transform_train = video.VideoGroupTrainTransform(size=(opt.input_size, opt.input_size), scale_ratios=[1.0, 0.875, 0.75, 0.66], mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
transform_test = video.VideoGroupValTransform(size=opt.input_size, mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
# Batch Size for Each GPU
per_device_batch_size = opt.per_device_batch_size
# Number of data loader workers
num_workers = opt.num_workers
# Calculate effective total batch size
batch_size = per_device_batch_size * num_gpus
# Set train=True for training data. Here we only use a subset of UCF101 for demonstration purpose.
# The subset has 101 training samples, one sample per class.
train_dataset = UCF101(setting=opt.train_setting, root=opt.train_dir, train=True,
new_width=opt.new_width, new_height=opt.new_height, new_length=opt.new_length,new_step=opt.new_step,
target_width=opt.input_size, target_height=opt.input_size,
num_segments=opt.num_segments, transform=transform_train)
val_dataset = UCF101(setting=opt.val_setting, root=opt.train_dir, train=False,
new_width=opt.new_width, new_height=opt.new_height, new_length=opt.new_length,new_step=opt.new_step,
target_width=opt.input_size, target_height=opt.input_size,
num_segments=opt.num_segments, transform=transform_test)
train_data = gluon.data.DataLoader(train_dataset, batch_size=batch_size, shuffle=True, num_workers=num_workers,
prefetch=int(opt.prefetch_ratio * num_workers), last_batch='rollover')
val_data = gluon.data.DataLoader(val_dataset, batch_size=batch_size, shuffle=False, num_workers=num_workers,
prefetch=int(opt.prefetch_ratio * num_workers), last_batch='discard')
train_data = gluon.data.DataLoader(train_dataset, batch_size=batch_size,
shuffle=True, num_workers=num_workers)
def get_data_loader(opt, batch_size, num_workers, logger, kvstore=None):
data_dir = opt.data_dir
val_data_dir = opt.val_data_dir
scale_ratios = [float(i) for i in opt.scale_ratios.split(',')]
input_size = opt.input_size
def batch_fn(batch, ctx):
data = split_and_load(batch[0],
ctx_list=ctx,
batch_axis=0,
even_split=False)
label = split_and_load(batch[1],
ctx_list=ctx,
batch_axis=0,
even_split=False)
return data, label
transform_train = video.VideoGroupTrainTransform(
size=(input_size, input_size),
scale_ratios=scale_ratios,
mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
transform_test = video.VideoGroupValTransform(size=input_size,
mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
if opt.dataset == 'kinetics400':
train_dataset = Kinetics400(setting=opt.train_list,
root=data_dir,
train=True,
new_width=opt.new_width,
new_height=opt.new_height,
new_length=opt.new_length,
new_step=opt.new_step,
target_width=input_size,
target_height=input_size,
video_loader=opt.video_loader,
use_decord=opt.use_decord,
num_segments=opt.num_segments,
transform=transform_train)
val_dataset = Kinetics400(setting=opt.val_list,
root=val_data_dir,
train=False,
new_width=opt.new_width,
new_height=opt.new_height,
new_length=opt.new_length,
new_step=opt.new_step,
target_width=input_size,
target_height=input_size,
video_loader=opt.video_loader,
use_decord=opt.use_decord,
num_segments=opt.num_segments,
transform=transform_test)
elif opt.dataset == 'ucf101':
train_dataset = UCF101(setting=opt.train_list,
root=data_dir,
train=True,
new_width=opt.new_width,
new_height=opt.new_height,
new_length=opt.new_length,
target_width=input_size,
target_height=input_size,
num_segments=opt.num_segments,
transform=transform_train)
val_dataset = UCF101(setting=opt.val_list,
root=data_dir,
train=False,
new_width=opt.new_width,
new_height=opt.new_height,
new_length=opt.new_length,
target_width=input_size,
target_height=input_size,
num_segments=opt.num_segments,
transform=transform_test)
elif opt.dataset == 'somethingsomethingv2':
train_dataset = SomethingSomethingV2(setting=opt.train_list,
root=data_dir,
train=True,
new_width=opt.new_width,
new_height=opt.new_height,
new_length=opt.new_length,
new_step=opt.new_step,
target_width=input_size,
target_height=input_size,
video_loader=opt.video_loader,
use_decord=opt.use_decord,
num_segments=opt.num_segments,
transform=transform_train)
val_dataset = SomethingSomethingV2(setting=opt.val_list,
root=data_dir,
train=False,
new_width=opt.new_width,
new_height=opt.new_height,
new_length=opt.new_length,
new_step=opt.new_step,
target_width=input_size,
target_height=input_size,
video_loader=opt.video_loader,
use_decord=opt.use_decord,
num_segments=opt.num_segments,
transform=transform_test)
elif opt.dataset == 'hmdb51':
train_dataset = HMDB51(setting=opt.train_list,
root=data_dir,
train=True,
new_width=opt.new_width,
new_height=opt.new_height,
new_length=opt.new_length,
new_step=opt.new_step,
target_width=input_size,
target_height=input_size,
video_loader=opt.video_loader,
use_decord=opt.use_decord,
num_segments=opt.num_segments,
transform=transform_train)
val_dataset = HMDB51(setting=opt.val_list,
root=data_dir,
train=False,
new_width=opt.new_width,
new_height=opt.new_height,
new_length=opt.new_length,
new_step=opt.new_step,
target_width=input_size,
target_height=input_size,
video_loader=opt.video_loader,
use_decord=opt.use_decord,
num_segments=opt.num_segments,
transform=transform_test)
else:
logger.info('Dataset %s is not supported yet.' % (opt.dataset))
logger.info('Load %d training samples and %d validation samples.' %
(len(train_dataset), len(val_dataset)))
if kvstore is not None:
train_data = gluon.data.DataLoader(
train_dataset,
batch_size=batch_size,
num_workers=num_workers,
sampler=SplitSampler(len(train_dataset),
num_parts=kvstore.num_workers,
part_index=kvstore.rank),
prefetch=int(opt.prefetch_ratio * num_workers),
last_batch='rollover')
val_data = gluon.data.DataLoader(
val_dataset,
batch_size=batch_size,
num_workers=num_workers,
sampler=SplitSampler(len(val_dataset),
num_parts=kvstore.num_workers,
part_index=kvstore.rank),
prefetch=int(opt.prefetch_ratio * num_workers),
last_batch='discard')
else:
train_data = gluon.data.DataLoader(train_dataset,
batch_size=batch_size,
shuffle=True,
num_workers=num_workers,
prefetch=int(opt.prefetch_ratio *
num_workers),
last_batch='rollover')
val_data = gluon.data.DataLoader(val_dataset,
batch_size=batch_size,
shuffle=False,
num_workers=num_workers,
prefetch=int(opt.prefetch_ratio *
num_workers),
last_batch='discard')
return train_data, val_data, batch_fn
def get_data_loader(opt, batch_size, num_workers, logger, kvstore=None):
data_dir = opt.data_dir
val_data_dir = opt.val_data_dir
scale_ratios = [float(i) for i in opt.scale_ratios.split(',')]
input_size = opt.input_size
default_mean = [0.485, 0.456, 0.406]
default_std = [0.229, 0.224, 0.225]
def batch_fn(batch, ctx):
data = split_and_load(batch[0],
ctx_list=ctx,
batch_axis=0,
even_split=False)
label = split_and_load(batch[1],
ctx_list=ctx,
batch_axis=0,
even_split=False)
return data, label
if opt.data_aug == 'v1':
# GluonCV style, not keeping aspect ratio, multi-scale crop
transform_train = video.VideoGroupTrainTransform(
size=(input_size, input_size),
scale_ratios=scale_ratios,
mean=default_mean,
std=default_std)
transform_test = video.VideoGroupValTransform(size=input_size,
mean=default_mean,
std=default_std)
elif opt.data_aug == 'v2':
# GluonCV style, keeping aspect ratio, multi-scale crop, same as mmaction style
transform_train = video.VideoGroupTrainTransformV2(
size=(input_size, input_size),
short_side=opt.new_height,
scale_ratios=scale_ratios,
mean=default_mean,
std=default_std)
transform_test = video.VideoGroupValTransformV2(
crop_size=(input_size, input_size),
short_side=opt.new_height,
mean=default_mean,
std=default_std)
elif opt.data_aug == 'v3':
# PySlowFast style, keeping aspect ratio, random short side scale jittering
transform_train = video.VideoGroupTrainTransformV3(
crop_size=(input_size, input_size),
min_size=opt.new_height,
max_size=opt.new_width,
mean=default_mean,
std=default_std)
transform_test = video.VideoGroupValTransformV2(
crop_size=(input_size, input_size),
short_side=opt.new_height,
mean=default_mean,
std=default_std)
elif opt.data_aug == 'v4':
# mmaction style, keeping aspect ratio, random crop and resize, only for SlowFast family models, similar to 'v3'
transform_train = video.VideoGroupTrainTransformV4(size=(input_size,
input_size),
mean=default_mean,
std=default_std)
transform_test = video.VideoGroupValTransformV2(
crop_size=(input_size, input_size),
short_side=opt.new_height,
mean=default_mean,
std=default_std)
else:
logger.info('Data augmentation %s is not supported yet.' %
(opt.data_aug))
if opt.dataset == 'kinetics400':
train_dataset = Kinetics400(
setting=opt.train_list,
root=data_dir,
train=True,
new_width=opt.new_width,
new_height=opt.new_height,
new_length=opt.new_length,
new_step=opt.new_step,
target_width=input_size,
target_height=input_size,
video_loader=opt.video_loader,
use_decord=opt.use_decord,
slowfast=opt.slowfast,
slow_temporal_stride=opt.slow_temporal_stride,
fast_temporal_stride=opt.fast_temporal_stride,
data_aug=opt.data_aug,
num_segments=opt.num_segments,
transform=transform_train)
val_dataset = Kinetics400(
setting=opt.val_list,
root=val_data_dir,
train=False,
new_width=opt.new_width,
new_height=opt.new_height,
new_length=opt.new_length,
new_step=opt.new_step,
target_width=input_size,
target_height=input_size,
video_loader=opt.video_loader,
use_decord=opt.use_decord,
slowfast=opt.slowfast,
slow_temporal_stride=opt.slow_temporal_stride,
fast_temporal_stride=opt.fast_temporal_stride,
data_aug=opt.data_aug,
num_segments=opt.num_segments,
transform=transform_test)
elif opt.dataset == 'ucf101':
train_dataset = UCF101(setting=opt.train_list,
root=data_dir,
train=True,
new_width=opt.new_width,
new_height=opt.new_height,
new_length=opt.new_length,
target_width=input_size,
target_height=input_size,
data_aug=opt.data_aug,
num_segments=opt.num_segments,
transform=transform_train)
val_dataset = UCF101(setting=opt.val_list,
root=data_dir,
train=False,
new_width=opt.new_width,
new_height=opt.new_height,
new_length=opt.new_length,
target_width=input_size,
target_height=input_size,
data_aug=opt.data_aug,
num_segments=opt.num_segments,
transform=transform_test)
elif opt.dataset == 'somethingsomethingv2':
train_dataset = SomethingSomethingV2(setting=opt.train_list,
root=data_dir,
train=True,
new_width=opt.new_width,
new_height=opt.new_height,
new_length=opt.new_length,
new_step=opt.new_step,
target_width=input_size,
target_height=input_size,
video_loader=opt.video_loader,
use_decord=opt.use_decord,
data_aug=opt.data_aug,
num_segments=opt.num_segments,
transform=transform_train)
val_dataset = SomethingSomethingV2(setting=opt.val_list,
root=data_dir,
train=False,
new_width=opt.new_width,
new_height=opt.new_height,
new_length=opt.new_length,
new_step=opt.new_step,
target_width=input_size,
target_height=input_size,
video_loader=opt.video_loader,
use_decord=opt.use_decord,
data_aug=opt.data_aug,
num_segments=opt.num_segments,
transform=transform_test)
elif opt.dataset == 'hmdb51':
train_dataset = HMDB51(setting=opt.train_list,
root=data_dir,
train=True,
new_width=opt.new_width,
new_height=opt.new_height,
new_length=opt.new_length,
new_step=opt.new_step,
target_width=input_size,
target_height=input_size,
video_loader=opt.video_loader,
use_decord=opt.use_decord,
data_aug=opt.data_aug,
num_segments=opt.num_segments,
transform=transform_train)
val_dataset = HMDB51(setting=opt.val_list,
root=data_dir,
train=False,
new_width=opt.new_width,
new_height=opt.new_height,
new_length=opt.new_length,
new_step=opt.new_step,
target_width=input_size,
target_height=input_size,
video_loader=opt.video_loader,
use_decord=opt.use_decord,
data_aug=opt.data_aug,
num_segments=opt.num_segments,
transform=transform_test)
elif opt.dataset == 'custom':
train_dataset = VideoClsCustom(
setting=opt.train_list,
root=data_dir,
train=True,
new_width=opt.new_width,
new_height=opt.new_height,
new_length=opt.new_length,
new_step=opt.new_step,
target_width=input_size,
target_height=input_size,
video_loader=opt.video_loader,
use_decord=opt.use_decord,
slowfast=opt.slowfast,
slow_temporal_stride=opt.slow_temporal_stride,
fast_temporal_stride=opt.fast_temporal_stride,
data_aug=opt.data_aug,
num_segments=opt.num_segments,
transform=transform_train)
val_dataset = VideoClsCustom(
setting=opt.val_list,
root=val_data_dir,
train=False,
new_width=opt.new_width,
new_height=opt.new_height,
new_length=opt.new_length,
new_step=opt.new_step,
target_width=input_size,
target_height=input_size,
video_loader=opt.video_loader,
use_decord=opt.use_decord,
slowfast=opt.slowfast,
slow_temporal_stride=opt.slow_temporal_stride,
fast_temporal_stride=opt.fast_temporal_stride,
data_aug=opt.data_aug,
num_segments=opt.num_segments,
transform=transform_test)
else:
logger.info('Dataset %s is not supported yet.' % (opt.dataset))
logger.info('Load %d training samples and %d validation samples.' %
(len(train_dataset), len(val_dataset)))
if kvstore is not None:
train_data = gluon.data.DataLoader(
train_dataset,
batch_size=batch_size,
num_workers=num_workers,
sampler=ShuffleSplitSampler(len(train_dataset),
num_parts=kvstore.num_workers,
part_index=kvstore.rank),
prefetch=int(opt.prefetch_ratio * num_workers),
last_batch='rollover')
val_data = gluon.data.DataLoader(
val_dataset,
batch_size=batch_size,
num_workers=num_workers,
sampler=ShuffleSplitSampler(len(val_dataset),
num_parts=kvstore.num_workers,
part_index=kvstore.rank),
prefetch=int(opt.prefetch_ratio * num_workers),
last_batch='discard')
else:
train_data = gluon.data.DataLoader(train_dataset,
batch_size=batch_size,
shuffle=True,
num_workers=num_workers,
prefetch=int(opt.prefetch_ratio *
num_workers),
last_batch='rollover')
val_data = gluon.data.DataLoader(val_dataset,
batch_size=batch_size,
shuffle=False,
num_workers=num_workers,
prefetch=int(opt.prefetch_ratio *
num_workers),
last_batch='discard')
return train_data, val_data, batch_fn
def main():
opt = parse_args()
print(opt)
# Garbage collection, default threshold is (700, 10, 10).
# Set threshold lower to collect garbage more frequently and release more CPU memory for heavy data loading.
gc.set_threshold(100, 5, 5)
# set env
num_gpus = opt.num_gpus
batch_size = opt.batch_size
batch_size *= max(1, num_gpus)
context = [mx.gpu(i)
for i in range(num_gpus)] if num_gpus > 0 else [mx.cpu()]
num_workers = opt.num_workers
print('Total batch size is set to %d on %d GPUs' % (batch_size, num_gpus))
# get data
if opt.ten_crop:
transform_test = transforms.Compose([
video.VideoTenCrop(opt.input_size),
video.VideoToTensor(),
video.VideoNormalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])
opt.num_crop = 10
elif opt.three_crop:
transform_test = transforms.Compose([
video.VideoThreeCrop(opt.input_size),
video.VideoToTensor(),
video.VideoNormalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])
opt.num_crop = 3
else:
transform_test = video.VideoGroupValTransform(
size=opt.input_size,
mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
opt.num_crop = 1
# get model
if opt.use_pretrained and len(opt.hashtag) > 0:
opt.use_pretrained = opt.hashtag
classes = opt.num_classes
model_name = opt.model
net = get_model(name=model_name,
nclass=classes,
pretrained=opt.use_pretrained,
num_segments=opt.num_segments,
num_crop=opt.num_crop)
net.cast(opt.dtype)
net.collect_params().reset_ctx(context)
if opt.mode == 'hybrid':
net.hybridize(static_alloc=True, static_shape=True)
if opt.resume_params is not '' and not opt.use_pretrained:
net.load_parameters(opt.resume_params, ctx=context)
print('Pre-trained model %s is successfully loaded.' %
(opt.resume_params))
else:
print('Pre-trained model is successfully loaded from the model zoo.')
if opt.dataset == 'ucf101':
val_dataset = UCF101(setting=opt.val_list,
root=opt.data_dir,
train=False,
new_width=opt.new_width,
new_height=opt.new_height,
new_length=opt.new_length,
target_width=opt.input_size,
target_height=opt.input_size,
test_mode=True,
num_segments=opt.num_segments,
transform=transform_test)
elif opt.dataset == 'kinetics400':
val_dataset = Kinetics400(setting=opt.val_list,
root=opt.data_dir,
train=False,
new_width=opt.new_width,
new_height=opt.new_height,
new_length=opt.new_length,
new_step=opt.new_step,
target_width=opt.input_size,
target_height=opt.input_size,
video_loader=opt.video_loader,
use_decord=opt.use_decord,
test_mode=True,
num_segments=opt.num_segments,
transform=transform_test)
elif opt.dataset == 'somethingsomethingv2':
val_dataset = SomethingSomethingV2(setting=opt.val_list,
root=opt.data_dir,
train=False,
new_width=opt.new_width,
new_height=opt.new_height,
new_length=opt.new_length,
new_step=opt.new_step,
target_width=opt.input_size,
target_height=opt.input_size,
video_loader=opt.video_loader,
use_decord=opt.use_decord,
num_segments=opt.num_segments,
transform=transform_test)
elif opt.dataset == 'hmdb51':
val_dataset = HMDB51(setting=opt.val_list,
root=opt.data_dir,
train=False,
new_width=opt.new_width,
new_height=opt.new_height,
new_length=opt.new_length,
new_step=opt.new_step,
target_width=opt.input_size,
target_height=opt.input_size,
video_loader=opt.video_loader,
use_decord=opt.use_decord,
num_segments=opt.num_segments,
transform=transform_test)
else:
logger.info('Dataset %s is not supported yet.' % (opt.dataset))
val_data = gluon.data.DataLoader(val_dataset,
batch_size=batch_size,
shuffle=False,
num_workers=num_workers,
prefetch=int(opt.prefetch_ratio *
num_workers),
last_batch='discard')
print('Load %d test samples in %d iterations.' %
(len(val_dataset), len(val_data)))
start_time = time.time()
acc_top1_val, acc_top5_val = test(context, val_data, opt, net)
end_time = time.time()
print('Test accuracy: acc-top1=%f acc-top5=%f' %
(acc_top1_val * 100, acc_top5_val * 100))
print('Total evaluation time is %4.2f minutes' %
((end_time - start_time) / 60))
def main(logger):
opt = parse_args()
print(opt)
# Garbage collection, default threshold is (700, 10, 10).
# Set threshold lower to collect garbage more frequently and release more CPU memory for heavy data loading.
gc.set_threshold(100, 5, 5)
# set env
num_gpus = opt.num_gpus
batch_size = opt.batch_size
context = [mx.cpu()]
if num_gpus > 0:
batch_size *= max(1, num_gpus)
context = [mx.gpu(i) for i in range(num_gpus)]
num_workers = opt.num_workers
print('Total batch size is set to %d on %d GPUs' % (batch_size, num_gpus))
# get data
image_norm_mean = [0.485, 0.456, 0.406]
image_norm_std = [0.229, 0.224, 0.225]
if opt.ten_crop:
transform_test = transforms.Compose([
video.VideoTenCrop(opt.input_size),
video.VideoToTensor(),
video.VideoNormalize(image_norm_mean, image_norm_std)
])
opt.num_crop = 10
elif opt.three_crop:
transform_test = transforms.Compose([
video.VideoThreeCrop(opt.input_size),
video.VideoToTensor(),
video.VideoNormalize(image_norm_mean, image_norm_std)
])
opt.num_crop = 3
else:
transform_test = video.VideoGroupValTransform(size=opt.input_size,
mean=image_norm_mean,
std=image_norm_std)
opt.num_crop = 1
if not opt.deploy:
# get model
if opt.use_pretrained and len(opt.hashtag) > 0:
opt.use_pretrained = opt.hashtag
classes = opt.num_classes
model_name = opt.model
# Currently, these is no hashtag for int8 models.
if opt.quantized:
model_name += '_int8'
opt.use_pretrained = True
net = get_model(name=model_name,
nclass=classes,
pretrained=opt.use_pretrained,
num_segments=opt.num_segments,
num_crop=opt.num_crop)
net.cast(opt.dtype)
net.collect_params().reset_ctx(context)
if opt.mode == 'hybrid':
net.hybridize(static_alloc=True, static_shape=True)
if opt.resume_params is not '' and not opt.use_pretrained:
net.load_parameters(opt.resume_params, ctx=context)
print('Pre-trained model %s is successfully loaded.' %
(opt.resume_params))
else:
print(
'Pre-trained model is successfully loaded from the model zoo.')
else:
model_name = 'deploy'
net = mx.gluon.SymbolBlock.imports(
'{}-symbol.json'.format(opt.model_prefix), ['data'],
'{}-0000.params'.format(opt.model_prefix))
net.hybridize(static_alloc=True, static_shape=True)
print("Successfully loaded model {}".format(model_name))
# dummy data for benchmarking performance
if opt.benchmark:
benchmarking(opt, net, context)
sys.exit()
if opt.dataset == 'ucf101':
val_dataset = UCF101(setting=opt.val_list,
root=opt.data_dir,
train=False,
new_width=opt.new_width,
new_height=opt.new_height,
new_length=opt.new_length,
target_width=opt.input_size,
target_height=opt.input_size,
test_mode=True,
num_segments=opt.num_segments,
transform=transform_test)
elif opt.dataset == 'kinetics400':
val_dataset = Kinetics400(
setting=opt.val_list,
root=opt.data_dir,
train=False,
new_width=opt.new_width,
new_height=opt.new_height,
new_length=opt.new_length,
new_step=opt.new_step,
target_width=opt.input_size,
target_height=opt.input_size,
video_loader=opt.video_loader,
use_decord=opt.use_decord,
slowfast=opt.slowfast,
slow_temporal_stride=opt.slow_temporal_stride,
fast_temporal_stride=opt.fast_temporal_stride,
test_mode=True,
num_segments=opt.num_segments,
num_crop=opt.num_crop,
transform=transform_test)
elif opt.dataset == 'somethingsomethingv2':
val_dataset = SomethingSomethingV2(setting=opt.val_list,
root=opt.data_dir,
train=False,
new_width=opt.new_width,
new_height=opt.new_height,
new_length=opt.new_length,
new_step=opt.new_step,
target_width=opt.input_size,
target_height=opt.input_size,
video_loader=opt.video_loader,
use_decord=opt.use_decord,
num_segments=opt.num_segments,
transform=transform_test)
elif opt.dataset == 'hmdb51':
val_dataset = HMDB51(setting=opt.val_list,
root=opt.data_dir,
train=False,
new_width=opt.new_width,
new_height=opt.new_height,
new_length=opt.new_length,
new_step=opt.new_step,
target_width=opt.input_size,
target_height=opt.input_size,
video_loader=opt.video_loader,
use_decord=opt.use_decord,
num_segments=opt.num_segments,
transform=transform_test)
else:
logger.info('Dataset %s is not supported yet.' % (opt.dataset))
val_data = gluon.data.DataLoader(val_dataset,
batch_size=batch_size,
shuffle=False,
num_workers=num_workers,
prefetch=int(opt.prefetch_ratio *
num_workers),
last_batch='discard')
print('Load %d test samples in %d iterations.' %
(len(val_dataset), len(val_data)))
# calibrate FP32 model into INT8 model
if opt.calibration:
calibration(net, val_data, opt, context, logger)
sys.exit()
start_time = time.time()
acc_top1_val, acc_top5_val = test(context, val_data, opt, net)
end_time = time.time()
print('Test accuracy: acc-top1=%f acc-top5=%f' %
(acc_top1_val * 100, acc_top5_val * 100))
print('Total evaluation time is %4.2f minutes' %
((end_time - start_time) / 60))
video.VideoNormalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])
##################################################################
# With the transform functions, we can define data loaders for our
# training datasets.
# Batch Size for Each GPU
per_device_batch_size = 5
# Number of data loader workers
num_workers = 8
# Calculate effective total batch size
batch_size = per_device_batch_size * num_gpus
# Set train=True for training the model. Here we set num_segments to 3 to enable TSN training.
train_dataset = UCF101(train=True, num_segments=3, transform=transform_train)
print('Load %d training samples.' % len(train_dataset))
train_data = gluon.data.DataLoader(train_dataset,
batch_size=batch_size,
shuffle=True,
num_workers=num_workers)
################################################################
# Optimizer, Loss and Metric
# --------------------------
# Learning rate decay factor
lr_decay = 0.1
# Epochs where learning rate decays
lr_decay_epoch = [30, 60, np.inf]
def main(logger):
opt = parse_args(parser)
print(opt)
assert not (os.path.isdir(opt.save_dir)), "already done this experiment..."
Path(opt.save_dir).mkdir(parents=True)
# Garbage collection, default threshold is (700, 10, 10).
# Set threshold lower to collect garbage more frequently and release more CPU memory for heavy data loading.
gc.set_threshold(100, 5, 5)
num_gpus = 1
context = [mx.gpu(i) for i in range(num_gpus)]
per_device_batch_size = 5
num_workers = 12
batch_size = per_device_batch_size * num_gpus
num_workers = opt.num_workers
print('Total batch size is set to %d on %d GPUs' % (batch_size, num_gpus))
# get data
default_mean = [0.485, 0.456, 0.406]
default_std = [0.229, 0.224, 0.225]
# if opt.ten_crop:
# if opt.data_aug == 'v1':
# transform_test = transforms.Compose([
# video.VideoTenCrop(opt.input_size),
# video.VideoToTensor(),
# video.VideoNormalize(default_mean, default_std)
# ])
# else:
# transform_test = transforms.Compose([
# video.ShortSideRescale(opt.input_size),
# video.VideoTenCrop(opt.input_size),
# video.VideoToTensor(),
# video.VideoNormalize(default_mean, default_std)
# ])
# opt.num_crop = 10
# elif opt.three_crop:
# if opt.data_aug == 'v1':
# transform_test = transforms.Compose([
# video.VideoThreeCrop(opt.input_size),
# video.VideoToTensor(),
# video.VideoNormalize(default_mean, default_std)
# ])
# else:
# transform_test = transforms.Compose([
# video.ShortSideRescale(opt.input_size),
# video.VideoThreeCrop(opt.input_size),
# video.VideoToTensor(),
# video.VideoNormalize(default_mean, default_std)
# ])
# opt.num_crop = 3
# else:
# if opt.data_aug == 'v1':
# transform_test = video.VideoGroupValTransform(size=opt.input_size, mean=default_mean, std=default_std)
# else:
# transform_test = video.VideoGroupValTransformV2(crop_size=(opt.input_size, opt.input_size), short_side=opt.input_size,
# mean=default_mean, std=default_std)
# opt.num_crop = 1
if not opt.deploy:
# get model
if opt.use_pretrained and len(opt.hashtag) > 0:
opt.use_pretrained = opt.hashtag
classes = opt.num_classes
model_name = opt.model
# Currently, these is no hashtag for int8 models.
if opt.quantized:
model_name += '_int8'
opt.use_pretrained = True
net = get_model(name=model_name,
nclass=classes,
pretrained=opt.use_pretrained,
num_segments=opt.num_segments,
num_crop=opt.num_crop)
net.cast(opt.dtype)
net.collect_params().reset_ctx(context)
resume_params = find_model_params(opt)
if opt.mode == 'hybrid':
net.hybridize(static_alloc=True, static_shape=True)
if resume_params is not '' and not opt.use_pretrained:
net.load_parameters(resume_params, ctx=context)
print('Pre-trained model %s is successfully loaded.' %
(resume_params))
else:
print(
'Pre-trained model is successfully loaded from the model zoo.')
else:
model_name = 'deploy'
net = mx.gluon.SymbolBlock.imports(
'{}-symbol.json'.format(opt.model_prefix), ['data'],
'{}-0000.params'.format(opt.model_prefix))
net.hybridize(static_alloc=True, static_shape=True)
print("Successfully loaded model {}".format(model_name))
# dummy data for benchmarking performance
if opt.benchmark:
benchmarking(opt, net, context)
sys.exit()
if opt.dataset == 'ucf101':
val_dataset = UCF101(setting=opt.val_list,
root=opt.data_dir,
train=False,
new_width=opt.new_width,
new_height=opt.new_height,
new_length=opt.new_length,
target_width=opt.input_size,
target_height=opt.input_size,
test_mode=True,
data_aug=opt.data_aug,
num_segments=opt.num_segments,
transform=transform_test)
elif opt.dataset == 'kinetics400':
val_dataset = Kinetics400(
setting=opt.val_list,
root=opt.data_dir,
train=False,
new_width=opt.new_width,
new_height=opt.new_height,
new_length=opt.new_length,
new_step=opt.new_step,
target_width=opt.input_size,
target_height=opt.input_size,
video_loader=opt.video_loader,
use_decord=opt.use_decord,
slowfast=opt.slowfast,
slow_temporal_stride=opt.slow_temporal_stride,
fast_temporal_stride=opt.fast_temporal_stride,
test_mode=True,
data_aug=opt.data_aug,
num_segments=opt.num_segments,
num_crop=opt.num_crop,
transform=transform_test)
elif opt.dataset == 'somethingsomethingv2':
val_dataset = SomethingSomethingV2(setting=opt.val_list,
root=opt.data_dir,
train=False,
new_width=opt.new_width,
new_height=opt.new_height,
new_length=opt.new_length,
new_step=opt.new_step,
target_width=opt.input_size,
target_height=opt.input_size,
video_loader=opt.video_loader,
use_decord=opt.use_decord,
data_aug=opt.data_aug,
num_segments=opt.num_segments,
transform=transform_test)
elif opt.dataset == 'hmdb51':
val_dataset = HMDB51(setting=opt.val_list,
root=opt.data_dir,
train=False,
new_width=opt.new_width,
new_height=opt.new_height,
new_length=opt.new_length,
new_step=opt.new_step,
target_width=opt.input_size,
target_height=opt.input_size,
video_loader=opt.video_loader,
use_decord=opt.use_decord,
data_aug=opt.data_aug,
num_segments=opt.num_segments,
transform=transform_test)
elif opt.dataset == 'custom':
transform_test = video.VideoGroupTrainTransform(
size=(224, 224),
scale_ratios=[1.0, 0.8],
mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
val_dataset = VideoClsCustom(
root=opt.val_data_dir,
setting=opt.val_list,
train=False,
new_length=32,
name_pattern='frame_%d.jpg',
transform=transform_test,
video_loader=False,
slowfast=True,
use_decord=True,
)
else:
logger.info('Dataset %s is not supported yet.' % (opt.dataset))
# val_data = gluon.data.DataLoader(val_dataset, batch_size=batch_size, shuffle=False, num_workers=num_workers,
# prefetch=int(opt.prefetch_ratio * num_workers), last_batch='discard')
val_data = gluon.data.DataLoader(val_dataset,
batch_size=batch_size,
shuffle=False,
num_workers=num_workers)
print('Load %d test samples in %d iterations.' %
(len(val_dataset), len(val_data)))
# calibrate FP32 model into INT8 model
if opt.calibration:
calibration(net, val_data, opt, context, logger)
sys.exit()
start_time = time.time()
acc_top1_val, acc_top5_val, true_labels, predicted_probabilities = test(
context, val_data, opt, net)
split_filename = os.path.split(opt.val_list)[1]
split = os.path.splitext(split_filename)[0]
#load encoder
encoder = joblib.load(opt.encoder_path)
#set-up metrics
classes = np.arange(len(encoder.classes_))
metrics_dict = {
"Accuracy": balanced_accuracy_score,
"Mcc": matthews_corrcoef,
"Precision_Avg": [precision_score, {
"average": "micro"
}],
"Recall_Avg": [recall_score, {
"average": "micro"
}],
"Precision_Class":
[precision_score, {
"labels": classes,
"average": None
}],
"Recall_Class": [recall_score, {
"labels": classes,
"average": None
}],
}
split_folder = os.path.join(opt.save_dir, split)
#set-up evaluator
evaluator = Evaluator_video(split_folder, encoder, true_labels,
predicted_probabilities, metrics_dict)
#compute report
report = get_split_report(evaluator)
#save report
save_results(report, split_folder)
print(f"Correctly process split {split}")
end_time = time.time()
print('Test accuracy: acc-top1=%f acc-top5=%f' %
(acc_top1_val * 100, acc_top5_val * 100))
print('Total evaluation time is %4.2f minutes' %
((end_time - start_time) / 60))
#########################################################################
# We first show an example that randomly reads 25 videos each time, randomly selects one frame per video and
# performs center cropping.
from gluoncv.data import UCF101
from mxnet.gluon.data import DataLoader
from mxnet.gluon.data.vision import transforms
from gluoncv.data.transforms import video
transform_train = transforms.Compose(
[video.VideoCenterCrop(size=224),
video.VideoToTensor()])
# Default location of the data is stored on ~/.mxnet/datasets/ucf101.
# You need to specify ``setting`` and ``root`` for UCF101 if you decoded the video frames into a different folder.
train_dataset = UCF101(train=True, transform=transform_train)
train_data = DataLoader(train_dataset, batch_size=25, shuffle=True)
#########################################################################
# We can see the shape of our loaded data as below. ``extra`` indicates if we select multiple crops or multiple segments
# from a video. Here, we only pick one frame per video, so the ``extra`` dimension is 1.
for x, y in train_data:
print('Video frame size (batch, extra, channel, height, width):', x.shape)
print('Video label:', y.shape)
break
#########################################################################
# Let's plot several training samples. index 0 is image, 1 is label
from gluoncv.utils import viz
viz.plot_image(train_dataset[7][0].squeeze().transpose(
(1, 2, 0)) * 255.0) # Basketball
def get_data_loader(opt, batch_size, num_workers, logger):
data_dir = opt.train_dir
scale_ratios = [1.0, 0.875, 0.75, 0.66]
input_size = opt.input_size
def batch_fn(batch, ctx):
data = split_and_load(batch[0],
ctx_list=ctx,
batch_axis=0,
even_split=False)
label = split_and_load(batch[1],
ctx_list=ctx,
batch_axis=0,
even_split=False)
return data, label
transform_train = video.VideoGroupTrainTransform(
size=(input_size, input_size),
scale_ratios=scale_ratios,
mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
transform_test = video.VideoGroupValTransform(size=input_size,
mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
if opt.dataset == 'ucf101':
train_dataset = UCF101(setting=opt.train_setting,
root=data_dir,
train=True,
new_width=opt.new_width,
new_height=opt.new_height,
new_length=opt.new_length_diff,
target_width=input_size,
target_height=input_size,
num_segments=opt.num_segments,
transform=transform_train)
val_dataset = UCF101(setting=opt.val_setting,
root=data_dir,
train=False,
new_width=opt.new_width,
new_height=opt.new_height,
new_length=opt.new_length_diff,
target_width=input_size,
target_height=input_size,
num_segments=opt.num_segments,
transform=transform_test)
else:
# logger.info('Dataset %s is not supported yet.' % (opt.dataset))
print('Dataset %s is not supported yet.' % (opt.dataset))
print('Load %d training samples and %d validation samples.' %
(len(train_dataset), len(val_dataset)))
# logger.info('Load %d training samples and %d validation samples.' % (len(train_dataset), len(val_dataset)))
train_data = gluon.data.DataLoader(train_dataset,
batch_size=batch_size,
shuffle=True,
num_workers=num_workers,
prefetch=int(opt.prefetch_ratio *
num_workers),
last_batch='rollover')
val_data = gluon.data.DataLoader(val_dataset,
batch_size=batch_size,
shuffle=False,
num_workers=num_workers,
prefetch=int(opt.prefetch_ratio *
num_workers),
last_batch='discard')
return train_data, val_data, batch_fn
