[人工智能]基于深度强化学习的绘画智能体 代码分析(二)

基于深度强化学习的绘画智能体 代码分析(二)

Github源码链接

critic.py

import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.nn.utils.weight_norm as weightNorm

from torch.autograd import Variable //Variable是Autograd的核心类，它封装了Tensor，并整合了反向传播的相关实现
import sys

def conv3x3(in_planes, out_planes, stride=1):  //定义3*3的二维卷积模板
    return weightNorm(nn.Conv2d(in_planes, out_planes, kernel_size=3, stride=stride, padding=1, bias=True))

device = torch.device("cuda" if torch.cuda.is_available() else "cpu") //这个device的用处是作为Tensor或者Model被分配到的位置。
coord = torch.zeros([1, 2, 64, 64]) //返回一个形状为[1, 2, 64, 64]的矩阵,里面的每一个值都是0的tensor
for i in range(64):
    for j in range(64):
        coord[0, 0, i, j] = i / 63. //目的是将矩阵的维度长度平均到1？？
        coord[0, 1, i, j] = j / 63.
        coord = coord.to(device) //这行代码的意思是将所有最开始读取数据时的tensor变量copy一份到device所指定的GPU上去，之后的运算都在GPU上进行。      

class TReLU(nn.Module):
    def __init__(self)://初始化
        super(TReLU, self).__init__()//就是对继承自父类nn.Module的属性进行初始化。而且是用nn.Module的初始化方法来初始化继承的属性。
        self.alpha = nn.Parameter(torch.FloatTensor(1), requires_grad=True) 
        
//torch.nn.Parameter()函数:含义是将一个固定不可训练的tensor转换成可以训练的类型parameter，并将这个parameter绑定到这个module里面(net.parameter()中就有这个绑定的parameter，所以在参数优化的时候可以进行优化的)，所以经过类型转换这个self.v变成了模型的一部分，成为了模型中根据训练可以改动的参数了。使用这个函数的目的也是想让某些变量在学习的过程中不断的修改其值以达到最优化。
//比如cnn输出4个东西，你又不想concate到到一起，你想用权重加法，权重又不想自己设定，想让网络自己学requires_grad=True这个很重要
        self.alpha.data.fill_(0)

def forward(self, x): //损失函数
        x = F.relu(x - self.alpha) + self.alpha
        return x

def cfg(depth):
    depth_lst = [18, 34, 50, 101, 152]
    assert (depth in depth_lst), "Error : Resnet depth should be either 18, 34, 50, 101, 152" //assert检查条件，不符合就终止程序，检测了Resnet 的层数？
    cf_dict = {
        '18': (BasicBlock, [2,2,2,2]),
        '34': (BasicBlock, [3,4,6,3]),
        '50': (Bottleneck, [3,4,6,3]), //Bottleneck是对于更深的网络，提出了另一种残差基础块
        '101':(Bottleneck, [3,4,23,3]),
        '152':(Bottleneck, [3,8,36,3]),
    }

    return cf_dict[str(depth)]

class BasicBlock(nn.Module): //定义BasicBlock内容
    expansion = 1 //expansion是BasicBlock和Bottleneck的核心区别之一
    

    def __init__(self, in_planes, planes, stride=1):
        super(BasicBlock, self).__init__()
        self.conv1 = conv3x3(in_planes, planes, stride) //两个3x3的卷积
        self.conv2 = conv3x3(planes, planes)

        self.shortcut = nn.Sequential()//输入和输出维度匹配的情况
        

//输入和输出维度不匹配的情况(需要借助conv+bn将输入尺寸降低)
        if stride != 1 or in_planes != self.expansion * planes:
            self.shortcut = nn.Sequential(
                weightNorm(nn.Conv2d(in_planes, self.expansion*planes, kernel_size=1, stride=stride, bias=True)),
            )
        self.relu_1 = TReLU()
        self.relu_2 = TReLU()

    def forward(self, x):
        out = self.relu_1(self.conv1(x))
        out = self.conv2(out)
        out += self.shortcut(x)
        out = self.relu_2(out)

        return out

class Bottleneck(nn.Module): //定义Bottleneck内容
    expansion = 4

    def __init__(self, in_planes, planes, stride=1):
        super(Bottleneck, self).__init__()
        
// weightNorm:对参数的规范化
        self.conv1 = weightNorm(nn.Conv2d(in_planes, planes, kernel_size=1, bias=True))
        self.conv2 = weightNorm(nn.Conv2d(planes, planes, kernel_size=3, stride=stride, padding=1, bias=True))
        self.conv3 = weightNorm(nn.Conv2d(planes, self.expansion*planes, kernel_size=1, bias=True))
        self.relu_1 = TReLU()
        self.relu_2 = TReLU()
        self.relu_3 = TReLU()

        self.shortcut = nn.Sequential()
        if stride != 1 or in_planes != self.expansion*planes:
            self.shortcut = nn.Sequential(
                weightNorm(nn.Conv2d(in_planes, self.expansion*planes, kernel_size=1, stride=stride, bias=True)),
            )

    def forward(self, x):
        out = self.relu_1(self.conv1(x))
        out = self.relu_2(self.conv2(out))
        out = self.conv3(out)
        out += self.shortcut(x)
        out = self.relu_3(out)

        return out

class ResNet_wobn(nn.Module): //定义整个残差网络
    def __init__(self, num_inputs, depth, num_outputs):
        super(ResNet_wobn, self).__init__()
        self.in_planes = 64

        block, num_blocks = cfg(depth)
        self.conv0 = conv3x3(num_inputs, 32, 2) // 64        
        self.layer1 = self._make_layer(block, 64, num_blocks[0], stride=2) // 32
        self.layer2 = self._make_layer(block, 128, num_blocks[1], stride=2) // 16
        self.layer3 = self._make_layer(block, 256, num_blocks[2], stride=2)
        self.layer4 = self._make_layer(block, 512, num_blocks[3], stride=1)
        self.conv4 = weightNorm(nn.Conv2d(512, 1, 1, 1, 0))
        self.relu_1 = TReLU()
        self.conv1 = weightNorm(nn.Conv2d(65 + 2, 64, 1, 1, 0))        
        self.conv2 = weightNorm(nn.Conv2d(64, 64, 1, 1, 0))
        self.conv3 = weightNorm(nn.Conv2d(64, 32, 1, 1, 0))
        self.relu_2 = TReLU()
        self.relu_3 = TReLU()
        self.relu_4 = TReLU()
        

//_make_layer方法作用是生成多个卷积层，形成一个大的模块。
    def _make_layer(self, block, planes, num_blocks, stride):
        strides = [stride] + [1]*(num_blocks-1)
        layers = []

        for stride in strides:
            layers.append(block(self.in_planes, planes, stride))
            self.in_planes = planes * block.expansion

        return nn.Sequential(*layers)

    def a2img(self, x):
        tmp = coord.expand(x.shape[0], 2, 64, 64) //coord_ 系列函数可以改变xy轴的位置，x.shape[0]读取矩阵第一维度的长度
        x = x.repeat(64, 64, 1, 1).permute(2, 3, 0, 1) //permute函数可以对任意高维矩阵进行转置
        x = self.relu_2(self.conv1(torch.cat([x, tmp], 1))) //torch.cat()是为了把函数torch.stack()得到tensor进行拼接(concatnate)而存在的。函数目的： 在给定维度上对输入的张量序列seq 进行连接操作。
        x = self.relu_3(self.conv2(x))
        x = self.relu_4(self.conv3(x))
        return x
        
    def forward(self, input):
        x, a = input
        a = self.a2img(a)
        x = self.relu_1(self.conv0(x))
        x = torch.cat([x, a], 1)        
        x = self.layer1(x)
        x = self.layer2(x)
        x = self.layer3(x)
        x = self.layer4(x)
        x = self.conv4(x)
        return x.view(x.size(0), 64)