def run(settings):
# Most common settings are assigned in the settings struct
settings.device = 'cuda'
settings.description = 'TransT with default settings.'
settings.batch_size = 24 # 38
settings.num_workers = min(settings.batch_size, 8)
settings.multi_gpu = False
settings.print_interval = 1
settings.normalize_mean = [0.485, 0.456, 0.406]
settings.normalize_std = [0.229, 0.224, 0.225]
settings.search_area_factor = 4.0
settings.template_area_factor = 2.0
settings.search_feature_sz = 32
settings.template_feature_sz = 16
settings.search_sz = settings.search_feature_sz * 8
settings.temp_sz = settings.template_feature_sz * 8
settings.center_jitter_factor = {'search': 3, 'template': 0}
settings.scale_jitter_factor = {'search': 0.25, 'template': 0}
settings.sequence_length = 8
settings.search_gap = 5
settings.init_ckpt = "pytracking/networks/transt.pth"
# Transformer
settings.position_embedding = 'sine'
settings.hidden_dim = 256
settings.dropout = 0.1
settings.nheads = 8
settings.dim_feedforward = 2048
settings.featurefusion_layers = 4
settings.sigma = 1 / 4 / 5.
settings.kernel = 4
settings.feature = 32 # 18
settings.output_sz = 256 # settings.feature * 16
settings.end_pad_if_even = False
settings.label_function_params = True
# Train datasets
# lasot_train = Lasot(settings.env.lasot_dir, split='train')
got10k_train = Got10k(settings.env.got10k_dir, split='vottrain')
# trackingnet_train = TrackingNet(settings.env.trackingnet_dir, set_ids=list(range(4)))
# coco_train = MSCOCOSeq(settings.env.coco_dir)
# The joint augmentation transform, that is applied to the pairs jointly
transform_joint = tfm.Transform(tfm.ToGrayscale(probability=0.05))
# The augmentation transform applied to the training set (individually to each image in the pair)
transform_train = tfm.Transform(tfm.ToTensorAndJitter(0.2),
tfm.Normalize(mean=settings.normalize_mean, std=settings.normalize_std))
# Data processing to do on the training pairs
data_processing_train = processing.TransTProcessing(search_area_factor=settings.search_area_factor,
template_area_factor = settings.template_area_factor,
search_sz=settings.search_sz,
temp_sz=settings.temp_sz,
center_jitter_factor=settings.center_jitter_factor,
scale_jitter_factor=settings.scale_jitter_factor,
mode='sequence',
transform=transform_train,
label_function_params=settings.label_function_params,
joint_transform=transform_joint)
# The sampler for training
dataset_train = sampler.TransTSampler([got10k_train], [1], samples_per_epoch=1000*settings.batch_size, max_gap=100, processing=data_processing_train, num_search_frames=settings.sequence_length, frame_sample_mode="causal")
# dataset_train = sampler.TransTSampler([lasot_train, got10k_train, coco_train, trackingnet_train], [1,1,1,1], samples_per_epoch=1000*settings.batch_size, max_gap=100, processing=data_processing_train)
# The loader for training
loader_train = LTRLoader('train', dataset_train, training=True, batch_size=settings.batch_size, num_workers=settings.num_workers,
shuffle=True, drop_last=True, stack_dim=0)
# Create network and actor
model = transt_models.transt_resnet50(settings)
# Wrap the network for multi GPU training
if settings.multi_gpu:
model = MultiGPU(model, dim=0)
objective = transt_models.transt_loss(settings)
n_parameters = sum(p.numel() for p in model.parameters() if p.requires_grad)
print('number of params:', n_parameters)
actor = actors.CircuitTranstActor(net=model, objective=objective)
# Optimizer
# Change learning rate forthe Q we have changed and the RNN and the readout
# q = self.mix_q(torch.cat([q, self.mix_norm(exc)], -1))
# self.class_embed_new = MLP(hidden_dim * 2, hidden_dim, num_classes + 1, 3)
# self.bbox_embed_new = MLP(hidden_dim * 2, hidden_dim, 4, 3)
param_dicts = [
# {"params": [p for n, p in model.named_parameters() if "backbone" not in n and p.requires_grad and "circuit" not in n and "mix" not in n and "new" not in n]},
# {
# "params": [p for n, p in model.named_parameters() if "backbone" in n and p.requires_grad],
# "lr": 1e-5,
# },
{
"params": [p for n, p in model.named_parameters() if "circuit" in n or "mix" in n or "new" in n], # or "class_embed" in n or "bbox_embed" in n],
"lr": 1e-3,
},
# {
# "params": [p for n, p in model.named_parameters() if "class_embed" in n or "bbox_embed" in n], # if "circuit" in n or "mix" in n or "new" in n or "class_embed" in n or "bbox_embed" in n],
# "lr": 1e-5,
# },
]
for n, p in model.named_parameters():
if "circuit" in n or "mix" in n or "new" in n or "class_embed" in n or "bbox_embed" in n or "decoder" in n:
pass
else:
p.requires_grad = False
print("Removing grad on {}".format(n))
optimizer = torch.optim.AdamW(param_dicts, lr=1e-5, # lr=1e-4,
weight_decay=1e-4)
lr_scheduler = torch.optim.lr_scheduler.StepLR(optimizer, 500)
# Create trainer
trainer = LTRTrainer(actor, [loader_train], optimizer, settings, lr_scheduler)
# Run training (set fail_safe=False if you are debugging)
trainer.train(1000, load_latest=True, fail_safe=True)
def run(settings):
settings.description = 'Default train settings for training full network'
settings.batch_size = 20
settings.num_workers = 8
settings.multi_gpu = True
settings.print_interval = 1
settings.normalize_mean = [102.9801, 115.9465, 122.7717]
settings.normalize_std = [1.0, 1.0, 1.0]
settings.feature_sz = (52, 30)
# Settings used for generating the image crop input to the network. See documentation of LWTLProcessing class in
# ltr/data/processing.py for details.
settings.output_sz = (settings.feature_sz[0] * 16,
settings.feature_sz[1] * 16
) # Size of input image crop
settings.search_area_factor = 5.0
settings.crop_type = 'inside_major'
settings.max_scale_change = None
settings.center_jitter_factor = {'train': 3, 'test': (5.5, 4.5)}
settings.scale_jitter_factor = {'train': 0.25, 'test': 0.5}
# Datasets
ytvos_train = YouTubeVOS(version="2019", multiobj=False, split='jjtrain')
davis_train = Davis(version='2017', multiobj=False, split='train')
ytvos_val = YouTubeVOS(version="2019", multiobj=False, split='jjvalid')
# Data transform
transform_joint = tfm.Transform(tfm.ToBGR(),
tfm.ToGrayscale(probability=0.05),
tfm.RandomHorizontalFlip(probability=0.5))
transform_train = tfm.Transform(
tfm.RandomAffine(p_flip=0.0,
max_rotation=15.0,
max_shear=0.0,
max_ar_factor=0.0,
max_scale=0.2,
pad_amount=0),
tfm.ToTensorAndJitter(0.2, normalize=False),
tfm.Normalize(mean=settings.normalize_mean,
std=settings.normalize_std))
transform_val = tfm.Transform(
tfm.ToTensorAndJitter(0.0, normalize=False),
tfm.Normalize(mean=settings.normalize_mean,
std=settings.normalize_std))
data_processing_train = processing.LWLProcessing(
search_area_factor=settings.search_area_factor,
output_sz=settings.output_sz,
center_jitter_factor=settings.center_jitter_factor,
scale_jitter_factor=settings.scale_jitter_factor,
mode='sequence',
crop_type=settings.crop_type,
max_scale_change=settings.max_scale_change,
transform=transform_train,
joint_transform=transform_joint,
new_roll=True)
data_processing_val = processing.LWLProcessing(
search_area_factor=settings.search_area_factor,
output_sz=settings.output_sz,
center_jitter_factor=settings.center_jitter_factor,
scale_jitter_factor=settings.scale_jitter_factor,
mode='sequence',
crop_type=settings.crop_type,
max_scale_change=settings.max_scale_change,
transform=transform_val,
joint_transform=transform_joint,
new_roll=True)
# Train sampler and loader
dataset_train = sampler.LWLSampler([ytvos_train, davis_train], [6, 1],
samples_per_epoch=settings.batch_size *
1000,
max_gap=100,
num_test_frames=3,
num_train_frames=1,
processing=data_processing_train)
dataset_val = sampler.LWLSampler([ytvos_val], [1],
samples_per_epoch=settings.batch_size *
100,
max_gap=100,
num_test_frames=3,
num_train_frames=1,
processing=data_processing_val)
loader_train = LTRLoader('train',
dataset_train,
training=True,
num_workers=settings.num_workers,
stack_dim=1,
batch_size=settings.batch_size)
loader_val = LTRLoader('val',
dataset_val,
training=False,
num_workers=settings.num_workers,
epoch_interval=5,
stack_dim=1,
batch_size=settings.batch_size)
# Network
net = lwl_networks.steepest_descent_resnet50(
filter_size=3,
num_filters=16,
optim_iter=5,
backbone_pretrained=True,
out_feature_dim=512,
frozen_backbone_layers=['conv1', 'bn1', 'layer1'],
label_encoder_dims=(16, 32, 64),
use_bn_in_label_enc=False,
clf_feat_blocks=0,
final_conv=True,
backbone_type='mrcnn')
base_net = network_loading.load_trained_network(
settings.env.workspace_dir,
'ltr/lwl/lwl_stage1/LWTLNet_ep0070.pth.tar')
net.load_state_dict(base_net.state_dict())
# Wrap the network for multi GPU training
if settings.multi_gpu:
net = MultiGPU(net, dim=1)
# Loss function
objective = {
'segm': LovaszSegLoss(per_image=False),
}
loss_weight = {'segm': 100.0}
actor = segm_actors.LWLActor(net=net,
objective=objective,
loss_weight=loss_weight,
num_refinement_iter=2,
disable_all_bn=True)
# Optimizer
optimizer = optim.Adam(
[{
'params': actor.net.target_model.filter_initializer.parameters(),
'lr': 5e-5
}, {
'params': actor.net.target_model.filter_optimizer.parameters(),
'lr': 2e-5
}, {
'params': actor.net.target_model.feature_extractor.parameters(),
'lr': 2e-5
}, {
'params': actor.net.decoder.parameters(),
'lr': 2e-5
}, {
'params': actor.net.label_encoder.parameters(),
'lr': 2e-5
}, {
'params': actor.net.feature_extractor.layer2.parameters(),
'lr': 2e-5
}, {
'params': actor.net.feature_extractor.layer3.parameters(),
'lr': 2e-5
}, {
'params': actor.net.feature_extractor.layer4.parameters(),
'lr': 2e-5
}],
lr=2e-4)
lr_scheduler = optim.lr_scheduler.MultiStepLR(optimizer, [25, 75],
gamma=0.2)
trainer = LTRTrainer(actor, [loader_train, loader_val], optimizer,
settings, lr_scheduler)
trainer.train(80, load_latest=True, fail_safe=True)
def run(settings):
settings.description = 'Default train settings for training VOS with box initialization.'
settings.batch_size = 8
settings.num_workers = 4
settings.multi_gpu = False
settings.print_interval = 1
settings.normalize_mean = [102.9801, 115.9465, 122.7717]
settings.normalize_std = [1.0, 1.0, 1.0]
settings.feature_sz = (52, 30)
settings.output_sz = (settings.feature_sz[0] * 16,
settings.feature_sz[1] * 16)
settings.search_area_factor = 5.0
settings.crop_type = 'inside_major'
settings.max_scale_change = None
settings.device = "cuda:0"
settings.center_jitter_factor = {'train': 3, 'test': (5.5, 4.5)}
settings.scale_jitter_factor = {'train': 0.25, 'test': 0.5}
settings.min_target_area = 500
ytvos_train = YouTubeVOS(version="2019", multiobj=False, split='jjtrain')
ytvos_valid = YouTubeVOS(version="2019", multiobj=False, split='jjvalid')
coco_train = MSCOCOSeq()
# Data transform
transform_joint = tfm.Transform(tfm.ToBGR(),
tfm.ToGrayscale(probability=0.05),
tfm.RandomHorizontalFlip(probability=0.5))
transform_train = tfm.Transform(
tfm.ToTensorAndJitter(0.2, normalize=False),
tfm.Normalize(mean=settings.normalize_mean,
std=settings.normalize_std))
transform_val = tfm.Transform(
tfm.ToTensorAndJitter(0.0, normalize=False),
tfm.Normalize(mean=settings.normalize_mean,
std=settings.normalize_std))
data_processing_train = processing.LWLProcessing(
search_area_factor=settings.search_area_factor,
output_sz=settings.output_sz,
center_jitter_factor=settings.center_jitter_factor,
scale_jitter_factor=settings.scale_jitter_factor,
mode='sequence',
crop_type=settings.crop_type,
max_scale_change=settings.max_scale_change,
transform=transform_train,
joint_transform=transform_joint,
new_roll=True)
data_processing_val = processing.LWLProcessing(
search_area_factor=settings.search_area_factor,
output_sz=settings.output_sz,
center_jitter_factor=settings.center_jitter_factor,
scale_jitter_factor=settings.scale_jitter_factor,
mode='sequence',
crop_type=settings.crop_type,
max_scale_change=settings.max_scale_change,
transform=transform_val,
joint_transform=transform_joint,
new_roll=True)
# Train sampler and loader
dataset_train = sampler.LWLSampler([ytvos_train, coco_train], [1, 1],
samples_per_epoch=settings.batch_size *
1000,
max_gap=100,
num_test_frames=1,
num_train_frames=1,
processing=data_processing_train)
dataset_val = sampler.LWLSampler([ytvos_valid], [1],
samples_per_epoch=settings.batch_size *
100,
max_gap=100,
num_test_frames=1,
num_train_frames=1,
processing=data_processing_val)
loader_train = LTRLoader('train',
dataset_train,
training=True,
num_workers=settings.num_workers,
stack_dim=1,
batch_size=settings.batch_size)
loader_val = LTRLoader('val',
dataset_val,
training=False,
num_workers=settings.num_workers,
epoch_interval=5,
stack_dim=1,
batch_size=settings.batch_size)
net = lwt_box_networks.steepest_descent_resnet50(
filter_size=3,
num_filters=16,
optim_iter=5,
backbone_pretrained=True,
out_feature_dim=512,
frozen_backbone_layers=['conv1', 'bn1', 'layer1'],
label_encoder_dims=(16, 32, 64),
use_bn_in_label_enc=False,
clf_feat_blocks=0,
final_conv=True,
backbone_type='mrcnn',
box_label_encoder_dims=(
64,
64,
),
final_bn=False)
base_net_weights = network_loading.load_trained_network(
settings.env.workspace_dir,
'ltr/lwl/lwl_stage2/LWTLNet_ep0080.pth.tar')
# Copy weights
net.feature_extractor.load_state_dict(
base_net_weights.feature_extractor.state_dict())
net.target_model.load_state_dict(
base_net_weights.target_model.state_dict())
net.decoder.load_state_dict(base_net_weights.decoder.state_dict())
net.label_encoder.load_state_dict(
base_net_weights.label_encoder.state_dict())
# Wrap the network for multi GPU training
if settings.multi_gpu:
net = MultiGPU(net, dim=1)
objective = {
'segm': LovaszSegLoss(per_image=False),
}
loss_weight = {
'segm': 100.0,
'segm_box': 10.0,
'segm_train': 10,
}
actor = lwtl_actors.LWLBoxActor(net=net,
objective=objective,
loss_weight=loss_weight)
# Optimizer
optimizer = optim.Adam([{
'params': actor.net.box_label_encoder.parameters(),
'lr': 1e-3
}],
lr=2e-4)
lr_scheduler = optim.lr_scheduler.StepLR(optimizer,
step_size=20,
gamma=0.2)
trainer = LTRTrainer(actor, [loader_train, loader_val], optimizer,
settings, lr_scheduler)
trainer.train(50, load_latest=True, fail_safe=True)
def run(settings):
settings.description = 'Default train settings for DiMP with ResNet50 as backbone.'
settings.batch_size = 10
settings.num_workers = 8
settings.multi_gpu = False
settings.print_interval = 1
settings.normalize_mean = [0.485, 0.456, 0.406]
settings.normalize_std = [0.229, 0.224, 0.225]
settings.search_area_factor = 5.0
settings.output_sigma_factor = 1/4
settings.target_filter_sz = 4
settings.feature_sz = 18
settings.output_sz = settings.feature_sz * 16
settings.center_jitter_factor = {'train': 3, 'test': 4.5}
settings.scale_jitter_factor = {'train': 0.25, 'test': 0.5}
settings.hinge_threshold = 0.05
# settings.print_stats = ['Loss/total', 'Loss/iou', 'ClfTrain/clf_ce', 'ClfTrain/test_loss']
# Train datasets
lasot_train = Lasot(settings.env.lasot_dir, split='train')
got10k_train = Got10k(settings.env.got10k_dir, split='vottrain')
trackingnet_train = TrackingNet(settings.env.trackingnet_dir, set_ids=list(range(4)))
coco_train = MSCOCOSeq(settings.env.coco_dir)
# Validation datasets
got10k_val = Got10k(settings.env.got10k_dir, split='votval')
# Data transform
transform_joint = tfm.Transform(tfm.ToGrayscale(probability=0.05))
transform_train = tfm.Transform(tfm.ToTensorAndJitter(0.2),
tfm.Normalize(mean=settings.normalize_mean, std=settings.normalize_std))
transform_val = tfm.Transform(tfm.ToTensor(),
tfm.Normalize(mean=settings.normalize_mean, std=settings.normalize_std))
# The tracking pairs processing module
output_sigma = settings.output_sigma_factor / settings.search_area_factor
proposal_params = {'min_iou': 0.1, 'boxes_per_frame': 8, 'sigma_factor': [0.01, 0.05, 0.1, 0.2, 0.3]}
label_params = {'feature_sz': settings.feature_sz, 'sigma_factor': output_sigma, 'kernel_sz': settings.target_filter_sz}
data_processing_train = processing.DiMPProcessing(search_area_factor=settings.search_area_factor,
output_sz=settings.output_sz,
center_jitter_factor=settings.center_jitter_factor,
scale_jitter_factor=settings.scale_jitter_factor,
mode='sequence',
proposal_params=proposal_params,
label_function_params=label_params,
transform=transform_train,
joint_transform=transform_joint)
data_processing_val = processing.DiMPProcessing(search_area_factor=settings.search_area_factor,
output_sz=settings.output_sz,
center_jitter_factor=settings.center_jitter_factor,
scale_jitter_factor=settings.scale_jitter_factor,
mode='sequence',
proposal_params=proposal_params,
label_function_params=label_params,
transform=transform_val,
joint_transform=transform_joint)
# Train sampler and loader
dataset_train = sampler.DiMPSampler([lasot_train, got10k_train, trackingnet_train, coco_train], [0.25,1,1,1],
samples_per_epoch=26000, max_gap=30, num_test_frames=3, num_train_frames=3,
processing=data_processing_train)
loader_train = LTRLoader('train', dataset_train, training=True, batch_size=settings.batch_size, num_workers=settings.num_workers,
shuffle=True, drop_last=True, stack_dim=1)
# Validation samplers and loaders
dataset_val = sampler.DiMPSampler([got10k_val], [1], samples_per_epoch=5000, max_gap=30,
num_test_frames=3, num_train_frames=3,
processing=data_processing_val)
loader_val = LTRLoader('val', dataset_val, training=False, batch_size=settings.batch_size, num_workers=settings.num_workers,
shuffle=False, drop_last=True, epoch_interval=5, stack_dim=1)
# Create network and actor
net = dimpnet.dimpnet50(filter_size=settings.target_filter_sz, backbone_pretrained=True, optim_iter=5,
clf_feat_norm=True, clf_feat_blocks=0, final_conv=True, out_feature_dim=512,
optim_init_step=0.9, optim_init_reg=0.1,
init_gauss_sigma=output_sigma * settings.feature_sz, num_dist_bins=100,
bin_displacement=0.1, mask_init_factor=3.0, target_mask_act='sigmoid', score_act='relu')
# Wrap the network for multi GPU training
if settings.multi_gpu:
net = MultiGPU(net, dim=1)
objective = {'iou': nn.MSELoss(), 'test_clf': ltr_losses.LBHinge(threshold=settings.hinge_threshold)}
loss_weight = {'iou': 1, 'test_clf': 100, 'test_init_clf': 100, 'test_iter_clf': 400}
actor = actors.DiMPActor(net=net, objective=objective, loss_weight=loss_weight)
# Optimizer
optimizer = optim.Adam([{'params': actor.net.classifier.filter_initializer.parameters(), 'lr': 5e-5},
{'params': actor.net.classifier.filter_optimizer.parameters(), 'lr': 5e-4},
{'params': actor.net.classifier.feature_extractor.parameters(), 'lr': 5e-5},
{'params': actor.net.bb_regressor.parameters()},
{'params': actor.net.feature_extractor.parameters(), 'lr': 2e-5}],
lr=2e-4)
lr_scheduler = optim.lr_scheduler.StepLR(optimizer, step_size=15, gamma=0.2)
trainer = LTRTrainer(actor, [loader_train, loader_val], optimizer, settings, lr_scheduler)
trainer.train(50, load_latest=True, fail_safe=True)
def run(settings):
settings.description = 'Default train settings with backbone weights fixed. We initialize the backbone ResNet with ' \
'pre-trained Mask-RCNN weights. These weights can be obtained from ' \
'https://drive.google.com/file/d/12pVHmhqtxaJ151dZrXN1dcgUa7TuAjdA/view?usp=sharing. ' \
'Download and save these weights in env_settings.pretrained_networks directory'
settings.batch_size = 20
settings.num_workers = 8
settings.multi_gpu = True
settings.print_interval = 1
settings.normalize_mean = [102.9801, 115.9465, 122.7717]
settings.normalize_std = [1.0, 1.0, 1.0]
settings.feature_sz = (52, 30)
# Settings used for generating the image crop input to the network. See documentation of LWTLProcessing class in
# ltr/data/processing.py for details.
settings.output_sz = (settings.feature_sz[0] * 16,
settings.feature_sz[1] * 16
) # Size of input image crop
settings.search_area_factor = 5.0
settings.crop_type = 'inside_major'
settings.max_scale_change = None
settings.center_jitter_factor = {'train': 3, 'test': (5.5, 4.5)}
settings.scale_jitter_factor = {'train': 0.25, 'test': 0.5}
# Datasets
ytvos_train = YouTubeVOS(version="2019", multiobj=False, split='jjtrain')
davis_train = Davis(version='2017', multiobj=False, split='train')
ytvos_val = YouTubeVOS(version="2019", multiobj=False, split='jjvalid')
# Data transform
transform_joint = tfm.Transform(tfm.ToBGR(),
tfm.ToGrayscale(probability=0.05),
tfm.RandomHorizontalFlip(probability=0.5))
transform_train = tfm.Transform(
tfm.RandomAffine(p_flip=0.0,
max_rotation=15.0,
max_shear=0.0,
max_ar_factor=0.0,
max_scale=0.2,
pad_amount=0),
tfm.ToTensorAndJitter(0.2, normalize=False),
tfm.Normalize(mean=settings.normalize_mean,
std=settings.normalize_std))
transform_val = tfm.Transform(
tfm.ToTensorAndJitter(0.0, normalize=False),
tfm.Normalize(mean=settings.normalize_mean,
std=settings.normalize_std))
data_processing_train = processing.LWLProcessing(
search_area_factor=settings.search_area_factor,
output_sz=settings.output_sz,
center_jitter_factor=settings.center_jitter_factor,
scale_jitter_factor=settings.scale_jitter_factor,
mode='sequence',
crop_type=settings.crop_type,
max_scale_change=settings.max_scale_change,
transform=transform_train,
joint_transform=transform_joint,
new_roll=True)
data_processing_val = processing.LWLProcessing(
search_area_factor=settings.search_area_factor,
output_sz=settings.output_sz,
center_jitter_factor=settings.center_jitter_factor,
scale_jitter_factor=settings.scale_jitter_factor,
mode='sequence',
crop_type=settings.crop_type,
max_scale_change=settings.max_scale_change,
transform=transform_val,
joint_transform=transform_joint,
new_roll=True)
# Train sampler and loader
dataset_train = sampler.LWLSampler([ytvos_train, davis_train], [6, 1],
samples_per_epoch=settings.batch_size *
1000,
max_gap=100,
num_test_frames=3,
num_train_frames=1,
processing=data_processing_train)
dataset_val = sampler.LWLSampler([ytvos_val], [1],
samples_per_epoch=settings.batch_size *
100,
max_gap=100,
num_test_frames=3,
num_train_frames=1,
processing=data_processing_val)
loader_train = LTRLoader('train',
dataset_train,
training=True,
num_workers=settings.num_workers,
stack_dim=1,
batch_size=settings.batch_size)
loader_val = LTRLoader('val',
dataset_val,
training=False,
num_workers=settings.num_workers,
epoch_interval=5,
stack_dim=1,
batch_size=settings.batch_size)
# Network
net = lwl_networks.steepest_descent_resnet50(filter_size=3,
num_filters=16,
optim_iter=5,
backbone_pretrained=True,
out_feature_dim=512,
frozen_backbone_layers=[
'conv1', 'bn1', 'layer1',
'layer2', 'layer3',
'layer4'
],
label_encoder_dims=(16, 32,
64),
use_bn_in_label_enc=False,
clf_feat_blocks=0,
final_conv=True,
backbone_type='mrcnn')
# Load pre-trained maskrcnn weights
weights_path = os.path.join(settings.env.pretrained_networks,
'e2e_mask_rcnn_R_50_FPN_1x_converted.pkl')
pretrained_weights = torch.load(weights_path)
net.feature_extractor.load_state_dict(pretrained_weights)
# Wrap the network for multi GPU training
if settings.multi_gpu:
net = MultiGPU(net, dim=1)
# Loss function
objective = {
'segm': LovaszSegLoss(per_image=False),
}
loss_weight = {'segm': 100.0}
actor = segm_actors.LWLActor(net=net,
objective=objective,
loss_weight=loss_weight,
num_refinement_iter=2,
disable_all_bn=True)
# Optimizer
optimizer = optim.Adam(
[{
'params': actor.net.target_model.filter_initializer.parameters(),
'lr': 5e-5
}, {
'params': actor.net.target_model.filter_optimizer.parameters(),
'lr': 1e-4
}, {
'params': actor.net.target_model.feature_extractor.parameters(),
'lr': 2e-5
}, {
'params': actor.net.decoder.parameters(),
'lr': 1e-4
}, {
'params': actor.net.label_encoder.parameters(),
'lr': 2e-4
}],
lr=2e-4)
lr_scheduler = optim.lr_scheduler.MultiStepLR(optimizer,
milestones=[
40,
],
gamma=0.2)
trainer = LTRTrainer(actor, [loader_train, loader_val], optimizer,
settings, lr_scheduler)
trainer.train(70, load_latest=True, fail_safe=True)
def run(settings):
# Most common settings are assigned in the settings struct
debug = False
if debug:
settings.batch_size = 4 # 8 # 4 # 120 # 70 # 38
settings.num_workers = 0 # 24 # 30 # 10 # 35 # 30 min(settings.batch_size, 16)
settings.multi_gpu = False # True # True # True # True # True # True # True
else:
settings.batch_size = 38 # 8 # 4 # 120 # 70 # 38
settings.num_workers = 20 # 24 # 30 # 10 # 35 # 30 min(settings.batch_size, 16)
settings.multi_gpu = True # True # True # True # True # True # True # True
settings.device = 'cuda'
settings.description = 'TransT with default settings.'
settings.print_interval = 1
settings.normalize_mean = [0.485, 0.456, 0.406]
settings.normalize_std = [0.229, 0.224, 0.225]
settings.search_area_factor = 6. # 4.0
settings.template_area_factor = 2.
settings.search_feature_sz = 32
settings.template_feature_sz = 16
settings.search_sz = settings.search_feature_sz * 8
settings.temp_sz = settings.template_feature_sz * 8
settings.center_jitter_factor = {'search': 2.0, 'template': 0} # 3
settings.scale_jitter_factor = {'search': 0.05, 'template': 0} # 0.25
settings.sequence_length = 34 # 30 # 64 NEXT # Same as PT
settings.rand = False
# settings.search_gap = 1 # Depreciated
settings.init_ckpt = "pytracking/networks/transt.pth"
# Transformer
settings.position_embedding = 'sine'
settings.hidden_dim = 256
settings.dropout = 0.1
settings.nheads = 8
settings.dim_feedforward = 2048
settings.featurefusion_layers = 4
# settings.sigma = 1 / 4 / 5.
# settings.kernel = 4
# settings.feature = 32 # 18
# settings.output_sz = 256 # settings.feature * 16
# settings.end_pad_if_even = False
# settings.label_function_params = True
settings.move_data_to_gpu = True
# Train datasets
lasot_train = Lasot(settings.env.lasot_dir, split='train')
got10k_train = Got10k(settings.env.got10k_dir, split='vottrain') # votval
trackingnet_train = TrackingNet(settings.env.trackingnet_dir, set_ids=list(range(4)))
# coco_train = MSCOCOSeq(settings.env.coco_dir)
# The joint augmentation transform, that is applied to the pairs jointly
transform_joint = tfm.Transform(tfm.ToGrayscale(probability=0.05))
# The augmentation transform applied to the training set (individually to each image in the pair)
# transform_train = tfm.Transform(tfm.ToTensorAndJitter(0.2),
# tfm.Normalize(mean=settings.normalize_mean, std=settings.normalize_std))
transform_train = tfm.Transform(tfm.ToTensorAndJitter(0.2),
# tfm.RandomHorizontalFlip(),
# tfm.RandomAffine(p_flip=0.5, max_scale=1.5),
# tfm.RandomBlur(1),
tfm.Normalize(mean=settings.normalize_mean, std=settings.normalize_std))
# Data processing to do on the training pairs
data_processing_train = processing.TransTProcessing(search_area_factor=settings.search_area_factor,
template_area_factor = settings.template_area_factor,
search_sz=settings.search_sz,
temp_sz=settings.temp_sz,
center_jitter_factor=settings.center_jitter_factor,
scale_jitter_factor=settings.scale_jitter_factor,
mode='sequence',
joint=False, # Whether or not to apply same transform to every image
transform=transform_train,
rand=settings.rand,
label_function_params=None, # settings.label_function_params,
joint_transform=transform_joint)
# The sampler for training
# dataset_train = sampler.TransTSampler([got10k_train], [1], samples_per_epoch=1000*settings.batch_size, max_gap=100, processing=data_processing_train, num_search_frames=settings.sequence_length, frame_sample_mode="rnn_causal")
# dataset_train = sampler.TransTSampler([got10k_train, trackingnet_train], [1, 1], samples_per_epoch=1000*settings.batch_size, max_gap=100, processing=data_processing_train, num_search_frames=settings.sequence_length, frame_sample_mode="interval")
dataset_train = sampler.TransTSampler([lasot_train, got10k_train, trackingnet_train], [1,1,1], samples_per_epoch=1000*settings.batch_size, max_gap=100, processing=data_processing_train)
# The loader for training
loader_train = LTRLoader('train', dataset_train, training=True, batch_size=settings.batch_size, num_workers=settings.num_workers,
shuffle=True, drop_last=True, stack_dim=0, pin_memory=settings.move_data_to_gpu == False)
# Create network and actor
model = transt_models.transt_resnet50(settings)
# Wrap the network for multi GPU training
if settings.multi_gpu:
model = MultiGPU(model, dim=0)
objective = transt_models.transt_loss(settings)
n_parameters = sum(p.numel() for p in model.parameters() if p.requires_grad)
print('number of params:', n_parameters)
actor = actors.TranstActor(net=model, objective=objective)
# Optimizer
# Change learning rate forthe Q we have changed and the RNN and the readout
# q = self.mix_q(torch.cat([q, self.mix_norm(exc)], -1))
# self.class_embed_new = MLP(hidden_dim * 2, hidden_dim, num_classes + 1, 3)
# self.bbox_embed_new = MLP(hidden_dim * 2, hidden_dim, 4, 3)
param_dicts = [
{"params": [p for n, p in model.named_parameters() if "backbone" not in n and p.requires_grad]},
{
"params": [p for n, p in model.named_parameters() if "backbone" in n and p.requires_grad],
"lr": 1e-5,
},
]
optimizer = torch.optim.AdamW(param_dicts, lr=1e-4,
weight_decay=1e-4)
lr_scheduler = torch.optim.lr_scheduler.StepLR(optimizer, 500)
# Create trainer
trainer = LTRTrainer(actor, [loader_train], optimizer, settings, lr_scheduler)
# Run training (set fail_safe=False if you are debugging)
trainer.train(1000, load_latest=True, fail_safe=True)
def run(settings):
settings.description = 'Default train settings for PrDiMP with ResNet18 as backbone.'
settings.batch_size = 26
settings.num_workers = 8
settings.multi_gpu = False
settings.print_interval = 1
settings.normalize_mean = [0.485, 0.456, 0.406]
settings.normalize_std = [0.229, 0.224, 0.225]
settings.search_area_factor = 5.0
settings.output_sigma_factor = 1/4
settings.target_filter_sz = 4
settings.feature_sz = 18
settings.output_sz = settings.feature_sz * 16
settings.center_jitter_factor = {'train': 3, 'test': 4.5}
settings.scale_jitter_factor = {'train': 0.25, 'test': 0.5}
settings.hinge_threshold = 0.05
settings.print_stats = ['Loss/total', 'Loss/bb_ce', 'ClfTrain/clf_ce']
# Train datasets
lasot_train = Lasot(settings.env.lasot_dir, split='train')
got10k_train = Got10k(settings.env.got10k_dir, split='vottrain')
trackingnet_train = TrackingNet(settings.env.trackingnet_dir, set_ids=list(range(4)))
coco_train = MSCOCOSeq(settings.env.coco_dir)
# Validation datasets
got10k_val = Got10k(settings.env.got10k_dir, split='votval')
# Data transform
transform_joint = tfm.Transform(tfm.ToGrayscale(probability=0.05))
transform_train = tfm.Transform(tfm.ToTensorAndJitter(0.2),
tfm.Normalize(mean=settings.normalize_mean, std=settings.normalize_std))
transform_val = tfm.Transform(tfm.ToTensor(),
tfm.Normalize(mean=settings.normalize_mean, std=settings.normalize_std))
# The tracking pairs processing module
output_sigma = settings.output_sigma_factor / settings.search_area_factor
proposal_params = {'boxes_per_frame': 128, 'gt_sigma': (0.05, 0.05), 'proposal_sigma': [(0.05, 0.05), (0.5, 0.5)]}
label_params = {'feature_sz': settings.feature_sz, 'sigma_factor': output_sigma, 'kernel_sz': settings.target_filter_sz}
label_density_params = {'feature_sz': settings.feature_sz, 'sigma_factor': output_sigma, 'kernel_sz': settings.target_filter_sz, 'normalize': True}
data_processing_train = processing.KLDiMPProcessing(search_area_factor=settings.search_area_factor,
output_sz=settings.output_sz,
center_jitter_factor=settings.center_jitter_factor,
scale_jitter_factor=settings.scale_jitter_factor,
mode='sequence',
proposal_params=proposal_params,
label_function_params=label_params,
label_density_params=label_density_params,
transform=transform_train,
joint_transform=transform_joint)
data_processing_val = processing.KLDiMPProcessing(search_area_factor=settings.search_area_factor,
output_sz=settings.output_sz,
center_jitter_factor=settings.center_jitter_factor,
scale_jitter_factor=settings.scale_jitter_factor,
mode='sequence',
proposal_params=proposal_params,
label_function_params=label_params,
label_density_params=label_density_params,
transform=transform_val,
joint_transform=transform_joint)
# Train sampler and loader
dataset_train = sampler.DiMPSampler([lasot_train, got10k_train, trackingnet_train, coco_train], [0.25,1,1,1],
samples_per_epoch=26000, max_gap=200, num_test_frames=3, num_train_frames=3,
processing=data_processing_train)
loader_train = LTRLoader('train', dataset_train, training=True, batch_size=settings.batch_size, num_workers=settings.num_workers,
shuffle=True, drop_last=True, stack_dim=1)
# Validation samplers and loaders
dataset_val = sampler.DiMPSampler([got10k_val], [1], samples_per_epoch=5000, max_gap=200,
num_test_frames=3, num_train_frames=3,
processing=data_processing_val)
loader_val = LTRLoader('val', dataset_val, training=False, batch_size=settings.batch_size, num_workers=settings.num_workers,
shuffle=False, drop_last=True, epoch_interval=5, stack_dim=1)
# Create network and actor
net = dimpnet.klcedimpnet18(filter_size=settings.target_filter_sz, backbone_pretrained=True, optim_iter=5,
clf_feat_norm=True, final_conv=True, optim_init_step=1.0, optim_init_reg=0.05, optim_min_reg=0.05,
gauss_sigma=output_sigma * settings.feature_sz, alpha_eps=0.05, normalize_label=True, init_initializer='zero')
# Wrap the network for multi GPU training
if settings.multi_gpu:
net = MultiGPU(net, dim=1)
objective = {'bb_ce': klreg_losses.KLRegression(), 'clf_ce': klreg_losses.KLRegressionGrid()}
loss_weight = {'bb_ce': 0.0025, 'clf_ce': 0.25, 'clf_ce_init': 0.25, 'clf_ce_iter': 1.0}
actor = tracking_actors.KLDiMPActor(net=net, objective=objective, loss_weight=loss_weight)
# Optimizer
optimizer = optim.Adam([{'params': actor.net.classifier.parameters(), 'lr': 1e-3},
{'params': actor.net.bb_regressor.parameters(), 'lr': 1e-3},
{'params': actor.net.feature_extractor.parameters()}],
lr=2e-4)
lr_scheduler = optim.lr_scheduler.StepLR(optimizer, step_size=15, gamma=0.2)
trainer = LTRTrainer(actor, [loader_train, loader_val], optimizer, settings, lr_scheduler)
trainer.train(50, load_latest=True, fail_safe=True)
def run(settings):
settings.move_data_to_gpu = False
settings.description = ''
settings.batch_size = 10 # 10
settings.test_sequence_length = 50
settings.num_workers = 4
settings.multi_gpu = True
settings.print_interval = 1
settings.normalize_mean = [0.485, 0.456, 0.406]
settings.normalize_std = [0.229, 0.224, 0.225]
settings.search_area_factor = 5.0
settings.output_sigma_factor = 1/4
settings.target_filter_sz = 4
settings.feature_sz = 18
settings.output_sz = settings.feature_sz * 16
settings.center_jitter_param = {'train_mode': 'uniform', 'train_factor': 3.0, 'train_limit_motion': False,
'test_mode': 'uniform', 'test_factor': 4.5, 'test_limit_motion': True}
settings.scale_jitter_param = {'train_factor': 0.25, 'test_factor': 0.3}
settings.hinge_threshold = 0.05
settings.print_stats = ['Loss/total', 'Loss/raw/test_clf', 'Loss/raw/test_clf_acc', 'Loss/raw/dimp_clf_acc',
'Loss/raw/is_target', 'Loss/raw/is_target_after_prop',
'Loss/raw/test_seq_acc',
'Loss/raw/dimp_seq_acc']
# lasot_train = Lasot(settings.env.lasot_dir, split='train')
got10k_train = Got10k(settings.env.got10k_dir, split='vottrain')
# trackingnet_train = TrackingNet(settings.env.trackingnet_dir, set_ids=[0, 1, 2, 3, 4])
# Validation datasets
got10k_val = Got10k(settings.env.got10k_dir, split='votval')
# Data transform
transform_joint = tfm.Transform(tfm.ToGrayscale(probability=0.05))
transform_train = tfm.Transform(tfm.ToTensorAndJitter(0.2),
tfm.Normalize(mean=settings.normalize_mean, std=settings.normalize_std))
transform_val = tfm.Transform(tfm.ToTensor(),
tfm.Normalize(mean=settings.normalize_mean, std=settings.normalize_std))
# The tracking pairs processing module
output_sigma = settings.output_sigma_factor / settings.search_area_factor
proposal_params = None
label_params = {'feature_sz': settings.feature_sz, 'sigma_factor': output_sigma,
'kernel_sz': settings.target_filter_sz,
'end_pad_if_even': True}
data_processing_train = processing.KYSProcessing(search_area_factor=settings.search_area_factor,
output_sz=settings.output_sz,
center_jitter_param=settings.center_jitter_param,
scale_jitter_param=settings.scale_jitter_param,
proposal_params=proposal_params,
label_function_params=label_params,
transform=transform_train,
joint_transform=transform_joint,
min_crop_inside_ratio=0.1)
data_processing_val = processing.KYSProcessing(search_area_factor=settings.search_area_factor,
output_sz=settings.output_sz,
center_jitter_param=settings.center_jitter_param,
scale_jitter_param=settings.scale_jitter_param,
proposal_params=proposal_params,
label_function_params=label_params,
transform=transform_val,
joint_transform=transform_joint,
min_crop_inside_ratio=0.1)
# Train sampler and loader
sequence_sample_info = {'num_train_frames': 3, 'num_test_frames': settings.test_sequence_length,
'max_train_gap': 30, 'allow_missing_target': True, 'min_fraction_valid_frames': 0.5,
'mode': 'Sequence'}
# dataset_train = sampler.KYSSampler([got10k_train, trackingnet_train, lasot_train],
# [0.3, 0.3, 0.25],
# samples_per_epoch=settings.batch_size * 150,
# sequence_sample_info=sequence_sample_info,
# processing=data_processing_train,
# sample_occluded_sequences=True)
dataset_train = sampler.KYSSampler([got10k_train],
[1],
samples_per_epoch=settings.batch_size * 150,
sequence_sample_info=sequence_sample_info,
processing=data_processing_train,
sample_occluded_sequences=True)
loader_train = LTRLoader('train', dataset_train, training=True, batch_size=settings.batch_size,
num_workers=settings.num_workers,
shuffle=True, drop_last=True, stack_dim=1)
# Validation samplers and loaders
dataset_val = sampler.KYSSampler([got10k_val], [1], samples_per_epoch=1000,
sequence_sample_info=sequence_sample_info, processing=data_processing_val,
sample_occluded_sequences=True)
loader_val = LTRLoader('val', dataset_val, training=False, batch_size=settings.batch_size,
num_workers=settings.num_workers,
shuffle=False, drop_last=True, epoch_interval=5, stack_dim=1)
# load base dimp
dimp_weights_path = os.path.join(settings.env.pretrained_networks, 'dimp50.pth')
base_net, _ = network_loading.load_network(checkpoint=dimp_weights_path)
net = kysnet_models.kysnet_res50(optim_iter=3, cv_kernel_size=3, cv_max_displacement=9,
cv_stride=1, init_gauss_sigma=output_sigma * settings.feature_sz,
train_feature_extractor=False, train_iounet=False, detach_length=0, state_dim=8,
representation_predictor_dims=(16,), conf_measure='entropy',
dimp_thresh=0.05)
# Move pre-trained dimp weights
net.backbone_feature_extractor.load_state_dict(base_net.feature_extractor.state_dict())
net.dimp_classifier.load_state_dict(base_net.classifier.state_dict())
net.bb_regressor.load_state_dict(base_net.bb_regressor.state_dict())
if settings.multi_gpu:
net = MultiGPU(net, dim=1)
# To be safe
for p in net.backbone_feature_extractor.parameters():
p.requires_grad_(False)
for p in net.dimp_classifier.parameters():
p.requires_grad_(False)
for p in net.bb_regressor.parameters():
p.requires_grad_(False)
objective = {'test_clf': ltr_losses.LBHingev2(threshold=settings.hinge_threshold, return_per_sequence=False),
'dimp_clf': ltr_losses.LBHingev2(threshold=settings.hinge_threshold, return_per_sequence=False),
'is_target': ltr_losses.IsTargetCellLoss(return_per_sequence=False),
'clf_acc': ltr_losses.TrackingClassificationAccuracy(threshold=0.25)}
loss_weight = {'test_clf': 1.0*500, 'test_clf_orig': 50, 'is_target': 0.1*500, 'is_target_after_prop': 0.1*500}
dimp_jitter_fn = DiMPScoreJittering(distractor_ratio=0.1, p_distractor=0.3, max_distractor_enhance_factor=1.3,
min_distractor_enhance_factor=0.8)
actor = actors.KYSActor(net=net, objective=objective, loss_weight=loss_weight,
dimp_jitter_fn=dimp_jitter_fn)
optimizer = optim.Adam([{'params': actor.net.predictor.parameters(), 'lr': 1e-2}],
lr=1e-2)
lr_scheduler = optim.lr_scheduler.StepLR(optimizer, step_size=15, gamma=0.2)
trainer = LTRTrainer(actor, [loader_train, loader_val], optimizer, settings, lr_scheduler)
trainer.train(40, load_latest=True, fail_safe=True)
def run(settings):
# Most common settings are assigned in the settings struct
settings.device = 'cuda'
settings.description = 'TransT with default settings.'
settings.batch_size = 4 # 38
settings.num_workers = 2 # min(settings.batch_size, 8)
settings.multi_gpu = False
settings.print_interval = 1
settings.normalize_mean = [0.485, 0.456, 0.406]
settings.normalize_std = [0.229, 0.224, 0.225]
settings.search_area_factor = 4.0
settings.template_area_factor = 2.0
settings.search_feature_sz = 32
settings.template_feature_sz = 16
settings.search_sz = settings.search_feature_sz * 8
settings.temp_sz = settings.template_feature_sz * 8
settings.center_jitter_factor = {'search': 0, 'template': 0} # 3
settings.scale_jitter_factor = {'search': 0., 'template': 0} # 0.25
settings.init_ckpt = "pytracking/networks/transt.pth"
# Transformer
settings.position_embedding = 'sine'
settings.hidden_dim = 256
settings.dropout = 0.1
settings.nheads = 8
settings.dim_feedforward = 2048
settings.featurefusion_layers = 4
# Train datasets
# lasot_train = Lasot(settings.env.lasot_dir, split='train')
# got10k_train = Got10k(settings.env.got10k_dir, split='vottrain')
got10k_train = Got10k(settings.env.got10k_dir, split='vottrain')
# trackingnet_train = TrackingNet(settings.env.trackingnet_dir, set_ids=list(range(4)))
# coco_train = MSCOCOSeq(settings.env.coco_dir)
# The joint augmentation transform, that is applied to the pairs jointly
transform_joint = tfm.Transform(tfm.ToGrayscale(probability=0.05))
# The augmentation transform applied to the training set (individually to each image in the pair)
transform_train = tfm.Transform(
tfm.ToTensorAndJitter(0.2),
tfm.Normalize(mean=settings.normalize_mean,
std=settings.normalize_std))
# Data processing to do on the training pairs
data_processing_train = processing.TransTProcessing(
search_area_factor=settings.search_area_factor,
template_area_factor=settings.template_area_factor,
search_sz=settings.search_sz,
temp_sz=settings.temp_sz,
center_jitter_factor=settings.center_jitter_factor,
scale_jitter_factor=settings.scale_jitter_factor,
mode='pair', # 'sequence',
transform=transform_train,
joint_transform=transform_joint)
# The sampler for training
dataset_train = sampler.TransTSampler([got10k_train], [1],
samples_per_epoch=1000 *
settings.batch_size,
max_gap=100,
processing=data_processing_train)
# dataset_train = sampler.TransTSampler([lasot_train, got10k_train, coco_train, trackingnet_train], [1,1,1,1], samples_per_epoch=1000*settings.batch_size, max_gap=100, processing=data_processing_train)
# The loader for training
loader_train = LTRLoader('train',
dataset_train,
training=True,
batch_size=settings.batch_size,
num_workers=settings.num_workers,
shuffle=True,
drop_last=True,
stack_dim=0)
# Create network and actor
model = transt_models.transt_resnet50(settings)
# Wrap the network for multi GPU training
if settings.multi_gpu:
model = MultiGPU(model, dim=0)
objective = transt_models.transt_loss(settings)
n_parameters = sum(p.numel() for p in model.parameters()
if p.requires_grad)
print('number of params:', n_parameters)
actor = actors.TranstActor(net=model, objective=objective)
# Optimizer
param_dicts = [
{
"params": [
p for n, p in model.named_parameters()
if "backbone" not in n and p.requires_grad
]
},
{
"params": [
p for n, p in model.named_parameters()
if "backbone" in n and p.requires_grad
],
"lr":
0.,
},
]
optimizer = torch.optim.AdamW(param_dicts, lr=0., weight_decay=0.)
lr_scheduler = torch.optim.lr_scheduler.StepLR(optimizer, 500)
# Create trainer
trainer = LTRTrainer(actor, [loader_train], optimizer, settings,
lr_scheduler)
# Run training (set fail_safe=False if you are debugging)
trainer.train(1000, load_latest=True, fail_safe=True)