def __init__(self, pretrained: bool = True, frame_num=4, resnet_3d_pretrained_model='r3d18_KM_200ep.pth'):
super().__init__()
resnet18_3d = generate_model(18)
if pretrained:
checkpoint = torch.load(resnet_3d_pretrained_model)
resnet18_3d.fc = nn.Linear(512, 1039)
resnet18_3d.load_state_dict(checkpoint['state_dict'])
self.resnet18_3d = resnet18_3d
resnet18_3d_last_dim = 512
cnn_last_dim = 1024
self.cnn_last = nn.Sequential(
nn.Linear(resnet18_3d_last_dim, cnn_last_dim),
nn.BatchNorm1d(cnn_last_dim),
nn.ReLU(inplace=True),
nn.Dropout()
)
comb_fc1_out = 512
self.comb_fc1 = nn.Sequential(
nn.Linear(cnn_last_dim * 2, comb_fc1_out),
# nn.Linear(resnet18_last_dim * 2, comb_fc1_out),
nn.BatchNorm1d(comb_fc1_out),
nn.ReLU(inplace=True),
nn.Dropout()
)
class_num = math.factorial(frame_num) // 2
self.combination_list = list(itertools.combinations(list(range(frame_num)), 2))
self.comb_fc2 = nn.Linear(512 * len(self.combination_list), class_num)
nn.init.kaiming_normal_(self.cnn_last[0].weight)
nn.init.kaiming_normal_(self.comb_fc1[0].weight)
nn.init.kaiming_normal_(self.comb_fc2.weight)
def __init__(self, pretrained: bool = True, resnet_3d_pretrained_model='r3d18_KM_200ep.pth'):
super().__init__()
resnet18_3d = generate_model(18)
if pretrained:
checkpoint = torch.load(resnet_3d_pretrained_model)
resnet18_3d.fc = nn.Linear(512, 1039)
resnet18_3d.load_state_dict(checkpoint['state_dict'])
self.resnet18_3d = resnet18_3d
resnet18_3d_last_dim = 512
# self.fc = nn.Linear(resnet18_3d_last_dim * 2, 2)
# nn.init.kaiming_normal_(self.fc.weight)
self.fc1 = nn.Linear(resnet18_3d_last_dim * 2, 4096)
self.relu1 = nn.ReLU(inplace=True)
self.fc2 = nn.Linear(4096, 4096)
self.relu2 = nn.ReLU(inplace=True)
self.fc3 = nn.Linear(4096, 2)
nn.init.kaiming_normal_(self.fc1.weight)
nn.init.kaiming_normal_(self.fc2.weight)
nn.init.kaiming_normal_(self.fc3.weight)
def __init__(self, class_num: int = 101, bidirectional: bool = True, pretrained: bool = True,
resnet_3d_pretrained_model='r3d18_KM_200ep.pth'):
super().__init__()
resnet18_3d = generate_model(18)
if pretrained:
checkpoint = torch.load(resnet_3d_pretrained_model)
resnet18_3d.fc = nn.Linear(512, 1039)
resnet18_3d.load_state_dict(checkpoint['state_dict'])
self.resnet18_3d = resnet18_3d
resnet18_3d_last_dim = 512
lstm_dim = 512
if bidirectional:
# self.rnn = nn.LSTM(resnet18_3d_last_dim, lstm_dim // 2, bidirectional=True, num_layers=2)
self.rnn = nn.GRU(resnet18_3d_last_dim, lstm_dim // 2, bidirectional=True, num_layers=2)
else:
# self.rnn = nn.LSTM(resnet18_3d_last_dim, lstm_dim, bidirectional=False, num_layers=2)
self.rnn = nn.GRU(resnet18_3d_last_dim, lstm_dim, bidirectional=False, num_layers=2)
self.fc = nn.Linear(lstm_dim, class_num)
nn.init.kaiming_normal_(self.fc.weight)
),
batch_size=batch_size,
shuffle=True)
test_loader = DataLoader(
VideoTestDataSet(
frame_num=frame_num,
path_list=generate_path_list(args, 'test'),
frame_interval=0
),
batch_size=batch_size,
shuffle=False)
train_iterate_len = len(train_loader)
test_iterate_len = len(test_loader)
# 初期設定
Net = generate_model(18)
if args.use_pretrained_model:
checkpoint = torch.load('r3d18_KM_200ep.pth')
Net.fc = nn.Linear(512, 1039)
Net.load_state_dict(checkpoint['state_dict'])
Net.fc = nn.Linear(512, args.class_num)
criterion = nn.CrossEntropyLoss() # Loss関数を定義
optimizer = torch.optim.Adam(Net.parameters(), lr=args.learning_rate) # 重み更新方法を定義
current_epoch = 0
# ログファイルの生成
if not args.no_reset_log_file:
with open(log_train_path, mode='w') as f:
f.write('epoch,loss,accuracy,time,learning_rate\n')
with open(log_test_path, mode='w') as f:
f.write('epoch,loss,accuracy,time,learning_rate\n')
def generate_model(opt):
assert opt.model in [
'resnet', 'resnet2p1d', 'preresnet', 'wideresnet', 'resnext',
'densenet'
]
if opt.model == 'resnet':
model = resnet_3d.generate_model(
model_depth=opt.model_depth, # 50
n_classes=opt.n_classes, # class num
n_input_channels=opt.n_input_channels, # 3
shortcut_type=opt.resnet_shortcut, # A or B default B
conv1_t_size=opt.conv1_t_size, # default 7
conv1_t_stride=opt.conv1_t_stride, # 步幅 default=1
no_max_pool=opt.no_max_pool,
widen_factor=opt.resnet_widen_factor) # default 1
# elif opt.model == 'resnet2p1d':
# model = resnet2p1d.generate_model(model_depth=opt.model_depth,
# n_classes=opt.n_classes,
# n_input_channels=opt.n_input_channels,
# shortcut_type=opt.resnet_shortcut,
# conv1_t_size=opt.conv1_t_size,
# conv1_t_stride=opt.conv1_t_stride,
# no_max_pool=opt.no_max_pool,
# widen_factor=opt.resnet_widen_factor)
# elif opt.model == 'wideresnet':
# model = wide_resnet.generate_model(
# model_depth=opt.model_depth,
# k=opt.wide_resnet_k,
# n_classes=opt.n_classes,
# n_input_channels=opt.n_input_channels,
# shortcut_type=opt.resnet_shortcut,
# conv1_t_size=opt.conv1_t_size,
# conv1_t_stride=opt.conv1_t_stride,
# no_max_pool=opt.no_max_pool)
# elif opt.model == 'resnext':
# model = resnext.generate_model(model_depth=opt.model_depth,
# cardinality=opt.resnext_cardinality,
# n_classes=opt.n_classes,
# n_input_channels=opt.n_input_channels,
# shortcut_type=opt.resnet_shortcut,
# conv1_t_size=opt.conv1_t_size,
# conv1_t_stride=opt.conv1_t_stride,
# no_max_pool=opt.no_max_pool)
# elif opt.model == 'preresnet':
# model = pre_act_resnet.generate_model(
# model_depth=opt.model_depth,
# n_classes=opt.n_classes,
# n_input_channels=opt.n_input_channels,
# shortcut_type=opt.resnet_shortcut,
# conv1_t_size=opt.conv1_t_size,
# conv1_t_stride=opt.conv1_t_stride,
# no_max_pool=opt.no_max_pool)
# elif opt.model == 'densenet':
# model = densenet.generate_model(model_depth=opt.model_depth,
# n_classes=opt.n_classes,
# n_input_channels=opt.n_input_channels,
# conv1_t_size=opt.conv1_t_size,
# conv1_t_stride=opt.conv1_t_stride,
# no_max_pool=opt.no_max_pool)
return model