DNN/Human Pose
HumanPose Optimizer setting
Small Octopus
2020. 1. 10. 11:25
Convolutional Pose Machine
CPM (https://github.com/namedBen/Convolutional-Pose-Machines-Pytorch/blob/master/config/config.yml)
def adjust_learning_rate(optimizer, iters, base_lr, policy_parameter, policy='step', multiple=[1]):
if policy == 'fixed':
lr = base_lr
elif policy == 'step':
lr = base_lr * (policy_parameter['gamma'] ** (iters // policy_parameter['step_size']))
elif policy == 'exp':
lr = base_lr * (policy_parameter['gamma'] ** iters)
elif policy == 'inv':
lr = base_lr * ((1 + policy_parameter['gamma'] * iters) ** (-policy_parameter['power']))
elif policy == 'multistep':
lr = base_lr
for stepvalue in policy_parameter['stepvalue']:
if iters >= stepvalue:
lr *= policy_parameter['gamma']
else:
break
elif policy == 'poly':
lr = base_lr * ((1 - iters * 1.0 / policy_parameter['max_iter']) ** policy_parameter['power'])
elif policy == 'sigmoid':
lr = base_lr * (1.0 / (1 + math.exp(-policy_parameter['gamma'] * (iters - policy_parameter['stepsize']))))
elif policy == 'multistep-poly':
lr = base_lr
stepstart = 0
stepend = policy_parameter['max_iter']
for stepvalue in policy_parameter['stepvalue']:
if iters >= stepvalue:
lr *= policy_parameter['gamma']
stepstart = stepvalue
else:
stepend = stepvalue
break
lr = max(lr * policy_parameter['gamma'], lr * (1 - (iters - stepstart) * 1.0 / (stepend - stepstart)) ** policy_parameter['power'])
for i, param_group in enumerate(optimizer.param_groups):
param_group['lr'] = lr * multiple[i]
return lr
def get_parameters(model, config, isdefault=True):
if isdefault:
return model.parameters(), [1.]
lr_1 = []
lr_2 = []
lr_4 = []
lr_8 = []
params_dict = dict(model.module.named_parameters())
for key, value in params_dict.items():
if ('model1_' not in key) and ('model0.' not in key):
if key[-4:] == 'bias':
lr_8.append(value)
else:
lr_4.append(value)
elif key[-4:] == 'bias':
lr_2.append(value)
else:
lr_1.append(value)
params = [{'params': lr_1, 'lr': config.base_lr},
{'params': lr_2, 'lr': config.base_lr * 2.},
{'params': lr_4, 'lr': config.base_lr * 4.},
{'params': lr_8, 'lr': config.base_lr * 8.}]
return params, [1., 2., 4., 8.]
params, multiple = get_parameters(model, config, False)
optimizer = torch.optim.SGD(params, config.base_lr, momentum=config.momentum,
weight_decay=config.weight_decay)
while iters < config.max_iter:
for i, (input, heatmap, centermap) in enumerate(train_loader):
learning_rate = adjust_learning_rate(optimizer, iters, config.base_lr, policy=config.lr_policy,
policy_parameter=config.policy_parameter, multiple=multiple)| optimizer: SGD |
| lr_policy: 'step' |
| policy_parameter: gamma: 0.333, step_size: 13275 |
| weight_decay: 0.0005 |
| momentum: 0.9 |
| max_iter: 62500 |
| batch_size: 16 |
| base_lr: 0.000004, 0.4e-5 |
| weight init: he |
Stacked Hourglass
SH(https://github.com/bearpaw/pytorch-pose/blob/master/example/main.py)
optimizer = torch.optim.RMSprop(model.parameters(),
lr=args.lr,
momentum=args.momentum,
weight_decay=args.weight_decay)
| optimizer: RMSprop |
| base_lr: 2.5e-4 |
| momentum: 0 |
| weight_decay: 0 |
| batchsize: 6 |
Pose ResNet50
| optimizer: Adam |
| base_lr: 1e-3 |
| momentum: 0 |
| weight_decay: 0 |
| batchsize: 32 |
| epoch: 140 |
| schedule: 90, 120 |
(https://github.com/princeton-vl/pytorch_stacked_hourglass/blob/master/task/pose.py)
| optimizer: Adam |
| base_lr: 1e-3 |
| decay_lr: 2e-4 |
| decay_iters: 1000 |
| batchsize: 6 |
Pytorch-Human-Pose-Estimation
Optimizer = getattr(torch.optim, self.opts.optimizer_type)(TrainableParams, lr = self.opts.LR, alpha = 0.99, eps = 1e-8)
for epoch in range(startepoch, endepoch+1):
adjust_learning_rate(self.optimizer, epoch, self.opts.dropLR, self.opts.dropMag)
def adjust_learning_rate(optimizer, epoch, dropLR, dropMag):
if epoch%dropLR==0:
lrfac = dropMag
else:
lrfac = 1
for i,param_group in enumerate(optimizer.param_groups):
if lrfac!=1:
print("Reducing learning rate of group %d from %f to %f"%(i,param_group['lr'],param_group['lr']*lrfac))
param_group['lr'] *= lrfac| model = StackedHourGlass |
| TargetType: heatmap |
| outputRes: 64 |
| nChannels: 256 |
| nStack: 2 |
| nModules: 2 |
| nReductions: 4 |
| data_loader_size = 6 |
| mini_batch_count: 1 |
| optimizer_type: RMSprop |
| LR: 2.5e-4 |
| dropLR: 50 |
| dropMag: 0.1 |
| valInterval: 4 |
| saveInterval: 8 |
| model = PyraNet |
| TargetType: heatmap |
| outputRes: 64 |
| nChannels: 256 |
| nStack: 2 |
| nModules: 2 |
| nReductions: 4 |
| baseWidth: 6 |
| cardinality:30 |
| data_loader_size = 6 |
| mini_batch_count: 1 |
| optimizer_type: RMSprop |
| LR: 2.5e-4 |
| dropLR: 50 |
| dropMag: 0.1 |
| valInterval: 4 |
| saveInterval: 8 |
| model = StackedHourGlass |
| TargetType: heatmap |
| outputRes: 64 |
| nChannels: 256 |
| nStack: 2 |
| nModules: 2 |
| nReductions: 4 |
| LRNSize : 1 |
| IterSize : 3 |
| data_loader_size = 6 |
| mini_batch_count: 1 |
| optimizer_type: RMSprop |
| LR: 2.5e-4 |
| dropLR: 50 |
| dropMag: 0.1 |
| valInterval: 4 |
| saveInterval: 8 |
| model = DeepPose |
| TargetType = direct |
| ### DeepPose |
| baseName: resnet34 |
| data_loader_size = 16 |
| mini_batch_count: 1 |
| optimizer_type: RMSprop |
| LR: 2.5e-4 |
| dropLR: 5 |
| dropMag: 0.7 |
| valInterval: 3 |
| saveInterval: 5 |
| outputRes = 32 |
| model = ChainedPredictions |
| TargetType = heatmap |
| outputRes = 32 |
| modelName: resnet34 |
| ohKernel: 1 |
| hhKernel: 1 |
| data_loader_size = 16 |
| mini_batch_count: 1 |
| optimizer_type: RMSprop |
| LR: 2.5e-4 |
| dropLR: 50 |
| dropMag: 0.1 |
| valInterval: 4 |
| saveInterval: 8 |
| outputRes = 32 |