Seresnext50转换不了

[saladjay]

在out = model.forward(inputs)这一步卡住了 layer0.conv1 conv14 was added to layers 2925302473856:conv_blob14 was added to blobs Add blob conv_blob14 : torch.Size([1, 64, 75, 75]) 2925302216384:blob3 getting 2925302473856:conv_blob14 getting layer0.bn1 2925302473856:conv_blob14 getting batch_norm11 was added to layers 2925302475944:batch_norm_blob11 was added to blobs Add blob batch_norm_blob11 : torch.Size([1, 64, 75, 75]) bn_scale11 was added to layers layer0.relu1 2925302475944:batch_norm_blob11 getting relu7 was added to layers 2925302475944:relu_blob7 was added to blobs Add blob relu_blob7 : torch.Size([1, 64, 75, 75]) 2925302475944:relu_blob7 getting layer0.pool max_pool3 was added to layers 2925302479320:max_pool_blob3 was added to blobs Add blob max_pool_blob3 : torch.Size([1, 64, 37, 37]) 2925302475944:relu_blob7 getting layer1.0.conv1 conv15 was added to layers 2925302509640:conv_blob15 was added to blobs Add blob conv_blob15 : torch.Size([1, 128, 37, 37]) 2925302479320:max_pool_blob3 getting 2925302509640:conv_blob15 getting layer1.0.bn1 2925302509640:conv_blob15 getting batch_norm12 was added to layers 2925302197992:batch_norm_blob12 was added to blobs Add blob batch_norm_blob12 : torch.Size([1, 128, 37, 37]) bn_scale12 was added to layers layer1.0.relu 2925302197992:batch_norm_blob12 getting relu8 was added to layers 2925302197992:relu_blob8 was added to blobs Add blob relu_blob8 : torch.Size([1, 128, 37, 37]) 2925302197992:relu_blob8 getting layer1.0.conv2 conv16 was added to layers 2925302511584:conv_blob16 was added to blobs Add blob conv_blob16 : torch.Size([1, 128, 37, 37]) 2925302197992:relu_blob8 getting 2925302511584:conv_blob16 getting layer1.0.bn2 2925302511584:conv_blob16 getting batch_norm13 was added to layers 2925302513456:batch_norm_blob13 was added to blobs Add blob batch_norm_blob13 : torch.Size([1, 128, 37, 37]) bn_scale13 was added to layers layer1.0.relu 2925302513456:batch_norm_blob13 getting relu9 was added to layers 2925302513456:relu_blob9 was added to blobs Add blob relu_blob9 : torch.Size([1, 128, 37, 37]) 2925302513456:relu_blob9 getting layer1.0.conv3 conv17 was added to layers 2925588200688:conv_blob17 was added to blobs Add blob conv_blob17 : torch.Size([1, 256, 37, 37]) 2925302513456:relu_blob9 getting 2925588200688:conv_blob17 getting layer1.0.bn3 2925588200688:conv_blob17 getting batch_norm14 was added to layers 2925588226840:batch_norm_blob14 was added to blobs Add blob batch_norm_blob14 : torch.Size([1, 256, 37, 37]) bn_scale14 was added to layers layer1.0.downsample.0 conv18 was added to layers 2925588226912:conv_blob18 was added to blobs Add blob conv_blob18 : torch.Size([1, 256, 37, 37]) 2925302479320:max_pool_blob3 getting 2925588226912:conv_blob18 getting layer1.0.downsample.1 2925588226912:conv_blob18 getting batch_norm15 was added to layers 2925302510648:batch_norm_blob15 was added to blobs Add blob batch_norm_blob15 : torch.Size([1, 256, 37, 37]) bn_scale15 was added to layers layer1.0.se_module.fc1 conv19 was added to layers 2925302217248:conv_blob19 was added to blobs Add blob conv_blob19 : torch.Size([1, 16, 1, 1])

---

KeyError Traceback (most recent call last)

in () ----> 1 out = model.forward(inputs) ~\PytorchToCaffe-master\model\models.py in forward(self, x) 386 387 def forward(self, x): --> 388 x = self.features(x) 389 x = self.logits(x) 390 return x ~\PytorchToCaffe-master\model\models.py in features(self, x) 371 def features(self, x): 372 x = self.layer0(x) --> 373 x = self.layer1(x) 374 x = self.layer2(x) 375 x = self.layer3(x) C:\ProgramData\Anaconda3\lib\site-packages\torch\nn\modules\module.py in __call__(self, *input, **kwargs) 487 result = self._slow_forward(*input, **kwargs) 488 else: --> 489 result = self.forward(*input, **kwargs) 490 for hook in self._forward_hooks.values(): 491 hook_result = hook(self, input, result) C:\ProgramData\Anaconda3\lib\site-packages\torch\nn\modules\container.py in forward(self, input) 90 def forward(self, input): 91 for module in self._modules.values(): ---> 92 input = module(input) 93 return input 94 C:\ProgramData\Anaconda3\lib\site-packages\torch\nn\modules\module.py in __call__(self, *input, **kwargs) 487 result = self._slow_forward(*input, **kwargs) 488 else: --> 489 result = self.forward(*input, **kwargs) 490 for hook in self._forward_hooks.values(): 491 hook_result = hook(self, input, result) ~\PytorchToCaffe-master\model\models.py in forward(self, x) 68 residual = self.downsample(x) 69 ---> 70 out = self.se_module(out) + residual 71 out = self.relu(out) 72 C:\ProgramData\Anaconda3\lib\site-packages\torch\nn\modules\module.py in __call__(self, *input, **kwargs) 487 result = self._slow_forward(*input, **kwargs) 488 else: --> 489 result = self.forward(*input, **kwargs) 490 for hook in self._forward_hooks.values(): 491 hook_result = hook(self, input, result) ~\PytorchToCaffe-master\model\models.py in forward(self, x) 40 module_input = x 41 x = self.avg_pool(x) ---> 42 x = self.fc1(x) 43 x = self.relu(x) 44 x = self.fc2(x) C:\ProgramData\Anaconda3\lib\site-packages\torch\nn\modules\module.py in __call__(self, *input, **kwargs) 487 result = self._slow_forward(*input, **kwargs) 488 else: --> 489 result = self.forward(*input, **kwargs) 490 for hook in self._forward_hooks.values(): 491 hook_result = hook(self, input, result) C:\ProgramData\Anaconda3\lib\site-packages\torch\nn\modules\conv.py in forward(self, input) 318 def forward(self, input): 319 return F.conv2d(input, self.weight, self.bias, self.stride, --> 320 self.padding, self.dilation, self.groups) 321 322 ~\PytorchToCaffe-master\pytorch_to_caffe.py in __call__(self, *args, **kwargs) 532 print(layer_names[layer]) 533 break --> 534 out=self.obj(self.raw,*args,**kwargs) 535 # if isinstance(out,Variable): 536 # out=[out] ~\PytorchToCaffe-master\pytorch_to_caffe.py in _conv2d(raw, input, weight, bias, stride, padding, dilation, groups) 101 log.add_blobs([x],name='conv_blob') 102 layer=caffe_net.Layer_param(name=name, type='Convolution', --> 103 bottom=[log.blobs(input)], top=[log.blobs(x)]) 104 layer.conv_param(x.size()[1],weight.size()[2:],stride=_pair(stride), 105 pad=_pair(padding),dilation=_pair(dilation),bias_term=bias is not None,groups=groups) ~\PytorchToCaffe-master\pytorch_to_caffe.py in blobs(self, var) 86 var=id(var) 87 if self.debug: ---> 88 print("{}:{} getting".format(var, self._blobs[var])) 89 try: 90 return self._blobs[var] ~\PytorchToCaffe-master\pytorch_to_caffe.py in __getitem__(self, key) 29 self.data[key]=value 30 def __getitem__(self, key): ---> 31 return self.data[key] 32 def __len__(self): 33 return len(self.data) KeyError: 2925302175928

[saladjay]

在转换torch.nn.AdaptiveAvgPool2d()的时候出错了 另外就是不能转换乘法，例如 module_input = x x = self.avg_pool(x) x = self.fc1(x) x = self.relu(x) x = self.fc2(x) x = self.sigmoid(x) return module_input * x 在最后一部会出错，同样的也是在log.blobs中没有要找的key-value

[saladjay]

在这种pytorch自动补全的地方会出错， 例子 a = torch.randn([1,3,5,5]) b1 = torch.ones([1,3,1,1]) b2 = torch.ones([1,3,5,5]) ab2可以正确的通过_mul转换成Eltwise的点乘layer, 而ab1则不行，ab1会提示 PytorchToCaffe-master\pytorch_to_caffe.py in _mul(input, args) 506 top_blobs = log.add_blobs([x], name='mul_blob') 507 layer = caffe_net.Layer_param(name=layer_name, type='Eltwise', --> 508 bottom=[log.blobs(input), log.blobs(args[0])], top=top_blobs) 509 layer.param.eltwise_param.operation = 0 # product is 1 510 log.cnet.add_layer(layer)

~\PytorchToCaffe-master\pytorch_to_caffe.py in blobs(self, var) 89 var=id(var) 90 if self.debug: ---> 91 print("{}:{} getting".format(var, self._blobs[var])) 92 try: 93 return self._blobs[var]

~\PytorchToCaffe-master\pytorch_to_caffe.py in getitem(self, key) 30 def getitem(self, key): 31 #if key in self.data.keys(): ---> 32 return self.data[key] 33 #else: 34 # return None KeyError: 1639445810200

在blob里没有这个key

[xiezheng-cs]

@saladjay 请问，怎么解决a*b1的问题 ？？

[xiezheng-cs]

OK，我知道怎么解决了。 https://github.com/shicai/SENet-Caffe/blob/master/se_resnet_50_v1_deploy.prototxt

[dreamhighchina]

@xiezheng-cs 怎么解决？？？能讲解下不大佬

[ischihmo]

问题：

a = torch.randn([1,3,5,5]) b1 = torch.ones([1,3,1,1]) b2 = torch.ones([1,3,5,5]) ab2可以正确的通过_mul转换成Eltwise的点乘layer, 而ab1则不行

思路： 参考 https://zhuanlan.zhihu.com/p/65459972 可知

对应的python代码是

def forward(self, x):
        b, c, _, _ = x.size()
        y = self.avg_pool(x).view(b, c)
        y = self.fc(y).view(b, c, 1, 1)
        return x * y.expand_as(x)

那么思路就有了

需要先将b1拉到和a一样的size，再做点乘

实现：

对应的操作就是caffe的scale层

[ShawKai666]

问题：

a = torch.randn([1,3,5,5]) b1 = torch.ones([1,3,1,1]) b2 = torch.ones([1,3,5,5]) ab2可以正确的通过_mul转换成Eltwise的点乘layer, 而ab1则不行

思路： 参考 https://zhuanlan.zhihu.com/p/65459972 可知

对应的python代码是

def forward(self, x):
        b, c, _, _ = x.size()
        y = self.avg_pool(x).view(b, c)
        y = self.fc(y).view(b, c, 1, 1)
        return x * y.expand_as(x)

那么思路就有了

需要先将b1拉到和a一样的size，再做点乘

实现：

对应的操作就是caffe的scale层

但是代码中的expand_as操作貌似不支持从(b, c, 1, 1)扩展到x.size

[174614361]

@VIRGoMerz 我自己仿照 _add 写 _expand_as 的转化时，还是出错了! 想请问下 bottom 那个地方怎么设置

[ischihmo]

@VIRGoMerz 我自己仿照 _add 写 _expand_as 的转化时，还是出错了! 想请问下 bottom 那个地方怎么设置

@ShockeenLee @174614361 这些层没有权重，所以我是直接跳过，生成caffemodel和prototxt之后，手动修改了prototxt