def make_kwargs(args, seed, loss, loss_params_fn, radius):
name = make_name(seed=seed, loss=loss, radius=radius)
kwargs = app.train_kwargs(args, name)
kwargs.update(
seed=seed,
model_params=dict(
use_desired_size=True,
target_size=64,
desired_template_scale=2.0,
desired_search_radius=1.0,
feature_arch='alexnet',
feature_arch_params=None,
feature_extra_conv_enable=False,
join_type='single',
join_arch='xcorr',
join_params=dict(use_batch_norm=True),
window_params=dict(normalize_method='mean',
window_profile='hann',
combine_method='mul'),
window_radius=1.0,
arg_max_eps=0.01,
wd=1e-4,
loss_params=loss_params_fn(radius),
),
)
return name, kwargs
def make_kwargs(args, seed, opt, opt_config, schedule, schedule_config, init):
name = make_name(seed=seed, opt=opt, schedule=schedule, init=init)
kwargs = app.train_kwargs(args, name)
# Note: Overrides the optimizer specified by args.
kwargs.update(opt_config)
# Note: Overrides the learning rate specified by args.
kwargs.update(lr_init=init, **schedule_config)
kwargs.update(
seed=seed,
model_params=dict(
use_desired_size=True,
target_size=64,
desired_template_scale=2.0,
desired_search_radius=1.0,
feature_arch='alexnet',
feature_arch_params=None,
join_type='single',
join_arch='xcorr',
join_params=dict(use_batch_norm=True),
window_params=dict(normalize_method='mean',
window_profile='hann',
combine_method='mul'),
window_radius=1.0,
arg_max_eps=0.01,
# TODO: Study weight decay and loss config.
wd=1e-4,
loss_params=args.loss_params,
),
)
return name, kwargs
def make_kwargs(args, feat, feat_config, weight, dims, dims_config, seed):
name = make_name(feat=feat, weight=weight, dims=dims, seed=seed)
kwargs = app.train_kwargs(args, name)
kwargs.update(
seed=seed,
model_params=dict(
target_size=args.target_size,
use_desired_size=False,
template_size=dims_config['template_size'],
search_size=dims_config['search_size'],
feature_arch=feat_config.arch,
feature_arch_params=feat_config.arch_params,
feature_extra_conv_enable=feat_config.extra_conv_enable,
feature_extra_conv_params=feat_config.extra_conv_params,
join_type='single',
join_arch='xcorr',
join_params=dict(
learn_spatial_weight=weight,
use_batch_norm=True,
),
window_params=dict(
normalize_method='mean',
window_profile='hann',
combine_method='mul',
),
window_radius=1.0,
arg_max_eps=args.arg_max_eps,
# TODO: Study weight decay and loss config.
wd=1e-4,
loss_params=args.loss_params,
),
)
return name, kwargs
def make_kwargs(args, seed):
name = 'seed_{}'.format(seed)
kwargs = app.train_kwargs(args)
kwargs['params_dict'] = app.train_params_kwargs(args)
kwargs['params_dict']['seed'] = seed
kwargs['params_dict']['model_params'] = args.model_params
kwargs['resume'] = args.resume
return name, kwargs
def make_kwargs(args, seed, join, join_config):
name = make_name(seed=seed, join=join)
kwargs = app.train_kwargs(args, name)
# Note: Overrides the learning rate specified by args.
kwargs.update(
seed=seed,
model_params=dict(
use_desired_size=True,
target_size=64,
desired_template_scale=2.0,
desired_search_radius=1.0,
feature_arch='alexnet',
feature_arch_params=None,
join_type='single',
join_arch=join_config['arch'],
join_params=join_config['params'],
window_params=dict(
normalize_method='mean',
window_profile='hann',
combine_method='mul',
),
window_radius=1.0,
arg_max_eps=0.01,
# TODO: Study weight decay and loss config.
wd=1e-4,
loss_params=dict(
method='sigmoid',
params=dict(
balanced=True,
pos_weight=1,
label_method='hard',
label_params=dict(positive_radius=0.3,
negative_radius=0.3),
),
),
),
)
return name, kwargs