def __init__(self,
ds_test,
batch_size=1,
num_workers=4,
loss='JointsMSELoss',
checkpoint_path="./weights/pose_hrnet_w48_384x288.pth",
model_c=48,
model_nof_joints=17,
model_bn_momentum=0.1,
flip_test_images=True,
device=None):
"""
Initializes a new Test object.
The HRNet model is initialized and the saved checkpoint is loaded.
The DataLoader and the loss function are defined.
Args:
ds_test (HumanPoseEstimationDataset): test dataset.
batch_size (int): batch size.
Default: 1
num_workers (int): number of workers for each DataLoader
Default: 4
loss (str): loss function. Valid values are 'JointsMSELoss' and 'JointsOHKMMSELoss'.
Default: "JointsMSELoss"
checkpoint_path (str): path to a previous checkpoint.
Default: None
model_c (int): hrnet parameters - number of channels.
Default: 48
model_nof_joints (int): hrnet parameters - number of joints.
Default: 17
model_bn_momentum (float): hrnet parameters - path to the pretrained weights.
Default: 0.1
flip_test_images (bool): flip images during validating.
Default: True
device (torch.device): device to be used (default: cuda, if available).
Default: None
"""
super(Test, self).__init__()
hrnet_m = 'HRNet'
hrnet_c = 48
hrnet_j = 17
hrnet_weights = "./weights/pose_hrnet_w48_384x288.pth"
hrnet_joints_set = "coco"
image_resolution = (384, 288)
single_person = False
disable_tracking = False
max_batch_size = 16
self.ds_test = ds_test
self.batch_size = batch_size
self.num_workers = num_workers
self.loss = loss
self.checkpoint_path = checkpoint_path
self.model_c = model_c
self.model_nof_joints = model_nof_joints
self.model_bn_momentum = model_bn_momentum
self.flip_test_images = flip_test_images
self.epoch = 0
# torch device
if device is not None:
self.device = device
else:
if torch.cuda.is_available():
self.device = torch.device('cuda:0')
else:
self.device = torch.device('cpu')
print(self.device)
#
# load model
#self.model = HRNet(c=self.model_c, nof_joints=self.model_nof_joints,
# bn_momentum=self.model_bn_momentum).to(self.device)
image_resolution = ast.literal_eval(image_resolution)
self.model = SimpleHRNet(hrnet_c,
hrnet_j,
hrnet_weights,
model_name=hrnet_m,
resolution=image_resolution,
multiperson=not single_person,
return_bounding_boxes=not disable_tracking,
max_batch_size=max_batch_size,
device=device)
#
# define loss
if self.loss == 'JointsMSELoss':
self.loss_fn = JointsMSELoss().to(self.device)
elif self.loss == 'JointsOHKMMSELoss':
self.loss_fn = JointsOHKMMSELoss().to(self.device)
else:
raise NotImplementedError
#
# load previous checkpoint
if os.path.basename(
self.checkpoint_path) == "pose_hrnet_w48_384x288.pth":
self.starting_epoch = 210 # 1?
#self.params = None
elif self.checkpoint_path is not None:
print('Loading checkpoint %s...' % self.checkpoint_path)
if os.path.isdir(self.checkpoint_path):
path = os.path.join(self.checkpoint_path,
'checkpoint_last.pth')
else:
path = self.checkpoint_path
self.starting_epoch, self.model, _, self.params = load_checkpoint(
path, self.model, device=self.device)
else:
raise ValueError('checkpoint_path is not defined')
#
# load test dataset
self.dl_test = DataLoader(self.ds_test,
batch_size=self.batch_size,
shuffle=False,
num_workers=self.num_workers)
self.len_dl_test = len(self.dl_test)
#print("print len_dl_test: ", self.len_dl_test)
#
# initialize variables
self.mean_loss_test = 0.
self.mean_acc_test = 0.
def __init__(self,
exp_name,
ds_train,
ds_val,
epochs=210,
batch_size=16,
num_workers=4,
loss='JointsMSELoss',
lr=0.001,
lr_decay=True,
lr_decay_steps=(170, 200),
lr_decay_gamma=0.1,
optimizer='Adam',
weight_decay=0.,
momentum=0.9,
nesterov=False,
pretrained_weight_path=None,
checkpoint_path=None,
log_path='./logs',
use_tensorboard=True,
model_c=48,
model_nof_joints=17,
model_bn_momentum=0.1,
flip_test_images=True,
device=None
):
"""
Initializes a new Train object.
The log folder is created, the HRNet model is initialized and optional pre-trained weights or saved checkpoints
are loaded.
The DataLoaders, the loss function, and the optimizer are defined.
Args:
exp_name (str): experiment name.
ds_train (HumanPoseEstimationDataset): train dataset.
ds_val (HumanPoseEstimationDataset): validation dataset.
epochs (int): number of epochs.
Default: 210
batch_size (int): batch size.
Default: 16
num_workers (int): number of workers for each DataLoader
Default: 4
loss (str): loss function. Valid values are 'JointsMSELoss' and 'JointsOHKMMSELoss'.
Default: "JointsMSELoss"
lr (float): learning rate.
Default: 0.001
lr_decay (bool): learning rate decay.
Default: True
lr_decay_steps (tuple): steps for the learning rate decay scheduler.
Default: (170, 200)
lr_decay_gamma (float): scale factor for each learning rate decay step.
Default: 0.1
optimizer (str): network optimizer. Valid values are 'Adam' and 'SGD'.
Default: "Adam"
weight_decay (float): weight decay.
Default: 0.
momentum (float): momentum factor.
Default: 0.9
nesterov (bool): Nesterov momentum.
Default: False
pretrained_weight_path (str): path to pre-trained weights (such as weights from pre-train on imagenet).
Default: None
checkpoint_path (str): path to a previous checkpoint.
Default: None
log_path (str): path where tensorboard data and checkpoints will be saved.
Default: "./logs"
use_tensorboard (bool): enables tensorboard use.
Default: True
model_c (int): hrnet parameters - number of channels.
Default: 48
model_nof_joints (int): hrnet parameters - number of joints.
Default: 17
model_bn_momentum (float): hrnet parameters - path to the pretrained weights.
Default: 0.1
flip_test_images (bool): flip images during validating.
Default: True
device (torch.device): device to be used (default: cuda, if available).
Default: None
"""
super(Train, self).__init__()
self.exp_name = exp_name
self.ds_train = ds_train
self.ds_val = ds_val
self.epochs = epochs
self.batch_size = batch_size
self.num_workers = num_workers
self.loss = loss
self.lr = lr
self.lr_decay = lr_decay
self.lr_decay_steps = lr_decay_steps
self.lr_decay_gamma = lr_decay_gamma
self.optimizer = optimizer
self.weight_decay = weight_decay
self.momentum = momentum
self.nesterov = nesterov
self.pretrained_weight_path = pretrained_weight_path
self.checkpoint_path = checkpoint_path
self.log_path = os.path.join(log_path, self.exp_name)
self.use_tensorboard = use_tensorboard
self.model_c = model_c
self.model_nof_joints = model_nof_joints
self.model_bn_momentum = model_bn_momentum
self.flip_test_images = flip_test_images
self.epoch = 0
# torch device
if device is not None:
self.device = device
else:
if torch.cuda.is_available():
self.device = torch.device('cuda:0')
else:
self.device = torch.device('cpu')
print(self.device)
os.makedirs(self.log_path, 0o755, exist_ok=False) # exist_ok=False to avoid overwriting
if self.use_tensorboard:
self.summary_writer = tb.SummaryWriter(self.log_path)
#
# write all experiment parameters in parameters.txt and in tensorboard text field
self.parameters = [x + ': ' + str(y) + '\n' for x, y in locals().items()]
with open(os.path.join(self.log_path, 'parameters.txt'), 'w') as fd:
fd.writelines(self.parameters)
if self.use_tensorboard:
self.summary_writer.add_text('parameters', '\n'.join(self.parameters))
#
# load model
self.model = HRNet(c=self.model_c, nof_joints=self.model_nof_joints,
bn_momentum=self.model_bn_momentum).to(self.device)
#
# define loss and optimizers
if self.loss == 'JointsMSELoss':
self.loss_fn = JointsMSELoss().to(self.device)
elif self.loss == 'JointsOHKMMSELoss':
self.loss_fn = JointsOHKMMSELoss().to(self.device)
else:
raise NotImplementedError
if optimizer == 'SGD':
self.optim = SGD(self.model.parameters(), lr=self.lr, weight_decay=self.weight_decay,
momentum=self.momentum, nesterov=self.nesterov)
elif optimizer == 'Adam':
self.optim = Adam(self.model.parameters(), lr=self.lr, weight_decay=self.weight_decay)
else:
raise NotImplementedError
#
# load pre-trained weights (such as those pre-trained on imagenet)
if self.pretrained_weight_path is not None:
self.model.load_state_dict(torch.load(self.pretrained_weight_path, map_location=self.device), strict=True)
print('Pre-trained weights loaded.')
#
# load previous checkpoint
if self.checkpoint_path is not None:
print('Loading checkpoint %s...' % self.checkpoint_path)
if os.path.isdir(self.checkpoint_path):
path = os.path.join(self.checkpoint_path, 'checkpoint_last.pth')
else:
path = self.checkpoint_path
self.starting_epoch, self.model, self.optim, self.params = load_checkpoint(path, self.model, self.optim,
self.device)
else:
self.starting_epoch = 0
if lr_decay:
self.lr_scheduler = MultiStepLR(self.optim, list(self.lr_decay_steps), gamma=self.lr_decay_gamma,
last_epoch=self.starting_epoch if self.starting_epoch else -1)
#
# load train and val datasets
self.dl_train = DataLoader(self.ds_train, batch_size=self.batch_size, shuffle=True,
num_workers=self.num_workers, drop_last=True)
self.len_dl_train = len(self.dl_train)
# dl_val = DataLoader(self.ds_val, batch_size=1, shuffle=False, num_workers=num_workers)
self.dl_val = DataLoader(self.ds_val, batch_size=self.batch_size, shuffle=False, num_workers=self.num_workers)
self.len_dl_val = len(self.dl_val)
#
# initialize variables
self.mean_loss_train = 0.
self.mean_acc_train = 0.
self.mean_loss_val = 0.
self.mean_acc_val = 0.
self.mean_mAP_val = 0.
self.best_loss = None
self.best_acc = None
self.best_mAP = None
def __init__(self,
ds_test,
batch_size=1,
num_workers=4,
loss='JointsMSELoss',
checkpoint_path=None,
model_c=48,
model_nof_joints=17,
model_bn_momentum=0.1,
flip_test_images=True,
device=None):
"""
Initializes a new Test object.
The HRNet model is initialized and the saved checkpoint is loaded.
The DataLoader and the loss function are defined.
Args:
ds_test (HumanPoseEstimationDataset): test dataset.
batch_size (int): batch size.
Default: 1
num_workers (int): number of workers for each DataLoader
Default: 4
loss (str): loss function. Valid values are 'JointsMSELoss' and 'JointsOHKMMSELoss'.
Default: "JointsMSELoss"
checkpoint_path (str): path to a previous checkpoint.
Default: None
model_c (int): hrnet parameters - number of channels.
Default: 48
model_nof_joints (int): hrnet parameters - number of joints.
Default: 17
model_bn_momentum (float): hrnet parameters - path to the pretrained weights.
Default: 0.1
flip_test_images (bool): flip images during validating.
Default: True
device (torch.device): device to be used (default: cuda, if available).
Default: None
"""
super(Test, self).__init__()
self.ds_test = ds_test
self.batch_size = batch_size
self.num_workers = num_workers
self.loss = loss
self.checkpoint_path = checkpoint_path
self.model_c = model_c
self.model_nof_joints = model_nof_joints
self.model_bn_momentum = model_bn_momentum
self.flip_test_images = flip_test_images
self.epoch = 0
# torch device
if device is not None:
self.device = device
else:
if torch.cuda.is_available():
self.device = torch.device('cuda:0')
else:
self.device = torch.device('cpu')
print(self.device)
#
# load model
self.model = HRNet(c=self.model_c,
nof_joints=self.model_nof_joints,
bn_momentum=self.model_bn_momentum).to(self.device)
#
# define loss
if self.loss == 'JointsMSELoss':
self.loss_fn = JointsMSELoss().to(self.device)
elif self.loss == 'JointsOHKMMSELoss':
self.loss_fn = JointsOHKMMSELoss().to(self.device)
else:
raise NotImplementedError
#
# load previous checkpoint
if self.checkpoint_path is not None:
print('Loading checkpoint %s...' % self.checkpoint_path)
if os.path.isdir(self.checkpoint_path):
path = os.path.join(self.checkpoint_path,
'checkpoint_last.pth')
else:
path = self.checkpoint_path
self.starting_epoch, self.model, _, self.params = load_checkpoint(
path, self.model, device=self.device)
else:
raise ValueError('checkpoint_path is not defined')
#
# load test dataset
self.dl_test = DataLoader(self.ds_test,
batch_size=self.batch_size,
shuffle=False,
num_workers=self.num_workers)
self.len_dl_test = len(self.dl_test)
#
# initialize variables
self.mean_loss_test = 0.
self.mean_acc_test = 0.
def __init__(self,
exp_name,
ds_train,
ds_val,
epochs=210,
batch_size=16,
num_workers=4,
loss='JointsMSELoss',
lr=0.001,
lr_decay=True,
lr_decay_steps=(170, 200),
lr_decay_gamma=0.1,
optimizer='Adam',
weight_decay=0.00001,
momentum=0.9,
nesterov=False,
pretrained_weight_path=None,
checkpoint_path=None,
log_path='./logs',
use_tensorboard=True,
model_c=48,
model_nof_joints=17,
model_bn_momentum=0.1,
flip_test_images=True,
device=None):
"""
Inicializa el nuevo objeto Train
Se crea el folder de logs, se inicializa el modelo HRNet y se determinan dimensiones pre entrenadas o puntos
de control guardos son cargados
"""
super(Train, self).__init__()
self.exp_name = exp_name
self.ds_train = ds_train
self.ds_val = ds_val
self.epochs = epochs
self.batch_size = batch_size
self.num_workers = num_workers
self.loss = loss
self.lr = lr
self.lr_decay = lr_decay
self.lr_decay_steps = lr_decay_steps
self.lr_decay_gamma = lr_decay_gamma
self.optimizer = optimizer
self.weight_decay = weight_decay
self.momentum = momentum
self.nesterov = nesterov
self.pretrained_weight_path = pretrained_weight_path
self.checkpoint_path = checkpoint_path
self.log_path = os.path.join(log_path, self.exp_name)
self.use_tensorboard = use_tensorboard
self.model_c = model_c
self.model_nof_joints = model_nof_joints
self.model_bn_momentum = model_bn_momentum
self.flip_test_images = flip_test_images
self.epoch = 0
# torch devz
if device is not None:
self.device = device
else:
if torch.cuda.is_available():
self.device = torch.device('cuda:0')
else:
self.device = torch.device('cpu')
print(self.device)
os.makedirs(self.log_path, 0o755, exist_ok=False)
if self.use_tensorboard:
self.summary_writer = tb.SummaryWriter(self.log_path)
#escribe todos los parametros experimentales en parameters.txt y en campos de texto de tensorboard
self.parameters = [
x + ': ' + str(y) + '\n' for x, y in locals().items()
]
with open(os.path.join(self.log_path, 'parameters.txt'), 'w') as fd:
fd.writelines(self.parameters)
if self.use_tensorboard:
self.summary_writer.add_text('parameters',
'\n'.join(self.parameters))
#
# Carga el modelo
self.model = HRNet(c=self.model_c,
nof_joints=self.model_nof_joints,
bn_momentum=self.model_bn_momentum).to(self.device)
if self.loss == 'JointsMSELoss':
self.loss_fn = JointsMSELoss().to(self.device)
elif self.loss == 'JointsOHKMMSELoss':
self.loss_fn = JointsOHKMMSELoss().to(self.device)
else:
raise NotImplementedError
if optimizer == 'SGD':
self.optim = SGD(self.model.parameters(),
lr=self.lr,
weight_decay=self.weight_decay,
momentum=self.momentum,
nesterov=self.nesterov)
elif optimizer == 'Adam':
self.optim = Adam(self.model.parameters(),
lr=self.lr,
weight_decay=self.weight_decay)
else:
raise NotImplementedError
# Carga las dimensiones preentrenadas
if self.pretrained_weight_path is not None:
self.model.load_state_dict(torch.load(self.pretrained_weight_path,
map_location=self.device),
strict=False)
#
# carga puntos de control previos
if self.checkpoint_path is not None:
print('Loading checkpoint %s...' % self.checkpoint_path)
if os.path.isdir(self.checkpoint_path):
path = os.path.join(self.checkpoint_path,
'checkpoint_last.pth')
else:
path = self.checkpoint_path
self.starting_epoch, self.model, self.optim, self.params = load_checkpoint(
path, self.model, self.optim, self.device)
else:
self.starting_epoch = 0
if lr_decay:
self.lr_scheduler = MultiStepLR(self.optim,
list(self.lr_decay_steps),
gamma=self.lr_decay_gamma,
last_epoch=self.starting_epoch)
# Carga el entrenamiento y los valores de los datasets
self.dl_train = DataLoader(self.ds_train,
batch_size=self.batch_size,
shuffle=True,
num_workers=self.num_workers,
drop_last=True)
self.len_dl_train = len(self.dl_train)
self.dl_val = DataLoader(self.ds_val,
batch_size=self.batch_size,
shuffle=False,
num_workers=self.num_workers)
self.len_dl_val = len(self.dl_val)
#
# inicializa las variables
self.mean_loss_train = 0.
self.mean_acc_train = 0.
self.mean_loss_val = 0.
self.mean_acc_val = 0.
self.mean_mAP_val = 0.
self.best_loss = None
self.best_acc = None
self.best_mAP = None
def __init__(self,
ds_test,
batch_size=1,
num_workers=4,
loss='JointsMSELoss',
checkpoint_path=None,
model_c=48,
model_nof_joints=17,
model_bn_momentum=0.1,
flip_test_images=True,
device=None):
super(Test, self).__init__()
self.ds_test = ds_test
self.batch_size = batch_size
self.num_workers = num_workers
self.loss = loss
self.checkpoint_path = checkpoint_path
self.model_c = model_c
self.model_nof_joints = model_nof_joints
self.model_bn_momentum = model_bn_momentum
self.flip_test_images = flip_test_images
self.epoch = 0
# dispositivo torch
if device is not None:
self.device = device
else:
if torch.cuda.is_available():
self.device = torch.device('cuda:0')
else:
self.device = torch.device('cpu')
print(self.device)
# cargando modelo
self.model = HRNet(c=self.model_c,
nof_joints=self.model_nof_joints,
bn_momentum=self.model_bn_momentum).to(self.device)
#Definiendo perdidas
if self.loss == 'JointsMSELoss':
self.loss_fn = JointsMSELoss().to(self.device)
elif self.loss == 'JointsOHKMMSELoss':
self.loss_fn = JointsOHKMMSELoss().to(self.device)
else:
raise NotImplementedError
#cargar punto de control anterior
if self.checkpoint_path is not None:
print('Loading checkpoint %s...' % self.checkpoint_path)
if os.path.isdir(self.checkpoint_path):
path = os.path.join(self.checkpoint_path,
'checkpoint_last.pth')
else:
path = self.checkpoint_path
self.starting_epoch, self.model, _, self.params = load_checkpoint(
path, self.model, device=self.device)
else:
raise ValueError('checkpoint_path is not defined')
#conjunto de datos de prueba de carga
self.dl_test = DataLoader(self.ds_test,
batch_size=self.batch_size,
shuffle=False,
num_workers=self.num_workers)
self.len_dl_test = len(self.dl_test)
#inicializar variables
self.mean_loss_test = 0.
self.mean_acc_test = 0.
