# 对tensor取数值，返回张量元素的值，对于tensor内有一个数值才可调用此方法，多个元素报错
tensor .item ()
# 和item方法很类似，但numpy方法可以将tensor转换成ndarray格式，不限制tensor内元素个数
tensor .numpy ()# tensor.eq()方法等同于tf.equal(),都是对tensor内的数据按元素对比（指定dim、axis），并且输出对应dim的布尔类型数据
torch .eq (tensor_a , tensor_b ) # --> def eq(self: Tensor, other: Tensor, *, out: Optional[Tensor]=None)
tensor_a .eq (tensor_b )
# tensor.equal()方法只输出两个tensor对比，是否相等
torch .equal (tensor_a , tensor_b ) # --> _bool:
tensor_a .equal (tensor_b )# tensor.numel()方法返回tensor占用内存大小
tensor_a .numel ()torch.set_default_tensor_type(torch.FloatTensor/torch.DoubleTensor) # torch.set_default_tensor_type()设定torch.Tensor()默认格式（如果不设置默认为FLostTensor）
torch .set_default_tensor_type ()# tensor.full()方法生成一个给定shape的全num型数组，给定输入需要是float不能是int
x = torch .full ((3 , 3 ), 5.3 )
x
tensor ([[5.3000 , 5.3000 , 5.3000 ],
[5.3000 , 5.3000 , 5.3000 ],
[5.3000 , 5.3000 , 5.3000 ]])torch.linspace(start, end, steps=None) # torch.linspace(start, end, steps=None)方法生成等分的数组
x = torch .linspace (0 , 3 , steps = 10 )
x
tensor ([0.0000 , 0.3333 , 0.6667 , 1.0000 , 1.3333 , 1.6667 , 2.0000 , 2.3333 , 2.6667 ,
3.0000 ])torch.logspace(start, end, steps=None) # torch.logspace(start: Number, end: Number, steps: _int=100, base: _float=10.0)生成以base为底的指数等分数组
x = torch .logspace (0 , 1 , steps = 10 , base = 2 )
x
tensor ([1.0000 , 1.0801 , 1.1665 , 1.2599 , 1.3608 , 1.4697 , 1.5874 , 1.7145 , 1.8517 ,
2.0000 ])# randperm(Number)生成基于number大小的随机乱序数，相当于np.random.permutation(),用于数据维度shuffle
x = torch .randperm (10 )
x
tensor ([3 , 6 , 7 , 5 , 9 , 2 , 4 , 0 , 8 , 1 ])tensor [0 , :, :, :] <= > tensor [0 , ...]
tensor [...] <= > tensor [:, :, :, :]
tensor [..., :2 ] <= > tensor [:, :, :, :2 ]tensor.view()和tensor.reshape() # 在pytorch0.3之前都是使用view函数对数组进行维度变换，但在0.4版本中为了保持与numpy一致，增加了reshape函数,两种函数用法完全相同
tensor_a .view (3 , 3 ) <= > tensor_a .reshape (3 ,3 )# dim -> 0 1 2 3
# -4 -3 -2 -1
# unsqueeze用于在axis增加维度
x = torch .rand ((2 , 3 , 28 , 28 )) # axis = 4的数据
y = x .unsqueeze (1 ) # 表示在axis=1的维度之前插入一个维度（pytorch常用dim，tensorflow常用axis，两者表达相同）
y .shape
torch .Size ([2 , 1 , 3 , 28 , 28 ])torch.cat([tensor_a, tensor_b], dim=Number) # cat方法与tf.concat(), np.concatante()用法一致都用于在原有维度上进行拼接，不会增加维度
x = torch .rand ((1 ,3 ,3 ))
y = torch .rand ((1 ,3 ,3 ))
torch .cat ([x , y ], dim = 0 )
- > tensor ([[[0.2267 , 0.5636 , 0.8269 ],
[0.8434 , 0.2561 , 0.7266 ],
[0.0459 , 0.9119 , 0.7341 ]],
[[0.4417 , 0.6696 , 0.7375 ],
[0.3352 , 0.4405 , 0.3100 ],
[0.2330 , 0.1244 , 0.1808 ]]])
torch .cat ([x , y ], dim = 0 ).shape
- > torch .Size ([2 , 3 , 3 ])torch.stack([tensor_a, tensor], dim=Number) # stack方法与tf.stack(),np.stack()用法一致，用于将两个张量在一个新的维度拼接，会增加维度
# dim位置表示新增加的维度在什么位置
# stack位置之后的数组大小一定要相同，发否则报错
x = torch .rand ((1 ,3 ,3 ))
y = torch .rand ((1 ,3 ,3 ))
torch .stack ([x , y ], dim = 1 ).shape
- > torch .Size ([1 , 2 , 3 , 3 ])
torch .stack ([x , y ], dim = 0 ).shape
- > torch .Size ([2 , 1 , 3 , 3 ])tensor.split(Number, dim=Number) # 方法按照给定dim将原张量按指定维度拆分成指定长度tensor，与tf.split()用法相同
# 示例为将z分为长度为2，2，1三个长度的张量
z = torch .rand ((5 ,3 ,3 ))
a ,b ,c = z .split ([2 ,2 ,1 ], dim = 0 )
a
tensor ([[[0.6907 , 0.1767 , 0.2757 ],
[0.0949 , 0.6095 , 0.0204 ],
[0.3591 , 0.7025 , 0.1655 ]],
[[0.8722 , 0.5272 , 0.8849 ],
[0.5268 , 0.4994 , 0.4219 ],
[0.1545 , 0.1390 , 0.5757 ]]])
b
tensor ([[[0.4904 , 0.2658 , 0.3442 ],
[0.5530 , 0.2156 , 0.0235 ],
[0.1533 , 0.5226 , 0.0291 ]],
[[0.1012 , 0.4376 , 0.9082 ],
[0.8089 , 0.9455 , 0.0567 ],
[0.8221 , 0.6370 , 0.9685 ]]])
c
tensor ([[[0.1732 , 0.1510 , 0.9870 ],
[0.7290 , 0.7948 , 0.9145 ],
[0.9889 , 0.4717 , 0.8628 ]]])tensor.chunk(num, dim=num) # 方法将原tensor按维度以给定数量拆分
z = torch .rand ((6 ,3 ,3 ))
a , b = z .chunk (2 , dim = 0 )
a .shape
- > torch .Size ([3 , 3 , 3 ])
b .shape
- > torch .Size ([3 , 3 , 3 ])# expand只能把维度为1的拓展成指定维度。如果哪个维度为-1，就是该维度不变
# 如果扩展非维度为1，则会报错
# expand使用内存较少，速度较快
x = torch .rand ((1 , 4 , 2 , 1 ))
x .expand (4 , 4 , 2 , 6 ).shape
- > torch .Size ([4 , 4 , 2 , 6 ])# torch.repeat()里面参数代表是重复多少次，就是复制多少次，比如下面2， 3， 1， 6代表复制2， 3， 1， 6次
# repeat中的数据不能为-1
x = torch .rand ((1 , 4 , 1 , 1 ))
x .repeat (2 , 3 , 1 , 6 ).shape
torch .Size ([2 , 12 , 1 , 6 ])# tensor.t()只能用于二维转置
x = torch .rand (4 ,3 )
x .t ()
tensor ([[0.3490 , 0.0071 , 0.6820 , 0.6466 ],
[0.8913 , 0.3535 , 0.5594 , 0.9340 ],
[0.7992 , 0.0214 , 0.9171 , 0.8219 ]])
x .t ().shape
torch .Size ([3 , 4 ])tensor.transpose(dim=Number, dim=Number) # transpose用于维度转换，但需要借助contiguous来复制内存从而不被原始数据混淆
x = torch .rand ((1 , 4 , 3 , 2 ))
y = x .transpose (1 , 3 ).contiguous ().view (1 , 4 * 3 * 2 ).view (1 , 2 , 3 , 4 ).contiguous ()
y
tensor ([[[[0.1190 , 0.2306 , 0.5631 , 0.4590 ],
[0.8008 , 0.9555 , 0.6722 , 0.7551 ],
[0.6167 , 0.3300 , 0.6375 , 0.1747 ]],
[[0.0554 , 0.7831 , 0.4375 , 0.2000 ],
[0.0568 , 0.7865 , 0.6037 , 0.5465 ],
[0.7036 , 0.4141 , 0.2600 , 0.1208 ]]]])
x
tensor ([[[[0.1190 , 0.0554 ],
[0.8008 , 0.0568 ],
[0.6167 , 0.7036 ]],
[[0.2306 , 0.7831 ],
[0.9555 , 0.7865 ],
[0.3300 , 0.4141 ]],
[[0.5631 , 0.4375 ],
[0.6722 , 0.6037 ],
[0.6375 , 0.2600 ]],
[[0.4590 , 0.2000 ],
[0.7551 , 0.5465 ],
[0.1747 , 0.1208 ]]]])
x .shape
torch .Size ([1 , 4 , 3 , 2 ])
y .shape
torch .Size ([1 , 2 , 3 , 4 ])
y = x .transpose (1 , 3 ).contiguous ().view (1 , 4 * 3 * 2 ).view (1 , 2 , 3 , 4 ).contiguous ().transpose (1 , 3 )
y
tensor ([[[[0.1190 , 0.0554 ],
[0.8008 , 0.0568 ],
[0.6167 , 0.7036 ]],
[[0.2306 , 0.7831 ],
[0.9555 , 0.7865 ],
[0.3300 , 0.4141 ]],
[[0.5631 , 0.4375 ],
[0.6722 , 0.6037 ],
[0.6375 , 0.2600 ]],
[[0.4590 , 0.2000 ],
[0.7551 , 0.5465 ],
[0.1747 , 0.1208 ]]]])
y .shape
torch .Size ([1 , 4 , 3 , 2 ])
# 只有在交换维度之后用view重排维度，之后才能将数据还原回去
# transpose、permute等维度变换操作后，tensor在内存中不再是连续存储的，而view操作要求tensor的内存连续存储，所以需要contiguous来返回一个contiguous copy # permute和transpose都可以进行维度交换，但permute在高维的功能性更强
# transpose只能对两个维度进行对调， 但permute可以一次完成这些操作
y = x .permute (0 , 2 , 3 , 1 )
y
tensor ([[[[0.4613 , 0.2058 , 0.1494 , 0.4891 ],
[0.9352 , 0.7827 , 0.4933 , 0.7655 ]],
[[0.5409 , 0.7702 , 0.9308 , 0.2221 ],
[0.5126 , 0.1622 , 0.0414 , 0.4646 ]],
[[0.3983 , 0.0387 , 0.1282 , 0.6299 ],
[0.1228 , 0.0965 , 0.4121 , 0.3827 ]]]])
# 如果是transpose需要两次才能完成操作 # 广播机制表示当维度不统一时，相当于unsqueeze去插入新的维度，并且用expand在已存在维度上复制数据
# 从而使两个张量维度相同
y
- > tensor ([[ 0 ],
[10 ],
[20 ],
[30 ]])
x
- > tensor ([0 , 1 , 2 ])
x + y
- > tensor ([[ 0 , 1 , 2 ],
[10 , 11 , 12 ],
[20 , 21 , 22 ],
[30 , 31 , 32 ]])torch.nn.init.kaiming_normal_(tensor) '''
根据He, K等人2015年在"深入研究了超越人类水平的性能：整流器在ImageNet分类"中描述的方法，采用正态分布，填充张量或变量。结果张量中的值采样自均值为0，标准差为sqrt(2/((1 + a^2) * fan_in))的正态分布。该方法也被称为He的初始化
'''
x = torch .rand ((200 , 784 ))
torch .nn .init .kaiming_normal_ (x )'''
class-style API
nn.ReLu() 对于这种类型为类风格API，不能进行直接调用，只能将类示例为方法再调用，通常用于Sequential模型构建(继承自nn.Module)，但是这种类型访问中间变量需要通过parameter参数间接访问
function-style API
F.relu() 方法风格API,直接访问调用，相比于上面的api，这种方法可以更直观更便捷的访问内部参数
'''
# 示例如下
# function-style
x = torch .rand ((28 , 28 ))
x = torch .nn .functional .relu (x , inplace = True )
# class-style
layer = torch .nn .ReLU ()
z = layer (x )'''
通过继承torch.nn.Module父类构建网络，构建完成的网络可以使用nn.Sequential对网络打包，从而达成更深层次的网络构建。
继承nn.Module构建网络需要重建__init__()和forward(), forward()方法的构建是继承网络中最为重要的一部分，它决定了网络的层次结构，以resnet网络构建举例.
通过在__init__实例化网络中间层的类，在forward中将网络进行拼接。
'''
class ResBlock (torch .nn .Module ):
def __init__ (self , in_Channel , out_Channel , stride = 1 ):
super (ResBlock , self ).__init__ ()
# 第一层卷积步长使用给定步长，以便将图像长宽按stride倍数减半
self .Conv1 = torch .nn .Conv2d (in_channels = in_Channel , out_channels = out_Channel , kernel_size = 3 , padding = 1 , stride = stride , bias = False )
self .BatchNorm1 = torch .nn .BatchNorm2d (out_Channel )
self .Conv2 = torch .nn .Conv2d (in_channels = out_Channel , out_channels = out_Channel , kernel_size = 3 , padding = 1 , stride = 1 , bias = False )
self .BatchNorm2 = torch .nn .BatchNorm2d (out_Channel )
# 为保证原输入通道数和输出通道数相同，使用kernal为1的卷积核改变输出channel5
if in_Channel != out_Channel or stride != 1 :
self .Conv3 = torch .nn .Conv2d (in_channels = in_Channel , out_channels = out_Channel , kernel_size = 1 , stride = stride , bias = False )
else :
self .Conv3 = None
def forward (self , input_ ):
fx = self .Conv1 (input_ )
fx = torch .nn .functional .relu (self .BatchNorm1 (fx ))
fx = self .Conv2 (fx )
fx = self .BatchNorm2 (fx )
# 如果输入输出channel不匹配
if self .Conv3 :
input_ = self .Conv3 (input_ )
return torch .nn .functional .relu (fx + input_ )
class ResNet (torch .nn .Module ):
def __init__ (self , channel_list , block_num ):
super (ResNet , self ).__init__ ()
self .image_channel = 3
self .resnet_block = channel_list
assert len (channel_list ) == block_num
self .net = self .net_initilize ()
self .layer1 = self .Block_Builder (64 , 64 , self .resnet_block [0 ], init = False )
self .layer2 = self .Block_Builder (64 , 128 , self .resnet_block [1 ], init = True ) # 下采样，使用步长为2
self .layer3 = self .Block_Builder (128 , 256 , self .resnet_block [2 ], init = True ) # 下采样，使用步长为2
self .layer4 = self .Block_Builder (256 , 512 , self .resnet_block [3 ], init = True , stride = 1 ) # 步长为1，不改变大小
self .Avgpool = torch .nn .AdaptiveAvgPool2d ((4 , 4 ))
self .fc = torch .nn .Sequential (
torch .nn .Linear (512 * 4 * 4 , 512 ),
torch .nn .Linear (512 , 256 ),
torch .nn .Linear (256 , 10 )
)
# x = torch.rand((3, 3, 32, 32))
# out = self.net(x)
# out = self.layer1(out)
# out = self.layer2(out)
# out = self.layer3(out)
# out = self.layer4(out)
# print(out.shape)
def forward (self , x ):
x = self .net (x )
x = self .layer1 (x )
x = self .layer2 (x )
x = self .layer3 (x )
x = self .layer4 (x )
x = self .Avgpool (x )
x = torch .flatten (x , start_dim = 1 , end_dim = - 1 )
x = self .fc (x )
return x
@staticmethod
def Block_Builder (inchannel , outchannel , num_block , init = False , stride = 2 ):
res_block = torch .nn .Sequential ()
for index in range (num_block ):
if init and index == 0 :
res_block .add_module ("init_layer" , ResBlock (inchannel , outchannel , stride = stride ))
else :
res_block .add_module (f"add_block{ index } , ResBlock (outchannel , outchannel , stride = 1 ))
return res_block
def net_initilize (self ):
net = torch .nn .Sequential (
torch .nn .Conv2d (in_channels = self .image_channel , out_channels = 64 , kernel_size = 7 , stride = 2 , padding = 3 ),
torch .nn .BatchNorm2d (64 ),
torch .nn .ReLU ())
# torch.nn.MaxPool2d(kernel_size=3, stride=2, padding=1))
return net 通过torch.save()保存训练数据（保存数据为字典形式，读取方式也为对应字典形式） # 通过torch.save方法保存checkpoints
torch .save ({'epochs' : epoch , 'model' : net .state_dict (), 'optimizer' : optim .state_dict ()}, '3in1.pt' )
# 通过torch.load方法读取checkpoints并从dict中读取参数到网络中
checkpoints = torch .load ('3in1.pt' )
checkpoints .keys ()
> >> dict_keys (['epochs' , 'model' , 'optimizer' ])
# 实例化网络
model = Data_dim_reduce ()
# 载入参数
model .load_state_dict (checkpoints ['model' ])
> >> < All keys matched successfully >
# 实例化优化器
optim = torch .optim .Adam (model .parameters ())
# 载入优化器参数
optim .load_state_dict (checkpoints ['optimizer' ])pytorch中clone和detach对数据的影响 '''
前几天在做DDQN代码移植的时候碰到一个问题，DDQN需要在迭代更新参数的过程中将参与训练的模型参数复制给用来减小自举影响的target模型，在复制的过程中，如果单纯的将target_model.state_dict(model.state_dict())通过这种方式赋值给target网络，由于.state_dict()方法的特殊性，导致target_model与model共享内存，从而使得DDQN自举过度，训练无法收敛；
所以考虑了deepcopy取state_dict的值，可以避免发生这种错误
今天在做yolo锚框的时候觉得可以使用另一种方法来解决这种问题
tensor.detach()用作将tensor与原计算图分离，内存位置仍然指向原tensor位置，但是require_grad置为Flase，也就意味着这个tensor不计算梯度，不具有grad
tensor.clone()用作将原tensor内容复制到新内存空间，保留新tensor的grad回溯
将此方法应用于之前参数复制问题上，实际可以通过model.state_dict.detach().clone()复制权重参数
''' '''
今天在用pytorch做多通道卷积举例的时候需要对卷积核里的参数进行自定义，发现pytorch可以通过实例化的调用直接更改权重，以下为代码示例
需要注意的一点是，当生成了和给定卷积权重相同shape的权重后，直接应用conv.weight = custom_weight是不行的，因为conv.weight限制了赋值变量类型，
变量类型也需要满足Parameter的数据类型，需要通过torch.nn.Parameter(custom_weight)改变数据类型
pytorch中torch.FloatTensor对应tensorflow中的tf.float32,
torch.LongTensor对应tensorflow中的tf.float64
'''
# -*- coding: utf-8 -*-
import torch
import numpy
x = torch .arange (0 , 9 ).reshape (3 , 3 )
y = torch .arange (1 , 10 ).reshape (3 , 3 )
# 定义多通道输入数组大小
input_array = torch .stack ((x , y ), dim = 0 )
input_array = input_array .unsqueeze (dim = 0 )
# noinspection PyTypeChecker
input_array = input_array .type (torch .FloatTensor )
print (f'输入卷积的数组大小：{ input_array .shape } )
# 实例化卷积方法， 禁用bias方便后续手算卷积
conv = torch .nn .Conv2d (in_channels = 2 , out_channels = 1 , kernel_size = 2 , padding = 0 , stride = 1 , bias = False )
# 定义卷积权重weight
w = torch .arange (0 , 4 ).reshape (2 , 2 )
w1 = torch .arange (1 , 5 ).reshape (2 , 2 )
w = torch .stack ((w , w1 ), dim = 0 )
w = w .unsqueeze (dim = 0 )
# noinspection PyTypeChecker
w = w .type (torch .FloatTensor )
print (f'卷积权重形状(shape): { w .shape } )
# 将我们自定义的卷积权重应用于conv方法
print (f'从nn类中实例化出来的Conv2D初始化参数为： { conv .weight } )
conv .weight = torch .nn .parameter .Parameter (w )
print (f'应用自定义权重之后的Conv2D参数为： { conv .weight } )
# 用以下方法自定义卷积权重一样可以，两个方法指向的代码相同
# conv.weight = torch.nn.Parameter(w)
out = conv (input_array )
print (out )Keras卷积padding为‘SAME’和‘VALID’时，对应Pytorch的padding方式 '''
# 实际卷积中，tensorflow和pytorch中给的padding凡是有确切数值的，padding都只代表一边的padding像素数
# keras卷积计算输出
# Valid
# 输出大小 = （输入大小 - kernal_size + strides）/ strides
# SAME
# 输出大小 = 输入大小 / strides
#Pytoch计算卷积输出
# 输出大小 = (输入大小 - kernal_size + padding*2 + strides) / strides
# 可以看出keras会自动平衡“(输入大小 - kernal_size + padding*2 + strides)”这一部分的大小，使其与原输入大小相同，
# 即，使得“kernal_size - padding*2 - strides”计算结果为0
'''
# 试验pytorch和keras相同输出
# keras喂入数据(需要满足tensorflow数据格式，即通道在最后一维)
x = np .random .normal (size = (1 , 32 , 32 , 3 ))
# 在keras和Pytorch中，通过卷积计算的输出图像如果不为整数，则只取整数部分，相当于留有未卷积的边
conv2_keras = tf .keras .layers .Conv2d (16 , (4 , 4 ), padding = "same" , strides = (4 , 4 ), input_shape = (3 , 32 , 32 ))
cone2_keras (x ).shape
> >> TensorShape ([1 , 8 , 8 , 16 ])
# Pytorch
# 喂入数据满足caffe数据格式，通道数在dim=1位置
x = torch .rand ((1 , 3 , 32 , 32 ))
conv2_torch = torch .nn .Conv2d (in_channels = 16 , out_channels = 32 , kernel_size = (4 , 4 ), stride = (4 , 4 ), padding = (0 , 0 ))
conv2_torch (x ).shape
> >> torch .Size ([1 , 16 , 8 , 8 ])
# keras会自动填充对应数据，使其满足输出大小，而填充的像素数可以通过上述公式计算出来，并且通过pytorch验证。 pytorch中optimizer.step(), loss.backward()以及optimizer.zero_grad()代码关系 # DDPG算法换用DDQN过程中，使用一个common网络对图像进行特征提取，再将提取的特征传入Actor和Critic网络。
# Actor和Critic更新的过程中，需要Critic_optimizer和Actor_optimizer不仅对自己的网络更新，
# 同时通过梯度回流到common网络一起进行更新。具体代码如下
# 优化器初始化代码
self .opt_actor = torch .optim .Adam (itertools .chain (self .common_model .parameters (),
self .actor_model .parameters ()),
lr = 1e-4 )
self .opt_critic = torch .optim .Adam (itertools .chain (self .common_model .parameters (),
self .critic_model .parameters ()),
lr = 1e-3 , weight_decay = 1e-2 )
# 经验池训练代码
def train_replay (self ):
if len (self .memory ) < self .train_start :
return
elif self .train_from_checkpoint :
if len (self .memory ) < self .train_from_checkpoint_start :
return
batch_size = min (self .batch_size , len (self .memory ))
minibatch = random .sample (self .memory , batch_size )
'''
torch.float64对应torch.DoubleTensor
torch.float32对应torch.FloatTensor
'''
# from_numpy会对numpy数据类型转换成双精度张量，而torch.Tensor不存在这种问题，torch.Tensor将数组转换成单精度张量
state_t , v_ego_t , action_t , reward_t , state_t1 , v_ego_t1 , terminal , step = zip (* minibatch )
state_t = Variable (torch .Tensor (state_t ).squeeze ().to (self .device ))
state_t1 = Variable (torch .Tensor (state_t1 ).squeeze ().to (self .device ))
v_ego_t = Variable (torch .Tensor (v_ego_t ).squeeze ().to (self .device ))
v_ego_t1 = Variable (torch .Tensor (v_ego_t1 ).squeeze ().to (device ))
reward_t = torch .Tensor (reward_t ).reshape (- 1 , 1 ).to (self .device )
action_t = torch .Tensor (action_t ).reshape (- 1 , 1 ).to (self .device )
self .opt_critic .zero_grad ()
feature_extraction_target = self .common_target_model (state_t1 , v_ego_t1 )
action_target = self .actor_target_model (feature_extraction_target )
td_target = reward_t + self .discount_factor * self .critic_target_model (feature_extraction_target , action_target .detach ())
feature_extraction = self .common_model (state_t , v_ego_t )
critic_loss_cal = self .loss_critic (self .critic_model (feature_extraction , action_t ), td_target .detach ())
critic_loss_cal .backward ()
self .opt_critic .step ()
self .history_loss_critic = critic_loss_cal .item ()
# 重置critic，actor优化器参数
self .opt_actor .zero_grad ()
self .opt_critic .zero_grad ()
acotr_feature_extraction = self .common_model (state_t , v_ego_t )
policy_actor = - self .critic_model (acotr_feature_extraction , self .actor_model (acotr_feature_extraction ))
policy_actor = policy_actor .mean ()
policy_actor .backward ()
self .opt_actor .step ()
self .history_loss_actor = policy_actor .item ()
soft_update_target_model (self .common_model , self .common_target_model , self .tua )
soft_update_target_model (self .actor_model , self .actor_target_model , self .tua )
soft_update_target_model (self .critic_model , self .critic_target_model , self .tua )pytorch 0.4版本中后Variable和Tensor合并之后的关系 import torch
from torch .autograd import Variable
x_tensor = torch .Tensor (3 ,4 )
x_var = Variable (x_tensor ,requires_grad = True )
'''
实际查看x_var和x_tensor的requires_grad_()查看对应tensor的grad发现Variable和Tensor创建的对象是一样的，也就是说在0.4版本之后不需要再对tensor做Variable操作，即可进行autograd追踪梯度。
''' YOLOv5 调用.show()方法后会在原图上叠加预测锚框 # ACC_RL中对yolo识别之后的数据处理
anchor_box = yolo (images )
position = np .array (xyxy2xyxy (anchor_box .pred [0 ]), dtype = 'uint8' )
if position .shape [0 ] != 0 :
ImageDraw .Draw (image ).rectangle ([(position [0 , 0 ], position [0 , 1 ]), (position [0 , 2 ], position [0 , 3 ])], outline = 'yellow' , width = 3 )
image = image .resize ((80 , 80 ), resample = Image .BILINEAR )
image = np .array (image )
# anchor_box = yolo(images)之后调用anchor_box.show()方法，相当于对原图进行ImageDraw,标记框会直接加在原图上