def train(images_folder,
num_refinement_stages,
base_lr,
batch_size,
batches_per_iter,
num_workers,
checkpoint_path,
weights_only,
from_mobilenet,
checkpoints_folder,
log_after,
checkpoint_after,
num_kps,
finetune=False):
net = SinglePersonPoseEstimationWithMobileNet(
num_refinement_stages=num_refinement_stages,
num_heatmaps=num_kps + 1).cuda()
stride = 8
sigma = 7
# num of kps is default 16 ,+bg=17
# the img size is arbitrary , flip may not need
data_flag = "real" if images_folder.split(
"/")[-1] == "data_lip" else "anime"
train_log = get_logger(checkpoints_folder, cmd_stream=True)
if data_flag == "real":
dataset = LipTrainDataset(images_folder,
stride,
sigma,
transform=transforms.Compose([
SinglePersonBodyMasking(),
ChannelPermutation(),
SinglePersonRotate(pad=(128, 128, 128),
max_rotate_degree=40),
SinglePersonCropPad(pad=(128, 128, 128),
crop_x=256,
crop_y=256),
SinglePersonFlip()
]))
else:
dataset = AnimeTrainDataset(
images_folder,
stride,
sigma,
transform=transforms.Compose([
SinglePersonBodyMasking(),
ChannelPermutation(),
SinglePersonRotate(pad=(128, 128, 128), max_rotate_degree=40),
SinglePersonCropPad(pad=(128, 128, 128),
crop_x=256,
crop_y=256)
]))
# b=32 default
train_loader = DataLoader(dataset,
batch_size=batch_size,
shuffle=True,
num_workers=num_workers)
backbone_p = [{
'params': get_parameters_conv(net.model, 'weight')
}, {
'params': get_parameters_conv_depthwise(net.model, 'weight'),
'weight_decay': 0
}, {
'params': get_parameters_bn(net.model, 'weight'),
'weight_decay': 0
}, {
'params': get_parameters_bn(net.model, 'bias'),
'lr': base_lr * 2,
'weight_decay': 0
}]
cpm_p = [{
'params': get_parameters_conv(net.cpm, 'weight'),
'lr': base_lr
}, {
'params': get_parameters_conv(net.cpm, 'bias'),
'lr': base_lr * 2,
'weight_decay': 0
}, {
'params': get_parameters_conv_depthwise(net.cpm, 'weight'),
'weight_decay': 0
}]
initial_p = [{
'params': get_parameters_conv(net.initial_stage, 'weight'),
'lr': base_lr
}, {
'params': get_parameters_conv(net.initial_stage, 'bias'),
'lr': base_lr * 2,
'weight_decay': 0
}, {
'params': get_parameters_bn(net.initial_stage, 'weight'),
'weight_decay': 0
}, {
'params': get_parameters_bn(net.initial_stage, 'bias'),
'lr': base_lr * 2,
'weight_decay': 0
}]
refine_p = [{
'params': get_parameters_conv(net.refinement_stages, 'weight'),
'lr': base_lr * 4
}, {
'params': get_parameters_conv(net.refinement_stages, 'bias'),
'lr': base_lr * 8,
'weight_decay': 0
}, {
'params': get_parameters_bn(net.refinement_stages, 'weight'),
'weight_decay': 0
}, {
'params': get_parameters_bn(net.refinement_stages, 'bias'),
'lr': base_lr * 2,
'weight_decay': 0
}]
opt_p = []
#TODO modify params needed update above and change the model structure.
if not finetune:
opt_p += backbone_p
opt_p += cpm_p
opt_p += initial_p
opt_p += refine_p
optimizer = optim.Adam(opt_p, lr=base_lr, weight_decay=5e-4)
num_iter = 0
current_epoch = 0
scheduler = optim.lr_scheduler.ReduceLROnPlateau(optimizer,
factor=0.1,
patience=5,
threshold=1e-2,
verbose=True)
if checkpoint_path:
checkpoint = torch.load(checkpoint_path)
if from_mobilenet:
load_from_mobilenet(net, checkpoint)
else:
load_state(net, checkpoint)
if not weights_only:
optimizer.load_state_dict(checkpoint['optimizer'])
scheduler.load_state_dict(checkpoint['scheduler'])
num_iter = checkpoint['iter']
num_iter = num_iter // log_after * log_after # round iterations, to print proper loss when resuming
current_epoch = checkpoint['current_epoch'] + 1
net = DataParallel(net, device_ids=[0])
net.train()
for epochId in range(current_epoch, 100):
train_log.debug('Epoch: {}'.format(epochId))
net.train()
total_losses = [0] * (num_refinement_stages + 1
) # heatmaps loss per stage
batch_per_iter_idx = 0
for batch_data in train_loader:
if batch_per_iter_idx == 0:
optimizer.zero_grad()
images = batch_data['image'].cuda()
keypoint_maps = batch_data['keypoint_maps'].cuda()
stages_output = net(images)
losses = []
# guess to update the init stage + refinement stages
for loss_idx in range(len(total_losses)):
losses.append(
l2_loss(stages_output[loss_idx], keypoint_maps,
images.shape[0]))
total_losses[loss_idx] += losses[-1].item() / batches_per_iter
loss = losses[0]
for loss_idx in range(1, len(losses)):
loss += losses[loss_idx]
loss /= batches_per_iter
loss.backward()
batch_per_iter_idx += 1
if batch_per_iter_idx == batches_per_iter:
optimizer.step()
batch_per_iter_idx = 0
num_iter += 1
else:
continue
#per 100 iter
if num_iter % log_after == 0:
train_log.debug('Iter: {}'.format(num_iter))
for loss_idx in range(len(total_losses)):
train_log.debug('\n'.join([
'stage{}_heatmaps_loss: {}'
]).format(loss_idx + 1,
total_losses[loss_idx] / log_after))
for loss_idx in range(len(total_losses)):
total_losses[loss_idx] = 0
snapshot_name = '{}/checkpoint_last_epoch.pth'.format(
checkpoints_folder)
torch.save(
{
'state_dict': net.module.state_dict(),
'optimizer': optimizer.state_dict(),
'scheduler': scheduler.state_dict(),
'iter': num_iter,
'current_epoch': epochId
}, snapshot_name)
if (epochId + 1) % checkpoint_after == 0:
snapshot_name = '{}/checkpoint_epoch_{}.pth'.format(
checkpoints_folder, epochId)
torch.save(
{
'state_dict': net.module.state_dict(),
'optimizer': optimizer.state_dict(),
'scheduler': scheduler.state_dict(),
'iter': num_iter,
'current_epoch': epochId
}, snapshot_name)
train_log.debug('Validation...')
net.eval()
eval_num = 1000
if data_flag == "real":
val_dataset = LipValDataset(images_folder, eval_num)
else:
val_dataset = AnimeValDataset(images_folder, eval_num)
predictions_name = '{}/val_results.csv'.format(checkpoints_folder)
evaluate(val_dataset, predictions_name, net, num_kps=num_kps)
pck = calc_pckh(val_dataset.labels_file_path,
predictions_name,
eval_num=eval_num)
val_loss = 100 - pck[-1][-1] # 100 - avg_pckh
train_log.debug('Val loss: {}'.format(val_loss))
scheduler.step(val_loss, epochId)
def train(images_folder, num_refinement_stages, base_lr, batch_size, batches_per_iter,
num_workers, checkpoint_path, weights_only, from_mobilenet, checkpoints_folder,
log_after, checkpoint_after):
dataset = CocoSingleTrainDataset(images_folder,
transform=transforms.Compose([
HalfBodyTransform(),
RandomScaleRotate(),
SinglePersonFlip(left_keypoints_indice=
CocoSingleTrainDataset.left_keypoints_indice,
right_keypoints_indice=
CocoSingleTrainDataset.right_keypoints_indice),
SinglePersonRandomAffineTransform(),
SinglePersonBodyMasking(),
Normalization(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
ChannelPermutation()
]))
net = SinglePersonPoseEstimationWithMobileNet(num_refinement_stages, num_heatmaps=dataset._num_keypoints,
mode='nearest').cuda()
train_loader = DataLoader(dataset, batch_size=batch_size, shuffle=True, num_workers=num_workers)
optimizer = optim.Adam(net.parameters(), lr=base_lr)
scheduler = optim.lr_scheduler.MultiStepLR(optimizer, [170, 200], 0.1)
num_iter = 0
current_epoch = 0
if checkpoint_path:
checkpoint = torch.load(checkpoint_path)
if from_mobilenet:
load_from_mobilenet(net, checkpoint)
else:
load_state(net, checkpoint)
if not weights_only:
optimizer.load_state_dict(checkpoint['optimizer'])
scheduler.load_state_dict(checkpoint['scheduler'])
num_iter = checkpoint['iter']
current_epoch = checkpoint['current_epoch']+1
net = DataParallel(net)
net.train()
for epochId in range(current_epoch, 210):
print('Epoch: {}'.format(epochId))
net.train()
total_losses = [0] * (num_refinement_stages + 1) # heatmaps loss per stage
batch_per_iter_idx = 0
for batch_data in train_loader:
if batch_per_iter_idx == 0:
optimizer.zero_grad()
images = batch_data['image'].float().cuda()
keypoint_maps = batch_data['keypoint_maps']
stages_output = net(images)
losses = []
for loss_idx in range(len(total_losses)):
losses.append(mse_loss(stages_output[loss_idx], keypoint_maps,
batch_data['keypoints'][:, 2::3].view(batch_data['keypoints'].shape[0], -1, 1)))
total_losses[loss_idx] += losses[-1].item() / batches_per_iter
loss = 0
for loss_idx in range(len(losses)):
loss += losses[loss_idx]
loss /= batches_per_iter
loss.backward()
batch_per_iter_idx += 1
if batch_per_iter_idx == batches_per_iter:
optimizer.step()
batch_per_iter_idx = 0
num_iter += 1
else:
continue
if num_iter % log_after == 0:
print('Iter: {}'.format(num_iter))
for loss_idx in range(len(total_losses)):
print('\n'.join(['stage{}_heatmaps_loss: {}']).format(
loss_idx + 1, total_losses[loss_idx] / log_after))
for loss_idx in range(len(total_losses)):
total_losses[loss_idx] = 0
snapshot_name = '{}/checkpoint_last_epoch.pth'.format(checkpoints_folder)
torch.save({'state_dict': net.module.state_dict(),
'optimizer': optimizer.state_dict(),
'scheduler': scheduler.state_dict(),
'iter': num_iter,
'current_epoch': epochId},
snapshot_name)
if (epochId + 1) % checkpoint_after == 0:
snapshot_name = '{}/checkpoint_epoch_{}.pth'.format(checkpoints_folder, epochId)
torch.save({'state_dict': net.module.state_dict(),
'optimizer': optimizer.state_dict(),
'scheduler': scheduler.state_dict(),
'iter': num_iter,
'current_epoch': epochId},
snapshot_name)
print('Validation...')
net.eval()
val_dataset = CocoSingleValDataset(images_folder, transform=transforms.Compose([
SinglePersonRandomAffineTransform(mode='val'),
Normalization(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])]))
predictions_name = '{}/val_results2.json'.format(checkpoints_folder)
val_loss = val(net, val_dataset, predictions_name, 'CocoSingle')
print('Val loss: {}'.format(val_loss))
scheduler.step()
def train(images_folder, num_refinement_stages, base_lr, batch_size,
batches_per_iter, num_workers, checkpoint_path, weights_only,
from_mobilenet, checkpoints_folder, log_after, checkpoint_after):
net = SinglePersonPoseEstimationWithMobileNet(num_refinement_stages,
num_heatmaps=18).cuda()
train_dataset = dtst_train(images_folder,
STRIDE,
SIGMA,
transform=transforms.Compose([
SinglePersonBodyMasking(),
ChannelPermutation(),
SinglePersonRotate(pad=(128, 128, 128),
max_rotate_degree=40),
SinglePersonCropPad(pad=(128, 128, 128),
crop_x=256,
crop_y=256),
SinglePersonFlip()
]))
train_loader = DataLoader(train_dataset,
batch_size=batch_size,
shuffle=True,
num_workers=num_workers)
val_dataset = dtst_val(images_folder, STRIDE, SIGMA)
val_loader = DataLoader(val_dataset,
batch_size=batch_size,
shuffle=False,
num_workers=num_workers)
optimizer = optim.Adam([
{
'params': get_parameters_conv(net.model, 'weight')
},
{
'params': get_parameters_conv_depthwise(net.model, 'weight'),
'weight_decay': 0
},
{
'params': get_parameters_bn(net.model, 'weight'),
'weight_decay': 0
},
{
'params': get_parameters_bn(net.model, 'bias'),
'lr': base_lr * 2,
'weight_decay': 0
},
{
'params': get_parameters_conv(net.cpm, 'weight'),
'lr': base_lr
},
{
'params': get_parameters_conv(net.cpm, 'bias'),
'lr': base_lr * 2,
'weight_decay': 0
},
{
'params': get_parameters_conv_depthwise(net.cpm, 'weight'),
'weight_decay': 0
},
{
'params': get_parameters_conv(net.initial_stage, 'weight'),
'lr': base_lr
},
{
'params': get_parameters_conv(net.initial_stage, 'bias'),
'lr': base_lr * 2,
'weight_decay': 0
},
{
'params': get_parameters_bn(net.initial_stage, 'weight'),
'weight_decay': 0
},
{
'params': get_parameters_bn(net.initial_stage, 'bias'),
'lr': base_lr * 2,
'weight_decay': 0
},
{
'params': get_parameters_conv(net.refinement_stages, 'weight'),
'lr': base_lr * 4
},
{
'params': get_parameters_conv(net.refinement_stages, 'bias'),
'lr': base_lr * 8,
'weight_decay': 0
},
{
'params': get_parameters_bn(net.refinement_stages, 'weight'),
'weight_decay': 0
},
{
'params': get_parameters_bn(net.refinement_stages, 'bias'),
'lr': base_lr * 2,
'weight_decay': 0
},
],
lr=base_lr,
weight_decay=5e-4)
num_iter = 0
current_epoch = 0
scheduler = optim.lr_scheduler.ReduceLROnPlateau(optimizer,
factor=0.1,
patience=5,
threshold=1e-2,
verbose=True)
if checkpoint_path:
checkpoint = torch.load(checkpoint_path)
if from_mobilenet:
load_from_mobilenet(net, checkpoint)
else:
load_state(net, checkpoint)
if not weights_only:
optimizer.load_state_dict(checkpoint['optimizer'])
scheduler.load_state_dict(checkpoint['scheduler'])
num_iter = checkpoint['iter']
num_iter = num_iter // log_after * log_after # round iterations, to print proper loss when resuming
current_epoch = checkpoint['current_epoch'] + 1
net = DataParallel(net)
net.train()
for epochId in range(current_epoch, 100):
print('Epoch: {}'.format(epochId))
N_losses = num_refinement_stages + 1
total_losses = [0] * N_losses # heatmaps loss per stage
for batch in train_loader:
images = batch['image'].cuda()
keypoint_maps = batch['keypoint_maps'].cuda()
stages_output = net(images)
losses = []
for loss_idx in range(N_losses):
loss = l2_loss(stages_output[loss_idx], keypoint_maps,
len(images))
losses.append(loss)
total_losses[loss_idx] += loss.item()
optimizer.zero_grad()
loss = losses[0]
for i in range(1, N_losses):
loss += losses[i]
loss.backward()
optimizer.step()
num_iter += 1
if num_iter % log_after == 0:
print('Iter: {}'.format(num_iter))
# for loss_idx in range(N_losses):
# print('\n'.join(['stage{}_heatmaps_loss: {}']).format(
# loss_idx + 1, total_losses[loss_idx] / log_after))
for loss_idx in range(N_losses):
total_losses[loss_idx] = 0
validate2(epochId, net, val_loader, scheduler)
snapshot_name = '{}/{}_epoch_last.pth'.format(checkpoints_folder,
DATASET)
torch.save(
{
'state_dict': net.module.state_dict(),
'optimizer': optimizer.state_dict(),
'scheduler': scheduler.state_dict(),
'iter': num_iter,
'current_epoch': epochId
}, snapshot_name)
if epochId % checkpoint_after == 0:
snapshot_name = '{}/{}_epoch_{}.pth'.format(
checkpoints_folder, DATASET, epochId)
torch.save(
{
'state_dict': net.module.state_dict(),
'optimizer': optimizer.state_dict(),
'scheduler': scheduler.state_dict(),
'iter': num_iter,
'current_epoch': epochId
}, snapshot_name)
validate2(epochID, net, val_loader, scheduler)
def train(
images_folder,
num_refinement_stages,
base_lr,
batch_size,
batches_per_iter,
num_workers,
checkpoint_path,
weights_only,
from_mobilenet,
checkpoints_folder,
log_after,
checkpoint_after,
):
net = SinglePersonPoseEstimationWithMobileNet(num_refinement_stages).cuda()
stride = 8
sigma = 7
dataset = LipTrainDataset(
images_folder,
stride,
sigma,
transform=transforms.Compose([
SinglePersonBodyMasking(),
ChannelPermutation(),
SinglePersonRotate(pad=(128, 128, 128), max_rotate_degree=40),
SinglePersonCropPad(pad=(128, 128, 128), crop_x=256, crop_y=256),
SinglePersonFlip(
left_keypoints_indice=LipTrainDataset.left_keypoints_indice,
right_keypoints_indice=LipTrainDataset.right_keypoints_indice,
),
]),
)
train_loader = DataLoader(dataset,
batch_size=batch_size,
shuffle=True,
num_workers=num_workers)
optimizer = optim.Adam(
[
{
"params": get_parameters_conv(net.model, "weight")
},
{
"params": get_parameters_conv_depthwise(net.model, "weight"),
"weight_decay": 0,
},
{
"params": get_parameters_bn(net.model, "weight"),
"weight_decay": 0
},
{
"params": get_parameters_bn(net.model, "bias"),
"lr": base_lr * 2,
"weight_decay": 0,
},
{
"params": get_parameters_conv(net.cpm, "weight"),
"lr": base_lr
},
{
"params": get_parameters_conv(net.cpm, "bias"),
"lr": base_lr * 2,
"weight_decay": 0,
},
{
"params": get_parameters_conv_depthwise(net.cpm, "weight"),
"weight_decay": 0,
},
{
"params": get_parameters_conv(net.initial_stage, "weight"),
"lr": base_lr
},
{
"params": get_parameters_conv(net.initial_stage, "bias"),
"lr": base_lr * 2,
"weight_decay": 0,
},
{
"params": get_parameters_bn(net.initial_stage, "weight"),
"weight_decay": 0,
},
{
"params": get_parameters_bn(net.initial_stage, "bias"),
"lr": base_lr * 2,
"weight_decay": 0,
},
{
"params": get_parameters_conv(net.refinement_stages, "weight"),
"lr": base_lr * 4,
},
{
"params": get_parameters_conv(net.refinement_stages, "bias"),
"lr": base_lr * 8,
"weight_decay": 0,
},
{
"params": get_parameters_bn(net.refinement_stages, "weight"),
"weight_decay": 0,
},
{
"params": get_parameters_bn(net.refinement_stages, "bias"),
"lr": base_lr * 2,
"weight_decay": 0,
},
],
lr=base_lr,
weight_decay=5e-4,
)
num_iter = 0
current_epoch = 0
scheduler = optim.lr_scheduler.ReduceLROnPlateau(optimizer,
factor=0.1,
patience=5,
threshold=1e-2,
verbose=True)
if checkpoint_path:
checkpoint = torch.load(checkpoint_path)
if from_mobilenet:
load_from_mobilenet(net, checkpoint)
else:
load_state(net, checkpoint)
if not weights_only:
optimizer.load_state_dict(checkpoint["optimizer"])
scheduler.load_state_dict(checkpoint["scheduler"])
num_iter = checkpoint["iter"]
current_epoch = checkpoint["current_epoch"] + 1
net = DataParallel(net)
net.train()
for epochId in range(current_epoch, 100):
print("Epoch: {}".format(epochId))
net.train()
total_losses = [0] * (num_refinement_stages + 1
) # heatmaps loss per stage
batch_per_iter_idx = 0
for batch_data in train_loader:
if batch_per_iter_idx == 0:
optimizer.zero_grad()
images = batch_data["image"].cuda()
keypoint_maps = batch_data["keypoint_maps"].cuda()
stages_output = net(images)
losses = []
for loss_idx in range(len(total_losses)):
losses.append(
l2_loss(stages_output[loss_idx], keypoint_maps,
images.shape[0]))
total_losses[loss_idx] += losses[-1].item() / batches_per_iter
loss = losses[0]
for loss_idx in range(1, len(losses)):
loss += losses[loss_idx]
loss /= batches_per_iter
loss.backward()
batch_per_iter_idx += 1
if batch_per_iter_idx == batches_per_iter:
optimizer.step()
batch_per_iter_idx = 0
num_iter += 1
else:
continue
if num_iter % log_after == 0:
print("Iter: {}".format(num_iter))
for loss_idx in range(len(total_losses)):
print("\n".join(["stage{}_heatmaps_loss: {}"]).format(
loss_idx + 1, total_losses[loss_idx] / log_after))
for loss_idx in range(len(total_losses)):
total_losses[loss_idx] = 0
snapshot_name = "{}/checkpoint_last_epoch.pth".format(
checkpoints_folder)
torch.save(
{
"state_dict": net.module.state_dict(),
"optimizer": optimizer.state_dict(),
"scheduler": scheduler.state_dict(),
"iter": num_iter,
"current_epoch": epochId,
},
snapshot_name,
)
if (epochId + 1) % checkpoint_after == 0:
snapshot_name = "{}/checkpoint_epoch_{}.pth".format(
checkpoints_folder, epochId)
torch.save(
{
"state_dict": net.module.state_dict(),
"optimizer": optimizer.state_dict(),
"scheduler": scheduler.state_dict(),
"iter": num_iter,
"current_epoch": epochId,
},
snapshot_name,
)
print("Validation...")
net.eval()
eval_num = 1000
val_dataset = LipValDataset(images_folder, eval_num)
predictions_name = "{}/val_results.csv".format(checkpoints_folder)
evaluate(val_dataset, predictions_name, net)
pck = calc_pckh(val_dataset.labels_file_path,
predictions_name,
eval_num=eval_num)
val_loss = 100 - pck[-1][-1] # 100 - avg_pckh
print("Val loss: {}".format(val_loss))
scheduler.step(val_loss, epochId)
def train(images_folder, num_refinement_stages, base_lr, batch_size,
batches_per_iter, num_workers, checkpoint_path, weights_only,
from_mobilenet, checkpoints_folder, log_after, checkpoint_after):
net = SinglePersonPoseEstimationWithMobileNet(num_refinement_stages).cuda()
stride = 8
sigma = 7
dataset = LipTrainDataset(images_folder,
stride,
sigma,
transform=transforms.Compose([
SinglePersonBodyMasking(),
ChannelPermutation(),
SinglePersonRotate(pad=(128, 128, 128),
max_rotate_degree=40),
SinglePersonCropPad(pad=(128, 128, 128),
crop_x=256,
crop_y=256),
SinglePersonFlip()
]))
train_loader = DataLoader(dataset,
batch_size=batch_size,
shuffle=True,
num_workers=num_workers)
optimizer = optim.Adam([
{
'params': get_parameters_conv(net.model, 'weight')
},
{
'params': get_parameters_conv_depthwise(net.model, 'weight'),
'weight_decay': 0
},
{
'params': get_parameters_bn(net.model, 'weight'),
'weight_decay': 0
},
{
'params': get_parameters_bn(net.model, 'bias'),
'lr': base_lr * 2,
'weight_decay': 0
},
{
'params': get_parameters_conv(net.cpm, 'weight'),
'lr': base_lr
},
{
'params': get_parameters_conv(net.cpm, 'bias'),
'lr': base_lr * 2,
'weight_decay': 0
},
{
'params': get_parameters_conv_depthwise(net.cpm, 'weight'),
'weight_decay': 0
},
{
'params': get_parameters_conv(net.initial_stage, 'weight'),
'lr': base_lr
},
{
'params': get_parameters_conv(net.initial_stage, 'bias'),
'lr': base_lr * 2,
'weight_decay': 0
},
{
'params': get_parameters_bn(net.initial_stage, 'weight'),
'weight_decay': 0
},
{
'params': get_parameters_bn(net.initial_stage, 'bias'),
'lr': base_lr * 2,
'weight_decay': 0
},
{
'params': get_parameters_conv(net.refinement_stages, 'weight'),
'lr': base_lr * 4
},
{
'params': get_parameters_conv(net.refinement_stages, 'bias'),
'lr': base_lr * 8,
'weight_decay': 0
},
{
'params': get_parameters_bn(net.refinement_stages, 'weight'),
'weight_decay': 0
},
{
'params': get_parameters_bn(net.refinement_stages, 'bias'),
'lr': base_lr * 2,
'weight_decay': 0
},
],
lr=base_lr,
weight_decay=5e-4)
num_iter = 0
current_epoch = 0
scheduler = optim.lr_scheduler.ReduceLROnPlateau(optimizer,
factor=0.1,
patience=5,
threshold=1e-2,
verbose=True)
if checkpoint_path:
checkpoint = torch.load(checkpoint_path)
if from_mobilenet:
load_from_mobilenet(net, checkpoint)
else:
load_state(net, checkpoint)
if not weights_only:
optimizer.load_state_dict(checkpoint['optimizer'])
scheduler.load_state_dict(checkpoint['scheduler'])
num_iter = checkpoint['iter']
num_iter = num_iter // log_after * log_after # round iterations, to print proper loss when resuming
current_epoch = checkpoint['current_epoch'] + 1
net = DataParallel(net)
net.train()
for epochId in range(current_epoch, 100):
print('Epoch: {}'.format(epochId))
net.train()
total_losses = [0] * (num_refinement_stages + 1
) # heatmaps loss per stage
batch_per_iter_idx = 0
for batch_data in train_loader:
if batch_per_iter_idx == 0:
optimizer.zero_grad()
images = batch_data['image'].cuda()
keypoint_maps = batch_data['keypoint_maps'].cuda()
stages_output = net(images)
losses = []
for loss_idx in range(len(total_losses)):
losses.append(
l2_loss(stages_output[loss_idx], keypoint_maps,
images.shape[0]))
total_losses[loss_idx] += losses[-1].item() / batches_per_iter
loss = losses[0]
for loss_idx in range(1, len(losses)):
loss += losses[loss_idx]
loss /= batches_per_iter
loss.backward()
batch_per_iter_idx += 1
if batch_per_iter_idx == batches_per_iter:
optimizer.step()
batch_per_iter_idx = 0
num_iter += 1
else:
continue
if num_iter % log_after == 0:
print('Iter: {}'.format(num_iter))
for loss_idx in range(len(total_losses)):
print('\n'.join(['stage{}_heatmaps_loss: {}']).format(
loss_idx + 1, total_losses[loss_idx] / log_after))
for loss_idx in range(len(total_losses)):
total_losses[loss_idx] = 0
snapshot_name = '{}/checkpoint_last_epoch.pth'.format(
checkpoints_folder)
torch.save(
{
'state_dict': net.module.state_dict(),
'optimizer': optimizer.state_dict(),
'scheduler': scheduler.state_dict(),
'iter': num_iter,
'current_epoch': epochId
}, snapshot_name)
if (epochId + 1) % checkpoint_after == 0:
snapshot_name = '{}/checkpoint_epoch_{}.pth'.format(
checkpoints_folder, epochId + 1)
torch.save(
{
'state_dict': net.module.state_dict(),
'optimizer': optimizer.state_dict(),
'scheduler': scheduler.state_dict(),
'iter': num_iter,
'current_epoch': epochId
}, snapshot_name)
print('Validation...')
net.eval()
eval_num = 1000
val_dataset = LipValDataset(images_folder, eval_num)
predictions_name = '{}/val_results.csv'.format(checkpoints_folder)
evaluate(val_dataset, predictions_name, net)
pck = calc_pckh(val_dataset.labels_file_path,
predictions_name,
eval_num=eval_num)
val_loss = 100 - pck[-1][-1] # 100 - avg_pckh
print('Val loss: {}'.format(val_loss))
scheduler.step(val_loss, epochId)