def get_data_loader(opt, batch_size, num_workers, logger):
data_dir = opt.data_dir
normalize = video.VideoNormalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
scale_ratios = [1.0, 0.875, 0.75, 0.66]
input_size = opt.input_size
def batch_fn(batch, ctx):
if opt.num_segments > 1:
data = split_and_load(batch[0], ctx_list=ctx, batch_axis=0, even_split=False, multiplier=opt.num_segments)
else:
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 = transforms.Compose([
video.VideoMultiScaleCrop(size=(input_size, input_size), scale_ratios=scale_ratios),
video.VideoRandomHorizontalFlip(),
video.VideoToTensor(),
normalize
])
transform_test = transforms.Compose([
video.VideoCenterCrop(size=input_size),
video.VideoToTensor(),
normalize
])
train_dataset = ucf101.classification.UCF101(setting=opt.train_list, root=data_dir, train=True,
new_width=opt.new_width, new_height=opt.new_height,
target_width=input_size, target_height=input_size,
num_segments=opt.num_segments, transform=transform_train)
val_dataset = ucf101.classification.UCF101(setting=opt.val_list, root=data_dir, train=False,
new_width=opt.new_width, new_height=opt.new_height,
target_width=input_size, target_height=input_size,
num_segments=opt.num_segments, transform=transform_test)
logger.info('Load %d training samples and %d validation samples.' % (len(train_dataset), len(val_dataset)))
if opt.num_segments > 1:
train_data = gluon.data.DataLoader(train_dataset, batch_size=batch_size, shuffle=True, num_workers=num_workers, batchify_fn=tsn_mp_batchify_fn)
val_data = gluon.data.DataLoader(val_dataset, batch_size=batch_size, shuffle=False, num_workers=num_workers, batchify_fn=tsn_mp_batchify_fn)
else:
train_data = gluon.data.DataLoader(train_dataset, batch_size=batch_size, shuffle=True, num_workers=num_workers)
val_data = gluon.data.DataLoader(val_dataset, batch_size=batch_size, shuffle=False, num_workers=num_workers)
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))
# want to randomly crop a video sequence, you need to make sure all the video
# frames in this sequence undergo the same cropping process. We provide a
# new set of transformation functions, working with multiple images.
# Please checkout the `video.py <../../../gluoncv/data/transforms/video.py>`_ for more details.
# Most video data augmentation strategies used here are introduced in [Wang15]_.
transform_train = transforms.Compose([
# Fix the input video frames size as 256×340 and randomly sample the cropping width and height from
# {256,224,192,168}. After that, resize the cropped regions to 224 × 224.
video.VideoMultiScaleCrop(size=(224, 224),
scale_ratios=[1.0, 0.875, 0.75, 0.66]),
# Randomly flip the video frames horizontally
video.VideoRandomHorizontalFlip(),
# Transpose the video frames from height*width*num_channels to num_channels*height*width
# and map values from [0, 255] to [0,1]
video.VideoToTensor(),
# Normalize the video frames with mean and standard deviation calculated across all images
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 = 25
# Number of data loader workers
num_workers = 8
# Calculate effective total batch size
batch_size = per_device_batch_size * num_gpus
def main(logger):
opt = parse_args()
logger.info(opt)
gc.set_threshold(100, 5, 5)
if not os.path.exists(opt.save_dir):
os.makedirs(opt.save_dir)
# set env
if opt.gpu_id == -1:
context = mx.cpu()
else:
gpu_id = opt.gpu_id
context = mx.gpu(gpu_id)
# get data preprocess
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
# 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,
feat_ext=True,
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 != '' and not opt.use_pretrained:
net.load_parameters(opt.resume_params, ctx=context)
logger.info('Pre-trained model %s is successfully loaded.' %
(opt.resume_params))
else:
logger.info(
'Pre-trained model is successfully loaded from the model zoo.')
logger.info("Successfully built model {}".format(model_name))
# get data
anno_file = opt.data_list
f = open(anno_file, 'r')
data_list = f.readlines()
logger.info('Load %d video samples.' % len(data_list))
# build a pseudo dataset instance to use its children class methods
video_utils = VideoClsCustom(root=opt.data_dir,
setting=opt.data_list,
num_segments=opt.num_segments,
num_crop=opt.num_crop,
new_length=opt.new_length,
new_step=opt.new_step,
new_width=opt.new_width,
new_height=opt.new_height,
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,
lazy_init=True)
start_time = time.time()
for vid, vline in enumerate(data_list):
video_path = vline.split()[0]
video_name = video_path.split('/')[-1]
if opt.need_root:
video_path = os.path.join(opt.data_dir, video_path)
video_data = read_data(opt, video_path, transform_test, video_utils)
video_input = video_data.as_in_context(context)
video_feat = net(video_input.astype(opt.dtype, copy=False))
feat_file = '%s_%s_feat.npy' % (model_name, video_name)
np.save(os.path.join(opt.save_dir, feat_file), video_feat.asnumpy())
if vid > 0 and vid % opt.log_interval == 0:
logger.info('%04d/%04d is done' % (vid, len(data_list)))
end_time = time.time()
logger.info('Total feature extraction time is %4.2f minutes' %
((end_time - start_time) / 60))
def main(logger):
opt = parse_args()
makedirs(opt.save_dir)
filehandler = logging.FileHandler(
os.path.join(opt.save_dir, opt.logging_file))
streamhandler = logging.StreamHandler()
logger = logging.getLogger('')
logger.setLevel(logging.INFO)
logger.addHandler(filehandler)
logger.addHandler(streamhandler)
logger.info(opt)
gc.set_threshold(100, 5, 5)
# set env
gpu_id = opt.gpu_id
context = mx.gpu(gpu_id)
# get data preprocess
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
# 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)
logger.info('Pre-trained model %s is successfully loaded.' %
(opt.resume_params))
else:
logger.info(
'Pre-trained model is successfully loaded from the model zoo.')
logger.info("Successfully built model {}".format(model_name))
# get data
anno_file = opt.data_list
f = open(anno_file, 'r')
data_list = f.readlines()
logger.info('Load %d video samples.' % len(data_list))
start_time = time.time()
for vid, vline in enumerate(data_list):
video_path = vline.split()[0]
video_name = video_path.split('/')[-1]
if opt.need_root:
video_path = os.path.join(opt.data_dir, video_path)
video_data = read_data(opt, video_path, transform_test)
video_input = video_data.as_in_context(context)
pred = net(video_input.astype(opt.dtype, copy=False))
if opt.save_logits:
logits_file = '%s_%s_logits.npy' % (model_name, video_name)
np.save(os.path.join(opt.save_dir, logits_file), pred.asnumpy())
pred_label = np.argmax(pred.asnumpy())
if opt.save_preds:
preds_file = '%s_%s_preds.npy' % (model_name, video_name)
np.save(os.path.join(opt.save_dir, preds_file), pred_label)
logger.info('%04d/%04d: %s is predicted to class %d' %
(vid, len(data_list), video_name, pred_label))
end_time = time.time()
logger.info('Total inference 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))
def main():
opt = parse_args()
makedirs(opt.save_dir)
filehandler = logging.FileHandler(
os.path.join(opt.save_dir, opt.logging_file))
streamhandler = logging.StreamHandler()
logger = logging.getLogger('')
logger.setLevel(logging.INFO)
logger.addHandler(filehandler)
logger.addHandler(streamhandler)
logger.info(opt)
gc.set_threshold(100, 5, 5)
# set env
if opt.gpu_id == -1:
context = mx.cpu()
else:
gpu_id = opt.gpu_id
context = mx.gpu(gpu_id)
# get data preprocess
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
# 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 != '' and not opt.use_pretrained:
net.load_parameters(opt.resume_params, ctx=context)
logger.info('Pre-trained model %s is successfully loaded.' %
(opt.resume_params))
else:
logger.info(
'Pre-trained model is successfully loaded from the model zoo.')
logger.info("Successfully built model {}".format(model_name))
# get classes list, if we are using a pretrained network from the model_zoo
classes = None
if opt.use_pretrained:
if "kinetics400" in model_name:
classes = Kinetics400Attr().classes
elif "ucf101" in model_name:
classes = UCF101Attr().classes
elif "hmdb51" in model_name:
classes = HMDB51Attr().classes
elif "sthsth" in model_name:
classes = SomethingSomethingV2Attr().classes
# get data
anno_file = opt.data_list
f = open(anno_file, 'r')
data_list = f.readlines()
logger.info('Load %d video samples.' % len(data_list))
# build a pseudo dataset instance to use its children class methods
video_utils = VideoClsCustom(root=opt.data_dir,
setting=opt.data_list,
num_segments=opt.num_segments,
num_crop=opt.num_crop,
new_length=opt.new_length,
new_step=opt.new_step,
new_width=opt.new_width,
new_height=opt.new_height,
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,
lazy_init=True)
start_time = time.time()
for vid, vline in enumerate(data_list):
video_path = vline.split()[0]
video_name = video_path.split('/')[-1]
if opt.need_root:
video_path = os.path.join(opt.data_dir, video_path)
video_data = read_data(opt, video_path, transform_test, video_utils)
video_input = video_data.as_in_context(context)
pred = net(video_input.astype(opt.dtype, copy=False))
if opt.save_logits:
logits_file = '%s_%s_logits.npy' % (model_name, video_name)
np.save(os.path.join(opt.save_dir, logits_file), pred.asnumpy())
pred_label = np.argmax(pred.asnumpy())
if opt.save_preds:
preds_file = '%s_%s_preds.npy' % (model_name, video_name)
np.save(os.path.join(opt.save_dir, preds_file), pred_label)
# Try to report a text label instead of the number.
if classes:
pred_label = classes[pred_label]
logger.info('%04d/%04d: %s is predicted to class %s' %
(vid, len(data_list), video_name, pred_label))
end_time = time.time()
logger.info('Total inference time is %4.2f minutes' %
((end_time - start_time) / 60))
def main():
opt = parse_args()
# 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 model
classes = opt.num_classes
model_name = opt.model
net = get_model(name=model_name,
nclass=classes,
pretrained=True,
tsn=opt.use_tsn)
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 '':
net.load_parameters(opt.resume_params, ctx=context)
print('Pre-trained model %s is successfully loaded' % (opt.resume_params))
# get data
normalize = video.VideoNormalize([0.485, 0.456, 0.406],
[0.229, 0.224, 0.225])
transform_test = transforms.Compose(
[video.VideoTenCrop(opt.input_size),
video.VideoToTensor(), normalize])
val_dataset = ucf101.classification.UCF101(setting=opt.val_list,
root=opt.data_dir,
train=False,
new_width=opt.new_width,
new_height=opt.new_height,
target_width=opt.input_size,
target_height=opt.input_size,
test_mode=True,
num_segments=opt.num_segments,
transform=transform_test)
val_data = gluon.data.DataLoader(val_dataset,
batch_size=batch_size,
shuffle=False,
num_workers=num_workers)
print('Load %d test samples.' % len(val_dataset))
# start evaluation
acc_top1 = mx.metric.Accuracy()
acc_top5 = mx.metric.TopKAccuracy(5)
"""Common practice during evaluation is to evenly sample 25 frames from a single video, and then perform 10-crop data augmentation.
This leads to 250 samples per video (750 channels). If this is too large to fit into one GPU, we can split it into multiple data bacthes.
`num_split_frames` has to be multiples of 3.
"""
num_data_batches = 10
num_split_frames = int(750 / num_data_batches)
def test(ctx, val_data):
acc_top1.reset()
acc_top5.reset()
for i, batch in enumerate(val_data):
outputs = []
for seg_id in range(num_data_batches):
bs = seg_id * num_split_frames
be = (seg_id + 1) * num_split_frames
new_batch = [batch[0][:, bs:be, :, :], batch[1]]
data, label = batch_fn(new_batch, ctx)
for gpu_id, X in enumerate(data):
X_reshaped = X.reshape(
(-1, 3, opt.input_size, opt.input_size))
pred = net(X_reshaped.astype(opt.dtype, copy=False))
if seg_id == 0:
outputs.append(pred)
else:
outputs[gpu_id] = nd.concat(outputs[gpu_id],
pred,
dim=0)
# Perform the mean operation on 250 samples of each video
for gpu_id, out in enumerate(outputs):
outputs[gpu_id] = nd.expand_dims(out.mean(axis=0), axis=0)
acc_top1.update(label, outputs)
acc_top5.update(label, outputs)
if i > 0 and i % opt.log_interval == 0:
print('%04d/%04d is done' % (i, len(val_data)))
_, top1 = acc_top1.get()
_, top5 = acc_top5.get()
return (top1, top5)
start_time = time.time()
acc_top1_val, acc_top5_val = test(context, val_data)
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():
opt = parse_args()
print(opt)
# 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 model
classes = opt.num_classes
model_name = opt.model
net = get_model(name=model_name,
nclass=classes,
pretrained=opt.use_pretrained)
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.')
# 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])
])
else:
transform_test = transforms.Compose([
video.VideoCenterCrop(opt.input_size),
video.VideoToTensor(),
video.VideoNormalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])
if opt.dataset == 'ucf101':
val_dataset = ucf101.classification.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.classification.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,
test_mode=True,
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))
print('Load %d test samples.' % len(val_dataset))
# start evaluation
acc_top1 = mx.metric.Accuracy()
acc_top5 = mx.metric.TopKAccuracy(5)
"""Common practice during evaluation is to evenly sample 25 frames from a single video, and then perform 10-crop data augmentation.
This leads to 250 samples per video (750 channels). If this is too large to fit into one GPU, we can split it into multiple data batches.
`num_data_batches` has to be set to a value as long as `num_split_frames` is multiples of 3.
For example, when `num_data_batches` is set to 10, `num_split_frames` will be 750/10=75, which is multiples of 3.
If you have enough GPU memory and prefer faster evaluation speed, you can set `num_data_batches` to 1.
"""
num_data_batches = 10
if opt.ten_crop:
num_frames = opt.num_segments * 10
else:
num_frames = opt.num_segments
num_split_frames = int(num_frames * 3 / num_data_batches)
def test(ctx, val_data):
acc_top1.reset()
acc_top5.reset()
for i, batch in enumerate(val_data):
outputs = []
for seg_id in range(num_data_batches):
bs = seg_id * num_split_frames
be = (seg_id + 1) * num_split_frames
if opt.input_5d:
new_batch = [batch[0][:, bs:be, :, :, :], batch[1]]
else:
new_batch = [batch[0][:, bs:be, :, :], batch[1]]
data, label = batch_fn(new_batch, ctx)
for gpu_id, X in enumerate(data):
if opt.input_5d:
new_X = X.reshape((-1, 3, opt.new_length,
opt.input_size, opt.input_size))
else:
new_X = X.reshape(
(-1, 3, opt.input_size, opt.input_size))
pred = net(new_X)
if seg_id == 0:
outputs.append(pred)
else:
outputs[gpu_id] = nd.concat(outputs[gpu_id],
pred,
dim=0)
# Perform the mean operation on 'num_frames' samples of each video
for gpu_id, out in enumerate(outputs):
outputs[gpu_id] = nd.expand_dims(out.mean(axis=0), axis=0)
acc_top1.update(label, outputs)
acc_top5.update(label, outputs)
mx.ndarray.waitall()
_, cur_top1 = acc_top1.get()
_, cur_top5 = acc_top5.get()
if i > 0 and i % opt.log_interval == 0:
print('%04d/%04d is done: acc-top1=%f acc-top5=%f' %
(i, len(val_data), cur_top1 * 100, cur_top5 * 100))
_, top1 = acc_top1.get()
_, top5 = acc_top5.get()
return (top1, top5)
start_time = time.time()
acc_top1_val, acc_top5_val = test(context, val_data)
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))