xxradon
commented
5 years ago 目前不支持双输入。------------------ 原始邮件 ------------------ 发件人: "lunalulu"notifications@github.com 发送时间: 2019年4月15日(星期一) 下午4:39 收件人: "xxradon/PytorchToCaffe"PytorchToCaffe@noreply.github.com; 抄送: "Subscribed"subscribed@noreply.github.com; 主题: [xxradon/PytorchToCaffe] 可以转换两个输入的网络吗？ (#12)
你好，请问本工程是否可以转换类似双流处理的网络，他的forward函数如下：
def forward(self, x,target=None): N, C, T, V, M = x.size() # N0, C1, T2, V3, M4 motion = x[:,:,1::,:,:]-x[:,:,0:-1,:,:] motion = motion.permute(0,1,4,2,3).contiguous().view(N,C*M,T-1,V) motion = F.upsample(motion, size=(T,V), mode='bilinear',align_corners=False).contiguous().view(N,C,M,T,V).permute(0,1,3,4,2) logits = [] for i in range(self.num_person): # position # N0,C1,T2,V3 point-level out = self.conv1(x[:,:,:,:,i]) out = self.conv2(out) # N0,V1,T2,C3, global level out = out.permute(0,3,2,1).contiguous() out = self.conv3(out) out_p = self.conv4(out) # motion # N0,T1,V2,C3 point-level out = self.conv1m(motion[:,:,:,:,i]) out = self.conv2m(out) # N0,V1,T2,C3, global level out = out.permute(0, 3, 2, 1).contiguous() out = self.conv3m(out) out_m = self.conv4m(out) # concat out = torch.cat((out_p,out_m),dim=1) out = self.conv5(out) out = self.conv6(out) logits.append(out) # max out logits out = torch.max(logits[0], logits[1]) out = out.view(out.size(0), -1) out = self.fc7(out) out = self.fc8(out) t = out print(t) assert not ((t != t).any())# find out nan in tensor assert not (t.abs().sum() == 0) # find out 0 tensor return out 第二个输入是根据第一个输入计算得到的，函数总的motion，我在转换的时候错误提示为： Traceback (most recent call last): File "hcn/convert.py", line 297, in pytorch_to_caffe.trans_net(model,x,name) File "./pytorch_to_caffe.py", line 625, in trans_net out = net.forward(input_var) File "hcn/convert.py", line 159, in forward motion = x[:,:,1::,:,:]-x[:,:,0:-1,:,:] File "./pytorch_to_caffe.py", line 476, in _sub b1 = log.blobs(input) File "./pytorch_to_caffe.py", line 91, in blobs print(self._blobs[var]) File "./pytorch_to_caffe.py", line 32, in getitem return self.data[key] KeyError: 140445193528256
请帮忙看下，是否支持此类模型的转换~非常感谢
— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub, or mute the thread.
lunalulu
mittlin
ycxia
xiyangyang123
你好，请问本工程是否可以转换类似双流处理的网络，他的forward函数如下： ` def forward(self, x,target=None): N, C, T, V, M = x.size() # N0, C1, T2, V3, M4 motion = x[:,:,1::,:,:]-x[:,:,0:-1,:,:] motion = motion.permute(0,1,4,2,3).contiguous().view(N,C*M,T-1,V) motion = F.upsample(motion, size=(T,V), mode='bilinear',align_corners=False).contiguous().view(N,C,M,T,V).permute(0,1,3,4,2)
第二个输入是根据第一个输入计算得到的，函数总的motion，我在转换的时候错误提示为：Traceback (most recent call last): File "hcn/convert.py", line 297, in