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(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
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=(input_size, input_size),
short_side=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)
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,
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)
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))