我对神经网络世界是陌生的,我正在尝试从this model实现一个CNN生成器以进行运动生成。我编写了以下代码,其中H_txt是一个字典,其中包含我的剪辑的名称和值(表示剪辑中显示的动作的向量)作为键,而z是维度为(1,256)的高斯白噪声。
N=32
class CNNGenerator(nn.Module):
def __init__(self, htxt=H_txt):
super(CNNGenerator, self).__init__()
self.htxt = htxt
self.conv1 = nn.Conv1d(1, 1, 3)
self.conv2 = nn.Conv1d(1, 1, 3)
self.conv3 = nn.Conv1d(1, 1, 3)
self.conv4 = nn.Conv1d(4, 4, 3)
self.conv5 = nn.Conv1d(2, 2, 3)
self.conv6 = nn.Conv1d(8, 8, 3)
self.conv7 = nn.Conv1d(4, 4, 3)
self.conv8 = nn.Conv1d(16, 16, 3)
self.conv9 = nn.Conv1d(8, 8, 3)
self.conv10 = nn.Conv1d(32, 32, 3)
self.conv11 = nn.Conv1d(16, 16, 3)
self.conv12 = nn.Conv1d(32, 32, 3)
self.conv13 = nn.Conv1d(1, 1, 3)
self.conv14 = nn.Conv1d(2, 2, 3)
self.conv15 = nn.Conv1d(2, 2, 3)
self.conv16 = nn.Conv1d(4, 4, 3)
self.conv17 = nn.Conv1d(4, 4, 3)
self.conv18 = nn.Conv1d(8, 8, 3)
self.conv19 = nn.Conv1d(8, 8, 3)
self.conv20 = nn.Conv1d(16, 16, 3)
self.conv21 = nn.Conv1d(16, 16, 3)
self.conv22 = nn.Conv1d(32, 32, 3)
self.conv23 = nn.Conv1d(32, 32, 3)
def forward(self, x):
x[0] = self.conv1(F.relu(self.conv2(z) + self.htxt))
x[1] = self.conv3(F.relu(self.conv4(z) + self.htxt))
x[2] = self.conv5(F.relu(self.conv6(z) + self.htxt))
x[3] = self.conv7(F.relu(self.conv8(z) + self.htxt))
x[4] = self.conv9(F.relu(self.conv10(z) + self.htxt))
x[5] = self.conv11(F.relu(self.conv12(z) + self.htxt))
h = np.zeros(np.log2(N))
h = x[0] #p6 l4
h[1] = nn.AdaptiveAvgPool1d(2*h[0]) + self.conv13(F.relu(nn.AdaptiveAvgPool1d(2*(self.conv14(h[0])+x[1]))))
h[2] = nn.AdaptiveAvgPool1d(2*h[1]) + self.conv15(F.relu(nn.AdaptiveAvgPool1d(2*(self.conv16(h[1])+x[2]))))
h[3] = nn.AdaptiveAvgPool1d(2*h[2]) + self.conv17(F.relu(nn.AdaptiveAvgPool1d(2*(self.conv18(h[2])+x[3]))))
h[4] = nn.AdaptiveAvgPool1d(2*h[3]) + self.conv19(F.relu(nn.AdaptiveAvgPool1d(2*(self.conv20(h[3])+x[4]))))
h[5] = nn.AdaptiveAvgPool1d(2*h[4]) + self.conv21(F.relu(nn.AdaptiveAvgPool1d(2*(self.conv22(h[4])+x[5]))))
A = self.conv23(h[np.log2(N)])
return A
def num_flat_features(self, x):
size = x.size()[1:] # all dimensions except the batch dimension
num_features = 1
for s in size:
num_features *= s
return num_features
net = CNNGenerator()
z= torch.randn(1, 1, 256) #k=256
out = net(z)
print(out)
运行代码时,出现以下错误消息,来自我的forward函数:
TypeError: unsupported operand type(s) for +: 'Tensor' and 'dict'
我还没有在线找到任何解决方案,所以我想知道是否可以将我的张量与字典相加?是否有可以将我的字典转换为张量的函数?
[我试图用np.asarray()将字典转换成数组,但收到一条错误消息,说我不能为此使用numpy函数。
非常感谢您阅读我的信息
您的问题缺少有关变量的详细信息,但是基于错误,我正在为您提供答案。您正在添加字典和张量,这会给您带来错误。如果要将字典的值添加到张量,则您必须将字典转换为张量。为什么还要直接添加字典,因为这里不需要键。如果要在沿所需轴将其转换为张量后将字典中的值连接起来,则需要使用torch.cat函数。要将字典中的值添加到张量中,请从dict中获取值,并通过类似这种方法将其转换为张量
torch.Tensor(list(htxt.values()))