def __init__(self):
"""
:param self.train_frames: how many frame are used for training
:param self.num_of_vid: number of videos
:param self.extractor: pretrained model for feature extraction
:param self.G: what I want to training model
:param self.optimizerG: optimizer (default: Adam)
:param self.train_data: training data loader
:param self.device: gpu or cpu device(default: gpu)
"""
self.train_frames = args.train_frames
self.num_of_vid = args.num_of_vid
print("===> Building model")
# Model Setting
# prtrained model setting
self.extractor = resnet101(num_classes=400,
shortcut_type='B',
cardinality=32,
sample_size=args.crop_size,
sample_duration=args.train_frames)
# load pretrained model
# eval 모드를 무조건 해야됨!
weight = get_pretrain_weight(pretrain_path, self.extractor)
self.extractor.load_state_dict(weight)
self.extractor.eval()
self.in_channel = args.in_channel
self.G = BaseNet(self.in_channel * self.num_of_vid, self.train_frames,
self.num_of_vid)
self.num_stage = args.num_stage
# optimizer
self.optimizerG = optim.Adam(self.G.parameters(),
lr=args.lr,
betas=(0.9, 0.999),
eps=1e-8)
# loss
self.CE_loss = nn.CrossEntropyLoss()
self.L1_loss = nn.L1Loss()
self.L2_loss = nn.MSELoss()
self.Softmax = nn.Softmax()
# data
self.train_data = train_data_loader
# self.test_data = test_data_loader
# training
self.device = device
self.epochs = args.epochs
self.avg_G_loss_arr = []
self.checkpoint = args.checkpoint
self.checkpoint_set()
# cuda
if torch.cuda.is_available():
self.extractor.cuda()
self.G.cuda()
def __init__(self):
self.train_frames = args.train_frames
print("===> Building model")
# Model Setting
# prtrained model setting
self.extractor = resnet101(num_classes=400,
shortcut_type='B',
cardinality=32,
sample_size=args.crop_size,
sample_duration=args.train_frames)
# load pretrained model
# eval 모드를 무조건 해야됨!
weight = get_pretrain_weight(pretrain_path, self.extractor)
self.extractor.load_state_dict(weight)
self.extractor.eval()
self.G = BaseNet(16384 * 4 * 4 * 4, 10)
# optimizer
self.optimizerG = optim.Adam(self.G.parameters(),
lr=args.lr,
betas=(0.9, 0.999),
eps=1e-8)
# loss
self.BCE_loss = nn.BCELoss()
self.L1_loss = nn.L1Loss()
self.L2_loss = nn.MSELoss()
# data
self.train_data = train_data_loader
# self.test_data = test_data_loader
# training
self.device = device
self.epochs = args.epochs
self.avg_G_loss_arr = []
self.checkpoint = args.checkpoint
# cuda
if torch.cuda.is_available():
self.extractor.cuda()
self.G.cuda()
# model save
self.save_mname = args.save_model_name
# checkpoint root
make_dirs(log_path)
self.log_dir = log_path + f'/{self.save_mname}'
# self.load_check_dir =
self.save_check_dir = make_dirs(self.log_dir) + '/' + 'checkpoint.pkl'
# CSV logging system
self.CSVlogger = LogCSV(log_dir=self.log_dir +
f"/{self.save_mname}_log.csv",
header=['epoch', 'avg_G_Loss', 'accuracy'])
fm_use=args.fm_use)
# I3D.replace_logits(157) # Charades: 157 only used for load trained model
# -----------load pre_train model -------- ------------
Pre_tn_model = './model/pretred_mod/rgb_imagenet.pt'
pretrained_dict = torch.load(Pre_tn_model)
model_dict = I3D.state_dict()
pretrained_dict = {k: v for k, v in pretrained_dict.items() if k in model_dict}
model_dict.update(pretrained_dict)
I3D.load_state_dict(model_dict)
I3D.replace_logits(157) # used for load Knitices model
Model = I3D
train_trans = transforms.Compose([videotransform.IMG_resize(args.in_fm_sz, args.in_fm_sz),
videotransform.ToTensor()]) # contain normalize
# -----------load pre_train model --------------------
elif args.arch == "R3D":
R3D = resnext.resnet101(num_classes=400, shortcut_type='B', cardinality=32, sample_size=args.in_fm_sz,
sample_duration=args.fm_use, phase='train', Dense_out=False)
R3D.replace_logits(157)
# -----------load pre_train model --------------------
# Pre_tn_model = './model/pretred_mod/resnext-101-kinetics.pth'
Pre_tn_model = './model/pretred_mod/resnext-101-64f-kinetics.pth'
# Pre_tn_model = './model/pretred_mod/R3D025_Charades_model.pth'
pretrained_dict = torch.load(Pre_tn_model)
model_dict = R3D.state_dict()
pretrained_dict = {k: v for k, v in pretrained_dict.items() if k in model_dict}
model_dict.update(pretrained_dict)
R3D.load_state_dict(model_dict)
Model = R3D
train_trans = transforms.Compose([videotransform.IMG_resize(args.in_fm_sz, args.in_fm_sz),
videotransform.Normalize_R3D([114.7748, 107.7354, 99.4750], [1, 1, 1])])
# -----------load pre_train model --------------------