代码阅读：models.py

1.对于图片（NxCxHxW）的2D卷积与反卷积

卷积

input1 = torch.randn(32, 3, 12, 12)
ngf = 8

downsample = nn.Conv2d(3, ngf*2, 4, stride=1, padding=0)
h = downsample(input1)
print(h.size())
>>>torch.Size([32, 16, 9, 9])

反卷积

input1 = torch.randn(32, 3, 12, 12)
ngf = 8

upsample = nn.ConvTranspose2d(ngf*2, ngf, 4, stride=1, padding=0)
output = upsample(h, output_size=input1.size())

print(output.size())
>>>torch.Size([32, 8, 12, 12])

归一化

m = nn.BatchNorm2d(50)
input = torch.randn(20, 50, 35, 45)
output = m(input)
print(output.size())
>>>torch.Size([20, 50, 35, 45])

BatchNorm2d（n）中的参数n与tensor（N x C x H x W）中的C必须相同

2.GeneratorVideo中的get_gru_initial_state函数

a = Variable(torch.FloatTensor(num_samples,dim_z_motion).normal_())
a = torch.FloatTensor(5,3).normal_()
a
>>>tensor([[-1.3049,  0.0021,  0.1509],
        [ 0.3527,  0.5780,  0.0437],
        [-0.2406, -0.8861,  0.1983],
        [ 1.0273,  1.0479, -0.1404],
        [ 0.5898,  0.0900,  0.1426]])

FloatTensor函数返回一个tensor张量，第一个参数为数组的个数，第二个参数为每个数组的维度。此处即为采样的个数和要构建的gru单元的motion的维度。

*normal_(mean=0, std=1, , generator=None) 函数则是根据参数mean和std从正态分布中取值
故此处函数的作用是返回一个num_sample X dim_z_motion 维度，从正态分布中取值的tensor作为gru的初始值
Variable 将tensor封装之后可以调用.backward()函数进行梯度计算

3.sample_z_categ

首先根据所要分配给类别的向量数，构建随机整数的向量作为类别初始值

构建一个全零向量，维度为采样视频数 x 类别向量数

根据初始类别和全零向量矩阵构建每个video的one_hot矩阵

为了给video中每帧图片类别值，将video的每行one_hot复制n份，n为video_length.

返回所有帧的类别、所有视频的类别

实验示例

a= np.random.randint(6, size=8)
b = np.zeros((8, 6), dtype=np.float32)
b[np.arange(8),a]  = 1
c = np.repeat(b, 3, axis=0)
print(a)
print(b)
print(c)
[0 2 0 3 5 3 1 0]

[[1. 0. 0. 0. 0. 0.]
 [0. 0. 1. 0. 0. 0.]
 [1. 0. 0. 0. 0. 0.]
 [0. 0. 0. 1. 0. 0.]
 [0. 0. 0. 0. 0. 1.]
 [0. 0. 0. 1. 0. 0.]
 [0. 1. 0. 0. 0. 0.]
 [1. 0. 0. 0. 0. 0.]]

[[1. 0. 0. 0. 0. 0.]
 [1. 0. 0. 0. 0. 0.]
 [1. 0. 0. 0. 0. 0.]
 [0. 0. 1. 0. 0. 0.]
 [0. 0. 1. 0. 0. 0.]
 [0. 0. 1. 0. 0. 0.]
 [1. 0. 0. 0. 0. 0.]
 [1. 0. 0. 0. 0. 0.]
 [1. 0. 0. 0. 0. 0.]
 [0. 0. 0. 1. 0. 0.]
 [0. 0. 0. 1. 0. 0.]
 [0. 0. 0. 1. 0. 0.]
 [0. 0. 0. 0. 0. 1.]
 [0. 0. 0. 0. 0. 1.]
 [0. 0. 0. 0. 0. 1.]
 [0. 0. 0. 1. 0. 0.]
 [0. 0. 0. 1. 0. 0.]
 [0. 0. 0. 1. 0. 0.]
 [0. 1. 0. 0. 0. 0.]
 [0. 1. 0. 0. 0. 0.]
 [0. 1. 0. 0. 0. 0.]
 [1. 0. 0. 0. 0. 0.]
 [1. 0. 0. 0. 0. 0.]
 [1. 0. 0. 0. 0. 0.]]

4.cat(tensor，dim)函数
可看作split函数的反向
将若干个数组拼合成一个
dim选择要合并的维度