Validation complete. {total_examples / (prev_time - start_time):>5.2f} img/s. '
f'Acc@1 {acc_1:>7.3f}, Acc@5 {acc_5:>7.3f}')
return dict(top1=float(acc_1), top5=float(acc_5))
def create_model(
name: str,
device: Union[str, torch.device],
pretrained=True
) -> Tuple[torch.nn.Module, TensorType[3, 1, 1], TensorType[3, 1, 1]]:
model = timm.create_model(name, pretrained=pretrained,
scriptable=True).to(device)
model.eval()
config = resolve_data_config({}, model=model)
mean = torch.tensor(config["mean"], device=device).view(3, 1, 1)
std = torch.tensor(config["std"], device=device).view(3, 1, 1)
return model, mean, std
def __init__(self, model_id: str):
self.hf_cfg, self.arch = load_model_config_from_hf(model_id)
self.model = timm.create_model(f"hf_hub:{model_id}", pretrained=True)
self.transform = create_transform(
**resolve_data_config(self.hf_cfg, model=self.model)
)
self.model.eval()
self.top_k = min(self.model.num_classes, 5)
self.labels = self.hf_cfg.get("labels", None)
if self.labels is None:
self.labels = [f"LABEL_{i}" for i in range(self.model.num_classes)]
def validate(args):
model = create_model(args.model, pretrained=True)
print(f'Created {args.model} model. Validating...')
eval_step = objax.Jit(
lambda images, labels: eval_forward(model, images, labels),
model.vars())
dataset = create_dataset('imagenet', args.data)
data_config = resolve_data_config(vars(args), model=model)
loader = create_loader(
dataset,
input_size=data_config['input_size'],
batch_size=args.batch_size,
use_prefetcher=False,
interpolation=data_config['interpolation'],
mean=data_config['mean'],
std=data_config['std'],
num_workers=8,
crop_pct=data_config['crop_pct'])
batch_time = AverageMeter()
correct_top1, correct_top5 = 0, 0
total_examples = 0
start_time = prev_time = time.time()
for batch_index, (images, labels) in enumerate(loader):
images = images.numpy()
labels = labels.numpy()
top1_count, top5_count = eval_step(images, labels)
correct_top1 += int(top1_count)
correct_top5 += int(top5_count)
total_examples += images.shape[0]
batch_time.update(time.time() - prev_time)
if batch_index % 20 == 0 and batch_index > 0:
print(
f'Test: [{batch_index:>4d}/{len(loader)}] '
f'Rate: {images.shape[0] / batch_time.val:>5.2f}/s ({images.shape[0] / batch_time.avg:>5.2f}/s) '
f'Acc@1: {100 * correct_top1 / total_examples:>7.3f} '
f'Acc@5: {100 * correct_top5 / total_examples:>7.3f}')
prev_time = time.time()
acc_1 = 100 * correct_top1 / total_examples
acc_5 = 100 * correct_top5 / total_examples
print(f'Validation complete. {total_examples / (prev_time - start_time):>5.2f} img/s. '
f'Acc@1 {acc_1:>7.3f}, Acc@5 {acc_5:>7.3f}')
return dict(top1=float(acc_1), top5=float(acc_5))
def main():
args = parser.parse_args()
setup_default_logging(log_path=args.log_path)
_logger = logging.getLogger('inference')
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
# create model
model = create_model(args.model,
num_classes=args.num_classes,
in_chans=3,
checkpoint_path=args.checkpoint)
data_config = resolve_data_config(vars(args),
model=model) # 默认配置 mean, std
img_transform = create_transform(data_config['input_size'],
is_training=False,
mean=data_config['mean'],
std=data_config['std'],
crop_pct=data_config['crop_pct'])
model = model.to(device)
model.eval()
# confusion matrix
conf_matrix = np.zeros((args.num_classes, args.num_classes))
# predict
class_dir = os.listdir(args.data)
for cl in class_dir:
imgs = os.listdir(os.path.join(args.data, cl))
for img in imgs:
im = Image.open(os.path.join(args.data, cl, img))
im1 = img_transform(im).unsqueeze(0)
im1 = im1.to(device)
outputs = model(im1)
idx = torch.max(outputs, 1)[1].tolist()[0]
probs = torch.nn.functional.softmax(outputs, dim=1)[0].tolist()
_logger.info('%s: %d, %.2f' %
(os.path.join(args.data, cl, img), idx, probs[idx]))
conf_matrix[int(cl), idx] += 1
_logger.info(conf_matrix)
def create_classification_model(self, model_name):
if str(model_name.value) == "efficientnet_b3a":
model_path = os.getenv("HOME") + '/ros2_models/effecient.pth'
model = create_model('efficientnet_b3a',
False,
model_path,
num_classes=6)
config = resolve_data_config({}, model=model)
self.transform = create_transform(**config)
# sd = torch.load(model_path, map_location=self.device)
# model = models.resnet18()
# model.fc = torch.nn.Linear(512, 6)
# model.load_state_dict(sd['state_dict'])
return model
print("Only valid model option currently is efficientnet_b3a")
def __init__(self,
save_path=None,
train_batch_size=256,
test_batch_size=512,
valid_size=None,
n_worker=32,
resize_scale=0.08,
distort_color=None,
image_size=224,
tf_preprocessing=False,
num_replicas=None,
rank=None,
use_prefetcher=False,
pin_memory=False,
fp16=False):
warnings.filterwarnings('ignore')
dataset = Dataset(os.path.join(save_path, "val"),
load_bytes=tf_preprocessing)
dummy_model = create_model('efficientnet_b0')
data_config = resolve_data_config({}, model=dummy_model)
test_loader = create_loader(dataset,
input_size=image_size,
batch_size=test_batch_size,
use_prefetcher=use_prefetcher,
interpolation=data_config['interpolation'],
mean=data_config['mean'],
std=data_config['std'],
num_workers=n_worker,
crop_pct=data_config['crop_pct'],
pin_memory=pin_memory,
fp16=fp16,
tf_preprocessing=None)
self.test = test_loader
#_entry('tv_resnext50_32x4d', , ), # same weights as torchvision
#_entry('wide_resnet50_2' , ), # same weights as torchvision
#_entry('wide_resnet101_2', , ), # same weights as torchvision
_entry('xception', 'Xception', '1610.02357'),
]
for m in model_list:
model_name = m['model']
# create model from name
model = create_model(model_name, pretrained=True)
param_count = sum([m.numel() for m in model.parameters()])
print('Model %s, %s created. Param count: %d' % (model_name, m['paper_model_name'], param_count))
# get appropriate transform for model's default pretrained config
data_config = resolve_data_config(m['args'], model=model, verbose=True)
if m['ttp']:
model = TestTimePoolHead(model, model.default_cfg['pool_size'])
data_config['crop_pct'] = 1.0
input_transform = create_transform(**data_config)
# Run the benchmark
ImageNet.benchmark(
model=model,
paper_model_name=m['paper_model_name'],
paper_arxiv_id=m['paper_arxiv_id'],
input_transform=input_transform,
batch_size=m['batch_size'],
num_gpu=NUM_GPU,
data_root=os.environ.get('IMAGENET_DIR', './imagenet')
)
def validate(args):
# might as well try to validate something
args.pretrained = args.pretrained or not args.checkpoint
args.prefetcher = not args.no_prefetcher
# create model
model = create_model(
args.model,
num_classes=args.num_classes,
in_chans=3,
pretrained=args.pretrained)
if args.checkpoint:
load_checkpoint(model, args.checkpoint, args.use_ema)
param_count = sum([m.numel() for m in model.parameters()])
logging.info('Model %s created, param count: %d' % (args.model, param_count))
data_config = resolve_data_config(vars(args), model=model)
model, test_time_pool = apply_test_time_pool(model, data_config, args)
if args.torchscript:
torch.jit.optimized_execution(True)
model = torch.jit.script(model)
if args.amp:
model = amp.initialize(model.cuda(), opt_level='O1')
else:
model = model.cuda()
if args.num_gpu > 1:
model = torch.nn.DataParallel(model, device_ids=list(range(args.num_gpu)))
criterion = nn.CrossEntropyLoss().cuda()
#from torchvision.datasets import ImageNet
#dataset = ImageNet(args.data, split='val')
if os.path.splitext(args.data)[1] == '.tar' and os.path.isfile(args.data):
dataset = DatasetTar(args.data, load_bytes=args.tf_preprocessing, class_map=args.class_map)
else:
dataset = Dataset(args.data, load_bytes=args.tf_preprocessing, class_map=args.class_map)
crop_pct = 1.0 if test_time_pool else data_config['crop_pct']
loader = create_loader(
dataset,
input_size=data_config['input_size'],
batch_size=args.batch_size,
use_prefetcher=args.prefetcher,
interpolation=data_config['interpolation'],
mean=data_config['mean'],
std=data_config['std'],
num_workers=args.workers,
crop_pct=crop_pct,
pin_memory=args.pin_mem,
tf_preprocessing=args.tf_preprocessing)
batch_time = AverageMeter()
losses = AverageMeter()
top1 = AverageMeter()
top5 = AverageMeter()
model.eval()
end = time.time()
with torch.no_grad():
for i, (input, target) in enumerate(loader):
if args.no_prefetcher:
target = target.cuda()
input = input.cuda()
if args.fp16:
input = input.half()
# compute output
output = model(input)
loss = criterion(output, target)
# measure accuracy and record loss
acc1, acc5 = accuracy(output.data, target, topk=(1, 2))
losses.update(loss.item(), input.size(0))
top1.update(acc1.item(), input.size(0))
top5.update(acc5.item(), input.size(0))
# measure elapsed time
batch_time.update(time.time() - end)
end = time.time()
if i % args.log_freq == 0:
logging.info(
'Test: [{0:>4d}/{1}] '
'Time: {batch_time.val:.3f}s ({batch_time.avg:.3f}s, {rate_avg:>7.2f}/s) '
'Loss: {loss.val:>7.4f} ({loss.avg:>6.4f}) '
'Acc@1: {top1.val:>7.3f} ({top1.avg:>7.3f}) '
'Acc@5: {top5.val:>7.3f} ({top5.avg:>7.3f})'.format(
i, len(loader), batch_time=batch_time,
rate_avg=input.size(0) / batch_time.avg,
loss=losses, top1=top1, top5=top5))
results = OrderedDict(
top1=round(top1.avg, 4), top1_err=round(100 - top1.avg, 4),
top5=round(top5.avg, 4), top5_err=round(100 - top5.avg, 4),
param_count=round(param_count / 1e6, 2),
img_size=data_config['input_size'][-1],
cropt_pct=crop_pct,
interpolation=data_config['interpolation'])
logging.info(' * Acc@1 {:.3f} ({:.3f}) Acc@5 {:.3f} ({:.3f})'.format(
results['top1'], results['top1_err'], results['top5'], results['top5_err']))
return results
def main():
setup_default_logging()
args, args_text = _parse_args()
args.prefetcher = not args.no_prefetcher
args.distributed = False
if 'WORLD_SIZE' in os.environ:
args.distributed = int(os.environ['WORLD_SIZE']) > 1
if args.distributed and args.num_gpu > 1:
logging.warning(
'Using more than one GPU per process in distributed mode is not allowed. Setting num_gpu to 1.'
)
args.num_gpu = 1
args.device = 'cuda:0'
args.world_size = 1
args.rank = 0 # global rank
if args.distributed:
args.num_gpu = 1
args.device = 'cuda:%d' % args.local_rank
torch.cuda.set_device(args.local_rank)
torch.distributed.init_process_group(backend='nccl',
init_method='env://')
args.world_size = torch.distributed.get_world_size()
args.rank = torch.distributed.get_rank()
assert args.rank >= 0
if args.distributed:
logging.info(
'Training in distributed mode with multiple processes, 1 GPU per process. Process %d, total %d.'
% (args.rank, args.world_size))
else:
logging.info('Training with a single process on %d GPUs.' %
args.num_gpu)
torch.manual_seed(args.seed + args.rank)
model = create_model(args.model,
pretrained=args.pretrained,
num_classes=args.num_classes,
drop_rate=args.drop,
drop_connect_rate=args.drop_connect,
global_pool=args.gp,
bn_tf=args.bn_tf,
bn_momentum=args.bn_momentum,
bn_eps=args.bn_eps,
checkpoint_path=args.initial_checkpoint)
if args.local_rank == 0:
logging.info('Model %s created, param count: %d' %
(args.model, sum([m.numel()
for m in model.parameters()])))
data_config = resolve_data_config(vars(args),
model=model,
verbose=args.local_rank == 0)
num_aug_splits = 0
if args.aug_splits > 0:
assert args.aug_splits > 1, 'A split of 1 makes no sense'
num_aug_splits = args.aug_splits
if args.split_bn:
assert num_aug_splits > 1 or args.resplit
model = convert_splitbn_model(model, max(num_aug_splits, 2))
if args.num_gpu > 1:
if args.amp:
logging.warning(
'AMP does not work well with nn.DataParallel, disabling. Use distributed mode for multi-GPU AMP.'
)
args.amp = False
model = nn.DataParallel(model,
device_ids=list(range(args.num_gpu))).cuda()
else:
model.cuda()
optimizer = create_optimizer(args, model)
use_amp = False
if has_apex and args.amp:
model, optimizer = amp.initialize(model, optimizer, opt_level='O1')
use_amp = True
if args.local_rank == 0:
logging.info('NVIDIA APEX {}. AMP {}.'.format(
'installed' if has_apex else 'not installed',
'on' if use_amp else 'off'))
# optionally resume from a checkpoint
resume_state = {}
resume_epoch = None
if args.resume:
resume_state, resume_epoch = resume_checkpoint(model, args.resume)
if resume_state and not args.no_resume_opt:
if 'optimizer' in resume_state:
if args.local_rank == 0:
logging.info('Restoring Optimizer state from checkpoint')
optimizer.load_state_dict(resume_state['optimizer'])
if use_amp and 'amp' in resume_state and 'load_state_dict' in amp.__dict__:
if args.local_rank == 0:
logging.info('Restoring NVIDIA AMP state from checkpoint')
amp.load_state_dict(resume_state['amp'])
del resume_state
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=args.resume)
if args.distributed:
if args.sync_bn:
assert not args.split_bn
try:
if has_apex:
model = convert_syncbn_model(model)
else:
model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(
model)
if args.local_rank == 0:
logging.info(
'Converted model to use Synchronized BatchNorm. WARNING: You may have issues if using '
'zero initialized BN layers (enabled by default for ResNets) while sync-bn enabled.'
)
except Exception as e:
logging.error(
'Failed to enable Synchronized BatchNorm. Install Apex or Torch >= 1.1'
)
if has_apex:
model = DDP(model, delay_allreduce=True)
else:
if args.local_rank == 0:
logging.info(
"Using torch DistributedDataParallel. Install NVIDIA Apex for Apex DDP."
)
model = DDP(model,
device_ids=[args.local_rank
]) # can use device str in Torch >= 1.1
# NOTE: EMA model does not need to be wrapped by DDP
lr_scheduler, num_epochs = create_scheduler(args, optimizer)
start_epoch = 0
if args.start_epoch is not None:
# a specified start_epoch will always override the resume epoch
start_epoch = args.start_epoch
elif resume_epoch is not None:
start_epoch = resume_epoch
if lr_scheduler is not None and start_epoch > 0:
lr_scheduler.step(start_epoch)
if args.local_rank == 0:
logging.info('Scheduled epochs: {}'.format(num_epochs))
train_dir = os.path.join(args.data, 'train')
if not os.path.exists(train_dir):
logging.error(
'Training folder does not exist at: {}'.format(train_dir))
exit(1)
dataset_train = Dataset(train_dir)
collate_fn = None
if args.prefetcher and args.mixup > 0:
assert not num_aug_splits # collate conflict (need to support deinterleaving in collate mixup)
collate_fn = FastCollateMixup(args.mixup, args.smoothing,
args.num_classes)
if num_aug_splits > 1:
dataset_train = AugMixDataset(dataset_train, num_splits=num_aug_splits)
loader_train = create_loader(
dataset_train,
input_size=data_config['input_size'],
batch_size=args.batch_size,
is_training=True,
use_prefetcher=args.prefetcher,
re_prob=args.reprob,
re_mode=args.remode,
re_count=args.recount,
re_split=args.resplit,
color_jitter=args.color_jitter,
auto_augment=args.aa,
num_aug_splits=num_aug_splits,
interpolation=args.train_interpolation,
mean=data_config['mean'],
std=data_config['std'],
num_workers=args.workers,
distributed=args.distributed,
collate_fn=collate_fn,
pin_memory=args.pin_mem,
)
eval_dir = os.path.join(args.data, 'val')
if not os.path.isdir(eval_dir):
eval_dir = os.path.join(args.data, 'validation')
if not os.path.isdir(eval_dir):
logging.error(
'Validation folder does not exist at: {}'.format(eval_dir))
exit(1)
dataset_eval = Dataset(eval_dir)
loader_eval = create_loader(
dataset_eval,
input_size=data_config['input_size'],
batch_size=4 * args.batch_size,
is_training=False,
use_prefetcher=args.prefetcher,
interpolation=data_config['interpolation'],
mean=data_config['mean'],
std=data_config['std'],
num_workers=args.workers,
distributed=args.distributed,
crop_pct=data_config['crop_pct'],
pin_memory=args.pin_mem,
)
if args.jsd:
assert num_aug_splits > 1 # JSD only valid with aug splits set
train_loss_fn = JsdCrossEntropy(num_splits=num_aug_splits,
smoothing=args.smoothing).cuda()
validate_loss_fn = nn.CrossEntropyLoss().cuda()
elif args.mixup > 0.:
# smoothing is handled with mixup label transform
train_loss_fn = SoftTargetCrossEntropy().cuda()
validate_loss_fn = nn.CrossEntropyLoss().cuda()
elif args.smoothing:
train_loss_fn = LabelSmoothingCrossEntropy(
smoothing=args.smoothing).cuda()
validate_loss_fn = nn.CrossEntropyLoss().cuda()
else:
train_loss_fn = nn.CrossEntropyLoss().cuda()
validate_loss_fn = train_loss_fn
eval_metric = args.eval_metric
best_metric = None
best_epoch = None
saver = None
output_dir = ''
if args.local_rank == 0:
output_base = args.output if args.output else './output'
exp_name = '-'.join([
datetime.now().strftime("%Y%m%d-%H%M%S"), args.model,
str(data_config['input_size'][-1])
])
output_dir = get_outdir(output_base, 'train', exp_name)
decreasing = True if eval_metric == 'loss' else False
saver = CheckpointSaver(checkpoint_dir=output_dir,
decreasing=decreasing)
with open(os.path.join(output_dir, 'args.yaml'), 'w') as f:
f.write(args_text)
try:
for epoch in range(start_epoch, num_epochs):
if args.distributed:
loader_train.sampler.set_epoch(epoch)
train_metrics = train_epoch(epoch,
model,
loader_train,
optimizer,
train_loss_fn,
args,
lr_scheduler=lr_scheduler,
saver=saver,
output_dir=output_dir,
use_amp=use_amp,
model_ema=model_ema)
if args.distributed and args.dist_bn in ('broadcast', 'reduce'):
if args.local_rank == 0:
logging.info(
"Distributing BatchNorm running means and vars")
distribute_bn(model, args.world_size, args.dist_bn == 'reduce')
eval_metrics = validate(model, loader_eval, validate_loss_fn, args)
if model_ema is not None and not args.model_ema_force_cpu:
if args.distributed and args.dist_bn in ('broadcast',
'reduce'):
distribute_bn(model_ema, args.world_size,
args.dist_bn == 'reduce')
ema_eval_metrics = validate(model_ema.ema,
loader_eval,
validate_loss_fn,
args,
log_suffix=' (EMA)')
eval_metrics = ema_eval_metrics
if lr_scheduler is not None:
# step LR for next epoch
lr_scheduler.step(epoch + 1, eval_metrics[eval_metric])
update_summary(epoch,
train_metrics,
eval_metrics,
os.path.join(output_dir, 'summary.csv'),
write_header=best_metric is None)
if saver is not None:
# save proper checkpoint with eval metric
save_metric = eval_metrics[eval_metric]
best_metric, best_epoch = saver.save_checkpoint(
model,
optimizer,
args,
epoch=epoch,
model_ema=model_ema,
metric=save_metric,
use_amp=use_amp)
except KeyboardInterrupt:
pass
if best_metric is not None:
logging.info('*** Best metric: {0} (epoch {1})'.format(
best_metric, best_epoch))
def main():
setup_default_logging()
args, args_text = _parse_args()
if args.log_wandb:
if has_wandb:
wandb.init(project=args.experiment, config=args)
else:
_logger.warning(
"You've requested to log metrics to wandb but package not found. "
"Metrics not being logged to wandb, try `pip install wandb`")
args.prefetcher = not args.no_prefetcher
args.distributed = False
if 'WORLD_SIZE' in os.environ:
args.distributed = int(os.environ['WORLD_SIZE']) > 1
args.device = 'cuda:0'
args.world_size = 1
args.rank = 0 # global rank
if args.distributed:
args.device = 'cuda:%d' % args.local_rank
torch.cuda.set_device(args.local_rank)
torch.distributed.init_process_group(backend='nccl',
init_method='env://')
args.world_size = torch.distributed.get_world_size()
args.rank = torch.distributed.get_rank()
_logger.info(
'Training in distributed mode with multiple processes, 1 GPU per process. Process %d, total %d.'
% (args.rank, args.world_size))
else:
_logger.info('Training with a single process on 1 GPUs.')
assert args.rank >= 0
# resolve AMP arguments based on PyTorch / Apex availability
use_amp = None
if args.amp:
# `--amp` chooses native amp before apex (APEX ver not actively maintained)
if has_native_amp:
args.native_amp = True
elif has_apex:
args.apex_amp = True
if args.apex_amp and has_apex:
use_amp = 'apex'
elif args.native_amp and has_native_amp:
use_amp = 'native'
elif args.apex_amp or args.native_amp:
_logger.warning(
"Neither APEX or native Torch AMP is available, using float32. "
"Install NVIDA apex or upgrade to PyTorch 1.6")
random_seed(args.seed, args.rank)
model_KD = None
if args.kd_model_path is not None:
model_KD = build_kd_model(args)
model = create_model(
args.model,
pretrained=args.pretrained,
num_classes=args.num_classes,
drop_rate=args.drop,
drop_connect_rate=args.drop_connect, # DEPRECATED, use drop_path
drop_path_rate=args.drop_path,
drop_block_rate=args.drop_block,
global_pool=args.gp,
bn_momentum=args.bn_momentum,
bn_eps=args.bn_eps,
scriptable=args.torchscript,
checkpoint_path=args.initial_checkpoint)
if args.num_classes is None:
assert hasattr(
model, 'num_classes'
), 'Model must have `num_classes` attr if not set on cmd line/config.'
args.num_classes = model.num_classes # FIXME handle model default vs config num_classes more elegantly
if args.local_rank == 0:
_logger.info(
f'Model {safe_model_name(args.model)} created, param count:{sum([m.numel() for m in model.parameters()])}'
)
data_config = resolve_data_config(vars(args),
model=model,
verbose=args.local_rank == 0)
# setup augmentation batch splits for contrastive loss or split bn
num_aug_splits = 0
if args.aug_splits > 0:
assert args.aug_splits > 1, 'A split of 1 makes no sense'
num_aug_splits = args.aug_splits
# enable split bn (separate bn stats per batch-portion)
if args.split_bn:
assert num_aug_splits > 1 or args.resplit
model = convert_splitbn_model(model, max(num_aug_splits, 2))
# move model to GPU, enable channels last layout if set
model.cuda()
if args.channels_last:
model = model.to(memory_format=torch.channels_last)
# setup synchronized BatchNorm for distributed training
if args.distributed and args.sync_bn:
assert not args.split_bn
if has_apex and use_amp == 'apex':
# Apex SyncBN preferred unless native amp is activated
model = convert_syncbn_model(model)
else:
model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(model)
if args.local_rank == 0:
_logger.info(
'Converted model to use Synchronized BatchNorm. WARNING: You may have issues if using '
'zero initialized BN layers (enabled by default for ResNets) while sync-bn enabled.'
)
if args.torchscript:
assert not use_amp == 'apex', 'Cannot use APEX AMP with torchscripted model'
assert not args.sync_bn, 'Cannot use SyncBatchNorm with torchscripted model'
model = torch.jit.script(model)
optimizer = create_optimizer_v2(model, **optimizer_kwargs(cfg=args))
# setup automatic mixed-precision (AMP) loss scaling and op casting
amp_autocast = suppress # do nothing
loss_scaler = None
if use_amp == 'apex':
model, optimizer = amp.initialize(model, optimizer, opt_level='O1')
loss_scaler = ApexScaler()
if args.local_rank == 0:
_logger.info('Using NVIDIA APEX AMP. Training in mixed precision.')
elif use_amp == 'native':
amp_autocast = torch.cuda.amp.autocast
loss_scaler = NativeScaler()
if args.local_rank == 0:
_logger.info(
'Using native Torch AMP. Training in mixed precision.')
else:
if args.local_rank == 0:
_logger.info('AMP not enabled. Training in float32.')
# optionally resume from a checkpoint
resume_epoch = None
if args.resume:
resume_epoch = resume_checkpoint(
model,
args.resume,
optimizer=None if args.no_resume_opt else optimizer,
loss_scaler=None if args.no_resume_opt else loss_scaler,
log_info=args.local_rank == 0)
# setup exponential moving average of model weights, SWA could be used here too
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 = ModelEmaV2(
model,
decay=args.model_ema_decay,
device='cpu' if args.model_ema_force_cpu else None)
if args.resume:
load_checkpoint(model_ema.module, args.resume, use_ema=True)
# setup distributed training
if args.distributed:
if has_apex and use_amp == 'apex':
# Apex DDP preferred unless native amp is activated
if args.local_rank == 0:
_logger.info("Using NVIDIA APEX DistributedDataParallel.")
model = ApexDDP(model, delay_allreduce=True)
else:
if args.local_rank == 0:
_logger.info("Using native Torch DistributedDataParallel.")
model = NativeDDP(model,
device_ids=[args.local_rank],
broadcast_buffers=not args.no_ddp_bb)
# NOTE: EMA model does not need to be wrapped by DDP
# setup learning rate schedule and starting epoch
lr_scheduler, num_epochs = create_scheduler(args, optimizer)
start_epoch = 0
if args.start_epoch is not None:
# a specified start_epoch will always override the resume epoch
start_epoch = args.start_epoch
elif resume_epoch is not None:
start_epoch = resume_epoch
if lr_scheduler is not None and start_epoch > 0:
lr_scheduler.step(start_epoch)
if args.local_rank == 0:
_logger.info('Scheduled epochs: {}'.format(num_epochs))
# create the train and eval datasets
dataset_train = create_dataset(args.dataset,
root=args.data_dir,
split=args.train_split,
is_training=True,
class_map=args.class_map,
download=args.dataset_download,
batch_size=args.batch_size,
repeats=args.epoch_repeats)
dataset_eval = create_dataset(args.dataset,
root=args.data_dir,
split=args.val_split,
is_training=False,
class_map=args.class_map,
download=args.dataset_download,
batch_size=args.batch_size)
# setup mixup / cutmix
collate_fn = None
mixup_fn = None
mixup_active = args.mixup > 0 or args.cutmix > 0. or args.cutmix_minmax is not None
if mixup_active:
mixup_args = dict(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.num_classes)
if args.prefetcher:
assert not num_aug_splits # collate conflict (need to support deinterleaving in collate mixup)
collate_fn = FastCollateMixup(**mixup_args)
else:
mixup_fn = Mixup(**mixup_args)
# wrap dataset in AugMix helper
if num_aug_splits > 1:
dataset_train = AugMixDataset(dataset_train, num_splits=num_aug_splits)
# create data loaders w/ augmentation pipeiine
train_interpolation = args.train_interpolation
if args.no_aug or not train_interpolation:
train_interpolation = data_config['interpolation']
loader_train = create_loader(
dataset_train,
input_size=data_config['input_size'],
batch_size=args.batch_size,
is_training=True,
use_prefetcher=args.prefetcher,
no_aug=args.no_aug,
re_prob=args.reprob,
re_mode=args.remode,
re_count=args.recount,
re_split=args.resplit,
scale=args.scale,
ratio=args.ratio,
hflip=args.hflip,
vflip=args.vflip,
color_jitter=args.color_jitter,
auto_augment=args.aa,
num_aug_repeats=args.aug_repeats,
num_aug_splits=num_aug_splits,
interpolation=train_interpolation,
mean=data_config['mean'],
std=data_config['std'],
num_workers=args.workers,
distributed=args.distributed,
collate_fn=collate_fn,
pin_memory=args.pin_mem,
use_multi_epochs_loader=args.use_multi_epochs_loader,
worker_seeding=args.worker_seeding,
)
loader_eval = create_loader(
dataset_eval,
input_size=data_config['input_size'],
batch_size=args.validation_batch_size or args.batch_size,
is_training=False,
use_prefetcher=args.prefetcher,
interpolation=data_config['interpolation'],
mean=data_config['mean'],
std=data_config['std'],
num_workers=args.workers,
distributed=args.distributed,
crop_pct=data_config['crop_pct'],
pin_memory=args.pin_mem,
)
# setup loss function
if args.jsd_loss:
assert num_aug_splits > 1 # JSD only valid with aug splits set
train_loss_fn = JsdCrossEntropy(num_splits=num_aug_splits,
smoothing=args.smoothing)
elif mixup_active:
# smoothing is handled with mixup target transform which outputs sparse, soft targets
if args.bce_loss:
train_loss_fn = BinaryCrossEntropy(
target_threshold=args.bce_target_thresh)
else:
train_loss_fn = SoftTargetCrossEntropy()
elif args.smoothing:
if args.bce_loss:
train_loss_fn = BinaryCrossEntropy(
smoothing=args.smoothing,
target_threshold=args.bce_target_thresh)
else:
train_loss_fn = LabelSmoothingCrossEntropy(
smoothing=args.smoothing)
else:
train_loss_fn = nn.CrossEntropyLoss()
train_loss_fn = train_loss_fn.cuda()
validate_loss_fn = nn.CrossEntropyLoss().cuda()
# setup checkpoint saver and eval metric tracking
eval_metric = args.eval_metric
best_metric = None
best_epoch = None
saver = None
output_dir = None
if args.rank == 0:
if args.experiment:
exp_name = args.experiment
else:
exp_name = '-'.join([
datetime.now().strftime("%Y%m%d-%H%M%S"),
safe_model_name(args.model),
str(data_config['input_size'][-1])
])
output_dir = get_outdir(
args.output if args.output else './output/train', exp_name)
decreasing = True if eval_metric == 'loss' else False
saver = CheckpointSaver(model=model,
optimizer=optimizer,
args=args,
model_ema=model_ema,
amp_scaler=loss_scaler,
checkpoint_dir=output_dir,
recovery_dir=output_dir,
decreasing=decreasing,
max_history=args.checkpoint_hist)
with open(os.path.join(output_dir, 'args.yaml'), 'w') as f:
f.write(args_text)
try:
for epoch in range(start_epoch, num_epochs):
if args.distributed and hasattr(loader_train.sampler, 'set_epoch'):
loader_train.sampler.set_epoch(epoch)
train_metrics = train_one_epoch(epoch,
model,
loader_train,
optimizer,
train_loss_fn,
args,
lr_scheduler=lr_scheduler,
saver=saver,
output_dir=output_dir,
amp_autocast=amp_autocast,
loss_scaler=loss_scaler,
model_ema=model_ema,
mixup_fn=mixup_fn,
model_KD=model_KD)
if args.distributed and args.dist_bn in ('broadcast', 'reduce'):
if args.local_rank == 0:
_logger.info(
"Distributing BatchNorm running means and vars")
distribute_bn(model, args.world_size, args.dist_bn == 'reduce')
eval_metrics = validate(model,
loader_eval,
validate_loss_fn,
args,
amp_autocast=amp_autocast)
if model_ema is not None and not args.model_ema_force_cpu:
if args.distributed and args.dist_bn in ('broadcast',
'reduce'):
distribute_bn(model_ema, args.world_size,
args.dist_bn == 'reduce')
ema_eval_metrics = validate(model_ema.module,
loader_eval,
validate_loss_fn,
args,
amp_autocast=amp_autocast,
log_suffix=' (EMA)')
eval_metrics = ema_eval_metrics
if lr_scheduler is not None:
# step LR for next epoch
lr_scheduler.step(epoch + 1, eval_metrics[eval_metric])
if output_dir is not None:
update_summary(epoch,
train_metrics,
eval_metrics,
os.path.join(output_dir, 'summary.csv'),
write_header=best_metric is None,
log_wandb=args.log_wandb and has_wandb)
if saver is not None:
# save proper checkpoint with eval metric
save_metric = eval_metrics[eval_metric]
best_metric, best_epoch = saver.save_checkpoint(
epoch, metric=save_metric)
except KeyboardInterrupt:
pass
if best_metric is not None:
_logger.info('*** Best metric: {0} (epoch {1})'.format(
best_metric, best_epoch))
def main():
setup_default_logging()
args, args_text = _parse_args()
args.prefetcher = not args.no_prefetcher
args.distributed = False
if 'WORLD_SIZE' in os.environ:
args.distributed = int(os.environ['WORLD_SIZE']) > 1
if args.distributed and args.num_gpu > 1:
_logger.warning(
'Using more than one GPU per process in distributed mode is not allowed.Setting num_gpu to 1.')
args.num_gpu = 1
args.device = 'cuda:0'
args.world_size = 1
args.rank = 0 # global rank
if args.distributed:
args.num_gpu = 1
args.device = 'cuda:%d' % args.local_rank
torch.cuda.set_device(args.local_rank)
torch.distributed.init_process_group(backend='nccl', init_method='env://')
args.world_size = torch.distributed.get_world_size()
args.rank = torch.distributed.get_rank()
assert args.rank >= 0
if args.distributed:
_logger.info('Training in distributed mode with multiple processes, 1 GPU per process. Process %d, total %d.'
% (args.rank, args.world_size))
else:
_logger.info('Training with a single process on %d GPUs.' % args.num_gpu)
torch.manual_seed(args.seed + args.rank)
model = create_model(
args.model,
pretrained=args.pretrained,
num_classes=args.num_classes,
drop_rate=args.drop,
drop_connect_rate=args.drop_connect, # DEPRECATED, use drop_path
drop_path_rate=args.drop_path,
drop_block_rate=args.drop_block,
global_pool=args.gp,
bn_tf=args.bn_tf,
bn_momentum=args.bn_momentum,
bn_eps=args.bn_eps,
checkpoint_path=args.initial_checkpoint)
if args.local_rank == 0:
_logger.info('Model %s created, param count: %d' %
(args.model, sum([m.numel() for m in model.parameters()])))
data_config = resolve_data_config(vars(args), model=model, verbose=args.local_rank == 0)
num_aug_splits = 0
if args.aug_splits > 0:
assert args.aug_splits > 1, 'A split of 1 makes no sense'
num_aug_splits = args.aug_splits
if args.split_bn:
assert num_aug_splits > 1 or args.resplit
model = convert_splitbn_model(model, max(num_aug_splits, 2))
use_amp = None
if args.amp:
# for backwards compat, `--amp` arg tries apex before native amp
if has_apex:
args.apex_amp = True
elif has_native_amp:
args.native_amp = True
if args.apex_amp and has_apex:
use_amp = 'apex'
elif args.native_amp and has_native_amp:
use_amp = 'native'
elif args.apex_amp or args.native_amp:
_logger.warning("Neither APEX or native Torch AMP is available, using float32. "
"Install NVIDA apex or upgrade to PyTorch 1.6")
if args.num_gpu > 1:
if use_amp == 'apex':
_logger.warning(
'Apex AMP does not work well with nn.DataParallel, disabling. Use DDP or Torch AMP.')
use_amp = None
model = nn.DataParallel(model, device_ids=list(range(args.num_gpu))).cuda()
assert not args.channels_last, "Channels last not supported with DP, use DDP."
else:
model.cuda()
if args.channels_last:
model = model.to(memory_format=torch.channels_last)
optimizer = create_optimizer(args, model)
amp_autocast = suppress # do nothing
loss_scaler = None
if use_amp == 'apex':
model, optimizer = amp.initialize(model, optimizer, opt_level='O1')
loss_scaler = ApexScaler()
if args.local_rank == 0:
_logger.info('Using NVIDIA APEX AMP. Training in mixed precision.')
elif use_amp == 'native':
amp_autocast = torch.cuda.amp.autocast
loss_scaler = NativeScaler()
if args.local_rank == 0:
_logger.info('Using native Torch AMP. Training in mixed precision.')
else:
if args.local_rank == 0:
_logger.info('AMP not enabled. Training in float32.')
# optionally resume from a checkpoint
resume_epoch = None
if args.resume:
resume_epoch = resume_checkpoint(
model, args.resume,
optimizer=None if args.no_resume_opt else optimizer,
loss_scaler=None if args.no_resume_opt else loss_scaler,
log_info=args.local_rank == 0)
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=args.resume)
if args.distributed:
if args.sync_bn:
assert not args.split_bn
try:
if has_apex and use_amp != 'native':
# Apex SyncBN preferred unless native amp is activated
model = convert_syncbn_model(model)
else:
model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(model)
if args.local_rank == 0:
_logger.info(
'Converted model to use Synchronized BatchNorm. WARNING: You may have issues if using '
'zero initialized BN layers (enabled by default for ResNets) while sync-bn enabled.')
except Exception as e:
_logger.error('Failed to enable Synchronized BatchNorm. Install Apex or Torch >= 1.1')
if has_apex and use_amp != 'native':
# Apex DDP preferred unless native amp is activated
if args.local_rank == 0:
_logger.info("Using NVIDIA APEX DistributedDataParallel.")
model = ApexDDP(model, delay_allreduce=True)
else:
if args.local_rank == 0:
_logger.info("Using native Torch DistributedDataParallel.")
model = NativeDDP(model, device_ids=[args.local_rank]) # can use device str in Torch >= 1.1
# NOTE: EMA model does not need to be wrapped by DDP
lr_scheduler, num_epochs = create_scheduler(args, optimizer)
start_epoch = 0
if args.start_epoch is not None:
# a specified start_epoch will always override the resume epoch
start_epoch = args.start_epoch
elif resume_epoch is not None:
start_epoch = resume_epoch
if lr_scheduler is not None and start_epoch > 0:
lr_scheduler.step(start_epoch)
if args.local_rank == 0:
_logger.info('Scheduled epochs: {}'.format(num_epochs))
train_dir = os.path.join(args.data, 'train')
if not os.path.exists(train_dir):
_logger.error('Training folder does not exist at: {}'.format(train_dir))
exit(1)
dataset_train = Dataset(train_dir)
collate_fn = None
mixup_fn = None
mixup_active = args.mixup > 0 or args.cutmix > 0. or args.cutmix_minmax is not None
if mixup_active:
mixup_args = dict(
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.num_classes)
if args.prefetcher:
assert not num_aug_splits # collate conflict (need to support deinterleaving in collate mixup)
collate_fn = FastCollateMixup(**mixup_args)
else:
mixup_fn = Mixup(**mixup_args)
if num_aug_splits > 1:
dataset_train = AugMixDataset(dataset_train, num_splits=num_aug_splits)
train_interpolation = args.train_interpolation
if args.no_aug or not train_interpolation:
train_interpolation = data_config['interpolation']
loader_train = create_loader(
dataset_train,
input_size=data_config['input_size'],
batch_size=args.batch_size,
is_training=True,
use_prefetcher=args.prefetcher,
no_aug=args.no_aug,
re_prob=args.reprob,
re_mode=args.remode,
re_count=args.recount,
re_split=args.resplit,
scale=args.scale,
ratio=args.ratio,
hflip=args.hflip,
vflip=args.vflip,
color_jitter=args.color_jitter,
auto_augment=args.aa,
num_aug_splits=num_aug_splits,
interpolation=train_interpolation,
mean=data_config['mean'],
std=data_config['std'],
num_workers=args.workers,
distributed=args.distributed,
collate_fn=collate_fn,
pin_memory=args.pin_mem,
use_multi_epochs_loader=args.use_multi_epochs_loader
)
eval_dir = os.path.join(args.data, 'val')
if not os.path.isdir(eval_dir):
eval_dir = os.path.join(args.data, 'validation')
if not os.path.isdir(eval_dir):
_logger.error('Validation folder does not exist at: {}'.format(eval_dir))
exit(1)
dataset_eval = Dataset(eval_dir)
loader_eval = create_loader(
dataset_eval,
input_size=data_config['input_size'],
batch_size=args.validation_batch_size_multiplier * args.batch_size,
is_training=False,
use_prefetcher=args.prefetcher,
interpolation=data_config['interpolation'],
mean=data_config['mean'],
std=data_config['std'],
num_workers=args.workers,
distributed=args.distributed,
crop_pct=data_config['crop_pct'],
pin_memory=args.pin_mem,
)
if args.jsd:
assert num_aug_splits > 1 # JSD only valid with aug splits set
train_loss_fn = JsdCrossEntropy(num_splits=num_aug_splits, smoothing=args.smoothing).cuda()
elif mixup_active:
# smoothing is handled with mixup target transform
train_loss_fn = SoftTargetCrossEntropy().cuda()
elif args.smoothing:
train_loss_fn = LabelSmoothingCrossEntropy(smoothing=args.smoothing).cuda()
else:
train_loss_fn = nn.CrossEntropyLoss().cuda()
validate_loss_fn = nn.CrossEntropyLoss().cuda()
eval_metric = args.eval_metric
best_metric = None
best_epoch = None
saver = None
output_dir = ''
if args.local_rank == 0:
output_base = args.output if args.output else './output'
exp_name = '-'.join([
datetime.now().strftime("%Y%m%d-%H%M%S"),
args.model,
str(data_config['input_size'][-1])
])
output_dir = get_outdir(output_base, 'train', exp_name)
decreasing = True if eval_metric == 'loss' else False
saver = CheckpointSaver(
model=model, optimizer=optimizer, args=args, model_ema=model_ema, amp_scaler=loss_scaler,
checkpoint_dir=output_dir, recovery_dir=output_dir, decreasing=decreasing)
with open(os.path.join(output_dir, 'args.yaml'), 'w') as f:
f.write(args_text)
try:
for epoch in range(start_epoch, num_epochs):
if args.distributed:
loader_train.sampler.set_epoch(epoch)
train_metrics = train_epoch(
epoch, model, loader_train, optimizer, train_loss_fn, args,
lr_scheduler=lr_scheduler, saver=saver, output_dir=output_dir,
amp_autocast=amp_autocast, loss_scaler=loss_scaler, model_ema=model_ema, mixup_fn=mixup_fn)
if args.distributed and args.dist_bn in ('broadcast', 'reduce'):
if args.local_rank == 0:
_logger.info("Distributing BatchNorm running means and vars")
distribute_bn(model, args.world_size, args.dist_bn == 'reduce')
eval_metrics = validate(model, loader_eval, validate_loss_fn, args, amp_autocast=amp_autocast)
if model_ema is not None and not args.model_ema_force_cpu:
if args.distributed and args.dist_bn in ('broadcast', 'reduce'):
distribute_bn(model_ema, args.world_size, args.dist_bn == 'reduce')
ema_eval_metrics = validate(
model_ema.ema, loader_eval, validate_loss_fn, args, amp_autocast=amp_autocast, log_suffix=' (EMA)')
eval_metrics = ema_eval_metrics
if lr_scheduler is not None:
# step LR for next epoch
lr_scheduler.step(epoch + 1, eval_metrics[eval_metric])
update_summary(
epoch, train_metrics, eval_metrics, os.path.join(output_dir, 'summary.csv'),
write_header=best_metric is None)
if saver is not None:
# save proper checkpoint with eval metric
save_metric = eval_metrics[eval_metric]
best_metric, best_epoch = saver.save_checkpoint(epoch, metric=save_metric)
# if saver.cmp(best_metric, save_metric):
# _logger.info(f"Metric is no longer improving [BEST: {best_metric}, CURRENT: {save_metric}]"
# f"\nFinishing training process")
# if epoch > 15:
# break
except KeyboardInterrupt:
pass
if best_metric is not None:
message = '*** Best metric: <{0:.2f}>, epoch: <{1}>, path: <{2}> ***'\
.format(best_metric, best_epoch, output_dir)
_logger.info(message)
print(message)
def validate(args):
# might as well try to validate something
args.pretrained = args.pretrained or not args.checkpoint
args.prefetcher = not args.no_prefetcher
if args.legacy_jit:
set_jit_legacy()
# create model
if 'inception' in args.model:
model = create_model(
args.model,
pretrained=args.pretrained,
num_classes=args.num_classes,
aux_logits=True, # ! add aux loss
in_chans=3,
scriptable=args.torchscript)
else:
model = create_model(args.model,
pretrained=args.pretrained,
num_classes=args.num_classes,
in_chans=3,
scriptable=args.torchscript)
# ! add more layer to classifier layer
if args.create_classifier_layerfc:
model.global_pool, model.classifier = create_classifier_layerfc(
model.num_features, model.num_classes)
if args.checkpoint:
load_checkpoint(model, args.checkpoint, args.use_ema)
param_count = sum([m.numel() for m in model.parameters()])
_logger.info('Model %s created, param count: %d' %
(args.model, param_count))
data_config = resolve_data_config(vars(args), model=model)
model, test_time_pool = apply_test_time_pool(model, data_config, args)
if args.torchscript:
torch.jit.optimized_execution(True)
model = torch.jit.script(model)
if args.amp:
model = amp.initialize(model.cuda(), opt_level='O1')
else:
model = model.cuda()
if args.num_gpu > 1:
model = torch.nn.DataParallel(model,
device_ids=list(range(args.num_gpu)))
if args.has_eval_label:
criterion = nn.CrossEntropyLoss().cuda() # ! don't have gold label
if os.path.splitext(args.data)[1] == '.tar' and os.path.isfile(args.data):
dataset = DatasetTar(args.data,
load_bytes=args.tf_preprocessing,
class_map=args.class_map)
else:
dataset = Dataset(args.data,
load_bytes=args.tf_preprocessing,
class_map=args.class_map,
args=args)
if args.valid_labels:
with open(args.valid_labels,
'r') as f: # @valid_labels is index numbering
valid_labels = {int(line.rstrip()) for line in f}
valid_labels = [i in valid_labels for i in range(args.num_classes)]
else:
valid_labels = None
if args.real_labels:
real_labels = RealLabelsImagenet(dataset.filenames(basename=True),
real_json=args.real_labels)
else:
real_labels = None
crop_pct = 1.0 if test_time_pool else data_config['crop_pct']
loader = create_loader(
dataset,
input_size=data_config['input_size'],
batch_size=args.batch_size,
use_prefetcher=args.prefetcher,
interpolation=data_config[
'interpolation'], # 'blank' is default Image.BILINEAR https://github.com/rwightman/pytorch-image-models/blob/470220b1f4c61ad7deb16dbfb8917089e842cd2a/timm/data/transforms.py#L43
mean=data_config['mean'],
std=data_config['std'],
num_workers=args.workers,
crop_pct=crop_pct,
pin_memory=args.pin_mem,
tf_preprocessing=args.tf_preprocessing,
auto_augment=args.aa,
scale=args.scale,
ratio=args.ratio,
hflip=args.hflip,
vflip=args.vflip,
color_jitter=args.color_jitter,
args=args)
batch_time = AverageMeter()
losses = AverageMeter()
top1 = AverageMeter()
topk = AverageMeter()
prediction = None # ! need to save output
true_label = None
model.eval()
with torch.no_grad():
# warmup, reduce variability of first batch time, especially for comparing torchscript vs non
input = torch.randn((args.batch_size, ) +
data_config['input_size']).cuda()
model(input)
end = time.time()
for batch_idx, (input,
target) in enumerate(loader): # ! not have real label
if args.has_eval_label: # ! just save true labels anyway... why not
if true_label is None: true_label = target.cpu().data.numpy()
else:
true_label = np.concatenate(
(true_label, target.cpu().data.numpy()), axis=0)
if args.no_prefetcher:
target = target.cuda()
input = input.cuda()
if args.fp16:
input = input.half()
# compute output
output = model(input)
if isinstance(output, (tuple, list)):
output = output[0] # ! some model returns both loss + aux loss
if valid_labels is not None:
output = output[:,
valid_labels] # ! keep only valid labels ? good to eval by class.
# ! save prediction, don't append too slow ... whatever ?
# ! are names of files also sorted ?
if prediction is None:
prediction = output.cpu().data.numpy() # batchsize x label
else: # stack
prediction = np.concatenate(
(prediction, output.cpu().data.numpy()), axis=0)
if real_labels is not None:
real_labels.add_result(output)
if args.has_eval_label:
# measure accuracy and record loss
loss = criterion(
output, target) # ! don't have gold standard on testset
acc1, acc5 = accuracy(output.data, target, topk=(1, args.topk))
losses.update(loss.item(), input.size(0))
top1.update(acc1.item(), input.size(0))
topk.update(acc5.item(), input.size(0))
# measure elapsed time
batch_time.update(time.time() - end)
end = time.time()
if args.has_eval_label and (batch_idx % args.log_freq == 0):
_logger.info(
'Test: [{0:>4d}/{1}] '
'Time: {batch_time.val:.3f}s ({batch_time.avg:.3f}s, {rate_avg:>7.2f}/s) '
'Loss: {loss.val:>7.4f} ({loss.avg:>6.4f}) '
'Acc@1: {top1.val:>7.3f} ({top1.avg:>7.3f}) '
'Acc@topk: {topk.val:>7.3f} ({topk.avg:>7.3f})'.format(
batch_idx,
len(loader),
batch_time=batch_time,
rate_avg=input.size(0) / batch_time.avg,
loss=losses,
top1=top1,
topk=topk))
if not args.has_eval_label:
top1a, topka = 0, 0 # just dummy, because we don't know ground labels
else:
if real_labels is not None:
# real labels mode replaces topk values at the end
top1a, topka = real_labels.get_accuracy(
k=1), real_labels.get_accuracy(k=args.topk)
else:
top1a, topka = top1.avg, topk.avg
results = OrderedDict(top1=round(top1a, 4),
top1_err=round(100 - top1a, 4),
topk=round(topka, 4),
topk_err=round(100 - topka, 4),
param_count=round(param_count / 1e6, 2),
img_size=data_config['input_size'][-1],
cropt_pct=crop_pct,
interpolation=data_config['interpolation'])
_logger.info(' * Acc@1 {:.3f} ({:.3f}) Acc@topk {:.3f} ({:.3f})'.format(
results['top1'], results['top1_err'], results['topk'],
results['topk_err']))
return results, prediction, true_label
def validate(args):
# might as well try to validate something
args.pretrained = args.pretrained or not args.checkpoint
args.prefetcher = not args.no_prefetcher
amp_autocast = suppress # do nothing
if args.amp:
if has_apex:
args.apex_amp = True
elif has_native_amp:
args.native_amp = True
else:
_logger.warning(
"Neither APEX or Native Torch AMP is available, using FP32.")
assert not args.apex_amp or not args.native_amp, "Only one AMP mode should be set."
if args.native_amp:
amp_autocast = torch.cuda.amp.autocast
if args.legacy_jit:
set_jit_legacy()
# create model
model = create_model(args.model,
pretrained=args.pretrained,
num_classes=args.num_classes,
in_chans=3,
global_pool=args.gp,
scriptable=args.torchscript)
if args.num_classes is None:
assert hasattr(
model, 'num_classes'
), 'Model must have `num_classes` attr if not set on cmd line/config.'
args.num_classes = model.num_classes
if args.checkpoint:
load_checkpoint(model, args.checkpoint, args.use_ema)
param_count = sum([m.numel() for m in model.parameters()])
_logger.info('Model %s created, param count: %d' %
(args.model, param_count))
data_config = resolve_data_config(vars(args), model=model)
model, test_time_pool = (
model, False) if args.no_test_pool else apply_test_time_pool(
model, data_config)
if args.torchscript:
torch.jit.optimized_execution(True)
model = torch.jit.script(model)
model = model.cuda()
if args.apex_amp:
model = amp.initialize(model, opt_level='O1')
if args.channels_last:
model = model.to(memory_format=torch.channels_last)
if args.num_gpu > 1:
model = torch.nn.DataParallel(model,
device_ids=list(range(args.num_gpu)))
criterion = nn.CrossEntropyLoss().cuda()
dataset = create_dataset(root=args.data,
name=args.dataset,
split=args.split,
load_bytes=args.tf_preprocessing,
class_map=args.class_map)
if args.valid_labels:
with open(args.valid_labels, 'r') as f:
valid_labels = {int(line.rstrip()) for line in f}
valid_labels = [i in valid_labels for i in range(args.num_classes)]
else:
valid_labels = None
if args.real_labels:
real_labels = RealLabelsImagenet(dataset.filenames(basename=True),
real_json=args.real_labels)
else:
real_labels = None
crop_pct = 1.0 if test_time_pool else data_config['crop_pct']
loader = create_loader(dataset,
input_size=data_config['input_size'],
batch_size=args.batch_size,
use_prefetcher=args.prefetcher,
interpolation=data_config['interpolation'],
mean=data_config['mean'],
std=data_config['std'],
num_workers=args.workers,
crop_pct=crop_pct,
pin_memory=args.pin_mem,
tf_preprocessing=args.tf_preprocessing)
batch_time = AverageMeter()
losses = AverageMeter()
top1 = AverageMeter()
top5 = AverageMeter()
model.eval()
with torch.no_grad():
# warmup, reduce variability of first batch time, especially for comparing torchscript vs non
input = torch.randn((args.batch_size, ) +
data_config['input_size']).cuda()
if args.channels_last:
input = input.contiguous(memory_format=torch.channels_last)
model(input)
end = time.time()
for batch_idx, (input, target) in enumerate(loader):
if args.no_prefetcher:
target = target.cuda()
input = input.cuda()
if args.channels_last:
input = input.contiguous(memory_format=torch.channels_last)
# compute output
with amp_autocast():
output = model(input)
if valid_labels is not None:
output = output[:, valid_labels]
loss = criterion(output, target)
if real_labels is not None:
real_labels.add_result(output)
# measure accuracy and record loss
acc1, acc5 = accuracy(output.detach(), target, topk=(1, 5))
losses.update(loss.item(), input.size(0))
top1.update(acc1.item(), input.size(0))
top5.update(acc5.item(), input.size(0))
# measure elapsed time
batch_time.update(time.time() - end)
end = time.time()
if batch_idx % args.log_freq == 0:
_logger.info(
'Test: [{0:>4d}/{1}] '
'Time: {batch_time.val:.3f}s ({batch_time.avg:.3f}s, {rate_avg:>7.2f}/s) '
'Loss: {loss.val:>7.4f} ({loss.avg:>6.4f}) '
'Acc@1: {top1.val:>7.3f} ({top1.avg:>7.3f}) '
'Acc@5: {top5.val:>7.3f} ({top5.avg:>7.3f})'.format(
batch_idx,
len(loader),
batch_time=batch_time,
rate_avg=input.size(0) / batch_time.avg,
loss=losses,
top1=top1,
top5=top5))
if real_labels is not None:
# real labels mode replaces topk values at the end
top1a, top5a = real_labels.get_accuracy(k=1), real_labels.get_accuracy(
k=5)
else:
top1a, top5a = top1.avg, top5.avg
results = OrderedDict(top1=round(top1a, 4),
top1_err=round(100 - top1a, 4),
top5=round(top5a, 4),
top5_err=round(100 - top5a, 4),
param_count=round(param_count / 1e6, 2),
img_size=data_config['input_size'][-1],
cropt_pct=crop_pct,
interpolation=data_config['interpolation'])
_logger.info(' * Acc@1 {:.3f} ({:.3f}) Acc@5 {:.3f} ({:.3f})'.format(
results['top1'], results['top1_err'], results['top5'],
results['top5_err']))
return results
def main():
setup_default_logging()
args = parser.parse_args()
args.prefetcher = not args.no_prefetcher
args.distributed = False
if 'WORLD_SIZE' in os.environ:
args.distributed = int(os.environ['WORLD_SIZE']) > 1
if args.distributed and args.num_gpu > 1:
logging.warning(
'Using more than one GPU per process in distributed mode is not allowed. Setting num_gpu to 1.'
)
args.num_gpu = 1
args.device = 'cuda:0'
args.world_size = 1
args.rank = 0 # global rank
if args.distributed:
args.num_gpu = 1
args.device = 'cuda:%d' % args.local_rank
torch.cuda.set_device(args.local_rank)
torch.distributed.init_process_group(backend='nccl',
init_method='env://')
args.world_size = torch.distributed.get_world_size()
args.rank = torch.distributed.get_rank()
assert args.rank >= 0
if args.distributed:
logging.info(
'Training in distributed mode with multiple processes, 1 GPU per process. Process %d, total %d.'
% (args.rank, args.world_size))
else:
logging.info('Training with a single process on %d GPUs.' %
args.num_gpu)
torch.manual_seed(args.seed + args.rank)
model = create_model(args.model,
pretrained=args.pretrained,
num_classes=args.num_classes,
drop_rate=args.drop,
global_pool=args.gp,
bn_tf=args.bn_tf,
bn_momentum=args.bn_momentum,
bn_eps=args.bn_eps,
drop_connect_rate=0.2,
checkpoint_path=args.initial_checkpoint,
args=args)
flops, params = get_model_complexity_info(
model, (3, 224, 224),
as_strings=True,
print_per_layer_stat=args.display_info)
print('Flops: ' + flops)
print('Params: ' + params)
if args.KD_train:
teacher_model = create_model("efficientnet_b7_dq",
pretrained=True,
num_classes=args.num_classes,
drop_rate=args.drop,
global_pool=args.gp,
bn_tf=args.bn_tf,
bn_momentum=args.bn_momentum,
bn_eps=args.bn_eps,
drop_connect_rate=0.2,
checkpoint_path=args.initial_checkpoint,
args=args)
flops_teacher, params_teacher = get_model_complexity_info(
teacher_model, (3, 224, 224),
as_strings=True,
print_per_layer_stat=False)
print("Using KD training...")
print("FLOPs of teacher model: ", flops_teacher)
print("Params of teacher model: ", params_teacher)
if args.local_rank == 0:
logging.info('Model %s created, param count: %d' %
(args.model, sum([m.numel()
for m in model.parameters()])))
data_config = resolve_data_config(model,
args,
verbose=args.local_rank == 0)
# optionally resume from a checkpoint
start_epoch = 0
optimizer_state = None
if args.resume:
optimizer_state, start_epoch = resume_checkpoint(
model, args.resume, args.start_epoch)
# import pdb;pdb.set_trace()
if args.num_gpu > 1:
if args.amp:
logging.warning(
'AMP does not work well with nn.DataParallel, disabling. Use distributed mode for multi-GPU AMP.'
)
args.amp = False
model = nn.DataParallel(model,
device_ids=list(range(args.num_gpu))).cuda()
if args.KD_train:
teacher_model = nn.DataParallel(teacher_model,
device_ids=list(range(
args.num_gpu))).cuda()
else:
model.cuda()
if args.KD_train:
teacher_model.cuda()
optimizer = create_optimizer(args, model)
if optimizer_state is not None:
optimizer.load_state_dict(optimizer_state)
use_amp = False
if has_apex and args.amp:
model, optimizer = amp.initialize(model, optimizer, opt_level='O1')
use_amp = True
if args.local_rank == 0:
logging.info('NVIDIA APEX {}. AMP {}.'.format(
'installed' if has_apex else 'not installed',
'on' if use_amp else 'off'))
model_ema = None
if args.model_ema:
# Important to create EMA model after cuda(), DP wrapper, and AMP but before SyncBN and DDP wrapper
# import pdb; pdb.set_trace()
model_ema = ModelEma(model,
decay=args.model_ema_decay,
device='cpu' if args.model_ema_force_cpu else '',
resume=args.resume)
if args.distributed:
if args.sync_bn:
try:
if has_apex:
model = convert_syncbn_model(model)
else:
model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(
model)
if args.local_rank == 0:
logging.info(
'Converted model to use Synchronized BatchNorm.')
except Exception as e:
logging.error(
'Failed to enable Synchronized BatchNorm. Install Apex or Torch >= 1.1'
)
if has_apex:
model = DDP(model, delay_allreduce=True)
else:
if args.local_rank == 0:
logging.info(
"Using torch DistributedDataParallel. Install NVIDIA Apex for Apex DDP."
)
model = DDP(model,
device_ids=[args.local_rank
]) # can use device str in Torch >= 1.1
# NOTE: EMA model does not need to be wrapped by DDP
lr_scheduler, num_epochs = create_scheduler(args, optimizer)
if start_epoch > 0:
lr_scheduler.step(start_epoch)
if args.local_rank == 0:
logging.info('Scheduled epochs: {}'.format(num_epochs))
train_dir = os.path.join(args.data, 'train')
if not os.path.exists(train_dir):
logging.error(
'Training folder does not exist at: {}'.format(train_dir))
exit(1)
dataset_train = Dataset(train_dir)
collate_fn = None
if args.prefetcher and args.mixup > 0:
collate_fn = FastCollateMixup(args.mixup, args.smoothing,
args.num_classes)
if args.auto_augment:
print('using auto data augumentation...')
loader_train = create_loader(
dataset_train,
input_size=data_config['input_size'],
batch_size=args.batch_size,
is_training=True,
use_prefetcher=args.prefetcher,
rand_erase_prob=args.reprob,
rand_erase_mode=args.remode,
interpolation=
'bicubic', # FIXME cleanly resolve this? data_config['interpolation'],
mean=data_config['mean'],
std=data_config['std'],
num_workers=args.workers,
distributed=args.distributed,
collate_fn=collate_fn,
use_auto_aug=args.auto_augment,
use_mixcut=args.mixcut,
)
eval_dir = os.path.join(args.data, 'val')
if not os.path.isdir(eval_dir):
logging.error(
'Validation folder does not exist at: {}'.format(eval_dir))
exit(1)
dataset_eval = Dataset(eval_dir)
loader_eval = create_loader(
dataset_eval,
input_size=data_config['input_size'],
batch_size=4 * args.batch_size,
is_training=False,
use_prefetcher=args.prefetcher,
interpolation=data_config['interpolation'],
mean=data_config['mean'],
std=data_config['std'],
num_workers=args.workers,
distributed=args.distributed,
)
if args.mixup > 0.:
# smoothing is handled with mixup label transform
train_loss_fn = SoftTargetCrossEntropy().cuda()
validate_loss_fn = nn.CrossEntropyLoss().cuda()
elif args.smoothing:
train_loss_fn = LabelSmoothingCrossEntropy(
smoothing=args.smoothing).cuda()
validate_loss_fn = nn.CrossEntropyLoss().cuda()
else:
train_loss_fn = nn.CrossEntropyLoss().cuda()
validate_loss_fn = train_loss_fn
if args.KD_train:
train_loss_fn = nn.KLDivLoss(reduction='batchmean').cuda()
eval_metric = args.eval_metric
best_metric = None
best_epoch = None
saver = None
output_dir = ''
if args.local_rank == 0:
output_base = args.output if args.output else './output'
exp_name = '-'.join([
datetime.now().strftime("%Y%m%d-%H%M%S"), args.model,
str(data_config['input_size'][-1])
])
output_dir = get_outdir(output_base, 'train', exp_name)
decreasing = True if eval_metric == 'loss' else False
saver = CheckpointSaver(checkpoint_dir=output_dir,
decreasing=decreasing)
try:
# import pdb;pdb.set_trace()
for epoch in range(start_epoch, num_epochs):
if args.distributed:
loader_train.sampler.set_epoch(epoch)
# import pdb; pdb.set_trace()
if args.KD_train:
train_metrics = train_epoch(epoch,
model,
loader_train,
optimizer,
train_loss_fn,
args,
lr_scheduler=lr_scheduler,
saver=saver,
output_dir=output_dir,
use_amp=use_amp,
model_ema=model_ema,
teacher_model=teacher_model)
else:
train_metrics = train_epoch(epoch,
model,
loader_train,
optimizer,
train_loss_fn,
args,
lr_scheduler=lr_scheduler,
saver=saver,
output_dir=output_dir,
use_amp=use_amp,
model_ema=model_ema)
# def __init__(self, model, bits_activations=8, bits_parameters=8, bits_accum=32,
# overrides=None, mode=LinearQuantMode.SYMMETRIC, clip_acts=ClipMode.NONE,
# per_channel_wts=False, model_activation_stats=None, fp16=False, clip_n_stds=None,
# scale_approx_mult_bits=None):
# import distiller
# import pdb; pdb.set_trace()
# quantizer = quantization.PostTrainLinearQuantizer.from_args(model, args)
# quantizer.prepare_model(distiller.get_dummy_input(input_shape=model.input_shape))
# quantizer = distiller.quantization.PostTrainLinearQuantizer(model, bits_activations=8, bits_parameters=8)
# quantizer.prepare_model()
# distiller.utils.assign_layer_fq_names(model)
# # msglogger.info("Generating quantization calibration stats based on {0} users".format(args.qe_calibration))
# collector = distiller.data_loggers.QuantCalibrationStatsCollector(model)
# with collector_context(collector):
# eval_metrics = validate(model, loader_eval, validate_loss_fn, args)
# # Here call your model evaluation function, making sure to execute only
# # the portion of the dataset specified by the qe_calibration argument
# yaml_path = './dir/quantization_stats.yaml'
# collector.save(yaml_path)
eval_metrics = validate(model, loader_eval, validate_loss_fn, args)
if model_ema is not None and not args.model_ema_force_cpu:
ema_eval_metrics = validate(model_ema.ema,
loader_eval,
validate_loss_fn,
args,
log_suffix=' (EMA)')
eval_metrics = ema_eval_metrics
if lr_scheduler is not None:
lr_scheduler.step(epoch, eval_metrics[eval_metric])
update_summary(epoch,
train_metrics,
eval_metrics,
os.path.join(output_dir, 'summary.csv'),
write_header=best_metric is None)
if saver is not None:
# save proper checkpoint with eval metric
save_metric = eval_metrics[eval_metric]
best_metric, best_epoch = saver.save_checkpoint(
model,
optimizer,
args,
epoch=epoch + 1,
model_ema=model_ema,
metric=save_metric)
except KeyboardInterrupt:
pass
if best_metric is not None:
logging.info('*** Best metric: {0} (epoch {1})'.format(
best_metric, best_epoch))
def validate(args):
# might as well try to validate something
args.pretrained = args.pretrained or not args.checkpoint
# create model
model = create_model(
args.model,
num_classes=args.num_classes,
in_chans=3,
pretrained=args.pretrained)
if args.checkpoint:
load_checkpoint(model, args.checkpoint, args.use_ema)
param_count = sum([m.numel() for m in model.parameters()])
logging.info('Model %s created, param count: %d' % (args.model, param_count))
data_config = resolve_data_config(model, args)
model, test_time_pool = apply_test_time_pool(model, data_config, args)
if args.num_gpu > 1:
model = torch.nn.DataParallel(model, device_ids=list(range(args.num_gpu))).cuda()
else:
model = model.cuda()
criterion = nn.CrossEntropyLoss().cuda()
loader = create_loader(
Dataset(args.data, load_bytes=args.tf_preprocessing),
input_size=data_config['input_size'],
batch_size=args.batch_size,
use_prefetcher=True,
interpolation=data_config['interpolation'],
mean=data_config['mean'],
std=data_config['std'],
num_workers=args.workers,
crop_pct=1.0 if test_time_pool else data_config['crop_pct'],
tf_preprocessing=args.tf_preprocessing)
batch_time = AverageMeter()
losses = AverageMeter()
top1 = AverageMeter()
top5 = AverageMeter()
model.eval()
end = time.time()
with torch.no_grad():
for i, (input, target) in enumerate(loader):
target = target.cuda()
input = input.cuda()
# compute output
output = model(input)
loss = criterion(output, target)
# measure accuracy and record loss
prec1, prec5 = accuracy(output.data, target, topk=(1, 5))
losses.update(loss.item(), input.size(0))
top1.update(prec1.item(), input.size(0))
top5.update(prec5.item(), input.size(0))
# measure elapsed time
batch_time.update(time.time() - end)
end = time.time()
if i % args.log_freq == 0:
logging.info(
'Test: [{0:>4d}/{1}] '
'Time: {batch_time.val:.3f} ({batch_time.avg:.3f}) '
'Loss: {loss.val:>7.4f} ({loss.avg:>6.4f}) '
'Prec@1: {top1.val:>7.4f} ({top1.avg:>7.4f}) '
'Prec@5: {top5.val:>7.4f} ({top5.avg:>7.4f})'.format(
i, len(loader), batch_time=batch_time,
rate_avg=input.size(0) / batch_time.avg,
loss=losses, top1=top1, top5=top5))
results = OrderedDict(
top1=round(top1.avg, 3), top1_err=round(100 - top1.avg, 3),
top5=round(top5.avg, 3), top5_err=round(100 - top5.avg, 3),
param_count=round(param_count / 1e6, 2))
logging.info(' * Prec@1 {:.3f} ({:.3f}) Prec@5 {:.3f} ({:.3f})'.format(
results['top1'], results['top1_err'], results['top5'], results['top5_err']))
return results
def main():
setup_default_logging()
args = parser.parse_args()
# might as well try to do something useful...
args.pretrained = args.pretrained or not args.checkpoint
# create model
model = create_model(args.model,
num_classes=args.num_classes,
in_chans=3,
pretrained=args.pretrained,
checkpoint_path=args.checkpoint)
_logger.info('Model %s created, param count: %d' %
(args.model, sum([m.numel() for m in model.parameters()])))
config = resolve_data_config(vars(args), model=model)
model, test_time_pool = (
model,
False) if args.no_test_pool else apply_test_time_pool(model, config)
if args.num_gpu > 1:
model = torch.nn.DataParallel(model,
device_ids=list(range(
args.num_gpu))).cuda()
else:
model = model.cuda()
loader = create_loader(
ImageDataset(args.data),
input_size=config['input_size'],
batch_size=args.batch_size,
use_prefetcher=True,
interpolation=config['interpolation'],
mean=config['mean'],
std=config['std'],
num_workers=args.workers,
crop_pct=1.0 if test_time_pool else config['crop_pct'])
model.eval()
k = min(args.topk, args.num_classes)
batch_time = AverageMeter()
end = time.time()
topk_ids = []
with torch.no_grad():
for batch_idx, (input, _) in enumerate(loader):
input = input.cuda()
labels = model(input)
topk = labels.topk(k)[1]
topk_ids.append(topk.cpu().numpy())
# measure elapsed time
batch_time.update(time.time() - end)
end = time.time()
if batch_idx % args.log_freq == 0:
_logger.info(
'Predict: [{0}/{1}] Time {batch_time.val:.3f} ({batch_time.avg:.3f})'
.format(batch_idx, len(loader), batch_time=batch_time))
topk_ids = np.concatenate(topk_ids, axis=0).squeeze()
with open(os.path.join(args.output_dir, 'topk_ids.csv'), 'w') as out_file:
filenames = loader.dataset.filenames(basename=True)
for filename, label in zip(filenames, topk_ids):
out_file.write('{0},{1},{2},{3},{4},{5}\n'.format(
filename, label[0], label[1], label[2], label[3], label[4]))
def validate(args):
_logger.info(f'\n\n ---------------EVALUATION {args.eps}------------------------------- \n\n')
_logger.info("Argument parser collected the following arguments:")
for arg in vars(args):
_logger.info(f" {arg}:{getattr(args, arg)}")
_logger.info("\n")
# might as well try to validate something
args.pretrained = args.pretrained or not args.checkpoint
args.prefetcher = not args.no_prefetcher
amp_autocast = suppress # do nothing
if args.amp:
if has_native_amp:
args.native_amp = True
elif has_apex:
args.apex_amp = True
else:
_logger.warning("Neither APEX or Native Torch AMP is available.")
assert not args.apex_amp or not args.native_amp, "Only one AMP mode should be set."
if args.native_amp:
amp_autocast = torch.cuda.amp.autocast
_logger.info('Validating in mixed precision with native PyTorch AMP.')
elif args.apex_amp:
_logger.info('Validating in mixed precision with NVIDIA APEX AMP.')
else:
_logger.info('Validating in float32. AMP not enabled.')
if args.legacy_jit:
set_jit_legacy()
# create model
model = create_model(
args.model,
pretrained=args.pretrained,
num_classes=args.num_classes,
in_chans=3,
global_pool=args.gp,
scriptable=args.torchscript)
if args.num_classes is None:
assert hasattr(model, 'num_classes'), 'Model must have `num_classes` attr if not set on cmd line/config.'
args.num_classes = model.num_classes
if args.checkpoint:
load_checkpoint(model, args.checkpoint, args.use_ema)
param_count = sum([m.numel() for m in model.parameters()])
_logger.info(
f'Model {args.model} created, param count: {param_count} ({(float(param_count)/(10.0**6)):.1f} M)'
)
data_config = resolve_data_config(vars(args), model=model, use_test_size=True, verbose=True)
test_time_pool = False
if not args.no_test_pool:
model, test_time_pool = apply_test_time_pool(model, data_config, use_test_size=True)
if args.torchscript:
torch.jit.optimized_execution(True)
model = torch.jit.script(model)
model = model.cuda()
if args.apex_amp:
model = amp.initialize(model, opt_level='O1')
if args.channels_last:
model = model.to(memory_format=torch.channels_last)
if args.num_gpu > 1:
model = torch.nn.DataParallel(model, device_ids=list(range(args.num_gpu)))
criterion = nn.CrossEntropyLoss().cuda()
dataset = create_dataset(
root=args.data_dir, name=args.dataset, split=args.split,
load_bytes=args.tf_preprocessing, class_map=args.class_map)
if args.valid_labels:
with open(args.valid_labels, 'r') as f:
valid_labels = {int(line.rstrip()) for line in f}
valid_labels = [i in valid_labels for i in range(args.num_classes)]
else:
valid_labels = None
if args.real_labels:
real_labels = RealLabelsImagenet(dataset.filenames(basename=True), real_json=args.real_labels)
else:
real_labels = None
crop_pct = 1.0 if test_time_pool else data_config['crop_pct']
loader = create_loader(
dataset,
input_size=data_config['input_size'],
batch_size=args.batch_size,
use_prefetcher=args.prefetcher,
interpolation=data_config['interpolation'],
mean=data_config['mean'],
std=data_config['std'],
num_workers=args.workers,
crop_pct=crop_pct,
pin_memory=args.pin_mem,
tf_preprocessing=args.tf_preprocessing)
batch_time = AverageMeter()
losses = AverageMeter()
top1 = AverageMeter()
top5 = AverageMeter()
top1_fgm_ae = AverageMeter()
top5_fgm_ae = AverageMeter()
top1_pgd_ae = AverageMeter()
top5_pgd_ae = AverageMeter()
model.eval()
#with torch.no_grad():# TODO Requires grad
# warmup, reduce variability of first batch time, especially for comparing torchscript vs non
input = torch.randn((args.batch_size,) + tuple(data_config['input_size'])).cuda()
if args.channels_last:
input = input.contiguous(memory_format=torch.channels_last)
model(input)
end = time.time()
for batch_idx, (input, target) in enumerate(loader):
if args.no_prefetcher:
target = target.cuda()
input = input.cuda()
if args.channels_last:
input = input.contiguous(memory_format=torch.channels_last)
# compute output
with amp_autocast():
output = model(input)
if valid_labels is not None:
output = output[:, valid_labels]
loss = criterion(output, target)
if real_labels is not None:
real_labels.add_result(output)
# TODO <---------------------
# Generate adversarial examples for current inputs
input_fgm_ae = fast_gradient_method(
model_fn=model,
x=input,
eps=args.eps,
norm=np.inf,
clip_min=None,
clip_max=None,
)
input_pgd_ae = projected_gradient_descent(
model_fn=model,
x=input,
eps=args.eps,
eps_iter=0.01,
nb_iter=40,
norm=np.inf,
clip_min=None,
clip_max=None,
)
# Predict with Adversarial Examples
with torch.no_grad():
with amp_autocast():
output_fgm_ae = model(input_fgm_ae)
output_pgd_ae = model(input_pgd_ae)
# measure accuracy and record loss
acc1, acc5 = accuracy(output.detach(), target, topk=(1, 5))
losses.update(loss.item(), input.size(0))
top1.update(acc1.item(), input.size(0))
top5.update(acc5.item(), input.size(0))
acc1_fgm_ae, acc5_fgm_ae = accuracy(output_fgm_ae.detach(), target, topk=(1, 5))
acc1_pgd_ae, acc5_pgd_ae = accuracy(output_pgd_ae.detach(), target, topk=(1, 5))
top1_fgm_ae.update(acc1_fgm_ae.item(), input.size(0))
top5_fgm_ae.update(acc5_fgm_ae.item(), input.size(0))
top1_pgd_ae.update(acc1_pgd_ae.item(), input.size(0))
top5_pgd_ae.update(acc5_pgd_ae.item(), input.size(0))
# measure elapsed time
batch_time.update(time.time() - end)
end = time.time()
if batch_idx % args.log_freq == 0:
_logger.info(
'Test: [{0:>4d}/{1}] '
'Time: {batch_time.val:.3f}s ({batch_time.avg:.3f}s, {rate_avg:>7.2f}/s) '
'Loss: {loss.val:>7.4f} ({loss.avg:>6.4f}) '
'Acc@1: {top1.val:>7.3f} ({top1.avg:>7.3f}) '
'Acc@5: {top5.val:>7.3f} ({top5.avg:>7.3f})'.format(
batch_idx, len(loader), batch_time=batch_time,
rate_avg=input.size(0) / batch_time.avg,
loss=losses, top1=top1, top5=top5))
if real_labels is not None:
raise NotImplementedError # TODO NOt modified for the adversarial examples mode
# real labels mode replaces topk values at the end
top1a, top5a = real_labels.get_accuracy(k=1), real_labels.get_accuracy(k=5)
else:
top1a, top5a = top1.avg, top5.avg
top1a_fgm_ae, top5a_fgm_ae = top1_fgm_ae.avg, top5_fgm_ae.avg
top1a_pgd_ae, top5a_pgd_ae = top1_pgd_ae.avg, top5_pgd_ae.avg
results = OrderedDict(
top1=round(top1a, 4), top1_err=round(100 - top1a, 4),
top5=round(top5a, 4), top5_err=round(100 - top5a, 4),
top1_fgm_ae=round(top1a_fgm_ae, 4),
top5_fgm_ae=round(top5a_fgm_ae, 4),
top1_pgd_ae=round(top1a_pgd_ae, 4),
top5_pgd_ae=round(top5a_pgd_ae, 4),
param_count=round(param_count / 1e6, 2),
img_size=data_config['input_size'][-1],
cropt_pct=crop_pct,
interpolation=data_config['interpolation'])
_logger.info(' * [Regular] Acc@1 {:.3f} ({:.3f}) Acc@5 {:.3f} ({:.3f})'.format(
results['top1'], results['top1_err'], results['top5'], results['top5_err']))
_logger.info(' * [FGM Adversarial Attack] Acc@1 {:.3f} Acc@5 {:.3f} '.format(
results['top1_fgm_ae'], results['top5_fgm_ae']))
_logger.info(' * [PGD Adversarial Attack] Acc@1 {:.3f} Acc@5 {:.3f} '.format(
results['top1_pgd_ae'], results['top5_pgd_ae']))
return results
def main():
setup_default_logging()
args, args_text = _parse_args()
args.prefetcher = not args.no_prefetcher
args.distributed = False
if 'WORLD_SIZE' in os.environ:
args.distributed = int(os.environ['WORLD_SIZE']) > 1
args.device = 'cuda:0'
args.world_size = 1
args.rank = 0 # global rank
if args.distributed:
args.device = 'cuda:%d' % args.local_rank
torch.cuda.set_device(args.local_rank)
torch.distributed.init_process_group(backend='nccl',
init_method='env://')
args.world_size = torch.distributed.get_world_size()
args.rank = torch.distributed.get_rank()
_logger.info(
'Training in distributed mode with multiple processes, 1 GPU per process. Process %d, total %d.'
% (args.rank, args.world_size))
else:
_logger.info('Training with a single process on 1 GPUs.')
assert args.rank >= 0
# resolve AMP arguments based on PyTorch / Apex availability
use_amp = None
if args.amp:
# for backwards compat, `--amp` arg tries apex before native amp
if has_apex:
args.apex_amp = True
elif has_native_amp:
args.native_amp = True
if args.apex_amp and has_apex:
use_amp = 'apex'
elif args.native_amp and has_native_amp:
use_amp = 'native'
elif args.apex_amp or args.native_amp:
_logger.warning(
"Neither APEX or native Torch AMP is available, using float32. "
"Install NVIDA apex or upgrade to PyTorch 1.6")
torch.manual_seed(args.seed + args.rank)
model = create_model(
args.model,
pretrained=args.pretrained,
num_classes=args.num_classes,
drop_rate=args.drop,
drop_connect_rate=args.drop_connect, # DEPRECATED, use drop_path
drop_path_rate=args.drop_path,
drop_block_rate=args.drop_block,
global_pool=args.gp,
bn_tf=args.bn_tf,
bn_momentum=args.bn_momentum,
bn_eps=args.bn_eps,
scriptable=args.torchscript,
use_cos_reg=args.cos_reg_component > 0,
checkpoint_path=args.initial_checkpoint)
with torch.cuda.device(0):
input = torch.randn(1, 3, 224, 224)
size_for_madd = 224 if args.img_size is None else args.img_size
# flops, params = get_model_complexity_info(model, (3, size_for_madd, size_for_madd), as_strings=True, print_per_layer_stat=True)
# print("=>Flops: " + flops)
# print("=>Params: " + params)
if args.local_rank == 0:
_logger.info('Model %s created, param count: %d' %
(args.model, sum([m.numel()
for m in model.parameters()])))
data_config = resolve_data_config(vars(args),
model=model,
verbose=args.local_rank == 0)
# setup augmentation batch splits for contrastive loss or split bn
num_aug_splits = 0
if args.aug_splits > 0:
assert args.aug_splits > 1, 'A split of 1 makes no sense'
num_aug_splits = args.aug_splits
# enable split bn (separate bn stats per batch-portion)
if args.split_bn:
assert num_aug_splits > 1 or args.resplit
model = convert_splitbn_model(model, max(num_aug_splits, 2))
# move model to GPU, enable channels last layout if set
model.cuda()
if args.channels_last:
model = model.to(memory_format=torch.channels_last)
# setup synchronized BatchNorm for distributed training
if args.distributed and args.sync_bn:
assert not args.split_bn
if has_apex and use_amp != 'native':
# Apex SyncBN preferred unless native amp is activated
model = convert_syncbn_model(model)
else:
model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(model)
if args.local_rank == 0:
_logger.info(
'Converted model to use Synchronized BatchNorm. WARNING: You may have issues if using '
'zero initialized BN layers (enabled by default for ResNets) while sync-bn enabled.'
)
if args.torchscript:
assert not use_amp == 'apex', 'Cannot use APEX AMP with torchscripted model'
assert not args.sync_bn, 'Cannot use SyncBatchNorm with torchscripted model'
model = torch.jit.script(model)
optimizer = create_optimizer(args, model)
# setup automatic mixed-precision (AMP) loss scaling and op casting
amp_autocast = suppress # do nothing
loss_scaler = None
if use_amp == 'apex':
model, optimizer = amp.initialize(model, optimizer, opt_level='O1')
loss_scaler = ApexScaler()
if args.local_rank == 0:
_logger.info('Using NVIDIA APEX AMP. Training in mixed precision.')
elif use_amp == 'native':
amp_autocast = torch.cuda.amp.autocast
loss_scaler = NativeScaler()
if args.local_rank == 0:
_logger.info(
'Using native Torch AMP. Training in mixed precision.')
else:
if args.local_rank == 0:
_logger.info('AMP not enabled. Training in float32.')
# optionally resume from a checkpoint
resume_epoch = None
if args.resume:
resume_epoch = resume_checkpoint(
model,
args.resume,
optimizer=None if args.no_resume_opt else optimizer,
loss_scaler=None if args.no_resume_opt else loss_scaler,
log_info=args.local_rank == 0)
# setup exponential moving average of model weights, SWA could be used here too
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 = ModelEmaV2(
model,
decay=args.model_ema_decay,
device='cpu' if args.model_ema_force_cpu else None)
if args.resume:
load_checkpoint(model_ema.module, args.resume, use_ema=True)
# setup distributed training
if args.distributed:
if has_apex and use_amp != 'native':
# Apex DDP preferred unless native amp is activated
if args.local_rank == 0:
_logger.info("Using NVIDIA APEX DistributedDataParallel.")
model = ApexDDP(model, delay_allreduce=True)
else:
if args.local_rank == 0:
_logger.info("Using native Torch DistributedDataParallel.")
model = NativeDDP(model, device_ids=[
args.local_rank
]) # can use device str in Torch >= 1.1
# NOTE: EMA model does not need to be wrapped by DDP
# setup learning rate schedule and starting epoch
lr_scheduler, num_epochs = create_scheduler(args, optimizer)
start_epoch = 0
if args.start_epoch is not None:
# a specified start_epoch will always override the resume epoch
start_epoch = args.start_epoch
elif resume_epoch is not None:
start_epoch = resume_epoch
if lr_scheduler is not None and start_epoch > 0:
lr_scheduler.step(start_epoch)
if args.local_rank == 0:
_logger.info('Scheduled epochs: {}'.format(num_epochs))
# create the train and eval datasets
train_dir = os.path.join(args.data, 'train')
if not os.path.exists(train_dir):
_logger.error(
'Training folder does not exist at: {}'.format(train_dir))
exit(1)
if args.use_lmdb:
dataset_train = ImageFolderLMDB('../dataset_lmdb/train')
else:
dataset_train = Dataset(train_dir)
# dataset_train = Dataset(train_dir)
eval_dir = os.path.join(args.data, 'val')
if not os.path.isdir(eval_dir):
eval_dir = os.path.join(args.data, 'validation')
if not os.path.isdir(eval_dir):
_logger.error(
'Validation folder does not exist at: {}'.format(eval_dir))
exit(1)
if args.use_lmdb:
dataset_eval = ImageFolderLMDB('../dataset_lmdb/val')
else:
dataset_eval = Dataset(eval_dir)
# dataset_eval = Dataset(eval_dir)
# setup mixup / cutmix
collate_fn = None
mixup_fn = None
mixup_active = args.mixup > 0 or args.cutmix > 0. or args.cutmix_minmax is not None
if mixup_active:
mixup_args = dict(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.num_classes)
if args.prefetcher:
assert not num_aug_splits # collate conflict (need to support deinterleaving in collate mixup)
collate_fn = FastCollateMixup(**mixup_args)
else:
mixup_fn = Mixup(**mixup_args)
# wrap dataset in AugMix helper
if num_aug_splits > 1:
dataset_train = AugMixDataset(dataset_train, num_splits=num_aug_splits)
# create data loaders w/ augmentation pipeiine
train_interpolation = args.train_interpolation
if args.no_aug or not train_interpolation:
train_interpolation = data_config['interpolation']
loader_train = create_loader(
dataset_train,
input_size=data_config['input_size'],
batch_size=args.batch_size,
is_training=True,
use_prefetcher=args.prefetcher,
no_aug=args.no_aug,
re_prob=args.reprob,
re_mode=args.remode,
re_count=args.recount,
re_split=args.resplit,
scale=args.scale,
ratio=args.ratio,
hflip=args.hflip,
vflip=args.vflip,
color_jitter=args.color_jitter,
auto_augment=args.aa,
num_aug_splits=num_aug_splits,
interpolation=train_interpolation,
mean=data_config['mean'],
std=data_config['std'],
num_workers=args.workers,
distributed=args.distributed,
collate_fn=collate_fn,
pin_memory=args.pin_mem,
use_multi_epochs_loader=args.use_multi_epochs_loader,
repeated_aug=args.use_repeated_aug,
world_size=args.world_size,
rank=args.rank)
loader_eval = create_loader(
dataset_eval,
input_size=data_config['input_size'],
batch_size=args.validation_batch_size_multiplier * args.batch_size,
is_training=False,
use_prefetcher=args.prefetcher,
interpolation=data_config['interpolation'],
mean=data_config['mean'],
std=data_config['std'],
num_workers=args.workers,
distributed=args.distributed,
crop_pct=data_config['crop_pct'],
pin_memory=args.pin_mem,
)
loader_cali = create_loader(
dataset_train,
input_size=data_config['input_size'],
batch_size=args.cali_batch_size,
is_training=False,
use_prefetcher=args.prefetcher,
no_aug=True,
re_prob=args.reprob,
re_mode=args.remode,
re_count=args.recount,
re_split=args.resplit,
scale=args.scale,
ratio=args.ratio,
hflip=args.hflip,
vflip=args.vflip,
color_jitter=args.color_jitter,
auto_augment=args.aa,
num_aug_splits=num_aug_splits,
interpolation=train_interpolation,
mean=data_config['mean'],
std=data_config['std'],
num_workers=args.workers,
distributed=args.distributed,
collate_fn=None,
pin_memory=args.pin_mem,
use_multi_epochs_loader=args.use_multi_epochs_loader,
repeated_aug=args.use_repeated_aug,
world_size=args.world_size,
rank=args.rank)
# setup loss function
if args.jsd:
assert num_aug_splits > 1 # JSD only valid with aug splits set
train_loss_fn = JsdCrossEntropy(num_splits=num_aug_splits,
smoothing=args.smoothing).cuda()
elif mixup_active:
# smoothing is handled with mixup target transform
if args.cos_reg_component > 0:
args.use_cos_reg_component = True
train_loss_fn = SoftTargetCrossEntropyCosReg(
n_comn=args.cos_reg_component).cuda()
else:
train_loss_fn = SoftTargetCrossEntropy().cuda()
args.use_cos_reg_component = False
elif args.smoothing:
train_loss_fn = LabelSmoothingCrossEntropy(
smoothing=args.smoothing).cuda()
else:
train_loss_fn = nn.CrossEntropyLoss().cuda()
validate_loss_fn = nn.CrossEntropyLoss().cuda()
# setup checkpoint saver and eval metric tracking
eval_metric = args.eval_metric
best_metric = None
best_epoch = None
saver = None
output_dir = ''
if args.local_rank == 0:
output_base = args.output if args.output else './output'
exp_name = '-'.join([
datetime.now().strftime("%Y%m%d-%H%M%S"), args.model,
str(data_config['input_size'][-1])
])
output_dir = get_outdir(output_base, 'train', exp_name)
code_dir = get_outdir(output_dir, 'code')
copy_tree(os.getcwd(), code_dir)
decreasing = True if eval_metric == 'loss' else False
saver = CheckpointSaver(model=model,
optimizer=optimizer,
args=args,
model_ema=model_ema,
amp_scaler=loss_scaler,
checkpoint_dir=output_dir,
recovery_dir=output_dir,
decreasing=decreasing)
with open(os.path.join(output_dir, 'args.yaml'), 'w') as f:
f.write(args_text)
try:
for epoch in range(start_epoch, num_epochs):
if args.distributed:
loader_train.sampler.set_epoch(epoch)
if not args.eval_only:
train_metrics = train_epoch(epoch,
model,
loader_train,
optimizer,
train_loss_fn,
args,
lr_scheduler=lr_scheduler,
saver=saver,
output_dir=output_dir,
amp_autocast=amp_autocast,
loss_scaler=loss_scaler,
model_ema=model_ema,
mixup_fn=mixup_fn)
if args.distributed and args.dist_bn in ('broadcast', 'reduce'):
if args.local_rank == 0:
_logger.info(
"Distributing BatchNorm running means and vars")
distribute_bn(model, args.world_size, args.dist_bn == 'reduce')
if args.max_iter > 0:
_ = validate(model,
loader_cali,
validate_loss_fn,
args,
amp_autocast=amp_autocast,
use_bn_calibration=True)
eval_metrics = validate(model,
loader_eval,
validate_loss_fn,
args,
amp_autocast=amp_autocast)
if model_ema is not None and not args.model_ema_force_cpu:
if args.distributed and args.dist_bn in ('broadcast',
'reduce'):
distribute_bn(model_ema, args.world_size,
args.dist_bn == 'reduce')
ema_eval_metrics = validate(model_ema.module,
loader_eval,
validate_loss_fn,
args,
amp_autocast=amp_autocast,
log_suffix=' (EMA)')
eval_metrics = ema_eval_metrics
if lr_scheduler is not None:
# step LR for next epoch
lr_scheduler.step(epoch + 1, eval_metrics[eval_metric])
if not args.eval_only:
update_summary(epoch,
train_metrics,
eval_metrics,
os.path.join(output_dir, 'summary.csv'),
write_header=best_metric is None)
if saver is not None:
# save proper checkpoint with eval metric
save_metric = eval_metrics[eval_metric]
best_metric, best_epoch = saver.save_checkpoint(
epoch, metric=save_metric)
if args.eval_only:
break
except KeyboardInterrupt:
pass
if best_metric is not None:
_logger.info('*** Best metric: {0} (epoch {1})'.format(
best_metric, best_epoch))
def main():
setup_default_logging()
args, args_text = _parse_args()
args.prefetcher = not args.no_prefetcher
args.distributed = False
if 'WORLD_SIZE' in os.environ:
args.distributed = int(os.environ['WORLD_SIZE']) > 1
args.device = 'cuda:0'
args.world_size = 1
args.rank = 0 # global rank
if args.distributed:
args.device = 'cuda:%d' % args.local_rank
torch.cuda.set_device(args.local_rank)
torch.distributed.init_process_group(backend='nccl', init_method='env://')
args.world_size = torch.distributed.get_world_size()
args.rank = torch.distributed.get_rank()
_logger.info('Training in distributed mode with multiple processes, 1 GPU per process. Process %d, total %d.'
% (args.rank, args.world_size))
else:
_logger.info('Training with a single process on 1 GPUs.')
assert args.rank >= 0
# resolve AMP arguments based on PyTorch / Apex availability
use_amp = None
if args.amp:
# for backwards compat, `--amp` arg tries apex before native amp
if has_apex:
args.apex_amp = True
elif has_native_amp:
args.native_amp = True
if args.apex_amp and has_apex:
use_amp = 'apex'
elif args.native_amp and has_native_amp:
use_amp = 'native'
elif args.apex_amp or args.native_amp:
_logger.warning("Neither APEX or native Torch AMP is available, using float32. "
"Install NVIDA apex or upgrade to PyTorch 1.6")
torch.manual_seed(args.seed + args.rank)
####################################################################################
# Start - SparseML optional load weights from SparseZoo
####################################################################################
if args.initial_checkpoint == "zoo":
# Load checkpoint from base weights associated with given SparseZoo recipe
if args.sparseml_recipe.startswith("zoo:"):
args.initial_checkpoint = Zoo.download_recipe_base_framework_files(
args.sparseml_recipe,
extensions=[".pth.tar", ".pth"]
)[0]
else:
raise ValueError(
"Attempting to load weights from SparseZoo recipe, but not given a "
"SparseZoo recipe stub. When initial-checkpoint is set to 'zoo'. "
"sparseml-recipe must start with 'zoo:' and be a SparseZoo model "
f"stub. sparseml-recipe was set to {args.sparseml_recipe}"
)
elif args.initial_checkpoint.startswith("zoo:"):
# Load weights from a SparseZoo model stub
zoo_model = Zoo.load_model_from_stub(args.initial_checkpoint)
args.initial_checkpoint = zoo_model.download_framework_files(extensions=[".pth"])
####################################################################################
# End - SparseML optional load weights from SparseZoo
####################################################################################
model = create_model(
args.model,
pretrained=args.pretrained,
num_classes=args.num_classes,
drop_rate=args.drop,
drop_connect_rate=args.drop_connect, # DEPRECATED, use drop_path
drop_path_rate=args.drop_path,
drop_block_rate=args.drop_block,
global_pool=args.gp,
bn_tf=args.bn_tf,
bn_momentum=args.bn_momentum,
bn_eps=args.bn_eps,
scriptable=args.torchscript,
checkpoint_path=args.initial_checkpoint)
if args.num_classes is None:
assert hasattr(model, 'num_classes'), 'Model must have `num_classes` attr if not set on cmd line/config.'
args.num_classes = model.num_classes # FIXME handle model default vs config num_classes more elegantly
if args.local_rank == 0:
_logger.info('Model %s created, param count: %d' %
(args.model, sum([m.numel() for m in model.parameters()])))
data_config = resolve_data_config(vars(args), model=model, verbose=args.local_rank == 0)
# setup augmentation batch splits for contrastive loss or split bn
num_aug_splits = 0
if args.aug_splits > 0:
assert args.aug_splits > 1, 'A split of 1 makes no sense'
num_aug_splits = args.aug_splits
# enable split bn (separate bn stats per batch-portion)
if args.split_bn:
assert num_aug_splits > 1 or args.resplit
model = convert_splitbn_model(model, max(num_aug_splits, 2))
# move model to GPU, enable channels last layout if set
model.cuda()
if args.channels_last:
model = model.to(memory_format=torch.channels_last)
# setup synchronized BatchNorm for distributed training
if args.distributed and args.sync_bn:
assert not args.split_bn
if has_apex and use_amp != 'native':
# Apex SyncBN preferred unless native amp is activated
model = convert_syncbn_model(model)
else:
model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(model)
if args.local_rank == 0:
_logger.info(
'Converted model to use Synchronized BatchNorm. WARNING: You may have issues if using '
'zero initialized BN layers (enabled by default for ResNets) while sync-bn enabled.')
if args.torchscript:
assert not use_amp == 'apex', 'Cannot use APEX AMP with torchscripted model'
assert not args.sync_bn, 'Cannot use SyncBatchNorm with torchscripted model'
model = torch.jit.script(model)
optimizer = create_optimizer(args, model)
# setup automatic mixed-precision (AMP) loss scaling and op casting
amp_autocast = suppress # do nothing
loss_scaler = None
if use_amp == 'apex':
model, optimizer = amp.initialize(model, optimizer, opt_level='O1')
loss_scaler = ApexScaler()
if args.local_rank == 0:
_logger.info('Using NVIDIA APEX AMP. Training in mixed precision.')
elif use_amp == 'native':
amp_autocast = torch.cuda.amp.autocast
loss_scaler = NativeScaler()
if args.local_rank == 0:
_logger.info('Using native Torch AMP. Training in mixed precision.')
else:
if args.local_rank == 0:
_logger.info('AMP not enabled. Training in float32.')
# optionally resume from a checkpoint
resume_epoch = None
if args.resume:
resume_epoch = resume_checkpoint(
model, args.resume,
optimizer=None if args.no_resume_opt else optimizer,
loss_scaler=None if args.no_resume_opt else loss_scaler,
log_info=args.local_rank == 0)
# setup exponential moving average of model weights, SWA could be used here too
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 = ModelEmaV2(
model, decay=args.model_ema_decay, device='cpu' if args.model_ema_force_cpu else None)
if args.resume:
load_checkpoint(model_ema.module, args.resume, use_ema=True)
# setup distributed training
if args.distributed:
if has_apex and use_amp != 'native':
# Apex DDP preferred unless native amp is activated
if args.local_rank == 0:
_logger.info("Using NVIDIA APEX DistributedDataParallel.")
model = ApexDDP(model, delay_allreduce=True)
else:
if args.local_rank == 0:
_logger.info("Using native Torch DistributedDataParallel.")
model = NativeDDP(model, device_ids=[args.local_rank]) # can use device str in Torch >= 1.1
# NOTE: EMA model does not need to be wrapped by DDP
# setup learning rate schedule and starting epoch
lr_scheduler, num_epochs = create_scheduler(args, optimizer)
start_epoch = 0
if args.start_epoch is not None:
# a specified start_epoch will always override the resume epoch
start_epoch = args.start_epoch
elif resume_epoch is not None:
start_epoch = resume_epoch
if lr_scheduler is not None and start_epoch > 0:
lr_scheduler.step(start_epoch)
# create the train and eval datasets
dataset_train = create_dataset(
args.dataset, root=args.data_dir, split=args.train_split, is_training=True, batch_size=args.batch_size)
dataset_eval = create_dataset(
args.dataset, root=args.data_dir, split=args.val_split, is_training=False, batch_size=args.batch_size)
# setup mixup / cutmix
collate_fn = None
mixup_fn = None
mixup_active = args.mixup > 0 or args.cutmix > 0. or args.cutmix_minmax is not None
if mixup_active:
mixup_args = dict(
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.num_classes)
if args.prefetcher:
assert not num_aug_splits # collate conflict (need to support deinterleaving in collate mixup)
collate_fn = FastCollateMixup(**mixup_args)
else:
mixup_fn = Mixup(**mixup_args)
# wrap dataset in AugMix helper
if num_aug_splits > 1:
dataset_train = AugMixDataset(dataset_train, num_splits=num_aug_splits)
# create data loaders w/ augmentation pipeiine
train_interpolation = args.train_interpolation
if args.no_aug or not train_interpolation:
train_interpolation = data_config['interpolation']
loader_train = create_loader(
dataset_train,
input_size=data_config['input_size'],
batch_size=args.batch_size,
is_training=True,
use_prefetcher=args.prefetcher,
no_aug=args.no_aug,
re_prob=args.reprob,
re_mode=args.remode,
re_count=args.recount,
re_split=args.resplit,
scale=args.scale,
ratio=args.ratio,
hflip=args.hflip,
vflip=args.vflip,
color_jitter=args.color_jitter,
auto_augment=args.aa,
num_aug_splits=num_aug_splits,
interpolation=train_interpolation,
mean=data_config['mean'],
std=data_config['std'],
num_workers=args.workers,
distributed=args.distributed,
collate_fn=collate_fn,
pin_memory=args.pin_mem,
use_multi_epochs_loader=args.use_multi_epochs_loader
)
loader_eval = create_loader(
dataset_eval,
input_size=data_config['input_size'],
batch_size=args.validation_batch_size_multiplier * args.batch_size,
is_training=False,
use_prefetcher=args.prefetcher,
interpolation=data_config['interpolation'],
mean=data_config['mean'],
std=data_config['std'],
num_workers=args.workers,
distributed=args.distributed,
crop_pct=data_config['crop_pct'],
pin_memory=args.pin_mem,
)
# setup loss function
if args.jsd:
assert num_aug_splits > 1 # JSD only valid with aug splits set
train_loss_fn = JsdCrossEntropy(num_splits=num_aug_splits, smoothing=args.smoothing).cuda()
elif mixup_active:
# smoothing is handled with mixup target transform
train_loss_fn = SoftTargetCrossEntropy().cuda()
elif args.smoothing:
train_loss_fn = LabelSmoothingCrossEntropy(smoothing=args.smoothing).cuda()
else:
train_loss_fn = nn.CrossEntropyLoss().cuda()
validate_loss_fn = nn.CrossEntropyLoss().cuda()
# setup checkpoint saver and eval metric tracking
eval_metric = args.eval_metric
best_metric = None
best_epoch = None
saver = None
output_dir = ''
if args.local_rank == 0:
output_base = args.output if args.output else './output'
exp_name = '-'.join([
datetime.now().strftime("%Y%m%d-%H%M%S"),
args.model,
str(data_config['input_size'][-1])
])
output_dir = get_outdir(output_base, 'train', exp_name)
decreasing = True if eval_metric == 'loss' else False
saver = CheckpointSaver(
model=model, optimizer=optimizer, args=args, model_ema=model_ema, amp_scaler=loss_scaler,
checkpoint_dir=output_dir, recovery_dir=output_dir, decreasing=decreasing, max_history=args.checkpoint_hist)
with open(os.path.join(output_dir, 'args.yaml'), 'w') as f:
f.write(args_text)
####################################################################################
# Start SparseML Integration
####################################################################################
sparseml_loggers = (
[PythonLogger(), TensorBoardLogger(log_path=output_dir)]
if output_dir
else None
)
manager = ScheduledModifierManager.from_yaml(args.sparseml_recipe)
optimizer = ScheduledOptimizer(
optimizer,
model,
manager,
steps_per_epoch=len(loader_train),
loggers=sparseml_loggers
)
# override lr scheduler if recipe makes any LR updates
if any("LearningRate" in str(modifier) for modifier in manager.modifiers):
_logger.info("Disabling timm LR scheduler, managing LR using SparseML recipe")
lr_scheduler = None
if manager.max_epochs:
_logger.info(
f"Overriding max_epochs to {manager.max_epochs} from SparseML recipe"
)
num_epochs = manager.max_epochs or num_epochs
####################################################################################
# End SparseML Integration
####################################################################################
if args.local_rank == 0:
_logger.info('Scheduled epochs: {}'.format(num_epochs))
try:
for epoch in range(start_epoch, num_epochs):
if args.distributed and hasattr(loader_train.sampler, 'set_epoch'):
loader_train.sampler.set_epoch(epoch)
train_metrics = train_one_epoch(
epoch, model, loader_train, optimizer, train_loss_fn, args,
lr_scheduler=lr_scheduler, saver=saver, output_dir=output_dir,
amp_autocast=amp_autocast, loss_scaler=loss_scaler, model_ema=model_ema, mixup_fn=mixup_fn)
if args.distributed and args.dist_bn in ('broadcast', 'reduce'):
if args.local_rank == 0:
_logger.info("Distributing BatchNorm running means and vars")
distribute_bn(model, args.world_size, args.dist_bn == 'reduce')
eval_metrics = validate(model, loader_eval, validate_loss_fn, args, amp_autocast=amp_autocast)
if model_ema is not None and not args.model_ema_force_cpu:
if args.distributed and args.dist_bn in ('broadcast', 'reduce'):
distribute_bn(model_ema, args.world_size, args.dist_bn == 'reduce')
ema_eval_metrics = validate(
model_ema.module, loader_eval, validate_loss_fn, args, amp_autocast=amp_autocast, log_suffix=' (EMA)')
eval_metrics = ema_eval_metrics
if lr_scheduler is not None:
# step LR for next epoch
lr_scheduler.step(epoch + 1, eval_metrics[eval_metric])
update_summary(
epoch, train_metrics, eval_metrics, os.path.join(output_dir, 'summary.csv'),
write_header=best_metric is None)
if saver is not None:
# save proper checkpoint with eval metric
save_metric = eval_metrics[eval_metric]
best_metric, best_epoch = saver.save_checkpoint(epoch, metric=save_metric)
#################################################################################
# Start SparseML ONNX Export
#################################################################################
if output_dir:
_logger.info(
f"training complete, exporting ONNX to {output_dir}/model.onnx"
)
exporter = ModuleExporter(model, output_dir)
exporter.export_onnx(torch.randn((1, *data_config["input_size"])))
#################################################################################
# End SparseML ONNX Export
#################################################################################
except KeyboardInterrupt:
pass
if best_metric is not None:
_logger.info('*** Best metric: {0} (epoch {1})'.format(best_metric, best_epoch))
def main():
args = parser.parse_args()
args.prefetcher = not args.no_prefetcher
args.distributed = False
if 'WORLD_SIZE' in os.environ:
args.distributed = int(os.environ['WORLD_SIZE']) > 1
if args.distributed and args.num_gpu > 1:
print(
'Using more than one GPU per process in distributed mode is not allowed. Setting num_gpu to 1.'
)
args.num_gpu = 1
args.device = 'cuda:0'
args.world_size = 1
args.rank = 0 # global rank
if args.distributed:
args.num_gpu = 1
args.device = 'cuda:%d' % args.local_rank
torch.cuda.set_device(args.local_rank)
torch.distributed.init_process_group(backend='nccl',
init_method='env://')
args.world_size = torch.distributed.get_world_size()
args.rank = torch.distributed.get_rank()
assert args.rank >= 0
if args.distributed:
print(
'Training in distributed mode with multiple processes, 1 GPU per process. Process %d, total %d.'
% (args.rank, args.world_size))
else:
print('Training with a single process on %d GPUs.' % args.num_gpu)
torch.manual_seed(args.seed + args.rank)
model = create_model(args.model,
pretrained=args.pretrained,
num_classes=args.num_classes,
drop_rate=args.drop,
global_pool=args.gp,
bn_tf=args.bn_tf,
bn_momentum=args.bn_momentum,
bn_eps=args.bn_eps,
checkpoint_path=args.initial_checkpoint)
print('Model %s created, param count: %d' %
(args.model, sum([m.numel() for m in model.parameters()])))
data_config = resolve_data_config(model,
args,
verbose=args.local_rank == 0)
# optionally resume from a checkpoint
start_epoch = 0
optimizer_state = None
if args.resume:
optimizer_state, start_epoch = resume_checkpoint(
model, args.resume, args.start_epoch)
if args.num_gpu > 1:
if args.amp:
print(
'Warning: AMP does not work well with nn.DataParallel, disabling. '
'Use distributed mode for multi-GPU AMP.')
args.amp = False
model = nn.DataParallel(model,
device_ids=list(range(args.num_gpu))).cuda()
else:
if args.distributed and args.sync_bn and has_apex:
model = convert_syncbn_model(model)
model.cuda()
optimizer = create_optimizer(args, model)
if optimizer_state is not None:
optimizer.load_state_dict(optimizer_state)
if has_apex and args.amp:
model, optimizer = amp.initialize(model, optimizer, opt_level='O1')
use_amp = True
print('AMP enabled')
else:
use_amp = False
print('AMP disabled')
model_ema = None
if args.model_ema:
model_ema = ModelEma(model,
decay=args.model_ema_decay,
device='cpu' if args.model_ema_force_cpu else '',
resume=args.resume)
if args.distributed:
model = DDP(model, delay_allreduce=True)
if model_ema is not None and not args.model_ema_force_cpu:
# must also distribute EMA model to allow validation
model_ema.ema = DDP(model_ema.ema, delay_allreduce=True)
model_ema.ema_has_module = True
lr_scheduler, num_epochs = create_scheduler(args, optimizer)
if start_epoch > 0:
lr_scheduler.step(start_epoch)
if args.local_rank == 0:
print('Scheduled epochs: ', num_epochs)
train_dir = os.path.join(args.data, 'train')
if not os.path.exists(train_dir):
print('Error: training folder does not exist at: %s' % train_dir)
exit(1)
dataset_train = Dataset(train_dir)
collate_fn = None
if args.prefetcher and args.mixup > 0:
collate_fn = FastCollateMixup(args.mixup, args.smoothing,
args.num_classes)
loader_train = create_loader(
dataset_train,
input_size=data_config['input_size'],
batch_size=args.batch_size,
is_training=True,
use_prefetcher=args.prefetcher,
rand_erase_prob=args.reprob,
rand_erase_mode=args.remode,
interpolation=
'random', # FIXME cleanly resolve this? data_config['interpolation'],
mean=data_config['mean'],
std=data_config['std'],
num_workers=args.workers,
distributed=args.distributed,
collate_fn=collate_fn,
)
eval_dir = os.path.join(args.data, 'validation')
if not os.path.isdir(eval_dir):
print('Error: validation folder does not exist at: %s' % eval_dir)
exit(1)
dataset_eval = Dataset(eval_dir)
loader_eval = create_loader(
dataset_eval,
input_size=data_config['input_size'],
batch_size=4 * args.batch_size,
is_training=False,
use_prefetcher=args.prefetcher,
interpolation=data_config['interpolation'],
mean=data_config['mean'],
std=data_config['std'],
num_workers=args.workers,
distributed=args.distributed,
)
if args.mixup > 0.:
# smoothing is handled with mixup label transform
train_loss_fn = SoftTargetCrossEntropy().cuda()
validate_loss_fn = nn.CrossEntropyLoss().cuda()
elif args.smoothing:
train_loss_fn = LabelSmoothingCrossEntropy(
smoothing=args.smoothing).cuda()
validate_loss_fn = nn.CrossEntropyLoss().cuda()
else:
train_loss_fn = nn.CrossEntropyLoss().cuda()
validate_loss_fn = train_loss_fn
eval_metric = args.eval_metric
best_metric = None
best_epoch = None
saver = None
output_dir = ''
if args.local_rank == 0:
output_base = args.output if args.output else './output'
exp_name = '-'.join([
datetime.now().strftime("%Y%m%d-%H%M%S"), args.model,
str(data_config['input_size'][-1])
])
output_dir = get_outdir(output_base, 'train', exp_name)
decreasing = True if eval_metric == 'loss' else False
saver = CheckpointSaver(checkpoint_dir=output_dir,
decreasing=decreasing)
try:
for epoch in range(start_epoch, num_epochs):
if args.distributed:
loader_train.sampler.set_epoch(epoch)
train_metrics = train_epoch(epoch,
model,
loader_train,
optimizer,
train_loss_fn,
args,
lr_scheduler=lr_scheduler,
saver=saver,
output_dir=output_dir,
use_amp=use_amp,
model_ema=model_ema)
eval_metrics = validate(model, loader_eval, validate_loss_fn, args)
if model_ema is not None and not args.model_ema_force_cpu:
ema_eval_metrics = validate(model_ema.ema,
loader_eval,
validate_loss_fn,
args,
log_suffix=' (EMA)')
eval_metrics = ema_eval_metrics
if lr_scheduler is not None:
lr_scheduler.step(epoch, eval_metrics[eval_metric])
update_summary(epoch,
train_metrics,
eval_metrics,
os.path.join(output_dir, 'summary.csv'),
write_header=best_metric is None)
if saver is not None:
# save proper checkpoint with eval metric
save_metric = eval_metrics[eval_metric]
best_metric, best_epoch = saver.save_checkpoint(
model,
optimizer,
args,
epoch=epoch + 1,
model_ema=model_ema,
metric=save_metric)
except KeyboardInterrupt:
pass
if best_metric is not None:
print('*** Best metric: {0} (epoch {1})'.format(
best_metric, best_epoch))
#wide_resnet50_2=_attrib(paper_model_name=, paper_arxiv_id=), # same weights as torchvision
#wide_resnet101_2=_attrib(paper_model_name=, paper_arxiv_id=), # same weights as torchvision
xception=_attrib(paper_model_name='Xception', paper_arxiv_id='1610.02357'),
)
model_names = list_models(pretrained=True)
for model_name in model_names:
if model_name not in model_map:
print('Skipping %s' % model_name)
continue
# create model from name
model = create_model(model_name, pretrained=True)
param_count = sum([m.numel() for m in model.parameters()])
print('Model %s created, param count: %d' % (model_name, param_count))
# get appropriate transform for model's default pretrained config
data_config = resolve_data_config(dict(), model=model, verbose=True)
input_transform = create_transform(**data_config)
# Run the benchmark
ImageNet.benchmark(
model=model,
paper_model_name=model_map[model_name]['paper_model_name'],
paper_arxiv_id=model_map[model_name]['paper_arxiv_id'],
input_transform=input_transform,
batch_size=model_map[model_name]['batch_size'],
num_gpu=NUM_GPU,
data_root=os.environ.get('IMAGENET_DIR', './imagenet'))
def validate(args):
# might as well try to validate something
args.pretrained = args.pretrained or not args.checkpoint
args.prefetcher = not args.no_prefetcher
# amp_autocast = suppress # do nothing
# if args.amp:
# if has_native_amp:
# args.native_amp = True
# elif has_apex:
# args.apex_amp = True
# else:
# _logger.warning("Neither APEX or Native Torch AMP is available.")
# assert not args.apex_amp or not args.native_amp, "Only one AMP mode should be set."
# if args.native_amp:
# amp_autocast = torch.cuda.amp.autocast
# _logger.info('Validating in mixed precision with native PyTorch AMP.')
# elif args.apex_amp:
# _logger.info('Validating in mixed precision with NVIDIA APEX AMP.')
# else:
# _logger.info('Validating in float32. AMP not enabled.')
if args.legacy_jit:
set_jit_legacy()
# create model
model = create_model(
args.model,
pretrained=args.pretrained,
num_classes=args.num_classes,
in_chans=3,
global_pool=args.gp,
scriptable=args.torchscript)
if args.num_classes is None:
assert hasattr(model, 'num_classes'), 'Model must have `num_classes` attr if not set on cmd line/config.'
args.num_classes = model.num_classes
if args.checkpoint:
load_checkpoint(model, args.checkpoint, args.use_ema)
param_count = sum([m.numel() for m in model.parameters()])
_logger.info('Model %s created, param count: %d' % (args.model, param_count))
data_config = resolve_data_config(vars(args), model=model, use_test_size=True)
test_time_pool = False
if not args.no_test_pool:
model, test_time_pool = apply_test_time_pool(model, data_config, use_test_size=True)
if args.torchscript:
torch.jit.optimized_execution(True)
model = torch.jit.script(model)
# model = model.cuda()
# if args.apex_amp:
# model = amp.initialize(model, opt_level='O1')
if args.channels_last:
model = model.to(memory_format=torch.channels_last)
# if args.num_gpu > 1:
# model = torch.nn.DataParallel(model, device_ids=list(range(args.num_gpu)))
# criterion = nn.CrossEntropyLoss().cuda()
criterion = nn.CrossEntropyLoss()
dataset = create_dataset(
root=args.data, name=args.dataset, split=args.split,
load_bytes=args.tf_preprocessing, class_map=args.class_map)
# added for post quantization calibration
calib_dataset = create_dataset(
root=args.data, name=args.dataset, split=args.split,
load_bytes=args.tf_preprocessing, class_map=args.class_map)
if args.valid_labels:
with open(args.valid_labels, 'r') as f:
valid_labels = {int(line.rstrip()) for line in f}
valid_labels = [i in valid_labels for i in range(args.num_classes)]
else:
valid_labels = None
if args.real_labels:
real_labels = RealLabelsImagenet(dataset.filenames(basename=True), real_json=args.real_labels)
else:
real_labels = None
crop_pct = 1.0 if test_time_pool else data_config['crop_pct']
loader = create_loader(
dataset,
input_size=data_config['input_size'],
batch_size=args.batch_size,
use_prefetcher=args.prefetcher,
interpolation=data_config['interpolation'],
mean=data_config['mean'],
std=data_config['std'],
num_workers=args.workers,
crop_pct=crop_pct,
pin_memory=args.pin_mem,
tf_preprocessing=args.tf_preprocessing)
#Also create loader for calibration dataset
calib_loader = create_loader(
calib_dataset,
input_size=data_config['input_size'],
batch_size=args.batch_size,
use_prefetcher=args.prefetcher,
interpolation=data_config['interpolation'],
mean=data_config['mean'],
std=data_config['std'],
num_workers=args.workers,
crop_pct=crop_pct,
pin_memory=args.pin_mem,
tf_preprocessing=args.tf_preprocessing)
batch_time = AverageMeter()
losses = AverageMeter()
top1 = AverageMeter()
top5 = AverageMeter()
print('Start calibration of quantization observers before post-quantization')
model_to_quantize = copy.deepcopy(model)
model_to_quantize.eval()
#post training static quantization
if args.quant_option == 'static':
qconfig_dict = {"": torch.quantization.default_static_qconfig}
model_to_quantize = copy.deepcopy(model_fp)
qconfig_dict = {"": torch.quantization.get_default_qconfig('qnnpack')}
model_to_quantize.eval()
# prepare
model_prepared = quantize_fx.prepare_fx(model_to_quantize, qconfig_dict)
# calibrate
with torch.no_grad():
# warmup, reduce variability of first batch time, especially for comparing torchscript vs non
input = torch.randn((args.batch_size,) + tuple(data_config['input_size']))
if args.channels_last:
input = input.contiguous(memory_format=torch.channels_last)
model(input)
end = time.time()
for batch_idx, (input, target) in enumerate(loader):
if args.channels_last:
input = input.contiguous(memory_format=torch.channels_last)
if valid_labels is not None:
output = output[:, valid_labels]
loss = criterion(output, target)
if real_labels is not None:
real_labels.add_result(output)
# measure accuracy and record loss
acc1, acc5 = accuracy(output.detach(), target, topk=(1, 5))
losses.update(loss.item(), input.size(0))
top1.update(acc1.item(), input.size(0))
top5.update(acc5.item(), input.size(0))
# measure elapsed time
batch_time.update(time.time() - end)
end = time.time()
if batch_idx % args.log_freq == 0:
_logger.info(
'Test: [{0:>4d}/{1}] '
'Time: {batch_time.val:.3f}s ({batch_time.avg:.3f}s, {rate_avg:>7.2f}/s) '
'Loss: {loss.val:>7.4f} ({loss.avg:>6.4f}) '
'Acc@1: {top1.val:>7.3f} ({top1.avg:>7.3f}) '
'Acc@5: {top5.val:>7.3f} ({top5.avg:>7.3f})'.format(
batch_idx, len(loader), batch_time=batch_time,
rate_avg=input.size(0) / batch_time.avg,
loss=losses, top1=top1, top5=top5))
# quantize
model_quantized = quantize_fx.convert_fx(model_prepared)
#post training dynamic/weight only quantization
elif args.quant_option == 'dynamic':
qconfig_dict = {"": torch.quantization.default_dynamic_qconfig}
# prepare
model_prepared = quantize_fx.prepare_fx(model_to_quantize, qconfig_dict)
# no calibration needed when we only have dynamici/weight_only quantization
# quantize
model_quantized = quantize_fx.convert_fx(model_prepared)
else:
_logger.warning("Invalid quantization option. Set option to default(static)")
#
# fusion
#
model_to_quantize = copy.deepcopy(model_fp)
model_fused = quantize_fx.fuse_fx(model_to_quantize)
model = model_fused
with torch.no_grad():
# warmup, reduce variability of first batch time, especially for comparing torchscript vs non
# input = torch.randn((args.batch_size,) + tuple(data_config['input_size'])).cuda()
input = torch.randn((args.batch_size,) + tuple(data_config['input_size']))
if args.channels_last:
input = input.contiguous(memory_format=torch.channels_last)
model(input)
end = time.time()
for batch_idx, (input, target) in enumerate(loader):
# if args.no_prefetcher:
# target = target.cuda()
# input = input.cuda()
if args.channels_last:
input = input.contiguous(memory_format=torch.channels_last)
# compute output
# with amp_autocast():
# output = model(input)
if valid_labels is not None:
output = output[:, valid_labels]
loss = criterion(output, target)
if real_labels is not None:
real_labels.add_result(output)
# measure accuracy and record loss
acc1, acc5 = accuracy(output.detach(), target, topk=(1, 5))
losses.update(loss.item(), input.size(0))
top1.update(acc1.item(), input.size(0))
top5.update(acc5.item(), input.size(0))
# measure elapsed time
batch_time.update(time.time() - end)
end = time.time()
if batch_idx % args.log_freq == 0:
_logger.info(
'Test: [{0:>4d}/{1}] '
'Time: {batch_time.val:.3f}s ({batch_time.avg:.3f}s, {rate_avg:>7.2f}/s) '
'Loss: {loss.val:>7.4f} ({loss.avg:>6.4f}) '
'Acc@1: {top1.val:>7.3f} ({top1.avg:>7.3f}) '
'Acc@5: {top5.val:>7.3f} ({top5.avg:>7.3f})'.format(
batch_idx, len(loader), batch_time=batch_time,
rate_avg=input.size(0) / batch_time.avg,
loss=losses, top1=top1, top5=top5))
if real_labels is not None:
# real labels mode replaces topk values at the end
top1a, top5a = real_labels.get_accuracy(k=1), real_labels.get_accuracy(k=5)
else:
top1a, top5a = top1.avg, top5.avg
results = OrderedDict(
top1=round(top1a, 4), top1_err=round(100 - top1a, 4),
top5=round(top5a, 4), top5_err=round(100 - top5a, 4),
param_count=round(param_count / 1e6, 2),
img_size=data_config['input_size'][-1],
cropt_pct=crop_pct,
interpolation=data_config['interpolation'])
_logger.info(' * Acc@1 {:.3f} ({:.3f}) Acc@5 {:.3f} ({:.3f})'.format(
results['top1'], results['top1_err'], results['top5'], results['top5_err']))
return results
