这一周将必看的torch文件全部看完!!!!
在构建神经网络架构时,都要继承自基类nn.Module。
class Module(object)
def __init__(self):
self._backend = thnn_backend
self._parameters = OrderedDict() # 字典,保存用户直接设置的parameters
self._buffers = OrderedDict() # 缓存
self._backward_hooks = OrderedDict() # 钩子技术,用来提取中间变量
self._forward_hooks = OrderedDict()
self._forward_pre_hooks = OrderedDict()
self._modules = OrderedDict()
self.training = True # BatchNorm和DropOut层在训练阶段和测试阶段中采取的策略不同,通过判断traing值决定前向传播的策略
接下来看看该类的比较常用的几个function:
def forward(self, *input):
# 继承该类的所有子类中都要重写该函数
raise NotImplementedError
增加子模块到当前模块中:
def add_module(self, name, module):
if not isinstance(module, Module) and module is not None:
raise TypeError("{} is not a Module subclass".format(
torch.typename(module)))
elif hasattr(self, name) and name not in self._modules:
raise KeyError("attribute '{}' already exists".format(name))
elif '.' in name:
raise KeyError("module name can't contain "."")
elif name == '':
raise KeyError("module name can't be empty string """)
self._modules[name] = module
该函数的最后一行self._modules[name] = module,而self._module是一个字典,保存的是子module。函数children可以查看直接子module。函数modules可以查看所有的子module。
def children(self):
for name, module in self.named_children():
yield module
def named_children(self):
memo = set()
for name, module in self._modules.items():
if module is not None and module not in memo:
memo.add(module)
yield name, module # yield同return差不多,但是它的type是一个生成器,可以节省时间
函数modules可以查看所有的子module:
def modules(self):
for name, module in self.named_modules():
yield module
def named_modules(self, memo=None, prefix=''):
if memo is None:
memo = set()
if self not in memo:
memo.add(self)
yield prefix, self
for name, module in self._modules.items():
if module is None:
continue
submodule_prefix = prefix + ('.' if prefix else '') + name
for m in module.named_modules(memo, submodule_prefix):
yield m
对于batchnorm, dropout, instancenorm等在训练和测试阶段行为差距较大的层,如果在测试时不将其training值设为False,则会有很大影响。推荐调用model.train()函数,它会将当前module及子module中的所有training属性都设为True。相应的,model.eval()函数会把training属性都设为False。
def train(self, mode=True):
self.training = mode
for module in self.children():
module.train(mode)
return self
def eval(self):
return self.train(False)
可以用named_parameters()来查询子module的名字和params的大小等信息:
def parameters(self):
for name, param in self.named_parameters():
yield param
def named_parameters(self, memo=None, prefix=''):
if memo is None:
memo = set()
for name, p in self._parameters.items():
if p is not None and p not in memo:
memo.add(p)
yield prefix + ('.' if prefix else '') + name, p
for mname, module in self.named_children():
submodule_prefix = prefix + ('.' if prefix else '') + mname
for name, p in module.named_parameters(memo, submodule_prefix):
yield name, p
使用方法:
for name, param in net.named_parameters():
print(name, param.size())
将模型放到GPU上运行也十分的简单,调用cuda函数就好,下面是在单GPU上运行:
def cuda(self, device=None):
return self._apply(lambda t: t.cuda(device))
model = model.cuda() # 将模型的所有参数转存到GPU上
input.cuda() # 将输入数据放置在GPU上
多GPU上并行计算时,pytorch提供了其他两个函数:nn.parallel.data_parallel()和class torch.nn.DataParallel(),这个以后再说。
加更:2020.04.10
self.load_state_dict()加载指定路径的模型。
def load_state_dict(self, state_dict, strict=True):
missing_keys = []
unexpected_keys = []
error_msgs = []
# copy state_dict so _load_from_state_dict can modify it
metadata = getattr(state_dict, '_metadata', None)
state_dict = state_dict.copy()
if metadata is not None:
state_dict._metadata = metadata
def load(module, prefix=''):
module._load_from_state_dict(
state_dict, prefix, strict, missing_keys, unexpected_keys, error_msgs)
for name, child in module._modules.items():
if child is not None:
load(child, prefix + name + '.')
load(self)
if strict:
error_msg = ''
if len(unexpected_keys) > 0:
error_msgs.insert(
0, 'Unexpected key(s) in state_dict: {}. '.format(
', '.join('"{}"'.format(k) for k in unexpected_keys)))
if len(missing_keys) > 0:
error_msgs.insert(
0, 'Missing key(s) in state_dict: {}. '.format(
', '.join('"{}"'.format(k) for k in missing_keys)))
if len(error_msgs) > 0:
raise RuntimeError('Error(s) in loading state_dict for {}:nt{}'.format(
self.__class__.__name__, "nt".join(error_msgs)))