'temporal':
TemporalCenterCrop(config.sample_duration),
'target':
ClassLabel()
} # 测试时的数据转换
print('==> Loading validation dataset........')
val_data = get_validation_set(config, validation_transforms['spatial'],
validation_transforms['temporal'],
validation_transforms['target'])
data_loader = DataLoader(val_data,
config.batch_size,
shuffle=True,
num_workers=config.num_workers,
pin_memory=True)
model = model_factory.get_model(config)
model.cuda()
print("==> Loading existing model '{}' ".format('lrcn'))
model_info = torch.load(
os.path.join('model/checkpoints/',
'{}_save_best.pth'.format(config.dataset)))
model.load_state_dict(model_info['state_dict'])
model.eval()
print('==> Starting test.......')
steps_in_epoch = int(np.ceil(len(data_loader.dataset) / config.batch_size))
print(steps_in_epoch)
accuracies = np.zeros(steps_in_epoch, np.float32)
epoch_start_time = time.time()
for step, (clips, targets) in enumerate(data_loader):
start_time = time.time()
from transforms.temporal_transforms import TemporalRandomCrop
from transforms.spatial_transforms import Compose, Normalize, RandomHorizontalFlip, MultiScaleRandomCrop, ToTensor, CenterCrop
'''--------------------------------------配置和日志设置------------------------------------------'''
config = parse_opts() # 配置解析
config = prepare_output_dirs(config) # 输出文件夹初始化
config = init_cropping_scales(config) # 裁剪配置
config = set_lr_scheduling_policy(config) # 学习率配置
# 均值和方差设置
mean = [0.485, 0.456, 0.406]
std = [0.229, 0.224, 0.225]
print_config(config) # 输出配置文件
write_config(config, os.path.join(config.save_dir, 'config.json')) # 写入json文件
'''---------------------------------------初始化模型-------------------------------------------'''
os.environ["CUDA_VISIBLE_DEVICES"] = '0' # 运行环境
model = model_factory.get_model(config) # 获取模型以及需要更新的参数
# 设置转换函数
norm_method = Normalize(mean, std)
train_transforms = {
'spatial':
Compose([
MultiScaleRandomCrop(config.scales, config.spatial_size),
RandomHorizontalFlip(),
ToTensor(255), norm_method
]),
'temporal':
TemporalRandomCrop(config.sample_duration),
'target':
ClassLabel()
} # 训练时的数据转换,255表示将数据转换到0~1
validation_transforms = {
from transforms.spatial_transforms import Compose, Normalize, RandomHorizontalFlip, MultiScaleRandomCrop, ToTensor, CenterCrop
'''--------------------------------------配置和日志设置------------------------------------------'''
config = parse_opts() # 配置解析
config = prepare_output_dirs(config) # 输出文件夹初始化
config = init_cropping_scales(config) # 裁剪配置
config = set_lr_scheduling_policy(config) # 学习率配置
# 均值和方差设置
mean = [0.39608, 0.38182, 0.35067]
std = [0.15199, 0.14856, 0.15698]
print_config(config) # 输出配置文件
write_config(config, os.path.join(config.save_dir, 'config.json')) # 写入json文件
'''---------------------------------------初始化模型-------------------------------------------'''
device = torch.device(config.device) # 运行环境
model, parameters = model_factory.get_model(config) # 获取模型以及需要更新的参数
# 设置转换函数
norm_method = Normalize(mean, std)
train_transforms = {
'spatial':
Compose([
MultiScaleRandomCrop(config.scales, config.spatial_size),
RandomHorizontalFlip(),
ToTensor(config.norm_value), norm_method
]),
'temporal':
TemporalRandomCrop(config.sample_duration),
'target':
ClassLabel()
} # 训练时的数据转换
validation_transforms = {
if not config.no_tensorboard:
from tensorboardX import SummaryWriter
writer = SummaryWriter(log_dir=config.log_dir)
else:
writer = None
####################################################################
####################################################################
# Initialize model
device = torch.device(config.device)
#torch.backends.cudnn.enabled = False
# Returns the network instance (I3D, 3D-ResNet etc.)
# Note: this also restores the weights and optionally replaces final layer
model, parameters = model_factory.get_model(config)
print('#' * 60)
if config.model == 'i3d':
param_names = [p['name'] for p in parameters]
print('Parameters to train:')
print(param_names)
print('#' * 60)
####################################################################
####################################################################
# Setup of data transformations
if config.no_dataset_mean and config.no_dataset_std:
# Just zero-center and scale to unit std
print('Data normalization: no dataset mean, no dataset std')