def get_loaders(opt):
""" Make dataloaders for train and validation sets
"""
# train loader
norm_method = Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
spatial_transform = Compose([
Scale((opt.sample_size, opt.sample_size)),
Resize(256),
CenterCrop(224),
ToTensor(), norm_method
])
temporal_transform = TemporalRandomCrop(25)
target_transform = ClassLabel()
training_data = get_training_set(opt, spatial_transform,
temporal_transform, target_transform)
train_loader = torch.utils.data.DataLoader(training_data,
batch_size=opt.batch_size,
shuffle=True,
num_workers=opt.num_workers,
pin_memory=True)
# validation loader
target_transform = ClassLabel()
temporal_transform = LoopPadding(25)
validation_data = get_validation_set(opt, spatial_transform,
temporal_transform, target_transform)
val_loader = torch.utils.data.DataLoader(validation_data,
batch_size=opt.batch_size,
shuffle=False,
num_workers=opt.num_workers,
pin_memory=True)
return train_loader, val_loader
def main():
opt = parse_opts()
if opt.root_path != '':
opt.video_path = os.path.join(opt.root_path, opt.video_path)
opt.annotation_path = os.path.join(opt.root_path, opt.annotation_path)
opt.result_path = os.path.join(opt.root_path, opt.result_path)
if opt.resume_path:
opt.resume_path = os.path.join(opt.root_path, opt.resume_path)
if opt.pretrain_path:
opt.pretrain_path = os.path.join(opt.root_path, opt.pretrain_path)
opt.scales = [opt.initial_scale]
for i in range(1, opt.n_scales):
opt.scales.append(opt.scales[-1] * opt.scale_step)
opt.arch = '{}-{}'.format(opt.model, opt.model_depth)
opt.mean = get_mean(opt.norm_value, dataset=opt.mean_dataset)
opt.std = get_std(opt.norm_value)
print(opt)
with open(os.path.join(opt.result_path, 'opts.json'), 'w') as opt_file:
json.dump(vars(opt), opt_file)
torch.manual_seed(opt.manual_seed)
model, _ = generate_model(opt)
if opt.no_mean_norm and not opt.std_norm:
norm_method = Normalize([0, 0, 0], [1, 1, 1])
elif not opt.std_norm:
norm_method = Normalize(opt.mean, [1, 1, 1])
else:
norm_method = Normalize(opt.mean, opt.std)
spatial_transform = Compose([
Scale(int(opt.sample_size / opt.scale_in_test)),
CornerCrop(opt.sample_size, opt.crop_position_in_test),
ToTensor(opt.norm_value), norm_method
])
temporal_transform = LoopPadding(opt.sample_duration)
target_transform = VideoID()
validation_data = get_validation_set(opt, spatial_transform,
temporal_transform, target_transform)
val_loader = torch.utils.data.DataLoader(validation_data,
batch_size=opt.batch_size,
shuffle=False,
num_workers=opt.n_threads,
pin_memory=True)
dump(val_loader, model, opt, validation_data.class_names)
def get_loaders(opt): """ Make dataloaders for train and validation sets """ # train loader opt.mean = get_mean(opt.norm_value, dataset=opt.mean_dataset) if opt.no_mean_norm and not opt.std_norm: norm_method = Normalize([0, 0, 0], [1, 1, 1]) elif not opt.std_norm: norm_method = Normalize(opt.mean, [1, 1, 1]) else: norm_method = Normalize(opt.mean, opt.std) spatial_transform = Compose([ # crop_method, Scale((opt.sample_size, opt.sample_size)), # RandomHorizontalFlip(), ToTensor(opt.norm_value), norm_method ]) temporal_transform = TemporalRandomCrop(16) target_transform = ClassLabel() training_data = get_training_set(opt, spatial_transform, temporal_transform, target_transform) train_loader = torch.utils.data.DataLoader( training_data, batch_size=opt.batch_size, shuffle=True, num_workers=opt.num_workers, pin_memory=True) # validation loader spatial_transform = Compose([ Scale((opt.sample_size, opt.sample_size)), # CenterCrop(opt.sample_size), ToTensor(opt.norm_value), norm_method ]) target_transform = ClassLabel() temporal_transform = LoopPadding(16) validation_data = get_validation_set( opt, spatial_transform, temporal_transform, target_transform) val_loader = torch.utils.data.DataLoader( validation_data, batch_size=opt.batch_size, shuffle=False, num_workers=opt.num_workers, pin_memory=True) return train_loader, val_loader
def get_valinfo(opt, norm_method):
spatial_transform = Compose([
Scale(opt.sample_size),
CenterCrop(opt.sample_size),
ToTensor(opt.norm_value), norm_method
])
temporal_transform = TemporalCenterCrop(opt.sample_duration)
target_transform = ClassLabel()
validation_data = get_validation_set(opt, spatial_transform,
temporal_transform, target_transform)
val_loader = torch.utils.data.DataLoader(validation_data,
batch_size=opt.batch_size,
shuffle=False,
num_workers=opt.n_threads,
pin_memory=True)
val_logger = Logger(os.path.join(opt.result_path, 'val.log'),
['epoch', 'loss', 'prec1', 'prec5'])
return validation_data, val_loader, val_logger
def main():
opt = parse_opts()
print(opt)
seed = 0
random.seed(seed)
np.random.seed(seed)
torch.manual_seed(seed)
# torch.backends.cudnn.deterministic = True
# torch.backends.cudnn.benchmark = False
# CUDA for PyTorch
use_cuda = torch.cuda.is_available()
device = torch.device(f"cuda:{opt.gpu}" if use_cuda else "cpu")
train_transform = transforms.Compose([
#transforms.RandomCrop(32, padding=3),
transforms.Resize((224, 224)),
transforms.RandomHorizontalFlip(),
transforms.ColorJitter(brightness=[0.5, 1]),
# transforms.RandomRotation(180),
GaussianNoise(0.5),
# transforms.RandomRotation(10),
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
])
test_transform = transforms.Compose([
#transforms.RandomCrop(32, padding=3),
transforms.Resize((224, 224)),
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
])
training_data = get_training_set(opt, train_transform)
training_data = ConcatDataset(training_data)
validation_data = get_validation_set(opt, test_transform)
validation_data = ConcatDataset(validation_data)
train_loader = torch.utils.data.DataLoader(training_data,
batch_size=opt.batch_size,
shuffle=True,
num_workers=0)
val_loader = torch.utils.data.DataLoader(validation_data,
batch_size=opt.batch_size,
shuffle=True,
num_workers=0)
print(f'Number of training examples: {len(train_loader.dataset)}')
print(f'Number of validation examples: {len(val_loader.dataset)}')
# tensorboard
summary_writer = tensorboardX.SummaryWriter(log_dir='tf_logs')
# define model
model = ResidualNet("ImageNet", opt.depth, opt.num_classes, "CBAM")
if opt.resume_path:
checkpoint = torch.load(opt.resume_path)
model.load_state_dict(checkpoint['model_state_dict'])
epoch = checkpoint['epoch']
print("Model Restored from Epoch {}".format(epoch))
opt.start_epoch = epoch + 1
model.to(device)
criterion = nn.CrossEntropyLoss()
optimizer = optim.Adam(model.parameters(), weight_decay=opt.wt_decay)
scheduler = lr_scheduler.ReduceLROnPlateau(optimizer,
'min',
patience=opt.lr_patience)
if opt.resume_path:
checkpoint = torch.load(opt.resume_path)
optimizer.load_state_dict(checkpoint['optimizer_state_dict'])
scheduler.load_state_dict(checkpoint['scheduler_state_dict'])
th = 100000
# start training
for epoch in range(opt.start_epoch, opt.epochs + 1):
# train, test model
train_loss, train_acc = train_epoch(model, train_loader, criterion,
optimizer, device, opt)
val_loss, val_acc = val_epoch(model, val_loader, criterion, device)
scheduler.step(val_loss)
lr = optimizer.param_groups[0]['lr']
# saving weights to checkpoint
if (epoch) % opt.save_interval == 0:
# write summary
summary_writer.add_scalar('losses/train_loss',
train_loss,
global_step=epoch)
summary_writer.add_scalar('losses/val_loss',
val_loss,
global_step=epoch)
summary_writer.add_scalar('acc/train_acc',
train_acc,
global_step=epoch)
summary_writer.add_scalar('acc/val_acc',
val_acc,
global_step=epoch)
summary_writer.add_scalar('lr_rate', lr, global_step=epoch)
state = {
'epoch': epoch,
'model_state_dict': model.state_dict(),
'optimizer_state_dict': optimizer.state_dict(),
'scheduler_state_dict': scheduler.state_dict()
}
if val_loss < th:
torch.save(
state,
os.path.join('./snapshots', f'{opt.dataset}_model.pth'))
print("Epoch {} model saved!\n".format(epoch))
th = val_loss
scheduler['fusion'] = lr_scheduler.MultiStepLR(
optimizer['fusion'],
[int(0.5 * opt.n_epochs) - 1,
int(0.75 * opt.n_epochs) - 1])
if not opt.no_val:
spatial_transform = Compose([
Scale([opt.sample_size, opt.sample_size]),
ToTensor(opt.norm_value), norm_method
])
temporal_transform = TemporalCenterCrop(opt.sample_duration)
target_transform = Label()
opt.dataset = 'diem'
validation_data_diem = get_validation_set(opt, spatial_transform,
temporal_transform,
target_transform)
opt.dataset = 'coutrot1'
validation_data_coutrot1 = get_validation_set(opt, spatial_transform,
temporal_transform,
target_transform)
opt.dataset = 'coutrot2'
validation_data_coutrot2 = get_validation_set(opt, spatial_transform,
temporal_transform,
target_transform)
opt.dataset = 'summe'
validation_data_summe = get_validation_set(opt, spatial_transform,
temporal_transform,
target_transform)
opt.dataset = 'etmd'
validation_data_etmd = get_validation_set(opt, spatial_transform,
weight_decay=opt.weight_decay)
scheduler = lr_scheduler.MultiStepLR(optimizer, milestones=[5,15], gamma=0.1)
if not opt.no_val:
spatial_transform = Compose([
Scale_longerside(opt.sample_size),
CenterCrop(opt.sample_size),
ToTensor(opt.norm_value), norm_method
])
target_transform = ClassLabel()
val_loader = {}
for j in range(0, len(opt.val_dataset)):
validation_data = get_validation_set(opt.val_dataset[j], spatial_transform, target_transform, opt)
val_loader[j] = torch.utils.data.DataLoader(
validation_data,
batch_size=opt.val_batch_size,
shuffle=False,
num_workers=opt.n_threads,
pin_memory=True)
if opt.validate_policy == '2stream':
val_logger = {}
for j in range(0, len(val_loader)):
val_logger[j] = Logger(
os.path.join(opt.result_path, 'val_'+opt.val_dataset[j]+'.log'),
['epoch', 'OBOA', 'MAE', 'MAE_std', 'MAEP', 'MAEN'])
if opt.nesterov:
dampening = 0
else:
dampening = opt.dampening
optimizer = optim.SGD(parameters,
lr=opt.learning_rate,
momentum=opt.momentum,
dampening=dampening,
weight_decay=opt.weight_decay,
nesterov=opt.nesterov)
scheduler = lr_scheduler.ReduceLROnPlateau(optimizer,
'min',
patience=opt.lr_patience)
if not opt.no_val:
validation_data = get_validation_set(opt)
val_loader = torch.utils.data.DataLoader(validation_data,
batch_size=opt.batch_size,
shuffle=False,
num_workers=opt.n_threads,
pin_memory=True)
val_logger = Logger(os.path.join(opt.result_path, 'val.log'),
['epoch', 'loss', 'acc'])
if opt.resume_path:
print('loading checkpoint {}'.format(opt.resume_path))
checkpoint = torch.load(opt.resume_path)
assert opt.arch == checkpoint['arch']
opt.begin_epoch = checkpoint['epoch']
model.load_state_dict(checkpoint['state_dict'])
def train_and_evaluate(net_name="aein_net", epochs = 250, train_batch_size=64, learning_rate=1e-3, optimizer="adam"):
train_set_size = 680
val_test_set_size = 340
input_width = input_height = 224
channel = 3
output_size = 17
train_set_1 = dataset.get_train_set(1)
train_set_1 = train_set_1.shuffle(buffer_size=10000)
train_set_1 = train_set_1.batch(train_batch_size)
validation_set_1 = dataset.get_validation_set(1)
validation_set_1 = validation_set_1.batch(68)
test_set_1 = dataset.get_test_set(1)
test_set_1 = test_set_1.batch(68)
X = tf.placeholder(tf.float32, [None, input_height, input_width, channel])
if net_name == 'aein_net':
y_pred = aein_net.aein_net(X)
elif net_name == 'alex_net':
y_pred = alex_net.alex_net(X)
elif net_name == 'vgg_16':
y_pred = vgg_16.vgg_16(X)
else:
y_pred = aein_net.aein_net(X)
y_true = tf.placeholder(tf.float32, [None, output_size])
loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=y_pred, labels=y_true))
if optimizer == 'adam':
optimizer = tf.train.AdamOptimizer(learning_rate)
elif optimizer == 'momentum':
optimizer = tf.train.MomentumOptimizer(learning_rate, 0.9)
elif optimizer == 'gradient_descent':
optimizer = tf.train.GradientDescentOptimizer(learning_rate)
elif optimizer == 'rmsprop':
optimizer = tf.train.RMSPropOptimizer(learning_rate)
else:
optimizer = tf.train.AdamOptimizer(learning_rate)
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
with tf.control_dependencies(update_ops):
train_op = optimizer.minimize(loss)
itr_trn_1 = train_set_1.make_initializable_iterator()
next_element_trn_1 = itr_trn_1.get_next()
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
start_train_time = time.time()
accuracies = []
losses = []
epochs_arr = []
epochs_10_arr = []
print("Start to train classifier, model: " + net_name)
for epoch in range(epochs):
print("------------------ Epoch %d starts ------------------" % (epoch + 1))
sess.run(itr_trn_1.initializer)
batch_idx = 1
while True:
try:
X_batch, Y_batch = sess.run(next_element_trn_1)
last_train_op_time = time.time()
_, l = sess.run([train_op, loss], feed_dict={X: X_batch, y_true: Y_batch})
losses.append(l)
epochs_arr.append(epoch)
print("epoch: %d, batch: %d, loss: %f, time cost: %f" % (
epoch + 1, batch_idx, l, time.time() - last_train_op_time))
batch_idx += 1
except tf.errors.OutOfRangeError: # this epoch ends
break
if (epoch + 1) % 10 == 0:
print()
print("------------------ evaluating accuracy on validation set ------------------")
accuracy = utils.evaluate(validation_set_1, sess, X, y_true, y_pred)
print("accuracy: %f" % accuracy)
print()
accuracies.append(accuracy)
epochs_10_arr.append(epoch + 1)
print("Classifier has been trained, total time: %f" % (time.time() - start_train_time))
print()
print("------------------ evaluating accuracy on test set ------------------")
print("accuracy: %f" % utils.evaluate(test_set_1, sess, X, y_true, y_pred))
print()
utils.save_result(net_name, epochs_arr, losses, epochs_10_arr, accuracies)
spatial_transforms[tokens[0]] = Compose([
CenterCrop(opt.sample_size),
ToTensor(opt.norm_value), norm_method
])
annotateData = pd.read_csv(opt.annotation_file, sep=',', header=0)
keys = annotateData[annotateData.Dataset == 'Test']['Location']
values = annotateData[annotateData.Dataset == 'Test']['MeanID']
annotationDictionary = dict(zip(keys, values))
temporal_transform = TemporalCenterCrop(opt.sample_duration)
#temporal_transform = LoopPadding(opt.sample_duration)
target_transform = ClassLabel()
validation_data = get_validation_set(opt, spatial_transforms,
temporal_transform,
target_transform,
annotationDictionary)
val_loader = torch.utils.data.DataLoader(validation_data,
batch_size=opt.batch_size,
shuffle=False,
num_workers=opt.n_threads,
pin_memory=True)
val_logger = Logger(os.path.join(opt.result_path, 'val.log'),
['epoch', 'loss', 'acc'])
if opt.resume_path:
print('loading checkpoint {}'.format(opt.resume_path))
checkpoint = torch.load(opt.resume_path)
assert opt.arch == checkpoint['arch']
opt.begin_epoch = checkpoint['epoch']
def create_dataloader(args):
if args.root_path != '':
args.video_path = os.path.join(args.root_path, args.video_path)
args.annotation_path = os.path.join(args.root_path,
args.annotation_path)
args.result_path = os.path.join(args.root_path, args.result_path)
if args.resume_path:
args.resume_path = os.path.join(args.root_path, args.resume_path)
if args.pretrain_path:
# args.pretrain_path = os.path.join(args.root_path, args.pretrain_path)
args.pretrain_path = os.path.abspath(args.pretrain_path)
args.scales = [args.initial_scale]
for i in range(1, args.n_scales):
args.scales.append(args.scales[-1] * args.scale_step)
args.mean = get_mean(args.norm_value, dataset=args.mean_dataset)
args.std = get_std(args.norm_value)
if args.no_mean_norm and not args.std_norm:
norm_method = Normalize([0, 0, 0], [1, 1, 1])
elif not args.std_norm:
norm_method = Normalize(args.mean, [1, 1, 1])
else:
norm_method = Normalize(args.mean, args.std)
assert args.train_crop in ['random', 'corner', 'center']
if args.train_crop == 'random':
crop_method = MultiScaleRandomCrop(args.scales, args.sample_size)
elif args.train_crop == 'corner':
crop_method = MultiScaleCornerCrop(args.scales, args.sample_size)
elif args.train_crop == 'center':
crop_method = MultiScaleCornerCrop(args.scales,
args.sample_size,
crop_positions=['c'])
spatial_transform = Compose([
crop_method,
RandomHorizontalFlip(),
ToTensor(args.norm_value), norm_method
])
temporal_transform = TemporalRandomCrop(args.sample_duration)
target_transform = ClassLabel()
training_data = get_training_set(args, spatial_transform,
temporal_transform, target_transform)
train_loader = torch.utils.data.DataLoader(training_data,
batch_size=args.batch_size,
shuffle=True,
num_workers=args.n_threads,
pin_memory=True)
spatial_transform = Compose([
# Scale(args.sample_size),
Scale(int(args.sample_size / args.scale_in_test)),
# CenterCrop(args.sample_size),
CornerCrop(args.sample_size, args.crop_position_in_test),
ToTensor(args.norm_value),
norm_method
])
temporal_transform = TemporalCenterCrop(args.sample_duration)
target_transform = ClassLabel()
validation_data = get_validation_set(args, spatial_transform,
temporal_transform, target_transform)
val_loader = torch.utils.data.DataLoader(validation_data,
batch_size=1,
shuffle=False,
num_workers=args.n_threads,
pin_memory=True)
return train_loader, val_loader
def main():
# parameters
# ???should we adjust the learning rate during the training process? ???
learn_rate = 1e-4
num_epochs = 600 # 最大迭代次数
max_patience = 60 # 停止训练的参数
result_rp = '../result/model/'
exp_name = 'P3D_saliency'
batch_size = 3
n_threads = 16
# 'Temporal duration of inputs'
sample_duration = 16
# the data for devison
norm_value = 255
# Height and width of inputs
sample_size = 224
# Number of validation samples for each activity
n_val_samples = 3
video_path = '/data1/guoxi/p3d_floder/resized_dataset/dataset/'
reference_path = '/data1/guoxi/p3d_floder/resized_dataset/reference_dataset/'
# video_path = '/data1/guoxi/p3d_floder/resized_dataset_for_test/dataset/'
# reference_path = '/data1/guoxi/p3d_floder/resized_dataset_for_test/reference_dataset/'
mean = [0.485, 0.456, 0.406]
std = [0.229, 0.224, 0.225]
norm_method = Normalize(mean, std)
RandomCrop_transform = Compose([ # 图像预处理一般用Compose把多个步骤整合到一起
joint_transforms.JointRandomCrop(
224
), # 在一个随机的位置进行裁剪 'JpegImageFile' object does not support indexing
])
RandomHorizontalFlip_transform = Compose([ # 图像预处理一般用Compose把多个步骤整合到一起
joint_transforms.JointRandomHorizontalFlip(), # 以0.5的概率水平翻转给定的PIL图像
])
RandomErase_transform = Compose(
[joint_transforms.RandomErase(probability=0.5, sh=0.4, r1=0.3)])
spatial_transform = Compose([ToTensor(norm_value=norm_value),
norm_method]) # 空间变换
target_transform = Compose([ToTensor(norm_value=norm_value)])
opt = [video_path, sample_duration]
train_data = dataset.get_training_set(opt, reference_path,
spatial_transform, target_transform,
RandomCrop_transform,
RandomHorizontalFlip_transform,
RandomErase_transform)
train_loader = torch.utils.data.DataLoader(
train_data,
batch_size=batch_size,
shuffle=True,
num_workers=n_threads,
pin_memory=True
) # 使用多进程加载的进程数 是否将数据保存在pin memory区，pin memory中的数据转到GPU会快一些
validation_data = dataset.get_validation_set(opt, reference_path,
spatial_transform,
target_transform, None, None,
None)
val_loader = torch.utils.data.DataLoader(validation_data,
batch_size=batch_size,
shuffle=False,
num_workers=n_threads,
pin_memory=True)
inputs, targets = next(iter(train_loader))
print('inputs.size(), inputs.min(), inputs.max()', inputs.size(),
inputs.min(), inputs.max())
print('targets.size(), targets.min(), targets.max():', targets.size(),
targets.min(), targets.max())
# every time we load weights, which may be slow
model_cla = network_cla.P3D199(pretrained=True, num_classes=400)
cla_dict = model_cla.state_dict()
model = network_seg.P3D199()
# model.apply(utils.weights_init)
seg_dict = model.state_dict()
pretrained_dict = {k: v for k, v in cla_dict.items() if k in seg_dict}
seg_dict.update(pretrained_dict)
model.load_state_dict(seg_dict)
model.cuda()
model = nn.DataParallel(model)
commen_layers = [
'conv1_custom', 'bn1', 'relu', 'maxpool', 'maxpool_2', 'layer1',
'layer2', 'layer3'
]
# seperate layers, to set different lr
param_exist = []
param_add = []
for k, (name, module) in enumerate(model.named_children()):
# existing layers
if name in commen_layers:
# print('existing layer: ', name)
for param in module.parameters():
param_exist.append(param)
# adding layers
else:
# print('adding layer: ', name)
for param in module.parameters():
param_add.append(param)
print(' + Number of params: {}'.format(
sum([p.data.nelement() for p in model.parameters()])))
optimizer = optim.Adam([{
'params': param_exist,
'lr': learn_rate * 0.1
}, {
'params': param_add
}])
criterion = nn.BCELoss().cuda()
exp_dir = result_rp + exp_name
##!!! existing directory will be removed
if os.path.exists(exp_dir):
shutil.rmtree(exp_dir)
exp = experiment.Experiment(exp_name, result_rp)
exp.init()
for epoch in range(num_epochs):
since = time.time()
### Train ###
trn_loss = utils.train(model, train_loader, optimizer, criterion)
print('Epoch {:d}: Train - Loss: {:.4f}'.format(epoch, trn_loss))
time_elapsed = time.time() - since
print('Train Time {:.0f}m {:.0f}s'.format(time_elapsed // 60,
time_elapsed % 60))
### Test ###
val_loss = utils.test(model, val_loader, criterion)
print('Val - Loss: {:.4f}'.format(val_loss))
time_elapsed = time.time() - since
print('Total Time {:.0f}m {:.0f}s\n'.format(time_elapsed // 60,
time_elapsed % 60))
### Save Metrics ###
exp.save_history('train', trn_loss)
exp.save_history('val', val_loss)
### Checkpoint ###
exp.save_weights(model, trn_loss, val_loss)
exp.save_optimizer(optimizer, val_loss)
## Early Stopping ##
if (epoch - exp.best_val_loss_epoch) > max_patience:
print(("Early stopping at epoch %d since no " +
"better loss found since epoch %.3").format(
epoch, exp.best_val_loss))
break
exp.epoch += 1
def main():
resnet_in = generate_model(opt)
resnet_in.module.fc = Identity()
model = ReNet34(resnet_in, encode_length=encode_length)
if opt.no_mean_norm and not opt.std_norm:
norm_method = Normalize([0, 0, 0], [1, 1, 1])
elif not opt.std_norm:
norm_method = Normalize(opt.mean, [1, 1, 1])
else:
norm_method = Normalize(opt.mean, opt.std)
if not opt.no_train:
assert opt.train_crop in ['random', 'corner', 'center']
if opt.train_crop == 'random':
crop_method = MultiScaleRandomCrop(opt.scales, opt.sample_size)
elif opt.train_crop == 'corner':
crop_method = MultiScaleCornerCrop(opt.scales, opt.sample_size)
elif opt.train_crop == 'center':
crop_method = MultiScaleCornerCrop(opt.scales,
opt.sample_size,
crop_positions=['c'])
## train loader
spatial_transform = Compose([
crop_method,
RandomHorizontalFlip(),
ToTensor(opt.norm_value), norm_method
])
temporal_transform = TemporalRandomCrop(opt.sample_duration)
target_transform = ClassLabel()
training_data = get_training_set(opt, spatial_transform,
temporal_transform, target_transform)
train_loader = torch.utils.data.DataLoader(training_data,
batch_size=opt.batch_size,
shuffle=True,
num_workers=opt.n_threads,
pin_memory=True)
## test loader
spatial_transform = Compose([
Scale(int(opt.sample_size / opt.scale_in_test)),
CornerCrop(opt.sample_size, opt.crop_position_in_test),
ToTensor(opt.norm_value), norm_method
])
temporal_transform = LoopPadding(opt.sample_duration)
target_transform = ClassLabel()
test_data = get_test_set(opt, spatial_transform, temporal_transform,
target_transform)
test_loader = torch.utils.data.DataLoader(test_data,
batch_size=opt.batch_size,
shuffle=False,
num_workers=opt.n_threads,
pin_memory=True)
## Database loader
spatial_transform = Compose([
Scale(int(opt.sample_size / opt.scale_in_test)),
CornerCrop(opt.sample_size, opt.crop_position_in_test),
ToTensor(opt.norm_value), norm_method
])
temporal_transform = LoopPadding(opt.sample_duration)
target_transform = ClassLabel()
validation_data = get_validation_set(opt, spatial_transform,
temporal_transform,
target_transform)
database_loader = torch.utils.data.DataLoader(
validation_data,
batch_size=opt.batch_size,
shuffle=False,
num_workers=opt.n_threads,
pin_memory=True)
if opt.nesterov:
dampening = 0
else:
dampening = opt.dampening
optimizer = optim.SGD(model.parameters(),
lr=opt.learning_rate,
momentum=opt.momentum,
dampening=dampening,
weight_decay=opt.weight_decay,
nesterov=opt.nesterov)
scheduler = lr_scheduler.ReduceLROnPlateau(optimizer,
'min',
patience=opt.lr_patience)
if opt.resume_path:
print('loading checkpoint {}'.format(opt.resume_path))
checkpoint = torch.load(opt.resume_path)
assert opt.arch == checkpoint['arch']
opt.begin_epoch = checkpoint['epoch']
model.load_state_dict(checkpoint['state_dict'])
if not opt.no_train:
optimizer.load_state_dict(checkpoint['optimizer'])
for state in optimizer.state.values():
for k, v in state.items():
if torch.is_tensor(v):
state[k] = v.cuda()
print('run')
for epoch in range(opt.begin_epoch, opt.n_epochs + 1):
model.cuda().train()
for i, (images, labels) in enumerate(train_loader):
images = Variable(images.cuda())
labels = Variable(labels.cuda().long())
# Forward + Backward + Optimize
optimizer.zero_grad()
x, _, b = model(images)
target_b = F.cosine_similarity(b[:int(labels.size(0) / 2)],
b[int(labels.size(0) / 2):])
target_x = F.cosine_similarity(x[:int(labels.size(0) / 2)],
x[int(labels.size(0) / 2):])
loss = F.mse_loss(target_b, target_x)
loss.backward()
optimizer.step()
scheduler.step()
# Test the Model
if (epoch + 1) % 10 == 0:
model.eval()
retrievalB, retrievalL, queryB, queryL = compress(
database_loader, test_loader, model)
result_map = calculate_top_map(qB=queryB,
rB=retrievalB,
queryL=queryL,
retrievalL=retrievalL,
topk=100)
print('--------mAP@100: {}--------'.format(result_map))
def objective(trial):
opt = parse_opts()
if trial:
opt.weight_decay = trial.suggest_uniform('weight_decay', 0.01, 0.1)
opt.learning_rate = trial.suggest_uniform('learning_rate', 1 - 5,
1 - 4)
if opt.root_path != '':
opt.video_path = os.path.join(opt.root_path, opt.video_path)
opt.annotation_path = os.path.join(opt.root_path, opt.annotation_path)
opt.result_path = os.path.join(opt.root_path, opt.result_path)
if opt.resume_path:
opt.resume_path = os.path.join(opt.root_path, opt.resume_path)
if opt.pretrain_path:
opt.pretrain_path = os.path.join(opt.root_path, opt.pretrain_path)
opt.scales = [opt.initial_scale]
for i in range(1, opt.n_scales):
opt.scales.append(opt.scales[-1] * opt.scale_step)
opt.arch = '{}-{}'.format(opt.model, opt.model_depth)
opt.mean = get_mean(opt.norm_value, dataset=opt.mean_dataset)
opt.std = get_std(opt.norm_value)
print(opt)
with open(os.path.join(opt.result_path, 'opts.json'), 'w') as opt_file:
json.dump(vars(opt), opt_file)
torch.manual_seed(opt.manual_seed)
model, parameters = generate_model(opt)
print(model)
criterion = nn.CrossEntropyLoss()
if not opt.no_cuda:
criterion = criterion.cuda()
if opt.no_mean_norm and not opt.std_norm:
norm_method = Normalize([0, 0, 0], [1, 1, 1])
elif not opt.std_norm:
norm_method = Normalize(opt.mean, [1, 1, 1])
else:
norm_method = Normalize(opt.mean, opt.std)
# norm_method = Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
if not opt.no_train:
assert opt.train_crop in ['random', 'corner', 'center']
if opt.train_crop == 'random':
crop_method = MultiScaleRandomCrop(opt.scales, opt.sample_size)
elif opt.train_crop == 'corner':
crop_method = MultiScaleCornerCrop(opt.scales, opt.sample_size)
elif opt.train_crop == 'center':
crop_method = MultiScaleCornerCrop(opt.scales,
opt.sample_size,
crop_positions=['c'])
spatial_transform = Compose([
crop_method,
RandomHorizontalFlip(),
ToTensor(opt.norm_value), norm_method
])
temporal_transform = TemporalRandomCrop(opt.sample_duration)
target_transform = ClassLabel()
training_data = get_training_set(opt, spatial_transform,
temporal_transform, target_transform)
train_loader = torch.utils.data.DataLoader(
training_data,
batch_size=opt.batch_size,
# sampler option is mutually exclusive with shuffle
shuffle=False,
sampler=ImbalancedDatasetSampler(training_data),
num_workers=opt.n_threads,
pin_memory=True)
train_logger = Logger(os.path.join(opt.result_path, 'train.log'),
['epoch', 'loss', 'acc', 'lr'])
train_batch_logger = Logger(
os.path.join(opt.result_path, 'train_batch.log'),
['epoch', 'batch', 'iter', 'loss', 'acc', 'lr'])
optimizer = optim.Adam(parameters,
lr=opt.learning_rate,
weight_decay=opt.weight_decay)
scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer,
verbose=True,
factor=0.1**0.5)
if not opt.no_val:
spatial_transform = Compose([
Scale(opt.sample_size),
CenterCrop(opt.sample_size),
ToTensor(opt.norm_value), norm_method
])
temporal_transform = LoopPadding(opt.sample_duration)
target_transform = ClassLabel()
validation_data = get_validation_set(opt, spatial_transform,
temporal_transform,
target_transform)
val_loader = torch.utils.data.DataLoader(
validation_data,
batch_size=opt.batch_size,
shuffle=False,
sampler=ImbalancedDatasetSampler(validation_data),
num_workers=opt.n_threads,
pin_memory=True)
val_logger = Logger(os.path.join(opt.result_path, 'val.log'),
['epoch', 'loss', 'acc'])
if opt.resume_path:
print('loading checkpoint {}'.format(opt.resume_path))
checkpoint = torch.load(opt.resume_path)
assert opt.arch == checkpoint['arch']
opt.begin_epoch = checkpoint['epoch']
model.load_state_dict(checkpoint['state_dict'])
if not opt.no_train:
optimizer.load_state_dict(checkpoint['optimizer'])
print('run')
writer = SummaryWriter(
comment=
f"_wd{opt.weight_decay}_lr{opt.learning_rate}_ft_begin{opt.ft_begin_index}_pretrain{not opt.pretrain_path == ''}"
)
for i in range(opt.begin_epoch, opt.n_epochs + 1):
if not opt.no_train:
epoch, losses_avg, accuracies_avg = train_epoch(
i, train_loader, model, criterion, optimizer, opt,
train_logger, train_batch_logger)
writer.add_scalar('loss/train', losses_avg, epoch)
writer.add_scalar('acc/train', accuracies_avg, epoch)
if not opt.no_val:
epoch, val_losses_avg, val_accuracies_avg = val_epoch(
i, val_loader, model, criterion, opt, val_logger)
writer.add_scalar('loss/val', val_losses_avg, epoch)
writer.add_scalar('acc/val', val_accuracies_avg, epoch)
if not opt.no_train and not opt.no_val:
scheduler.step(val_losses_avg)
print('=' * 100)
if opt.test:
spatial_transform = Compose([
Scale(int(opt.sample_size / opt.scale_in_test)),
CornerCrop(opt.sample_size, opt.crop_position_in_test),
ToTensor(opt.norm_value), norm_method
])
temporal_transform = LoopPadding(opt.sample_duration)
target_transform = VideoID()
test_data = get_test_set(opt, spatial_transform, temporal_transform,
target_transform)
test_loader = torch.utils.data.DataLoader(test_data,
batch_size=opt.batch_size,
shuffle=False,
num_workers=opt.n_threads,
pin_memory=True)
test.test(test_loader, model, opt, test_data.class_names)
writer.close()
return val_losses_avg
label_length = combined_dataset.labeled_length
unlabel_length = combined_dataset.unlabeled_length
assert label_length + unlabel_length == len(
combined_dataset), 'Fatal Error!'
print('\nTotal: {} labeled samples and {} unlabeled samples'.format(
label_length, unlabel_length))
val_sp_transform = Compose([
Scale(args.sample_size),
CenterCrop(args.sample_size),
ToTensor(args.norm_value),
norm_method,
])
val_tp_transform = TemporalCenterCrop(args.sample_duration,
args.downsample)
val_dataset = get_validation_set(args, val_sp_transform, val_tp_transform,
target_transform[0])
# generate loader
label_indices = torch.from_numpy(np.arange(label_length))
unlabel_indices = torch.from_numpy(
np.arange(label_length, len(combined_dataset)))
"""A batchsampler yielding a batch of indices with 4 parts,
1&2 from primary indices(kinetics),
3&4 from secondary indices(ucf),
1 and 2 are same indices (they are same samples), but each is augmented differently (spatially and temporally)
3 and 4 same as above
batch_size should be 64 so the 1&3 will be fed to student model (with batchsize 128)
2&4 will be fed to teacher model (with batchsize 128)
"""
sampler = FourStreamSampler(unlabel_indices, label_indices,
args.batch_size, args.batch_size)
def main_worker(gpu, ngpus_per_node, opt, test_results=None):
opt.gpu = gpu
# suppress printing if not master
if opt.multiprocessing_distributed and opt.gpu != 0:
def print_pass(*args):
pass
builtins.print = print_pass
if opt.gpu is not None:
print("Use GPU: {} for training".format(opt.gpu))
if opt.distributed:
if opt.dist_url == "env://" and opt.rank == -1:
opt.rank = int(os.environ["RANK"])
if opt.multiprocessing_distributed:
# For multiprocessing distributed training, rank needs to be the
# global rank among all the processes
opt.rank = opt.rank * ngpus_per_node + gpu
dist.init_process_group(backend=opt.dist_backend,
init_method=opt.dist_url,
world_size=opt.world_size,
rank=opt.rank)
opt.batch_size = int(opt.batch_size / ngpus_per_node)
opt.n_threads = int(
(opt.n_threads + ngpus_per_node - 1) / ngpus_per_node)
if opt.rank % ngpus_per_node == 0:
if not os.path.exists(opt.result_path):
os.makedirs(opt.result_path)
opt.arch = '{}-{}'.format(opt.model, opt.model_depth)
with open(os.path.join(opt.result_path, 'opts.json'), 'w') as opt_file:
json.dump(vars(opt), opt_file)
if not opt.no_train:
training_data = get_training_set(opt)
opt.N_data = len(training_data)
if opt.distributed:
train_sampler = torch.utils.data.distributed.DistributedSampler(
training_data)
else:
train_sampler = None
train_loader = torch.utils.data.DataLoader(
training_data,
batch_size=opt.batch_size,
shuffle=(train_sampler is None),
num_workers=opt.n_threads,
pin_memory=True,
drop_last=True,
sampler=train_sampler)
model, parameters = generate_model(opt)
if opt.phase == 'finetuning':
criterion = nn.CrossEntropyLoss().cuda(opt.gpu)
elif opt.phase == 'pretraining':
criterion = NCECriterion(len(training_data)).cuda(opt.gpu)
else:
raise NotImplementedError('not implement {} phase'.format(
opt.phase))
train_logger = Logger(
os.path.join(opt.result_path, 'train.log.rank{}'.format(opt.rank)),
['epoch', 'loss', 'acc', 'lr'])
train_batch_logger = Logger(
os.path.join(opt.result_path,
'train_batch.log.{}'.format(opt.rank)),
['epoch', 'batch', 'iter', 'loss', 'acc', 'lr'])
optimizer = optim.SGD(parameters,
lr=opt.learning_rate,
momentum=opt.momentum,
weight_decay=opt.weight_decay)
if opt.phase == 'finetuning':
scheduler = lr_scheduler.ReduceLROnPlateau(
optimizer,
'max',
patience=opt.lr_patience,
min_lr=1e-6,
factor=opt.lr_factor)
if not opt.no_val:
validation_data = get_validation_set(opt)
if opt.distributed:
val_sampler = torch.utils.data.distributed.DistributedSampler(
validation_data)
else:
val_sampler = None
val_loader = torch.utils.data.DataLoader(validation_data,
batch_size=opt.batch_size,
shuffle=(val_sampler is None),
num_workers=opt.n_threads,
pin_memory=True,
drop_last=True,
sampler=val_sampler)
val_logger = Logger(
os.path.join(opt.result_path, 'val.log.rank{}'.format(opt.rank)),
['epoch', 'acc1', 'acc5'] if opt.phase == 'finetuning' else
['epoch', 'recall@1', 'recall@10'])
if opt.test:
model, parameters = generate_model(opt)
test_data = get_test_set(opt)
idx_to_labels = test_data.get_idx_to_label()
if opt.distributed:
test_sampler = torch.utils.data.distributed.DistributedSampler(
test_data, shuffle=False)
else:
test_sampler = None
test_loader = torch.utils.data.DataLoader(
test_data,
batch_size=opt.batch_size,
shuffle=(test_sampler is None),
num_workers=opt.n_threads,
pin_memory=True,
drop_last=False,
sampler=test_sampler)
if opt.resume_path:
print('==>loading checkpoint {}'.format(opt.resume_path))
if opt.gpu is None:
checkpoint = torch.load(opt.resume_path)
else:
# Map model to be loaded to specified single gpu.
loc = 'cuda:{}'.format(opt.gpu)
checkpoint = torch.load(opt.resume_path, map_location=loc)
opt.begin_epoch = checkpoint['epoch']
model.load_state_dict(checkpoint['state_dict'])
if not opt.no_train:
optimizer.load_state_dict(checkpoint['optimizer'])
else:
opt.begin_epoch = 1
torch.backends.cudnn.benchmark = True
if opt.rank % ngpus_per_node == 0:
summary_writer = SummaryWriter(log_dir=opt.result_path)
else:
summary_writer = None
for i in range(opt.begin_epoch, opt.n_epochs + 1):
if not opt.no_train:
if opt.distributed:
train_sampler.set_epoch(i)
train_epoch(i, train_loader, model, criterion, optimizer, opt,
train_logger, train_batch_logger, summary_writer)
if not opt.no_val:
if opt.phase == 'finetuning':
val_acc = val_finetune_epoch(i, val_loader, model, opt,
val_logger, summary_writer)
elif opt.phase == 'pretraining':
val_acc = val_pretrain_epoch(i, val_loader, model, opt,
val_logger, summary_writer)
if not opt.no_train and not opt.no_val:
if opt.phase == 'finetuning':
scheduler.step(val_acc)
elif opt.phase == 'pretraining':
adjuest_learning_rate(optimizer, i, opt)
if opt.test:
test.test(test_loader, model, opt, idx_to_labels, test_results)
if opt.multiprocessing_distributed and opt.gpu == 0:
result_json = {}
finish_procs = 0
while (finish_procs < ngpus_per_node):
rst = test_results.get()
if rst == -1:
finish_procs += 1
else:
result_json[rst[0]] = rst[1]
with open(
os.path.join(opt.result_path,
'{}.json'.format(opt.test_subset)), 'w') as f:
json.dump({'results': result_json}, f)
'min',
patience=opt.lr_patience)
if not opt.no_val:
spatial_transform = Compose([
Scale(opt.sample_size),
CenterCrop(opt.sample_size),
ToTensor(opt.norm_value), norm_method
])
#temporal_transform = LoopPadding(opt.sample_duration)
temporal_transform = TemporalCenterCrop(opt.sample_duration,
opt.downsample)
target_transform = ClassLabel()
validation_data1 = get_validation_set(opt, 'all', opt.view,
opt.image_type,
spatial_transform,
temporal_transform,
target_transform)
#validation_data2 = get_validation_set(
# opt, 'N', spatial_transform, temporal_transform, target_transform)
#print ("@@@@@@@@@@@@@@@@@@@2",validation_data1.__len__())
val_loader1 = torch.utils.data.DataLoader(validation_data1,
batch_size=10,
shuffle=False,
num_workers=opt.n_threads,
pin_memory=True,
drop_last=True)
#val_loader2 = torch.utils.data.DataLoader(
# validation_data2,
# batch_size=8,
# shuffle=True,
def main():
opt = parse_opts()
print(opt)
seed = 0
random.seed(seed)
np.random.seed(seed)
torch.manual_seed(seed)
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
# CUDA for PyTorch
use_cuda = torch.cuda.is_available()
device = torch.device(f"cuda:{opt.gpu}" if use_cuda else "cpu")
train_transform = transforms.Compose([
#transforms.RandomCrop(32, padding=3),
transforms.Resize((opt.img_H, opt.img_W)),
transforms.RandomHorizontalFlip(),
transforms.RandomRotation(10),
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[
0.229, 0.224, 0.225])
])
test_transform = transforms.Compose([
#transforms.RandomCrop(32, padding=3),
transforms.Resize((opt.img_H, opt.img_W)),
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[
0.229, 0.224, 0.225])
])
training_data = get_training_set(opt, train_transform)
validation_data = get_validation_set(opt, test_transform)
n_train_examples = int(len(training_data)*0.8)
n_valid_examples = len(training_data) - n_train_examples
# split data
training_data, validation_data = torch.utils.data.random_split(training_data, [n_train_examples, n_valid_examples])
train_loader = torch.utils.data.DataLoader(training_data,
batch_size=opt.batch_size,
shuffle=True,
num_workers=1)
val_loader = torch.utils.data.DataLoader(validation_data,
batch_size=opt.batch_size,
shuffle=True,
num_workers=1)
print(f'Number of training examples: {len(train_loader.dataset)}')
print(f'Number of validation examples: {len(val_loader.dataset)}')
# tensorboard
summary_writer = tensorboardX.SummaryWriter(log_dir='tf_logs')
# define model
model = resnet18(num_classes=opt.num_classes)
# if torch.cuda.device_count() > 1:
# print("Let's use", torch.cuda.device_count(), "GPUs!")
# model = nn.DataParallel(model)
model = model.to(device)
if opt.nesterov:
dampening = 0
else:
dampening = opt.dampening
#define optimizer and criterion
# optimizer = optim.Adam(model.parameters())
# optimizer = optim.SGD(
# model.parameters(),
# lr=opt.learning_rate,
# momentum=opt.momentum,
# dampening=dampening,
# weight_decay=opt.weight_decay,
# nesterov=opt.nesterov)
# scheduler = lr_scheduler.ReduceLROnPlateau(optimizer, 'min', patience=opt.lr_patience)
# criterion = nn.CrossEntropyLoss()
# define optimizer and criterion
optimizer = optim.Adam(model.parameters())
# loss function
criterion = BCEWithLogitsLoss()
# resume model, optimizer if already exists
if opt.resume_path:
checkpoint = torch.load(opt.resume_path)
model.load_state_dict(checkpoint['model_state_dict'])
optimizer.load_state_dict(checkpoint['optimizer_state_dict'])
epoch = checkpoint['epoch']
print("Model Restored from Epoch {}".format(epoch))
start_epoch = epoch + 1
else:
start_epoch = 1
# start training
#th = 10000
for epoch in range(start_epoch, opt.epochs+1):
val_loss, val_mAP = val_epoch(model, val_loader, criterion, device, opt)
ToTensor(opt.norm_value) #, norm_method
])
else:
normalize = Normalize(mean=input_config['mean'],
std=input_config['std'])
spatial_transform = Compose([
Scale((256, 256)),
CenterCrop(opt.sample_size),
ToTensor(),
# normalize,
])
temporal_transform = LoopPadding(opt.sample_duration)
opt.compressed = 1
opt.residual_only = 1
residual_data = get_validation_set(opt, spatial_transform,
temporal_transform)
opt.residual_only = 0
compressed_data = get_validation_set(opt, spatial_transform,
temporal_transform)
opt.compressed = 0 ###super important this will take data from normal rgb frames
standard_data = get_validation_set(opt, spatial_transform,
temporal_transform)
residual_data_loader = torch.utils.data.DataLoader(
compressed_data,
batch_size=opt.batch_size,
shuffle=False,
num_workers=opt.n_threads,
pin_memory=True)
compressed_data_loader = torch.utils.data.DataLoader(
compressed_data,
weight_decay=opt.weight_decay,
nesterov=opt.nesterov)
scheduler = lr_scheduler.ReduceLROnPlateau(optimizer,
'min',
patience=opt.lr_patience)
if not opt.no_val:
spatial_transform = Compose([
Scale(opt.sample_size),
CenterCrop(opt.sample_size),
ToTensor(opt.norm_value), norm_method
])
temporal_transform = LoopPadding(opt.sample_duration)
target_transform = ClassLabel()
validation_data = get_validation_set(opt,
spatial_transform,
temporal_transform,
target_transform,
image_type=opt.image_type)
val_loader = torch.utils.data.DataLoader(validation_data,
batch_size=opt.batch_size,
shuffle=False,
num_workers=opt.n_threads,
pin_memory=True)
val_logger = Logger(os.path.join(opt.result_path, 'val.log'),
['epoch', 'loss', 'acc'])
if opt.resume_path:
print('loading checkpoint {}'.format(opt.resume_path))
checkpoint = torch.load(opt.resume_path)
assert opt.arch == checkpoint['arch']
[15, 25, 40, 45, 50, 55, 60],
gamma=0.1)
if not opt.no_val:
##--------------------------------------------------------------------------------------------
if opt.model == 'I3D':
spatial_transform = Compose([
Scale((256, 256)),
CenterCrop(224),
ToTensor(opt.norm_value), norm_method
])
temporal_transform = LoopPadding(0)
target_transform = ClassLabel()
validation_data = get_validation_set(opt, spatial_transform,
temporal_transform,
target_transform)
val_loader = torch.utils.data.DataLoader(validation_data,
batch_size=1,
shuffle=False,
num_workers=opt.n_threads,
pin_memory=True)
elif opt.model == 'resnet_50':
spatial_transform = Compose([
Scale(256),
CenterCrop(256),
ToTensor(opt.norm_value), norm_method
])
temporal_transform = TemporalCenterCrop(opt.sample_duration, 1)
target_transform = ClassLabel()
validation_data = get_validation_set(opt, spatial_transform,
def main():
opt = parse_opts()
ecd_name, cls_name = opt.model_name.split('-')
ecd_model = get_encoder_net(ecd_name)
cls_model = get_end_net(cls_name)
cfg.encoder_model = ecd_name
cfg.classification_model = cls_name
if opt.debug:
cfg.debug = opt.debug
else:
if opt.tensorboard == 'TEST':
cfg.tensorboard = opt.model_name
else:
cfg.tensorboard = opt.tensorboard
cfg.flag = opt.flag
model = cls_model(cfg,
encoder=CNNencoder(
cfg,
ecd_model(pretrained=True, path=opt.encoder_model)))
cfg.video_path = os.path.join(cfg.root_path, cfg.video_path)
cfg.annotation_path = os.path.join(cfg.root_path, cfg.annotation_path)
cfg.list_all_member()
torch.manual_seed(cfg.manual_seed)
print('##########################################')
print('####### model 仅支持单GPU')
print('##########################################')
model = model.cuda()
print(model)
criterion = nn.CrossEntropyLoss()
if cfg.cuda:
criterion = criterion.cuda()
norm_method = Normalize([0, 0, 0], [1, 1, 1])
print('##########################################')
print('####### train')
print('##########################################')
assert cfg.train_crop in ['random', 'corner', 'center']
if cfg.train_crop == 'random':
crop_method = (cfg.scales, cfg.sample_size)
elif cfg.train_crop == 'corner':
crop_method = MultiScaleCornerCrop(cfg.scales, cfg.sample_size)
elif cfg.train_crop == 'center':
crop_method = MultiScaleCornerCrop(cfg.scales,
cfg.sample_size,
crop_positions=['c'])
spatial_transform = Compose([
crop_method,
RandomHorizontalFlip(),
ToTensor(cfg.norm_value), norm_method
])
temporal_transform = TemporalRandomCrop(cfg.sample_duration)
target_transform = ClassLabel()
training_data = get_training_set(cfg, spatial_transform,
temporal_transform, target_transform)
train_loader = torch.utils.data.DataLoader(training_data,
batch_size=cfg.batch_size,
shuffle=True,
num_workers=cfg.n_threads,
drop_last=False,
pin_memory=True)
optimizer = model.get_optimizer(lr1=cfg.lr, lr2=cfg.lr2)
scheduler = lr_scheduler.ReduceLROnPlateau(optimizer,
'min',
patience=cfg.lr_patience)
print('##########################################')
print('####### val')
print('##########################################')
spatial_transform = Compose([
Scale(cfg.sample_size),
CenterCrop(cfg.sample_size),
ToTensor(cfg.norm_value), norm_method
])
temporal_transform = LoopPadding(cfg.sample_duration)
target_transform = ClassLabel()
validation_data = get_validation_set(cfg, spatial_transform,
temporal_transform, target_transform)
val_loader = torch.utils.data.DataLoader(validation_data,
batch_size=cfg.batch_size,
shuffle=False,
num_workers=cfg.n_threads,
drop_last=False,
pin_memory=True)
print('##########################################')
print('####### run')
print('##########################################')
if cfg.debug:
logger = None
else:
path = get_log_dir(cfg.logdir, name=cfg.tensorboard, flag=cfg.flag)
logger = Logger(logdir=path)
cfg.save_config(path)
for i in range(cfg.begin_epoch, cfg.n_epochs + 1):
train_epoch(i, train_loader, model, criterion, optimizer, cfg, logger)
validation_loss = val_epoch(i, val_loader, model, criterion, cfg,
logger)
scheduler.step(validation_loss)
scheduler = lr_scheduler.ReduceLROnPlateau(optimizer,
'min',
patience=opt.lr_patience)
if not opt.no_val:
spatial_transform = Compose([
Scale(opt.sample_size),
CenterCrop(opt.sample_size),
ToTensor(opt.norm_value), norm_method
])
#temporal_transform = LoopPadding(opt.sample_duration)
temporal_transform = TemporalCenterCrop(opt.sample_duration,
opt.downsample)
target_transform = ClassLabel()
validation_data1 = get_validation_set(opt, 'all', 'top',
spatial_transform,
temporal_transform,
target_transform)
#validation_data2 = get_validation_set(
# opt, 'N', spatial_transform, temporal_transform, target_transform)
#print ("@@@@@@@@@@@@@@@@@@@2",validation_data1.__len__())
val_loader1 = torch.utils.data.DataLoader(validation_data1,
batch_size=8,
shuffle=False,
num_workers=opt.n_threads,
pin_memory=True,
drop_last=True)
#val_loader2 = torch.utils.data.DataLoader(
# validation_data2,
# batch_size=8,
# shuffle=True,
# num_workers=opt.n_threads,
lr=opt.learning_rate,
momentum=opt.momentum,
dampening=dampening,
weight_decay=opt.weight_decay,
nesterov=opt.nesterov)
scheduler = lr_scheduler.ReduceLROnPlateau(
optimizer, 'min', patience=opt.lr_patience)
if not opt.no_val:
spatial_transform = Compose([
Scale(opt.sample_size),
CenterCrop(opt.sample_size),
ToTensor(opt.norm_value), norm_method
])
temporal_transform = LoopPadding(opt.sample_duration)
target_transform = ClassLabel()
validation_data = get_validation_set(
opt, spatial_transform, temporal_transform, target_transform)
val_loader = torch.utils.data.DataLoader(
validation_data,
batch_size=opt.batch_size,
shuffle=False,
num_workers=opt.n_threads,
pin_memory=True)
val_logger = Logger(
os.path.join(opt.result_path, 'val.log'), ['epoch', 'loss', 'acc'])
if opt.resume_path:
print('loading checkpoint {}'.format(opt.resume_path))
checkpoint = torch.load(opt.resume_path)
assert opt.arch == checkpoint['arch']
opt.begin_epoch = checkpoint['epoch']
dampening=dampening,
weight_decay=opt.weight_decay,
nesterov=opt.nesterov)
scheduler = lr_scheduler.ReduceLROnPlateau(optimizer,
'min',
patience=opt.lr_patience)
if not opt.no_val:
spatial_transform = Compose([
Scale(opt.sample_size),
CenterCrop(opt.sample_size),
ToTensor(opt.norm_value), norm_method
])
temporal_transform = LoopPadding(opt.sample_duration)
target_transform = ClassLabel()
validation_data = get_validation_set(opt, spatial_transform,
temporal_transform,
target_transform)
val_loader = torch.utils.data.DataLoader(validation_data,
batch_size=opt.batch_size,
shuffle=False,
num_workers=opt.n_threads,
pin_memory=True)
val_logger = Logger(os.path.join(opt.result_path, 'val.log'),
['epoch', 'loss', 'acc'])
if opt.resume_path:
print('loading checkpoint {}') # .format(opt.resume_path))
checkpoint = torch.load(opt.resume_path)
assert opt.arch == checkpoint['arch']
opt.begin_epoch = checkpoint['epoch']
def main():
opt = parse_opts()
# Path configurations
opt.annotation_path = os.path.join(opt.annotation_directory,
opt.annotation_path)
save_result_dir_name = \
os.path.join(opt.result_path,
get_prefix() + '_{}{}_{}_epochs'.format(opt.model, opt.model_depth, opt.n_epochs))
if not os.path.exists(save_result_dir_name):
os.mkdir(save_result_dir_name)
opt.result_path = os.path.join(opt.result_path, save_result_dir_name)
# For data generator
opt.scales = [opt.initial_scale]
for epoch in range(1, opt.n_scales):
opt.scales.append(opt.scales[-1] * opt.scale_step)
opt.arch = '{}-{}'.format(opt.model, opt.model_depth)
# Model
model, parameters = generate_model(opt)
# print(model)
# Loss function
criterion = nn.CrossEntropyLoss()
if not opt.no_cuda:
criterion = criterion.cuda()
# Normalizing
if not opt.no_mean_norm:
opt.mean = get_mean(opt.norm_value, dataset=opt.mean_dataset)
opt.std = get_std(opt.norm_value, dataset=opt.std_dataset)
norm_method = Normalize(opt.mean, opt.std)
else:
norm_method = Normalize([0, 0, 0], [1, 1, 1])
print(opt)
with open(os.path.join(opt.result_path, 'opts.json'), 'w') as opt_file:
json.dump(vars(opt), opt_file)
# **************************** TRAINING CONFIGURATIONS ************************************
assert opt.train_crop in ['corner', 'center']
if opt.train_crop == 'corner':
crop_method = MultiScaleCornerCrop(opt.scales, opt.sample_size)
elif opt.train_crop == 'center':
crop_method = MultiScaleCornerCrop(opt.scales,
opt.sample_size,
crop_positions=['c'])
# Пространственное преобразование
spatial_transform = Compose([
crop_method,
#RandomHorizontalFlip(),
ToTensor(opt.norm_value),
norm_method
])
# Временное преобразование
temporal_transform = TemporalRandomCrop(opt.sample_duration)
# Целевое преобразование
target_transform = ClassLabel()
train_loader_list = []
if not opt.no_cross_validation:
annotation_list = os.listdir(opt.annotation_directory)
for annotation in annotation_list:
opt.annotation_path = os.path.join(opt.annotation_directory,
annotation)
training_data = get_training_set(opt, spatial_transform,
temporal_transform,
target_transform)
train_loader = torch.utils.data.DataLoader(
training_data,
batch_size=opt.batch_size,
shuffle=True,
num_workers=opt.n_threads,
pin_memory=True)
train_loader_list.append(train_loader)
else:
training_data = get_training_set(opt, spatial_transform,
temporal_transform, target_transform)
train_loader = torch.utils.data.DataLoader(training_data,
batch_size=opt.batch_size,
shuffle=True,
num_workers=opt.n_threads,
pin_memory=True)
train_loader_list.append(train_loader)
train_logger = Logger(os.path.join(opt.result_path, 'train.log'),
['epoch', 'loss', 'acc', 'lr'])
train_batch_logger = Logger(
os.path.join(opt.result_path, 'train_batch.log'),
['epoch', 'batch', 'iter', 'loss', 'acc', 'lr'])
optimizer = optim.SGD(parameters,
lr=opt.learning_rate,
momentum=opt.momentum,
dampening=opt.dampening,
weight_decay=opt.weight_decay)
scheduler = lr_scheduler.ReduceLROnPlateau(optimizer,
'min',
patience=opt.lr_patience)
# ***************************** VALIDATION CONFIGURATIONS *********************************
spatial_transform = Compose([
Scale(opt.sample_size),
CenterCrop(opt.sample_size),
ToTensor(opt.norm_value), norm_method
])
temporal_transform = LoopPadding(opt.sample_duration)
target_transform = ClassLabel()
val_loader_list = []
if not opt.no_cross_validation:
annotation_list = os.listdir(opt.annotation_directory)
for annotation in annotation_list:
opt.annotation_path = os.path.join(opt.annotation_directory,
annotation)
validation_data = get_validation_set(opt, spatial_transform,
temporal_transform,
target_transform)
val_loader = torch.utils.data.DataLoader(validation_data,
batch_size=opt.batch_size,
shuffle=False,
num_workers=opt.n_threads,
pin_memory=True)
val_loader_list.append(val_loader)
else:
validation_data = get_validation_set(opt, spatial_transform,
temporal_transform,
target_transform)
val_loader = torch.utils.data.DataLoader(validation_data,
batch_size=opt.batch_size,
shuffle=False,
num_workers=opt.n_threads,
pin_memory=True)
val_loader_list.append(val_loader)
val_logger = Logger(os.path.join(opt.result_path, 'val.log'),
['epoch', 'loss', 'acc'])
# **************************************** TRAINING ****************************************
epoch_avg_time = AverageMeter()
train_loss_list = []
train_acc_list = []
valid_acc_list = []
best_accuracy = 0
current_train_data = 0
current_valid_data = 0
opt.frequence_cross_validation = round(opt.n_epochs /
opt.n_cross_validation_sets + 0.5)
for epoch in range(opt.begin_epoch, opt.n_epochs + 1):
epoch_start_time = time.time()
print('Epoch #' + str(epoch))
# optimizer = regulate_learning_rate(optimizer, epoch, opt.frequence_regulate_lr)
train_loader = train_loader_list[current_train_data]
if not opt.no_cross_validation and epoch % opt.frequence_cross_validation == 0:
print('\t##### Cross-validation: switch training data #####')
current_train_data = (current_train_data +
1) % len(train_loader_list)
train_loader = train_loader_list[current_train_data]
train_loss, train_acc = train_epoch(epoch, train_loader, model,
criterion, optimizer, opt,
train_logger, train_batch_logger)
val_loader = val_loader_list[current_valid_data]
if not opt.no_cross_validation and epoch % opt.frequence_cross_validation == 0:
print('\t##### Cross-validation: switch validation data #####')
current_valid_data = (current_valid_data +
1) % len(val_loader_list)
val_loader = val_loader_list[current_valid_data]
validation_acc = val_epoch(epoch, val_loader, model, criterion, opt,
val_logger)
train_loss_list.append(train_loss)
train_acc_list.append(train_acc)
valid_acc_list.append(validation_acc)
# Save model with best accuracy
if validation_acc > best_accuracy:
best_accuracy = validation_acc
save_file_path = os.path.join(opt.result_path, 'best_model.pth')
states = {
'epoch': epoch + 1,
'arch': opt.arch,
'state_dict': model.state_dict(),
'optimizer': optimizer.state_dict()
}
torch.save(states, save_file_path)
epoch_end_time = time.time() - epoch_start_time
epoch_avg_time.update(epoch_end_time)
print('\tTime left: ' +
str(round(epoch_avg_time.avg *
(opt.n_epochs - epoch) / 60, 1)) + ' minutes')
# ******************************* SAVING RESULTS OF TRAINING ******************************
save_pictures(np.linspace(1, opt.n_epochs, opt.n_epochs),
train_loss_list, 'red', 'Loss',
os.path.join(opt.result_path, 'train_loss.png'))
save_pictures(np.linspace(1, opt.n_epochs, opt.n_epochs), train_acc_list,
'blue', 'Accuracy',
os.path.join(opt.result_path, 'train_accuracy.png'))
save_pictures(np.linspace(1, opt.n_epochs, opt.n_epochs), valid_acc_list,
'blue', 'Accuracy',
os.path.join(opt.result_path, 'validation_accuracy.png'))
