def init_environment(self):
# super().init_environment() # 先执行一次父类的该方法，这里实际上就是在父类方法的基础上进行扩展
#父类方法，直接抄过来方便查阅
self.io = torchlight.IO( #class 'torchlight.io.IO',这里的torchlight.IO就是一个容器
self.arg.work_dir, #存储结果的路径，默认为'./work_dir/tmp'
save_log=self.arg.save_log, #是否保存日志，默认true
print_log=self.arg.print_log) #是否打印日志，默认true
self.io.save_arg(self.arg)
# gpu
if self.arg.use_gpu: #如果指定了使用GPU，就默认使用0号GPU，否则使用CPU
gpus = torchlight.visible_gpu(
self.arg.device) #返回的是list(range(len(gpus)))
torchlight.occupy_gpu(gpus)
self.gpus = gpus
print('现在使用的gpu为：', self.gpus)
self.dev = "cuda:0"
else:
self.dev = "cpu"
#子类方法
#定义了几个字典对象
self.result = dict()
self.iter_info = dict()
self.epoch_info = dict()
self.meta_info = dict(epoch=0, iter=0)
def init_io(self):
self.io = torchlight.IO(self.arg.work_dir,
self.arg.session,
save_log=self.arg.save_log,
print_log=self.arg.print_log)
if self.arg.phase == 'train':
self.io.save_arg(self.arg)
def __init__(self,
args,
ftype,
data_loader,
C,
T,
V,
num_classes,
graph_dict,
n_z=32,
device='cuda:0'):
self.args = args
self.ftype = ftype
self.data_loader = data_loader
self.num_classes = num_classes
self.result = dict()
self.iter_info = dict()
self.epoch_info = dict()
self.meta_info = dict(epoch=0, iter=0)
self.device = device
self.io = torchlight.IO(self.args.work_dir,
save_log=self.args.save_log,
print_log=self.args.print_log)
# model
self.C = C
self.T = T
self.V = V
self.n_z = n_z
if not os.path.isdir(self.args.work_dir):
os.mkdir(self.args.work_dir)
self.model = CVAE.CVAE(C, T, V, self.n_z, num_classes, graph_dict)
self.model.cuda('cuda:0')
self.model.apply(weights_init)
self.loss = vae_loss
self.best_loss = math.inf
self.loss_updated = False
self.step_epochs = [
math.ceil(float(self.args.num_epoch * x)) for x in self.args.step
]
self.best_epoch = 0
self.mean = 0.
self.lsig = 1.
# optimizer
if self.args.optimizer == 'SGD':
self.optimizer = optim.SGD(self.model.parameters(),
lr=self.args.base_lr,
momentum=0.9,
nesterov=self.args.nesterov,
weight_decay=self.args.weight_decay)
elif self.args.optimizer == 'Adam':
self.optimizer = optim.Adam(self.model.parameters(),
lr=self.args.base_lr,
weight_decay=self.args.weight_decay)
else:
raise ValueError()
self.lr = self.args.base_lr
def init_environment(self):
self.io = torchlight.IO(self.arg.work_dir,
save_log=self.arg.save_log,
print_log=self.arg.print_log)
self.io.save_arg(self.arg)
# gpu
gpus = torchlight.visible_gpu(self.arg.device)
torchlight.occupy_gpu(gpus)
self.gpus = gpus
self.dev = "cuda:0"
def __init__(self,
args,
data_loader,
C,
F,
num_classes,
graph_dict,
device='cuda:0'):
self.args = args
self.data_loader = data_loader
self.num_classes = num_classes
self.result = dict()
self.iter_info = dict()
self.epoch_info = dict()
self.meta_info = dict(epoch=0, iter=0)
self.device = device
self.io = torchlight.IO(self.args.work_dir,
save_log=self.args.save_log,
print_log=self.args.print_log)
# model
if not os.path.isdir(self.args.work_dir):
os.mkdir(self.args.work_dir)
self.model = classifier.Classifier(C, F, num_classes, graph_dict)
self.model.cuda('cuda:0')
self.model.apply(weights_init)
self.loss = nn.CrossEntropyLoss()
self.best_loss = math.inf
self.step_epochs = [
math.ceil(float(self.args.num_epoch * x)) for x in self.args.step
]
self.best_epoch = 0
self.best_accuracy = np.zeros((1, np.max(self.args.show_topk)))
self.accuracy_updated = False
# optimizer
if self.args.optimizer == 'SGD':
self.optimizer = optim.SGD(self.model.parameters(),
lr=self.args.base_lr,
momentum=0.9,
nesterov=self.args.nesterov,
weight_decay=self.args.weight_decay)
elif self.args.optimizer == 'Adam':
self.optimizer = optim.Adam(self.model.parameters(),
lr=self.args.base_lr,
weight_decay=self.args.weight_decay)
else:
raise ValueError()
self.lr = self.args.base_lr
def init_environment(self):
self.io = torchlight.IO(self.arg.work_dir,
save_log=self.arg.save_log,
print_log=self.arg.print_log)
self.io.save_arg(self.arg)
print('Mid of init enviroment')
# gpu
if self.arg.use_gpu:
gpus = torchlight.visible_gpu(self.arg.device)
torchlight.occupy_gpu(gpus)
self.gpus = gpus
self.dev = "cuda:0"
else:
self.dev = "cpu"
def init_environment(self):
self.io = torchlight.IO( #class 'torchlight.io.IO',这里的torchlight.IO就是一个容器
self.arg.work_dir, #存储结果的路径，默认为'./work_dir/tmp'
save_log=self.arg.save_log, #是否保存日志，默认true
print_log=self.arg.print_log) #是否打印日志，默认true
self.io.save_arg(self.arg)
# gpu
if self.arg.use_gpu: #如果指定了使用GPU，就默认使用0号GPU，否则使用CPU
gpus = torchlight.visible_gpu(self.arg.device)
torchlight.occupy_gpu(gpus)
self.gpus = gpus
self.dev = "cuda:0"
else:
self.dev = "cpu"
def __init__(self, args, data_loader, device='cuda:0'):
self.args = args
# TBD:The data loader class has to be implemented to load the train, test and evaluation images
self.data_loader = data_loader
self.result = dict()
self.iter_info = dict()
self.epoch_info = dict()
self.meta_info = dict(epoch=0, iter=0)
self.device = device
self.io = torchlight.IO(
"./model_output/",
save_log=self.args.save_log,
print_log=self.args.print_log)
# Initialize timer to see the time taken for each batch
self.io.init_timer("batch_processing_time")
#
# model
self.model = classifier1.DepthPredictionNet()
self.model.cuda('cuda:0')
# TBD: Apart from the Resnet layers, all other layers has to be initialized with the weights as mentioned in the
# paper, For now we can go ahead and try the default initialization which can be later modified
# [Resolved: Doing at Model level]
self.loss = ReverseHubberLoss()
self.best_loss = math.inf
#self.step_epochs = [math.ceil(float(self.args.num_epoch * x)) for x in self.args.step]
self.best_epoch = None
self.best_mean_error = np.zeros((1, np.max(self.args.topk)))
self.mean_error_updated = False
# optimizer
if self.args.optimizer == 'SGD':
self.optimizer = optim.SGD(
self.model.parameters(),
lr=self.args.base_lr,
momentum=0.9,
nesterov=self.args.nesterov,
weight_decay=self.args.weight_decay)
elif self.args.optimizer == 'Adam':
self.optimizer = optim.Adam(
self.model.parameters(),
lr=self.args.base_lr,
weight_decay=self.args.weight_decay)
else:
raise ValueError()
self.lr = self.args.base_lr
self.mean_loss_per_lr_step = 0
def init_environment(self):
self.save_dir = os.path.join(self.arg.work_dir, self.arg.max_hop_dir,
self.arg.lamda_act_dir)
self.io = torchlight.IO(self.save_dir,
save_log=self.arg.save_log,
print_log=self.arg.print_log)
self.io.save_arg(self.arg)
# gpu
if self.arg.use_gpu:
gpus = torchlight.visible_gpu(self.arg.device)
torchlight.occupy_gpu(gpus)
self.gpus = gpus
self.dev = "cuda:0"
else:
self.dev = "cpu"
def init_environment(self):
self.save_dir = os.path.join(self.arg.work_dir,
self.arg.fusion_layer_dir,
self.arg.learning_rate_dir,
self.arg.lamda_dir,
self.arg.crossw_dir,
self.arg.note)
self.io = torchlight.IO(self.save_dir, save_log=self.arg.save_log, print_log=self.arg.print_log)
self.io.save_arg(self.arg)
if self.arg.use_gpu:
gpus = torchlight.visible_gpu(self.arg.device)
torchlight.occupy_gpu(gpus)
self.gpus = gpus
self.dev = "cuda:0"
else:
self.dev = "cpu"
