def train(args, epoch, model, data_loader, optimizer):
"""
TRAINING PROCEDURE
Parameters:
-----------
- args: various arguments
- epoch: number of epochs
- model: specified model to test
- data_loader: specified train data_loader
- optimizer: specified optimizer to use
Returns:
--------
- average_loss: average loss per batch
"""
statistics = []
total_loss = 0
model.train()
title = 'Training Epoch {}'.format(epoch)
progress = tqdm(tools.IteratorTimer(data_loader),
ncols=120,
total=len(data_loader),
smoothing=.9,
miniters=1,
leave=True,
desc=title)
sys.stdout.flush()
for batch_idx, (data, target) in enumerate(progress):
#data, target = data.to(args.device), target.to(args.device)
optimizer.zero_grad()
d = model(data[0].to(args.device), im_2=data[1].to(args.device))
loss = _apply_loss(d, target).mean()
loss.backward()
optimizer.step()
total_loss += loss.item()
assert not np.isnan(total_loss)
# Print out statistics
statistics.append(loss.item())
title = '{} Epoch {}'.format('Training', epoch)
progress.set_description(title + '\tLoss:\t' + str(statistics[-1]))
sys.stdout.flush()
progress.close()
return total_loss / float(batch_idx + 1)
def test(args, epoch, model, data_loader):
"""
TESTING PROCEDURE
Parameters:
-----------
- args: various arguments
- epoch: number of epochs
- model: specified model to test
- data_loader: specified test data_loader
Returns:
--------
- average_loss: average loss per batch
- pck: Percentage of Correct Keypoints metric
"""
statistics = []
total_loss = 0
model.eval()
title = 'Validating Epoch {}'.format(epoch)
progress = tqdm(tools.IteratorTimer(data_loader),
ncols=120,
total=len(data_loader),
smoothing=.9,
miniters=1,
leave=True,
desc=title)
predictions = []
gt = []
sys.stdout.flush()
with torch.no_grad():
for batch_idx, (data, target) in enumerate(progress):
d = model(data[0].to(args.device), im_2=data[1].to(args.device))
loss = _apply_loss(d, target).mean()
total_loss += loss.item()
predictions.extend(d.numpy())
gt.extend(target.numpy())
# Print out statistics
statistics.append(loss.item())
title = '{} Epoch {}'.format('Validating', epoch)
progress.set_description(title + '\tLoss:\t' + str(statistics[-1]))
sys.stdout.flush()
progress.close()
pck = tools.calc_pck(np.asarray(predictions), np.asarray(gt))
print('PCK for epoch %d is %f' % (epoch, pck))
return total_loss / float(batch_idx + 1), pck
def train(args,
epoch,
start_iteration,
data_loader,
model,
optimizer,
loss,
logger,
is_validate=False,
offset=0):
statistics = []
total_loss = 0
gpu_mem = tools.gpumemusage()
if is_validate:
model.eval()
title = 'Validating {} Epoch {}'.format(gpu_mem, epoch)
args.validation_n_batches = np.inf if args.validation_n_batches < 0 else args.validation_n_batches
progress = tqdm(tools.IteratorTimer(data_loader),
ncols=100,
total=np.minimum(len(data_loader),
args.validation_n_batches),
leave=True,
position=offset,
desc=title)
else:
model.train()
title = 'Training {} Epoch {}'.format(tools.gpumemusage(), epoch)
args.train_n_batches = np.inf if args.train_n_batches < 0 else args.train_n_batches
progress = tqdm(tools.IteratorTimer(data_loader),
ncols=120,
total=np.minimum(len(data_loader),
args.train_n_batches),
smoothing=.9,
miniters=1,
leave=True,
position=offset,
desc=title)
last_log_time = progress._time()
for batch_idx, (data, target) in enumerate(progress):
data, target = [Variable(d, volatile=is_validate) for d in data], [
Variable(t, volatile=is_validate) for t in target
]
if args.cuda:
data, target = [d.cuda(async=True) for d in data
], [t.cuda(async=True) for t in target]
optimizer.zero_grad() if not is_validate else None
output = model(data[0])
loss_labels, loss_values = loss(output, target[0])
loss_val = loss_values[0]
total_loss += loss_val.data[0]
loss_values = [v.data[0] for v in loss_values]
assert not np.isnan(total_loss)
if not is_validate and args.fp16:
loss_val.backward()
if args.gradient_clip:
torch.nn.utils.clip_grad_norm(model.parameters(),
args.gradient_clip)
params = list(model.parameters())
for i in range(len(params)):
param_copy[i].grad = params[i].grad.clone().type_as(
params[i]).detach()
param_copy[i].grad.mul_(1. / args.loss_scale)
optimizer.step()
for i in range(len(params)):
params[i].data.copy_(param_copy[i].data)
elif not is_validate:
loss_val.backward()
if args.gradient_clip:
torch.nn.utils.clip_grad_norm(model.parameters(),
args.gradient_clip)
optimizer.step()
# Update hyperparameters if needed
global_iteration = start_iteration + batch_idx
if not is_validate:
tools.update_hyperparameter_schedule(args, epoch,
global_iteration,
optimizer)
loss_labels.append('lr')
loss_values.append(optimizer.param_groups[0]['lr'])
loss_labels.append('load')
loss_values.append(progress.iterable.last_duration)
# Print out statistics
statistics.append(loss_values)
title = '{} {} Epoch {}'.format(
'Validating' if is_validate else 'Training',
tools.gpumemusage(), epoch)
progress.set_description(
title + ' ' +
tools.format_dictionary_of_losses(loss_labels, statistics[-1]))
if ((((global_iteration + 1) % args.log_frequency) == 0
and not is_validate) or
(is_validate and batch_idx == args.validation_n_batches - 1)):
global_iteration = global_iteration if not is_validate else start_iteration
logger.add_scalar(
'batch logs per second',
len(statistics) / (progress._time() - last_log_time),
global_iteration)
last_log_time = progress._time()
all_losses = np.array(statistics)
for i, key in enumerate(loss_labels):
logger.add_scalar('average batch ' + key,
all_losses[:,
i].mean(), global_iteration)
logger.add_histogram(key, all_losses[:, i],
global_iteration)
# Reset Summary
statistics = []
if (is_validate and (batch_idx == args.validation_n_batches)):
break
if ((not is_validate) and (batch_idx == (args.train_n_batches))):
break
progress.close()
return total_loss / float(batch_idx + 1), (batch_idx + 1)
def main(args):
dataset = args.dataset
bsize = args.batch_size
root = args.data_root
cache_root = args.cache
prediction_root = args.pre
train_root = root + dataset + '/train'
val_root = root + dataset + '/val' # validation dataset
# mkdir( path [,mode] )：创建一个目录，可以是相对或者绝对路径，mode的默认模式是0777。
# 如果目录有多级，则创建最后一级。如果最后一级目录的上级目录有不存在的，则会抛出一个OSError。
# makedirs( path [,mode] )：创建递归的目录树，可以是相对或者绝对路径，mode的默认模式是
# 0777。如果子目录创建失败或者已经存在，会抛出一个OSError的异常，Windows上Error 183即为
# 目录已经存在的异常错误。如果path只有一级，与mkdir相同。
check_root_opti = cache_root + '/opti' # save checkpoint parameters
if not os.path.exists(check_root_opti):
os.makedirs(check_root_opti)
check_root_feature = cache_root + '/feature' # save checkpoint parameters
if not os.path.exists(check_root_feature):
os.makedirs(check_root_feature)
# 获取调整后的数据集
train_loader = torch.utils.data.DataLoader(
MyData(train_root, transform=True),
batch_size=bsize, shuffle=True, num_workers=4, pin_memory=True
)
val_loader = torch.utils.data.DataLoader(
MyTestData(val_root, transform=True),
batch_size=bsize, shuffle=True, num_workers=4, pin_memory=True
)
model = densenet169(pretrained=True, new_block=RCL_Module).cuda()
criterion = nn.BCELoss()
optimizer_feature = torch.optim.Adam(model.parameters(), lr=args.lr)
# http://www.spytensor.com/index.php/archives/32/
scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(
optimizer_feature, 'max', verbose=1, patience=10
)
progress = tqdm(
range(args.start_epoch, args.total_epochs + 1), miniters=1,
ncols=100, desc='Overall Progress', leave=True, position=0
)
offset = 1
best = 0
result = {'epoch': [], 'F_measure': [], 'MAE': []}
for epoch in progress:
# ===============================TRAIN=================================
title = 'Training Epoch {}'.format(epoch)
progress_epoch = tqdm(
tools.IteratorTimer(train_loader), ncols=120,
total=len(train_loader), smoothing=0.9, miniters=1,
leave=True, position=offset, desc=title
)
train(model, progress_epoch, criterion, optimizer_feature, epoch, args)
# ==============================TEST===================================
if epoch % args.val_rate == 0:
epoch, F_measure, mae = test(
model, val_loader, epoch,
prediction_root, check_root_feature, check_root_opti, val_root
)
result['epoch'].append(int(epoch))
result['F_measure'].append(round(float(F_measure), 3))
result['MAE'].append(round(float(mae), 3))
df = pd.DataFrame(result).set_index('epoch')
df.to_csv('./lart/result.csv')
if epoch == 0:
best = F_measure - mae
elif (F_measure - mae) > best:
best = F_measure - mae
# 存储最好的权重和偏置
filename = ('%s/feature-best.pth' % check_root_feature)
torch.save(model.state_dict(), filename)
# 存储最好的优化器状态
filename_opti = ('%s/opti-best.pth' % check_root_opti)
torch.save(optimizer_feature.state_dict(), filename_opti)
# 只在验证期间考虑更改学习率
scheduler.step(best)
def train(args,
epoch,
start_iteration,
data_loader,
model,
optimizer,
logger,
is_validate=False,
offset=0):
statistics = []
total_loss = 0
if is_validate:
model.eval()
title = 'Validating Epoch {}'.format(epoch)
args.validation_n_batches = len(
data_loader
) - 1 if args.validation_n_batches < 0 else args.validation_n_batches
progress = tqdm(tools.IteratorTimer(data_loader),
ncols=100,
total=np.minimum(len(data_loader),
args.validation_n_batches),
leave=True,
position=offset,
desc=title)
else:
model.train()
title = 'Training Epoch {}'.format(epoch)
args.train_n_batches = len(
data_loader
) - 1 if args.train_n_batches < 0 else args.train_n_batches
progress = tqdm(tools.IteratorTimer(data_loader),
ncols=120,
total=np.minimum(len(data_loader),
args.train_n_batches),
smoothing=.9,
miniters=1,
leave=True,
position=offset,
desc=title)
last_log_time = progress._time()
for batch_idx, (data, target) in enumerate(progress):
data, target = [Variable(d, volatile=is_validate) for d in data], [
Variable(t, volatile=is_validate) for t in target
]
if args.cuda and args.number_gpus == 1:
data, target = [d.cuda(async=True) for d in data
], [t.cuda(async=True) for t in target]
optimizer.zero_grad() if not is_validate else None
losses = model(data[0], target[0])
losses = [torch.mean(loss_value) for loss_value in losses]
loss_val = losses[1] # Collect first loss for weight update
total_loss += loss_val.data[0]
loss_values = [v.data[0] for v in losses]
# gather loss_labels, direct return leads to recursion limit error as it looks for variables to gather'
loss_labels = list(model.module.loss.loss_labels)
assert not np.isnan(total_loss)
if not is_validate and args.fp16:
loss_val.backward()
if args.gradient_clip:
torch.nn.utils.clip_grad_norm(model.parameters(),
args.gradient_clip)
params = list(model.parameters())
for i in range(len(params)):
param_copy[i].grad = params[i].grad.clone().type_as(
params[i]).detach()
param_copy[i].grad.mul_(1. / args.loss_scale)
optimizer.step()
for i in range(len(params)):
params[i].data.copy_(param_copy[i].data)
elif not is_validate:
loss_val.backward()
if args.gradient_clip:
torch.nn.utils.clip_grad_norm(model.parameters(),
args.gradient_clip)
optimizer.step()
# Update hyperparameters if needed
global_iteration = start_iteration + batch_idx
if not is_validate:
tools.update_hyperparameter_schedule(args, epoch,
global_iteration,
optimizer)
loss_labels.append('lr')
loss_values.append(optimizer.param_groups[0]['lr'])
loss_labels.append('load')
loss_values.append(progress.iterable.last_duration)
# Print out statistics
statistics.append(loss_values)
title = '{} Epoch {}'.format(
'Validating' if is_validate else 'Training', epoch)
if (type(loss_labels[0]) is list) or (type(loss_labels[0]) is
tuple):
progress.set_description(title + ' ' +
tools.format_dictionary_of_losses(
loss_labels[0], statistics[-1]))
else:
progress.set_description(title + ' ' +
tools.format_dictionary_of_losses(
loss_labels, statistics[-1]))
if ((((global_iteration + 1) % args.log_frequency) == 0
and not is_validate) or
(is_validate and batch_idx == args.validation_n_batches - 1)):
global_iteration = global_iteration if not is_validate else start_iteration
logger.add_scalar(
'batch logs per second',
len(statistics) / (progress._time() - last_log_time),
global_iteration)
last_log_time = progress._time()
all_losses = np.array(statistics)
for i, key in enumerate(loss_labels[0] if (
type(loss_labels[0]) is list) or (
type(loss_labels[0]) is tuple) else loss_labels):
logger.add_scalar('average batch ' + str(key),
all_losses[:,
i].mean(), global_iteration)
#logger.add_histogram(str(key), all_losses[:, i], global_iteration)
if is_validate:
_, output = model(data[0], target[0], inference=True)
render_flow = output[0].data.cpu().numpy().transpose(
1, 2, 0)
ground_truth = target[0][0].data.cpu().numpy().transpose(
1, 2, 0)
render_img = tools.flow_to_image(render_flow).transpose(
2, 0, 1)
true_img = tools.flow_to_image(ground_truth).transpose(
2, 0, 1)
render_img = torch.Tensor(render_img) / 255.0
true_img = torch.Tensor(true_img) / 255.0
input_img = data[0][0, :, 0, :, :].data.cpu() / 255.0
logger.add_image('renderimg',
torchvision.utils.make_grid(render_img),
global_iteration)
logger.add_image('ground_truth',
torchvision.utils.make_grid(true_img),
global_iteration)
logger.add_image('input_img',
torchvision.utils.make_grid(input_img),
global_iteration)
# Reset Summary
statistics = []
if (is_validate and (batch_idx == args.validation_n_batches)):
break
if ((not is_validate) and (batch_idx == (args.train_n_batches))):
break
progress.close()
return total_loss / float(batch_idx + 1), (batch_idx + 1)
def train(input_args,
train_epoch,
start_iteration,
files_loader,
model,
model_optimizer,
logger,
is_validate=False,
offset=0):
statistics = []
total_loss = 0
if is_validate:
model.eval()
title = 'Validating Epoch {}'.format(train_epoch)
input_args.validation_n_batches = np.inf if input_args.validation_n_batches < 0 else input_args.validation_n_batches
file_progress = tqdm(tools.IteratorTimer(files_loader),
ncols=100,
total=np.minimum(
len(files_loader),
input_args.validation_n_batches),
leave=True,
position=offset,
desc=title)
else:
model.train()
title = 'Training Epoch {}'.format(train_epoch)
input_args.train_n_batches = np.inf if input_args.train_n_batches < 0 else input_args.train_n_batches
file_progress = tqdm(tools.IteratorTimer(files_loader),
ncols=120,
total=np.minimum(len(files_loader),
input_args.train_n_batches),
smoothing=.9,
miniters=1,
leave=True,
position=offset,
desc=title)
last_log_time = file_progress._time()
for batch_idx, (data_file) in enumerate(file_progress):
video_dataset = datasets_video.VideoFileDataJIT(
input_args, data_file[0])
video_loader = DataLoader(video_dataset,
batch_size=args.effective_batch_size,
shuffle=True,
**gpuargs)
global_iteration = start_iteration + batch_idx
# note~ for debugging purposes
# video_frame_progress = tqdm(tools.IteratorTimer(video_loader), ncols=120,
# total=len(video_loader), smoothing=0.9, miniters=1,
# leave=True, desc=data_file[0])
for i_batch, (data, target) in enumerate(video_loader):
data, target = [Variable(d)
for d in data], [Variable(t) for t in target]
if input_args.cuda and input_args.number_gpus == 1:
data, target = [d.cuda(async=True) for d in data
], [t.cuda(async=True) for t in target]
model_optimizer.zero_grad() if not is_validate else None
losses = model(data[0], target[0])
losses = [torch.mean(loss_value) for loss_value in losses]
loss_val = losses[0] # Collect first loss for weight update
total_loss += loss_val.data
loss_values = [v.data for v in losses]
# gather loss_labels, direct return leads to recursion limit error as it looks for variables to gather'
loss_labels = list(model.module.loss.loss_labels)
assert not np.isnan(total_loss.cpu().numpy())
if not is_validate and input_args.fp16:
loss_val.backward()
if input_args.gradient_clip:
torch.nn.utils.clip_grad_norm(model.parameters(),
input_args.gradient_clip)
params = list(model.parameters())
for i in range(len(params)):
param_copy[i].grad = params[i].grad.clone().type_as(
params[i]).detach()
param_copy[i].grad.mul_(1. / input_args.loss_scale)
model_optimizer.step()
for i in range(len(params)):
params[i].data.copy_(param_copy[i].data)
elif not is_validate:
loss_val.backward()
if input_args.gradient_clip:
torch.nn.utils.clip_grad_norm(model.parameters(),
input_args.gradient_clip)
model_optimizer.step()
# Update hyperparameters if needed
if not is_validate:
tools.update_hyperparameter_schedule(
input_args, train_epoch, global_iteration,
model_optimizer)
loss_labels.append('lr')
loss_values.append(model_optimizer.param_groups[0]['lr'])
loss_labels.append('load')
loss_values.append(file_progress.iterable.last_duration)
# Print out statistics
statistics.append(loss_values)
title = '{} Epoch {}'.format(
'Validating' if is_validate else 'Training', train_epoch)
file_progress.set_description(
title + ' ' + tools.format_dictionary_of_losses(
tools.flatten_list(loss_labels), statistics[-1]))
if ((((global_iteration + 1) % input_args.log_frequency) == 0
and not is_validate)
or (is_validate
and batch_idx == input_args.validation_n_batches - 1)):
global_iteration = global_iteration if not is_validate else start_iteration
logger.add_scalar(
'batch logs per second',
len(statistics) / (file_progress._time() - last_log_time),
global_iteration)
last_log_time = file_progress._time()
all_losses = np.array(statistics)
for i, key in enumerate(tools.flatten_list(loss_labels)):
if isinstance(all_losses[:, i].item(), torch.Tensor):
average_batch = all_losses[:, i].item().mean()
else:
average_batch = all_losses[:, i].item()
logger.add_scalar('average batch ' + str(key),
average_batch, global_iteration)
logger.add_histogram(str(key), all_losses[:, i],
global_iteration)
# Reset Summary
statistics = []
if is_validate and (batch_idx == input_args.validation_n_batches):
break
if (not is_validate) and (batch_idx
== (input_args.train_n_batches)):
break
file_progress.close()
return total_loss / float(batch_idx + 1), (batch_idx + 1)
def train(args, epoch, start_iteration, data_loader, model, optimizer, logger, is_validate=False, offset=0):
statistics = []
total_loss = 0
if is_validate:
model.eval()
title = "Validating Epoch {}".format(epoch)
args.validation_n_batches = np.inf if args.validation_n_batches < 0 else args.validation_n_batches
progress = tqdm(
tools.IteratorTimer(data_loader),
ncols=100,
total=np.minimum(len(data_loader), args.validation_n_batches),
leave=True,
position=offset,
desc=title,
)
else:
model.train()
title = "Training Epoch {}".format(epoch)
args.train_n_batches = np.inf if args.train_n_batches < 0 else args.train_n_batches
progress = tqdm(
tools.IteratorTimer(data_loader),
ncols=120,
total=np.minimum(len(data_loader), args.train_n_batches),
smoothing=0.9,
miniters=1,
leave=True,
position=offset,
desc=title,
)
last_log_time = progress._time()
for batch_idx, (data, target) in enumerate(progress):
data, target = [Variable(d) for d in data], [Variable(t) for t in target]
if args.cuda and args.number_gpus == 1:
data, target = [d.cuda(non_blocking=True) for d in data], [t.cuda(non_blocking=True) for t in target]
optimizer.zero_grad() if not is_validate else None
losses = model(data[0], target[0])
losses = [torch.mean(loss_value) for loss_value in losses]
loss_val = losses[0] # Collect first loss for weight update
total_loss += loss_val.item()
loss_values = [v.item() for v in losses]
# gather loss_labels, direct return leads to recursion limit error as it looks for variables to gather'
loss_labels = list(model.module.loss.loss_labels)
assert not np.isnan(total_loss)
if not is_validate and args.fp16:
loss_val.backward()
if args.gradient_clip:
torch.nn.utils.clip_grad_norm(model.parameters(), args.gradient_clip)
params = list(model.parameters())
for i in range(len(params)):
param_copy[i].grad = params[i].grad.clone().type_as(params[i]).detach()
param_copy[i].grad.mul_(1.0 / args.loss_scale)
optimizer.step()
for i in range(len(params)):
params[i].data.copy_(param_copy[i].data)
elif not is_validate:
loss_val.backward()
if args.gradient_clip:
torch.nn.utils.clip_grad_norm(model.parameters(), args.gradient_clip)
optimizer.step()
# Update hyperparameters if needed
global_iteration = start_iteration + batch_idx
if not is_validate:
tools.update_hyperparameter_schedule(args, epoch, global_iteration, optimizer)
loss_labels.append("lr")
loss_values.append(optimizer.param_groups[0]["lr"])
loss_labels.append("load")
loss_values.append(progress.iterable.last_duration)
# Print out statistics
statistics.append(loss_values)
title = "{} Epoch {}".format("Validating" if is_validate else "Training", epoch)
progress.set_description(title + " " + tools.format_dictionary_of_losses(loss_labels, statistics[-1]))
if (((global_iteration + 1) % args.log_frequency) == 0 and not is_validate) or (
is_validate and batch_idx == args.validation_n_batches - 1
):
global_iteration = global_iteration if not is_validate else start_iteration
logger.add_scalar(
"batch logs per second", len(statistics) / (progress._time() - last_log_time), global_iteration
)
last_log_time = progress._time()
all_losses = np.array(statistics)
for i, key in enumerate(loss_labels):
logger.add_scalar("average batch " + str(key), all_losses[:, i].mean(), global_iteration)
logger.add_histogram(str(key), all_losses[:, i], global_iteration)
# Reset Summary
statistics = []
if is_validate and (batch_idx == args.validation_n_batches):
break
if (not is_validate) and (batch_idx == (args.train_n_batches)):
break
progress.close()
return total_loss / float(batch_idx + 1), (batch_idx + 1)
def main(args):
dataset = args.dataset
bsize = args.batch_size
root = args.data_root
cache_root = args.cache
prediction_root = args.pre
train_root = root + dataset + '/train'
val_root = root + dataset + '/val' # validation dataset
check_root_opti = cache_root + '/opti' # save checkpoint parameters
if not os.path.exists(check_root_opti):
os.mkdir(check_root_opti)
check_root_feature = cache_root + '/feature' # save checkpoint parameters
if not os.path.exists(check_root_feature):
os.mkdir(check_root_feature)
train_loader = torch.utils.data.DataLoader(MyData(train_root,
transform=True),
batch_size=bsize,
shuffle=True,
num_workers=4,
pin_memory=True)
val_loader = torch.utils.data.DataLoader(MyTestData(val_root,
transform=True),
batch_size=bsize,
shuffle=True,
num_workers=4,
pin_memory=True)
model = Feature(RCL_Module)
model.cuda()
criterion = nn.BCELoss()
optimizer_feature = torch.optim.Adam(model.parameters(), lr=args.lr)
train_losses = []
progress = tqdm(range(args.start_epoch, args.total_epochs + 1),
miniters=1,
ncols=100,
desc='Overall Progress',
leave=True,
position=0)
offset = 1
best = 0
evaluation = []
result = {'epoch': [], 'F_measure': [], 'MAE': []}
for epoch in progress:
if (epoch != 0):
print("\nloading parameters")
model.load_state_dict(
torch.load(check_root_feature + '/feature-current.pth'))
optimizer_feature.load_state_dict(
torch.load(check_root_opti + '/opti-current.pth'))
#
title = 'Training Epoch {}'.format(epoch)
progress_epoch = tqdm(tools.IteratorTimer(train_loader),
ncols=120,
total=len(train_loader),
smoothing=.9,
miniters=1,
leave=True,
position=offset,
desc=title)
for ib, (input, gt) in enumerate(progress_epoch):
inputs = Variable(input).cuda()
gt = Variable(gt.unsqueeze(1)).cuda()
gt_28 = functional.interpolate(gt, size=28, mode='bilinear')
gt_56 = functional.interpolate(gt, size=56, mode='bilinear')
gt_112 = functional.interpolate(gt, size=112, mode='bilinear')
msk1, msk2, msk3, msk4, msk5 = model.forward(inputs)
loss = criterion(msk1, gt_28) + criterion(msk2, gt_28) + criterion(
msk3, gt_56) + criterion(msk4, gt_112) + criterion(msk5, gt)
model.zero_grad()
loss.backward()
optimizer_feature.step()
train_losses.append(round(float(loss.data.cpu()), 3))
title = '{} Epoch {}/{}'.format('Training', epoch,
args.total_epochs)
progress_epoch.set_description(title + ' ' + 'loss:' +
str(loss.data.cpu().numpy()))
filename = ('%s/feature-current.pth' % (check_root_feature))
filename_opti = ('%s/opti-current.pth' % (check_root_opti))
torch.save(model.state_dict(), filename)
torch.save(optimizer_feature.state_dict(), filename_opti)
#--------------------------validation on the test set every n epoch--------------
if (epoch % args.val_rate == 0):
fileroot = ('%s/feature-current.pth' % (check_root_feature))
model.load_state_dict(torch.load(fileroot))
val_output_root = (prediction_root + '/epoch_current')
if not os.path.exists(val_output_root):
os.mkdir(val_output_root)
print("\ngenerating output images")
for ib, (input, img_name, _) in enumerate(val_loader):
inputs = Variable(input).cuda()
_, _, _, _, output = model.forward(inputs)
output = functional.sigmoid(output)
out = output.data.cpu().numpy()
for i in range(len(img_name)):
imsave(os.path.join(val_output_root, img_name[i] + '.png'),
out[i, 0],
cmap='gray')
print("\nevaluating mae....")
F_measure, mae = get_FM(salpath=val_output_root + '/',
gtpath=val_root + '/gt/')
evaluation.append([int(epoch), float(F_measure), float(mae)])
result['epoch'].append(int(epoch))
result['F_measure'].append(round(float(F_measure), 3))
result['MAE'].append(round(float(mae), 3))
df = pd.DataFrame(result).set_index('epoch')
df.to_csv('./result.csv')
if (epoch == 0): best = F_measure - mae
elif ((F_measure - mae) > best):
best = F_measure - mae
filename = ('%s/feature-best.pth' % (check_root_feature))
filename_opti = ('%s/opti-best.pth' % (check_root_opti))
torch.save(model.state_dict(), filename)
torch.save(optimizer_feature.state_dict(), filename_opti)
def train(args,
epoch,
start_iteration,
data_loader,
model,
optimizer,
logger,
is_validate=False,
offset=0):
statistics = []
all_gradient_norms = []
total_loss = 0
if is_validate:
model.eval()
title = 'Validating Epoch {}'.format(epoch)
args.validation_n_batches = np.inf if args.validation_n_batches < 0 else args.validation_n_batches
progress = tqdm(tools.IteratorTimer(data_loader),
ncols=200,
total=np.minimum(len(data_loader),
args.validation_n_batches),
leave=True,
position=offset,
desc=title)
else:
model.train()
title = 'Training Epoch {}'.format(epoch)
args.train_n_batches = np.inf if args.train_n_batches < 0 else args.train_n_batches
progress = tqdm(tools.IteratorTimer(data_loader),
ncols=200,
total=np.minimum(len(data_loader),
args.train_n_batches),
smoothing=.9,
miniters=1,
leave=True,
position=offset,
desc=title)
last_log_time = progress._time()
for batch_idx, (data, target) in enumerate(progress):
data, target = [Variable(d)
for d in data], [Variable(t) for t in target]
if args.cuda and args.number_gpus == 1:
data, target = [d.cuda(non_blocking=True) for d in data
], [t.cuda(non_blocking=True) for t in target]
optimizer.zero_grad() if not is_validate else None
losses, flow = model(data[0], target[0])
#print('Losses shape {} {}'.format(losses[0].shape, losses[1].shape))
losses = [torch.mean(loss_value) for loss_value in losses]
loss_val = losses[0] # Collect first loss for weight update
total_loss += loss_val.item()
loss_values = [v.item() for v in losses]
loss_labels = list(model.module.loss.loss_labels)
assert not np.isnan(total_loss)
if not is_validate and args.fp16:
loss_val.backward()
if args.gradient_clip:
torch.nn.utils.clip_grad_norm(model.parameters(),
args.gradient_clip)
params = list(model.parameters())
for i in range(len(params)):
param_copy[i].grad = params[i].grad.clone().type_as(
params[i]).detach()
param_copy[i].grad.mul_(1. / args.loss_scale)
optimizer.step()
for i in range(len(params)):
params[i].data.copy_(param_copy[i].data)
elif not is_validate:
loss_val.backward()
if args.gradient_clip:
gradient_norm = torch.nn.utils.clip_grad_norm(
model.parameters(), args.gradient_clip)
all_gradient_norms.append(gradient_norm)
optimizer.step()
# Update hyperparameters if needed
global_iteration = start_iteration + batch_idx
if not is_validate:
tools.update_hyperparameter_schedule(args, epoch,
global_iteration,
optimizer)
loss_labels.append('lr')
loss_values.append(optimizer.param_groups[0]['lr'])
loss_labels.append('load')
loss_values.append(progress.iterable.last_duration)
# Print out statistics
statistics.append(loss_values)
title = '{} Epoch {}'.format(
'Validating' if is_validate else 'Training', epoch)
progress.set_description(
title + ' ' +
tools.format_dictionary_of_losses(loss_labels, statistics[-1]))
if ((((global_iteration + 1) % args.log_frequency) == 0
and not is_validate) or
(is_validate and batch_idx == args.validation_n_batches - 1)):
global_iteration = global_iteration if not is_validate else start_iteration
logger.add_scalar(
'batch logs per second',
len(statistics) / (progress._time() - last_log_time),
global_iteration)
last_log_time = progress._time()
all_losses = np.array(statistics)
for i, key in enumerate(loss_labels):
logger.add_scalar('average batch ' + str(key),
all_losses[:,
i].mean(), global_iteration)
logger.add_histogram(str(key), all_losses[:, i],
global_iteration)
if args.gradient_clip:
logger.add_scalar('average batch gradient_norm',
np.array(all_gradient_norms).mean(),
global_iteration)
all_gradient_norms = []
# Returns multiscale flow, get largest scale and first element in batch
if args.multiframe or args.multiframe_two_output:
flow = flow_utils.flow_postprocess(flow)[0][0]
num_flows = len(args.frame_weights)
flows_scaled = [
cv2.resize(flow[:, :, i:i + 2], None, fx=4.0, fy=4.0)
for i in range(0, 2 * num_flows, 2)
]
target = target[0].detach().cpu().numpy()
target_flow = np.transpose(target[0], (1, 2, 3, 0))
results_images = [
visualize_results(flows_scaled[i], target_flow[i],
data[0][0] if i == 0 else None)
for i in range(0, num_flows)
]
for i in range(0, num_flows):
logger.add_image('flow{} and target'.format(i),
ToTensor()(results_images[i]),
global_iteration)
else:
flow = flow_utils.flow_postprocess(flow)[0][0]
flow_scaled = cv2.resize(flow, None, fx=4.0, fy=4.0)
target_flow = flow_utils.flow_postprocess(target)[0][0]
results_image = visualize_results(flow_scaled, target_flow,
data[0][0])
logger.add_image('flow and target',
ToTensor()(results_image),
global_iteration)
# logger.add_histogram('flow_values', flow[0], global_iteration)
# Reset Summary
statistics = []
if (is_validate and (batch_idx == args.validation_n_batches)):
break
if ((not is_validate) and (batch_idx == (args.train_n_batches))):
break
progress.close()
return total_loss / float(batch_idx + 1), (batch_idx + 1)
def train(args,
epoch,
start_iteration,
data_loader,
model,
optimizer,
logger,
is_validate=False,
offset=0):
statistics = []
total_loss = 0
if is_validate:
model.eval()
title = 'Validating Epoch {}'.format(epoch)
#print("validation_n_batches", args.validation_n_batches)
args.validation_n_batches = np.inf if args.validation_n_batches < 0 else args.validation_n_batches
#print("validation_n_batches", args.validation_n_batches)
progress = tqdm(tools.IteratorTimer(data_loader),
ncols=100,
total=np.minimum(len(data_loader),
args.validation_n_batches),
leave=True,
position=offset,
desc=title)
else:
model.train()
title = 'Training Epoch {}'.format(epoch)
args.train_n_batches = np.inf if args.train_n_batches < 0 else args.train_n_batches
progress = tqdm(tools.IteratorTimer(data_loader),
ncols=120,
total=np.minimum(len(data_loader),
args.train_n_batches),
smoothing=.9,
miniters=1,
leave=True,
position=offset,
desc=title)
last_log_time = progress._time()
for batch_idx, (data, target) in enumerate(progress):
data, target = [Variable(d)
for d in data], [Variable(t) for t in target]
if args.cuda and args.number_gpus == 1:
data, target = [d.cuda(async=True) for d in data
], [t.cuda(async=True) for t in target]
optimizer.zero_grad() if not is_validate else None
#print("this is data type",data[0].type())
#print("\n")
#print("this is target type",target[0].type())
#print("\n")
losses = model(data[0], target[0])
losses = [torch.mean(loss_value)
for loss_value in losses] # taking mean of batches
loss_val = losses[
0] # Collect first loss for weight update #take first loss, second is EPE
total_loss += loss_val.data.cpu()
loss_values = [v.data.cpu() for v in losses] #collect loss values
# gather loss_labels, direct return leads to recursion limit error as it looks for variables to gather'
#loss_labels = [y for x in model.module.loss.loss_labels for y in x] #list(model.module.loss.loss_labels)
loss_labels = list(model.module.loss.loss_labels)
assert not np.isnan(total_loss.cpu())
if not is_validate and args.fp16:
loss_val.backward()
if args.gradient_clip:
torch.nn.utils.clip_grad_norm(model.parameters(),
args.gradient_clip)
params = list(model.parameters())
for i in range(len(params)):
param_copy[i].grad = params[i].grad.clone().type_as(
params[i]).detach()
param_copy[i].grad.mul_(1. / args.loss_scale)
optimizer.step()
for i in range(len(params)):
params[i].data.copy_(param_copy[i].data)
elif not is_validate:
loss_val.backward()
if args.gradient_clip:
torch.nn.utils.clip_grad_norm(model.parameters(),
args.gradient_clip)
optimizer.step()
# Update hyperparameters if needed
global_iteration = start_iteration + batch_idx
if not is_validate:
tools.update_hyperparameter_schedule(args, epoch,
global_iteration,
optimizer)
loss_labels.append('lr')
loss_values.append(optimizer.param_groups[0]['lr'])
loss_labels.append('load')
loss_values.append(progress.iterable.last_duration) #add load
#if is_validate:
# print("this is EPE length", len(loss_values[:,1]))
# Print out statistics
#if is_validate:
# print(statistics)
statistics.append(loss_values)
#if is_validate:
# print(statistics)
title = '{} Epoch {}'.format(
'Validating' if is_validate else 'Training', epoch)
progress.set_description(
title + ' ' +
tools.format_dictionary_of_losses(loss_labels, statistics[-1]))
#if is_validate:
# print(batch_idx)
# args.log_frequency == 1 by default
if ((((global_iteration + 1) % args.log_frequency) == 0
and not is_validate) or is_validate
and batch_idx == min(args.validation_n_batches,
len(data_loader) - 1)):
#if ((((global_iteration + 1) % args.log_frequency) == 0 and not is_validate) or (is_validate and batch_idx == args.validation_n_batches - 1)):
global_iteration = global_iteration if not is_validate else start_iteration
logger.add_scalar(
'batch logs per second',
len(statistics) / (progress._time() - last_log_time),
global_iteration)
last_log_time = progress._time()
all_losses = np.array(statistics)
#if is_validate:
# print(all_losses)
for i, key in enumerate(loss_labels):
logger.add_scalar('average batch ' + str(key),
all_losses[:,
i].mean(), global_iteration)
logger.add_histogram(str(key), all_losses[:, i],
global_iteration)
# Reset Summary
statistics = []
if (is_validate and (batch_idx == args.validation_n_batches)):
break
if ((not is_validate) and (batch_idx == (args.train_n_batches))):
break
progress.close()
return total_loss / float(batch_idx + 1), (batch_idx + 1)
def train(args,
epoch,
data_loader,
model,
optimizer,
is_validate=False,
offset=0):
total_loss = 0
if is_validate:
model.eval()
title = 'Validating Epoch {}'.format(epoch)
args.validation_n_batches = np.inf if args.validation_n_batches < 0 else args.validation_n_batches
progress = tqdm(tools.IteratorTimer(data_loader),
ncols=100,
total=np.minimum(len(data_loader),
args.validation_n_batches),
leave=True,
position=offset,
desc=title)
else:
model.train()
title = 'Training Epoch {}'.format(epoch)
args.train_n_batches = np.inf if args.train_n_batches < 0 else args.train_n_batches
progress = tqdm(tools.IteratorTimer(data_loader),
ncols=120,
total=np.minimum(len(data_loader),
args.train_n_batches),
smoothing=.9,
miniters=1,
leave=True,
position=offset,
desc=title)
def torch2numpy(i):
return i[0].numpy()
for batch_idx, datas in enumerate(progress):
data = np.array([list(map(down_scailing, d)) for d in datas])
target = np.array([list(map(torch2numpy, d))[1] for d in datas])
high_frames = np.array([list(map(torch2numpy, d)) for d in datas])
# if args.cuda and args.number_gpus >= 1:
# data, target, high_frames = [d.cuda for d in data], [t.cuda for t in target], [hf.cuda for hf in high_frames]
estimated_image = None
for x, y in zip(data, target):
optimizer.zero_grad() if not is_validate else None
output, losses = model(x, y, high_frames, estimated_image)
estimated_image = output
loss_val = torch.mean(losses)
total_loss += loss_val.item()
if not is_validate:
loss_val.backward()
optimizer.step()
title = '{} Epoch {}'.format(
'Validating' if is_validate else 'Training', epoch)
progress.set_description(title)
if (is_validate and (batch_idx == args.validation_n_batches)) or \
((not is_validate) and (batch_idx == (args.train_n_batches))):
progress.close()
break
return total_loss / float(batch_idx + 1), (batch_idx + 1)
def train(args,
epoch,
start_iteration,
data_loader,
model,
optimizer,
logger,
is_validate=False,
offset=0):
#print(str(model))
statistics = []
total_loss = 0
debug = False
if is_validate:
model.eval()
title = 'Validating Epoch {}'.format(epoch)
args.validation_n_batches = np.inf if args.validation_n_batches < 0 else args.validation_n_batches
progress = tqdm(tools.IteratorTimer(data_loader),
ncols=100,
total=np.minimum(len(data_loader),
args.validation_n_batches),
leave=True,
position=offset,
desc=title)
else:
model.train()
title = 'Training Epoch {}'.format(epoch)
args.train_n_batches = np.inf if args.train_n_batches < 0 else args.train_n_batches
progress = tqdm(tools.IteratorTimer(data_loader),
ncols=120,
total=np.minimum(len(data_loader),
args.train_n_batches),
smoothing=.9,
miniters=1,
leave=True,
position=offset,
desc=title)
last_log_time = progress._time()
for batch_idx, (data, target, cdm) in enumerate(progress):
data, target, cdm = [
Variable(d, volatile=is_validate) for d in data
], [Variable(t, volatile=is_validate) for t in target
], [Variable(q, volatile=is_validate) for q in cdm]
if args.cuda and args.number_gpus == 1:
data, target, cdm = [d.cuda(async=True) for d in data
], [t.cuda(async=True) for t in target
], [q.cuda(async=True) for q in cdm]
if debug:
print(
'****************************************************************'
)
print('data_0')
print(data[0])
print('target_0')
print(target[0])
print('cdm')
print(type(cdm))
temp1 = cdm[0].data.cpu().numpy()
print(np.max(temp1))
print(temp1.shape)
print(
'****************************************************************'
)
optimizer.zero_grad() if not is_validate else None
losses = model(data[0], target[0])
losses = [torch.mean(loss_value) for loss_value in losses]
loss_val = losses[0] # Collect first loss for weight update
#A[batch_idx] = loss_val.data[0]
#np.savetxt('test_loss.out', np.array(A) , delimiter=',' , newline='\r\n' )
total_loss += loss_val.data[0]
loss_values = [v.data[0] for v in losses]
# gather loss_labels, direct return leads to recursion limit error as it looks for variables to gather'
loss_labels = list(model.module.loss.loss_labels)
assert not np.isnan(total_loss)
if not is_validate and args.fp16:
loss_val.backward()
if args.gradient_clip:
torch.nn.utils.clip_grad_norm(model.parameters(),
args.gradient_clip)
params = list(model.parameters())
for i in range(len(params)):
param_copy[i].grad = params[i].grad.clone().type_as(
params[i]).detach()
param_copy[i].grad.mul_(1. / args.loss_scale)
optimizer.step()
for i in range(len(params)):
params[i].data.copy_(param_copy[i].data)
elif not is_validate:
loss_val.backward()
if args.gradient_clip:
torch.nn.utils.clip_grad_norm(model.parameters(),
args.gradient_clip)
optimizer.step()
# Update hyperparameters if needed
global_iteration = start_iteration + batch_idx
if not is_validate:
tools.update_hyperparameter_schedule(args, epoch,
global_iteration,
optimizer)
loss_labels.append('lr')
loss_values.append(optimizer.param_groups[0]['lr'])
loss_labels.append('load')
loss_values.append(progress.iterable.last_duration)
# Print out statistics
statistics.append(loss_values)
title = '{} Epoch {}'.format(
'Validating' if is_validate else 'Training', epoch)
progress.set_description(
title + ' ' +
tools.format_dictionary_of_losses(loss_labels, statistics[-1]))
if ((((global_iteration + 1) % args.log_frequency) == 0
and not is_validate) or
(is_validate and batch_idx == args.validation_n_batches - 1)):
global_iteration = global_iteration if not is_validate else start_iteration
logger.add_scalar(
'batch logs per second',
len(statistics) / (progress._time() - last_log_time),
global_iteration)
last_log_time = progress._time()
all_losses = np.array(statistics)
for i, key in enumerate(loss_labels):
logger.add_scalar('average batch ' + str(key),
all_losses[:,
i].mean(), global_iteration)
logger.add_histogram(str(key), all_losses[:, i],
global_iteration)
# Reset Summary
statistics = []
if (is_validate and (batch_idx == args.validation_n_batches)):
break
if ((not is_validate) and (batch_idx == (args.train_n_batches))):
break
progress.close()
return total_loss / float(batch_idx + 1), (batch_idx + 1)
def train(args,
epoch,
start_iteration,
data_loader,
model,
optimizer,
scheduler,
logger,
is_validate=False,
offset=0,
max_flows_to_show=8):
running_statistics = None # Initialize below when the first losses are collected
all_losses = None # Initialize below when the first losses are collected
total_loss = 0
if is_validate:
model.eval()
title = 'Validating Epoch {}'.format(epoch)
args.validation_n_batches = np.inf if args.validation_n_batches < 0 else args.validation_n_batches
progress = tqdm(tools.IteratorTimer(data_loader),
ncols=100,
total=np.minimum(len(data_loader),
args.validation_n_batches),
leave=True,
position=offset,
desc=title)
else:
model.train()
title = 'Training Epoch {}'.format(epoch)
args.train_n_batches = np.inf if args.train_n_batches < 0 else args.train_n_batches
progress = tqdm(tools.IteratorTimer(data_loader),
ncols=120,
total=np.minimum(len(data_loader),
args.train_n_batches),
smoothing=.9,
miniters=1,
leave=True,
position=offset,
desc=title)
def convert_flow_to_image(flow_converter, flows_viz):
imgs = []
for flow_pair in flows_viz:
for flow in flow_pair:
flow = flow.numpy().transpose((1, 2, 0))
img = flow_converter._flowToColor(flow)
imgs.append(torch.from_numpy(img.transpose((2, 0, 1))))
epe_img = torch.sqrt(
torch.sum(torch.pow(flow_pair[0] - flow_pair[1], 2),
dim=0))
max_epe = torch.max(epe_img)
if max_epe == 0:
max_epe = torch.ones(1)
normalized_epe_img = epe_img / max_epe
normalized_epe_img = (255 * normalized_epe_img).type(
torch.uint8)
normalized_epe_img = torch.stack(
(normalized_epe_img, normalized_epe_img,
normalized_epe_img),
dim=0)
imgs.append(normalized_epe_img)
saturated_epe_img = torch.min(epe_img,
5.0 * torch.ones_like(epe_img))
saturated_epe_img = (51 * saturated_epe_img).type(torch.uint8)
saturated_epe_img = torch.stack(
(saturated_epe_img, saturated_epe_img, saturated_epe_img),
dim=0)
imgs.append(saturated_epe_img)
return imgs
max_iters = min(len(data_loader),
(args.validation_n_batches if
(is_validate and args.validation_n_batches > 0) else
len(data_loader)),
(args.train_n_batches if
(not is_validate
and args.train_n_batches > 0) else len(data_loader)))
if is_validate:
flow_converter = f2i.Flow()
collect_flow_interval = int(
np.ceil(float(max_iters) / max_flows_to_show))
flows_viz = []
last_log_batch_idx = 0
last_log_time = progress._time()
for batch_idx, (data, target) in enumerate(progress):
global_iteration = start_iteration + batch_idx
data, target = [Variable(d)
for d in data], [Variable(t) for t in target]
if args.cuda and args.number_gpus == 1:
data, target = [d.cuda()
for d in data], [t.cuda() for t in target]
optimizer.zero_grad() if not is_validate else None
losses, output = model(data[0], target[0], inference=True)
losses = [torch.mean(loss_value) for loss_value in losses]
loss_val = losses[0] # Collect first loss for weight update
total_loss += loss_val.item()
loss_values = [v.item() for v in losses]
if is_validate and batch_idx % collect_flow_interval == 0:
flows_viz.append(
(target[0][0].detach().cpu(), output[0].detach().cpu()))
if is_validate and args.validation_log_images and batch_idx == (
max_iters - 1):
imgs = convert_flow_to_image(flow_converter, flows_viz)
imgs = torchvision_utils.make_grid(imgs,
nrow=4,
normalize=False,
scale_each=False)
logger.add_image('target/predicted flows', imgs,
global_iteration)
# gather loss_labels, direct return leads to recursion limit error as it looks for variables to gather'
loss_labels = list(model.module.loss.loss_labels)
assert not np.isnan(total_loss)
if not is_validate and args.fp16:
loss_val.backward()
if args.gradient_clip:
torch.nn.utils.clip_grad_norm(model.parameters(),
args.gradient_clip)
params = list(model.parameters())
for i in range(len(params)):
param_copy[i].grad = params[i].grad.clone().type_as(
params[i]).detach()
param_copy[i].grad.mul_(1. / args.loss_scale)
optimizer.step()
for i in range(len(params)):
params[i].data.copy_(param_copy[i].data)
elif not is_validate:
loss_val.backward()
if args.gradient_clip:
torch.nn.utils.clip_grad_norm(model.parameters(),
args.gradient_clip)
optimizer.step()
# Update hyperparameters if needed
if not is_validate:
scheduler.step()
loss_labels.append('lr')
loss_values.append(optimizer.param_groups[0]['lr'])
loss_labels.append('load')
loss_values.append(progress.iterable.last_duration)
if running_statistics is None:
running_statistics = np.array(loss_values)
all_losses = np.zeros((len(data_loader), len(loss_values)),
np.float32)
else:
running_statistics += np.array(loss_values)
all_losses[batch_idx] = loss_values.copy()
title = '{} Epoch {}'.format(
'Validating' if is_validate else 'Training', epoch)
progress.set_description(title + ' ' +
tools.format_dictionary_of_losses(
loss_labels, running_statistics /
(batch_idx + 1)))
if ((((global_iteration + 1) % args.log_frequency) == 0
and not is_validate) or (batch_idx == max_iters - 1)):
global_iteration = global_iteration if not is_validate else start_iteration
logger.add_scalar('batch logs per second',
(batch_idx - last_log_batch_idx) /
(progress._time() - last_log_time),
global_iteration)
last_log_time = progress._time()
last_log_batch_idx = batch_idx
for i, key in enumerate(loss_labels):
logger.add_scalar('average batch ' + str(key),
all_losses[:batch_idx + 1,
i].mean(), global_iteration)
logger.add_histogram(str(key), all_losses[:batch_idx + 1,
i],
global_iteration)
if (is_validate and (batch_idx == args.validation_n_batches)):
break
if ((not is_validate) and (batch_idx == (args.train_n_batches))):
break
progress.close()
return total_loss / float(batch_idx + 1), (batch_idx + 1)
def main(args):
dataset = args.dataset
bsize = args.batch_size
root = args.data_root
cache_root = args.cache
prediction_root = args.pre
train_root = root + dataset + '/Train'
val_root = root + dataset + '/Test' # validation dataset
# mkdir( path [,mode] )：创建一个目录，可以是相对或者绝对路径，mode的默认模式是0777。
# 如果目录有多级，则创建最后一级。如果最后一级目录的上级目录有不存在的，则会抛出一个OSError。
# makedirs( path [,mode] )：创建递归的目录树，可以是相对或者绝对路径，mode的默认模式是
# 0777。如果子目录创建失败或者已经存在，会抛出一个OSError的异常，Windows上Error 183即为
# 目录已经存在的异常错误。如果path只有一级，与mkdir相同。
check_root_opti = cache_root + '/opti' # save checkpoint parameters
if not os.path.exists(check_root_opti):
os.makedirs(check_root_opti)
check_root_feature = cache_root + '/feature' # save checkpoint parameters
if not os.path.exists(check_root_feature):
os.makedirs(check_root_feature)
check_root_model = cache_root + '/model' # save checkpoint parameters
if not os.path.exists(check_root_model):
os.makedirs(check_root_model)
# 获取调整后的数据集
train_loader = torch.utils.data.DataLoader(
MyData(train_root, transform=True),
batch_size=bsize, shuffle=True, num_workers=4, pin_memory=True
)
val_loader = torch.utils.data.DataLoader(
MyTestData(val_root, transform=True),
batch_size=bsize, shuffle=True, num_workers=4, pin_memory=True
)
model = Vgg(RCL_Module)
model.cuda()
criterion = nn.BCELoss()
optimizer_feature = torch.optim.Adam(model.parameters(), lr=args.lr)
# http://www.spytensor.com/index.php/archives/32/
scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(
optimizer_feature, 'max', verbose=1, patience=10
)
progress = tqdm(
range(args.start_epoch, args.total_epochs + 1), miniters=1,
ncols=100, desc='Overall Progress', leave=True, position=0
)
offset = 1
best = 0
evaluation = []
result = {'epoch': [], 'F_measure': [], 'MAE': []}
for epoch in progress:
# ===============================TRAIN=================================
# if epoch != 0:
# print("\nloading parameters")
# # 载入上一次的训练结果(权重和偏置项), 进一步的训练
# model.load_state_dict(
# torch.load(check_root_feature + '/feature-current.pth')
# )
# # 载入优化器状态
# optimizer_feature.load_state_dict(
# torch.load(check_root_opti + '/opti-current.pth')
# )
title = 'Training Epoch {}'.format(epoch)
progress_epoch = tqdm(
tools.IteratorTimer(train_loader), ncols=120,
total=len(train_loader), smoothing=0.9, miniters=1,
leave=True, position=offset, desc=title
)
# 一个周期内部进行迭代计算
for ib, (input_, gt) in enumerate(progress_epoch):
# 获取对应的5个掩膜预测结果
inputs = Variable(input_).cuda()
msk1, msk2, msk3, msk4, msk5 = model.forward(inputs)
gt = Variable(gt.unsqueeze(1)).cuda()
gt_28 = functional.interpolate(gt, size=28, mode='bilinear')
gt_56 = functional.interpolate(gt, size=56, mode='bilinear')
gt_112 = functional.interpolate(gt, size=112, mode='bilinear')
loss = criterion(msk1, gt_28) + criterion(msk2, gt_28) \
+ criterion(msk3, gt_56) + criterion(msk4, gt_112) \
+ criterion(msk5, gt)
model.zero_grad()
loss.backward()
optimizer_feature.step()
title = '{} Epoch {}/{}'.format(
'Training', epoch, args.total_epochs
)
progress_epoch.set_description(
title + ' ' + 'loss:' + str(loss.data.cpu().numpy())
)
# 存储一个epoch后的模型(权重和偏置项), 以便后期使用
filename = ('%s/feature-current.pth' % check_root_feature)
torch.save(model.state_dict(), filename)
# 存储优化器状态
filename_opti = ('%s/opti-current.pth' % check_root_opti)
torch.save(optimizer_feature.state_dict(), filename_opti)
# ==============================TEST===================================
if epoch % args.val_rate == 0:
fileroot = ('%s/feature-current.pth' % check_root_feature)
# 基于torch.save(model.state_dict(), filename)存储方法的对应的恢复方法
model.load_state_dict(torch.load(fileroot))
val_output_root = (prediction_root + '/epoch_current')
if not os.path.exists(val_output_root):
os.makedirs(val_output_root)
print("\ngenerating output images")
for ib, (input_, img_name, _) in enumerate(val_loader):
inputs = Variable(input_).cuda()
_, _, _, _, output = model.forward(inputs)
out = output.data.cpu().numpy()
for i in range(len(img_name)):
print(out[i])
imsave(os.path.join(val_output_root, img_name[i] + '.png'),
out[i, 0], cmap='gray')
print("\nevaluating mae....")
# mean = np.array([0.485, 0.456, 0.406])
# std = np.array([0.229, 0.224, 0.225])
# img = Image.open("./data/ILSVRC2012_test_00000004_224x224.jpg")
# img = np.array(img)
# img = img.astype(np.float64) / 255
# img -= mean
# img /= std
# img = img.transpose(2, 0, 1)
# img = np.array(img)[np.newaxis, :, :, :].astype(np.float32)
# img = torch.from_numpy(img).float()
# inputs = Variable(img).cuda()
# _, _, _, _, output = model.forward(inputs)
# out = output.data.cpu().numpy()
# print(out)
# imsave(os.path.join(val_output_root, 'caffe2_test' + '.png'),
# out[0, 0], cmap='gray')
# 计算F测度和平均绝对误差
F_measure, mae = get_FM(
salpath=val_output_root + '/', gtpath=val_root + '/masks/'
)
evaluation.append([int(epoch), float(F_measure), float(mae)])
result['epoch'].append(int(epoch))
result['F_measure'].append(round(float(F_measure), 3))
result['MAE'].append(round(float(mae), 3))
df = pd.DataFrame(result).set_index('epoch')
df.to_csv('./result.csv')
if epoch == 0:
best = F_measure - mae
elif (F_measure - mae) > best:
best = F_measure - mae
# 存储最好的权重和偏置
filename = ('%s/feature-best.pth' % check_root_feature)
torch.save(model.state_dict(), filename)
# 存储最好的优化器状态
filename_opti = ('%s/opti-best.pth' % check_root_opti)
torch.save(optimizer_feature.state_dict(), filename_opti)
# # 存储最好的完整网络
# filename_opti = ('%s/model-best.pth' % check_root_model)
# torch.save(model, filename_opti)
print("完成一次保存")
# 只在验证期间考虑更改学习率
scheduler.step(best)
print("完成一次测试")
