# Inconsistent tensor sizes at the last layer of SegNet

**URL:** <https://discuss.pytorch.org/t/inconsistent-tensor-sizes-at-the-last-layer-of-segnet/16388>\
**Category:** vision\
**Created:** [April 13, 2018, 8:10pm UTC](https://discuss.pytorch.org/t/inconsistent-tensor-sizes-at-the-last-layer-of-segnet/16388 "2018-04-13T20:10:50Z")\
**Posts on this page:** 20\
**Page:** 1

<div class="post-metadata">

**Author:** ![Shounak\_Kundu](https://discuss.pytorch.org/letter_avatar_proxy/v4/letter/s/a88e4f/32.png) [@Shounak\_Kundu](https://discuss.pytorch.org/u/Shounak_Kundu)\
**Post date:** [April 13, 2018, 8:10pm UTC](https://discuss.pytorch.org/t/inconsistent-tensor-sizes-at-the-last-layer-of-segnet/16388/1 "2018-04-13T20:10:50Z")

</div>

Hi , I am using SegNet for image segmentation.

But , I am getting error :

```auto
Iter:1
x torch.Size([1, 3, 256, 256])
conv1 torch.Size([1, 64, 256, 256])
layer1 torch.Size([1, 256, 256, 256])
layer2 torch.Size([1, 512, 256, 256])
layer3 torch.Size([1, 1024, 256, 256])
layer4 torch.Size([1, 2048, 256, 256])
/opt/anaconda/lib/python3.6/site-packages/torch/nn/modules/upsampling.py:180: UserWarning: nn.UpsamplingBilinear2d is deprecated. Use nn.Upsample instead.
  warnings.warn("nn.UpsamplingBilinear2d is deprecated. Use nn.Upsample instead.")
---------------------------------------------------------------------------
RuntimeError Traceback (most recent call last)
<ipython-input-75-a24d68a5b61a> in <module>()
      8 targets = Variable(labels)
      9 
---> 10 outputs = model(inputs)
     11 optimizer.zero_grad()
     12 print("outputs size ==> ",outputs.size())

/opt/anaconda/lib/python3.6/site-packages/torch/nn/modules/module.py in __call__ (self, *input, **kwargs)
    323 for hook in self._forward_pre_hooks.values():
    324 hook(self, input)
--> 325 result = self.forward(*input, **kwargs)
    326 for hook in self._forward_hooks.values():
    327 hook_result = hook(self, input, result)

<ipython-input-7-8f80fe0445f5> in forward(self, x)
     60 self.layer5c(x),
     61 self.layer5d(x),
---> 62 ], 1))
     63 
     64 print('final', x.size())

RuntimeError: inconsistent tensor sizes at /opt/conda/conda-bld/pytorch_1513368888240/work/torch/lib/TH/generic/THTensorMath.c:2864

```

This is my network code :

```auto
class PSPNet(nn.Module):

    def __init__ (self, num_classes):
        #super(PSPNet,self). __init__ ()
        super(). __init__ ()

        resnet = models.resnet101(pretrained=True)
        
        self.conv1 = resnet.conv1
        self.layer1 = resnet.layer1
        self.layer2 = resnet.layer2
        self.layer3 = resnet.layer3
        self.layer4 = resnet.layer4
        
        
        for m in self.modules():
            if isinstance(m, nn.Conv2d):
                m.stride = 1
                m.requires_grad = False
            if isinstance(m, nn.BatchNorm2d):
                m.requires_grad = False
        
        
        self.layer5a = PSPDec(2048, 512, 60)
        self.layer5b = PSPDec(2048, 512, 30)
        self.layer5c = PSPDec(2048, 512, 20)
        self.layer5d = PSPDec(2048, 512, 10)
        
        self.final = nn.Sequential(
            nn.Conv2d(2048, 512, 3, padding=1, bias=False),
            nn.BatchNorm2d(512, momentum=.95),
            nn.ReLU(inplace=True),
            nn.Dropout(.1),
            nn.Conv2d(512, num_classes, 1),
        )

    def forward(self, x):
        
        print('x', x.size())
        
        x = self.conv1(x)
        print('conv1', x.size())
        
        x = self.layer1(x)
        print('layer1', x.size())
        
        x = self.layer2(x)
        print('layer2', x.size())
        
        x = self.layer3(x)
        print('layer3', x.size())
        
        x = self.layer4(x)
        print('layer4', x.size())
        
        x = self.final(torch.cat([
            x,
            self.layer5a(x),
            self.layer5b(x),
            self.layer5c(x),
            self.layer5d(x),
        ], 1))
        
        print('final', x.size())

        return F.upsample_bilinear(final, x.size()[2:])

```

And here I am running it :

```auto

for epoch in range(1, num_epochs+1):
    epoch_loss = []
    iteration=1
    for step, (images, labels) in enumerate(trainLoader):
        print("Iter:"+str(iteration))
        iteration=iteration+1
        inputs = Variable(images)
        targets = Variable(labels)
        
        outputs = model(inputs)
        optimizer.zero_grad()
        print("outputs size ==> ",outputs.size())
        print("targets[:, 0] size ==> ",targets[:, 0].size())

 
        loss = criterion(outputs, targets[:, 0])
        loss.backward()
        optimizer.step()
        epoch_loss.append(loss.data[0])
        
        average = sum(epoch_loss) / len(epoch_loss)
        
        print("loss: "+str(average)+" epoch: "+str(epoch)+", step: "+str(step))

```

The error is coming from the argument 1 of

```auto
x = self.final(torch.cat([
            x,
            self.layer5a(x),
            self.layer5b(x),
            self.layer5c(x),
            self.layer5d(x),
        ], **1** ))

```

Can anyone please help me to solve this problem ?  
Thanks in advance.

---

<div class="post-metadata">

**Author:** ![tom](https://discuss.pytorch.org/user_avatar/discuss.pytorch.org/tom/32/3162_2.png) [@tom](https://discuss.pytorch.org/u/tom)\
**Post date:** [April 13, 2018, 8:36pm UTC](https://discuss.pytorch.org/t/inconsistent-tensor-sizes-at-the-last-layer-of-segnet/16388/2 "2018-04-13T20:36:57Z")

</div>

Probably padding/shape in the PSPDec?  
Note that you don’t print the sizes of the layers being `cat`ted.

Best regards

Thomas

---

<div class="post-metadata">

**Author:** ![Shounak\_Kundu](https://discuss.pytorch.org/letter_avatar_proxy/v4/letter/s/a88e4f/32.png) [@Shounak\_Kundu](https://discuss.pytorch.org/u/Shounak_Kundu)\
**Post date:** [April 14, 2018, 8:17am UTC](https://discuss.pytorch.org/t/inconsistent-tensor-sizes-at-the-last-layer-of-segnet/16388/3 "2018-04-14T08:17:43Z")

</div>

Hi @tom,

thanks a lot for your reply.

Here is the code for padding/shape in PSPDec :

```auto
class PSPDec(nn.Module):

    def __init__ (self, in_features, out_features, downsize, upsize=18):
        super(PSPDec,self). __init__ ()

        self.features = nn.Sequential(
            nn.AvgPool2d(downsize, stride=downsize),
            nn.Conv2d(in_features, out_features, 1, bias=False),
            nn.BatchNorm2d(out_features,momentum=.95),
            nn.ReLU(inplace=True),
            nn.UpsamplingBilinear2d(upsize)
        )

    def forward(self, x):
        return self.features(x)

```

Also , I have printed the layers being `cat`ed , here they are :

```auto
 print('x', x.size())
        
        x = self.conv1(x)
        print('conv1', x.size())
        
        x = self.layer1(x)
        print('layer1', x.size())
        
        x = self.layer2(x)
        print('layer2', x.size())
        
        x = self.layer3(x)
        print('layer3', x.size())
        
        x = self.layer4(x)
        print('layer4', x.size())
        
        t5a = self.layer5a(x)
        print('layer5a',t5a.size())
        
        
        t5b = self.layer5b(x)
        print('layer5b',t5b.size())
        
        t5c = self.layer5c(x)
        print('layer5c',t5c.size())
        
        t5d = self.layer5d(x)
        print('layer5d',t5d.size())
        
        
        x = self.final(torch.cat([
            x,
            t5a,
            t5b,
            t5c,
            t5d,
        ], 1))
        
        print('final', x.size())

        return F.upsample_bilinear(final, x.size()[2:])

```

Here is the result :

```auto

x torch.Size([1, 3, 256, 256])
conv1 torch.Size([1, 64, 256, 256])
layer1 torch.Size([1, 256, 256, 256])
layer2 torch.Size([1, 512, 256, 256])
layer3 torch.Size([1, 1024, 256, 256])
layer4 torch.Size([1, 2048, 256, 256])
/opt/anaconda/lib/python3.6/site-packages/torch/nn/modules/upsampling.py:180: UserWarning: nn.UpsamplingBilinear2d is deprecated. Use nn.Upsample instead.
  warnings.warn("nn.UpsamplingBilinear2d is deprecated. Use nn.Upsample instead.")
layer5a torch.Size([1, 512, 18, 18])
layer5b torch.Size([1, 512, 18, 18])
layer5c torch.Size([1, 512, 18, 18])
layer5d torch.Size([1, 512, 18, 18])
---------------------------------------------------------------------------
RuntimeError Traceback (most recent call last)
<ipython-input-24-7d73f3a481fa> in <module>()
      8 targets = Variable(labels)
      9 
---> 10 outputs = model(inputs)
     11 optimizer.zero_grad()
     12 print("outputs size ==> ",outputs.size())

/opt/anaconda/lib/python3.6/site-packages/torch/nn/modules/module.py in __call__ (self, *input, **kwargs)
    323 for hook in self._forward_pre_hooks.values():
    324 hook(self, input)
--> 325 result = self.forward(*input, **kwargs)
    326 for hook in self._forward_hooks.values():
    327 hook_result = hook(self, input, result)

<ipython-input-7-86307f5840a1> in forward(self, x)
     74 t5c,
     75 t5d,
---> 76 ], 1))
     77 
     78 print('final', x.size())

RuntimeError: inconsistent tensor sizes at /opt/conda/conda-bld/pytorch_1513368888240/work/torch/lib/TH/generic/THTensorMath.c:2864

```

Can you please help me ?

Thanks a lot again.

---

<div class="post-metadata">

**Author:** ![tom](https://discuss.pytorch.org/user_avatar/discuss.pytorch.org/tom/32/3162_2.png) [@tom](https://discuss.pytorch.org/u/tom)\
**Post date:** [April 14, 2018, 8:25am UTC](https://discuss.pytorch.org/t/inconsistent-tensor-sizes-at-the-last-layer-of-segnet/16388/4 "2018-04-14T08:25:39Z")

</div>

The x and the t5x don’t have compatible dimensions 256 vs 18 in w and h. Which dimensions did you expect them to have? You could pool x to 18 or use upsize 256 (though that sounds a lot to me compared to 18).

Best regards

Thomas

---

<div class="post-metadata">

**Author:** ![Shounak\_Kundu](https://discuss.pytorch.org/letter_avatar_proxy/v4/letter/s/a88e4f/32.png) [@Shounak\_Kundu](https://discuss.pytorch.org/u/Shounak_Kundu)\
**Post date:** [April 19, 2018, 5:29am UTC](https://discuss.pytorch.org/t/inconsistent-tensor-sizes-at-the-last-layer-of-segnet/16388/5 "2018-04-19T05:29:28Z")

</div>

Hi @tom, sorry for late reply. Thanks a lot for your help.

I downsized x from 256 to 18. But, still having another issue :

This is the error :

```auto
Iter:1
x torch.Size([1, 3, 256, 256])
conv1 torch.Size([1, 64, 256, 256])
layer1 torch.Size([1, 256, 256, 256])
layer2 torch.Size([1, 512, 256, 256])
layer3 torch.Size([1, 1024, 256, 256])
layer4 torch.Size([1, 2048, 256, 256])
After downsample x ==> torch.Size([1, 2048, 18, 18])
/opt/anaconda/lib/python3.6/site-packages/torch/nn/modules/upsampling.py:180: UserWarning: nn.UpsamplingBilinear2d is deprecated. Use nn.Upsample instead.
  warnings.warn("nn.UpsamplingBilinear2d is deprecated. Use nn.Upsample instead.")
layer5a torch.Size([1, 512, 18, 18])
layer5b torch.Size([1, 512, 18, 18])
layer5c torch.Size([1, 512, 18, 18])
layer5d torch.Size([1, 512, 18, 18])
---------------------------------------------------------------------------
RuntimeError Traceback (most recent call last)
<ipython-input-32-7d73f3a481fa> in <module>()
      8 targets = Variable(labels)
      9 
---> 10 outputs = model(inputs)
     11 optimizer.zero_grad()
     12 print("outputs size ==> ",outputs.size())

/opt/anaconda/lib/python3.6/site-packages/torch/nn/modules/module.py in __call__ (self, *input, **kwargs)
    323 for hook in self._forward_pre_hooks.values():
    324 hook(self, input)
--> 325 result = self.forward(*input, **kwargs)
    326 for hook in self._forward_hooks.values():
    327 hook_result = hook(self, input, result)

<ipython-input-12-f3975da59387> in forward(self, x)
     85 t5c,
     86 t5d,
---> 87 ], 1))
     88 
     89 print('final', x.size())

/opt/anaconda/lib/python3.6/site-packages/torch/nn/modules/module.py in __call__ (self, *input, **kwargs)
    323 for hook in self._forward_pre_hooks.values():
    324 hook(self, input)
--> 325 result = self.forward(*input, **kwargs)
    326 for hook in self._forward_hooks.values():
    327 hook_result = hook(self, input, result)

/opt/anaconda/lib/python3.6/site-packages/torch/nn/modules/container.py in forward(self, input)
     65 def forward(self, input):
     66 for module in self._modules.values():
---> 67 input = module(input)
     68 return input
     69 

/opt/anaconda/lib/python3.6/site-packages/torch/nn/modules/module.py in __call__ (self, *input, **kwargs)
    323 for hook in self._forward_pre_hooks.values():
    324 hook(self, input)
--> 325 result = self.forward(*input, **kwargs)
    326 for hook in self._forward_hooks.values():
    327 hook_result = hook(self, input, result)

/opt/anaconda/lib/python3.6/site-packages/torch/nn/modules/conv.py in forward(self, input)
    275 def forward(self, input):
    276 return F.conv2d(input, self.weight, self.bias, self.stride,
--> 277 self.padding, self.dilation, self.groups)
    278 
    279 

/opt/anaconda/lib/python3.6/site-packages/torch/nn/functional.py in conv2d(input, weight, bias, stride, padding, dilation, groups)
     88 _pair(0), groups, torch.backends.cudnn.benchmark,
     89 torch.backends.cudnn.deterministic, torch.backends.cudnn.enabled)
---> 90 return f(input, weight, bias)
     91 
     92 

 

RuntimeError: Given groups=1, weight[512, 2048, 3, 3], so expected input[1, 4096, 18, 18] to have 2048 channels, but got 4096 channels instead

​

​

```

Here is my code for final layer :

```auto
self.final = nn.Sequential(
            nn.Conv2d(2048, 512, 3, padding=1, bias=False),
            nn.BatchNorm2d(512, momentum=.95),
            nn.ReLU(inplace=True),
            nn.Dropout(.1),
            nn.Conv2d(512, num_classes, 1),
        )

```

Could you please help ?

Thanks in advance .

---

<div class="post-metadata">

**Author:** ![ptrblck](https://discuss.pytorch.org/user_avatar/discuss.pytorch.org/ptrblck/32/1823_2.png) [@ptrblck](https://discuss.pytorch.org/u/ptrblck)\
**Post date:** [April 19, 2018, 7:46am UTC](https://discuss.pytorch.org/t/inconsistent-tensor-sizes-at-the-last-layer-of-segnet/16388/6 "2018-04-19T07:46:00Z")

</div>

This time the channels of the input to `final` and the layer definition do not match.  
`final` takes 2048 `in_channels`, while the `torch.cat` operation concatenates `t5x`, where each one has `512` channels.  
You could try to set the `in_channels` in `final` to 4096.

---

<div class="post-metadata">

**Author:** ![Shounak\_Kundu](https://discuss.pytorch.org/letter_avatar_proxy/v4/letter/s/a88e4f/32.png) [@Shounak\_Kundu](https://discuss.pytorch.org/u/Shounak_Kundu)\
**Post date:** [April 19, 2018, 8:04am UTC](https://discuss.pytorch.org/t/inconsistent-tensor-sizes-at-the-last-layer-of-segnet/16388/7 "2018-04-19T08:04:47Z")

</div>

Hi @ptrblck,  
First of all , thanks a lot for your help.  
I changed the code as you mentioned and that error is gone. 🙂

However , I am getting error in

`---> 90 return F.upsample_bilinear(self.final, x.size()[2:])`  
in this line

The error is :  
`AttributeError: 'Sequential' object has no attribute 'dim'`

Could you please help ?

Thanks in advance.

---

<div class="post-metadata">

**Author:** ![ptrblck](https://discuss.pytorch.org/user_avatar/discuss.pytorch.org/ptrblck/32/1823_2.png) [@ptrblck](https://discuss.pytorch.org/u/ptrblck)\
**Post date:** [April 19, 2018, 8:13am UTC](https://discuss.pytorch.org/t/inconsistent-tensor-sizes-at-the-last-layer-of-segnet/16388/8 "2018-04-19T08:13:07Z")

</div>

`final` is the name of a `Sequential` Module. I guess you would like to upsample its output?  
If so, just pass `x` to `F.upsample_bilinear`.

---

<div class="post-metadata">

**Author:** ![Shounak\_Kundu](https://discuss.pytorch.org/letter_avatar_proxy/v4/letter/s/a88e4f/32.png) [@Shounak\_Kundu](https://discuss.pytorch.org/u/Shounak_Kundu)\
**Post date:** [April 19, 2018, 9:09am UTC](https://discuss.pytorch.org/t/inconsistent-tensor-sizes-at-the-last-layer-of-segnet/16388/9 "2018-04-19T09:09:28Z")

</div>

Hi @ptrblck,

Thanks a lot for your help.

I am now passing x to `F.upsample_bilinear` , which is throwing me this error :

```auto
TypeError Traceback (most recent call last)
<ipython-input-25-7d73f3a481fa> in <module>()
      8 targets = Variable(labels)
      9 
---> 10 outputs = model(inputs)
     11 optimizer.zero_grad()
     12 print("outputs size ==> ",outputs.size())

/opt/anaconda/lib/python3.6/site-packages/torch/nn/modules/module.py in __call__ (self, *input, **kwargs)
    323 for hook in self._forward_pre_hooks.values():
    324 hook(self, input)
--> 325 result = self.forward(*input, **kwargs)
    326 for hook in self._forward_hooks.values():
    327 hook_result = hook(self, input, result)

<ipython-input-8-a6972b090565> in forward(self, x)
     88 print('final', x.size())
     89 
---> 90 return F.upsample_bilinear(x)
     91 #return F.upsample_bilinear(self.final, x.size()[2:])

/opt/anaconda/lib/python3.6/site-packages/torch/nn/functional.py in upsample_bilinear(input, size, scale_factor)
   1422 # DeprecationWarning is ignored by default
   1423 warnings.warn("nn.functional.upsample_bilinear is deprecated. Use nn.functional.upsample instead.")
-> 1424 return upsample(input, size, scale_factor, mode='bilinear')
   1425 
   1426 

/opt/anaconda/lib/python3.6/site-packages/torch/nn/functional.py in upsample(input, size, scale_factor, mode)
   1373 raise NotImplementedError("Got 4D input, but linear mode needs 3D input")
   1374 elif input.dim() == 4 and mode == 'bilinear':
-> 1375 return _functions.thnn.UpsamplingBilinear2d.apply(input, _pair(size), scale_factor)
   1376 elif input.dim() == 4 and mode == 'trilinear':
   1377 raise NotImplementedError("Got 4D input, but trilinear mode needs 5D input")

/opt/anaconda/lib/python3.6/site-packages/torch/nn/_functions/thnn/upsampling.py in forward(ctx, input, size, scale_factor)
    275 output,
    276 ctx.output_size[0],
--> 277 ctx.output_size[1],
    278 )
    279 return output

TypeError: FloatSpatialUpSamplingBilinear_updateOutput received an invalid combination of arguments - got (int, torch.FloatTensor, torch.FloatTensor, NoneType, NoneType), but expected (int state, torch.FloatTensor input, torch.FloatTensor output, int outputHeight, int outputWidth)

```

---

<div class="post-metadata">

**Author:** ![ptrblck](https://discuss.pytorch.org/user_avatar/discuss.pytorch.org/ptrblck/32/1823_2.png) [@ptrblck](https://discuss.pytorch.org/u/ptrblck)\
**Post date:** [April 19, 2018, 9:13am UTC](https://discuss.pytorch.org/t/inconsistent-tensor-sizes-at-the-last-layer-of-segnet/16388/10 "2018-04-19T09:13:24Z")

</div>

You didn’t specify the `size` argument. The error points to `return F.upsample_bilinear(x)`.  
Try something like `F.upsample_bilinear(x, [20, 20])`.  
I’m not sure, what your first try is supposed to do, since you are passing the same size:

```auto
F.upsample_bilinear(x, x.size()[2:])

```

---

<div class="post-metadata">

**Author:** ![Shounak\_Kundu](https://discuss.pytorch.org/letter_avatar_proxy/v4/letter/s/a88e4f/32.png) [@Shounak\_Kundu](https://discuss.pytorch.org/u/Shounak_Kundu)\
**Post date:** [April 19, 2018, 9:31am UTC](https://discuss.pytorch.org/t/inconsistent-tensor-sizes-at-the-last-layer-of-segnet/16388/11 "2018-04-19T09:31:44Z")

</div>

Hi @ptrblck,

Thanks a lot for your help. I understood my mistake and corrected it  
I just did :

`return F.upsample_bilinear(x,[256,256])`  
But , now I am getting error in evaluation of loss function.

```auto
Iter:1
x torch.Size([1, 3, 256, 256])
conv1 torch.Size([1, 64, 256, 256])
layer1 torch.Size([1, 256, 256, 256])
layer2 torch.Size([1, 512, 256, 256])
layer3 torch.Size([1, 1024, 256, 256])
layer4 torch.Size([1, 2048, 256, 256])
After downsample x ==> torch.Size([1, 2048, 18, 18])
/opt/anaconda/lib/python3.6/site-packages/torch/nn/modules/upsampling.py:180: UserWarning: nn.UpsamplingBilinear2d is deprecated. Use nn.Upsample instead.
  warnings.warn("nn.UpsamplingBilinear2d is deprecated. Use nn.Upsample instead.")
layer5a torch.Size([1, 512, 18, 18])
layer5b torch.Size([1, 512, 18, 18])
layer5c torch.Size([1, 512, 18, 18])
layer5d torch.Size([1, 512, 18, 18])
final torch.Size([1, 2, 18, 18])
/opt/anaconda/lib/python3.6/site-packages/torch/nn/functional.py:1423: UserWarning: nn.functional.upsample_bilinear is deprecated. Use nn.functional.upsample instead.
  warnings.warn("nn.functional.upsample_bilinear is deprecated. Use nn.functional.upsample instead.")
outputs size ==> torch.Size([1, 2, 256, 256])
targets[:, 0] size ==> torch.Size([1, 256, 256])
/opt/anaconda/lib/python3.6/site-packages/ipykernel_launcher.py:13: UserWarning: Implicit dimension choice for log_softmax has been deprecated. Change the call to include dim=X as an argument.
  del sys.path[0]
---------------------------------------------------------------------------
RuntimeError Traceback (most recent call last)
<ipython-input-22-7d73f3a481fa> in <module>()
     14 
     15 
---> 16 loss = criterion(outputs, targets[:, 0])
     17 loss.backward()
     18 optimizer.step()

/opt/anaconda/lib/python3.6/site-packages/torch/nn/modules/module.py in __call__ (self, *input, **kwargs)
    323 for hook in self._forward_pre_hooks.values():
    324 hook(self, input)
--> 325 result = self.forward(*input, **kwargs)
    326 for hook in self._forward_hooks.values():
    327 hook_result = hook(self, input, result)

<ipython-input-1-62b1bfa509f1> in forward(self, outputs, targets)
     11 
     12 def forward(self, outputs, targets):
---> 13 return self.loss(F.log_softmax(outputs), targets)

/opt/anaconda/lib/python3.6/site-packages/torch/nn/modules/module.py in __call__ (self, *input, **kwargs)
    323 for hook in self._forward_pre_hooks.values():
    324 hook(self, input)
--> 325 result = self.forward(*input, **kwargs)
    326 for hook in self._forward_hooks.values():
    327 hook_result = hook(self, input, result)

/opt/anaconda/lib/python3.6/site-packages/torch/nn/modules/loss.py in forward(self, input, target)
    145 _assert_no_grad(target)
    146 return F.nll_loss(input, target, self.weight, self.size_average,
--> 147 self.ignore_index, self.reduce)
    148 
    149 

/opt/anaconda/lib/python3.6/site-packages/torch/nn/functional.py in nll_loss(input, target, weight, size_average, ignore_index, reduce)
   1049 return torch._C._nn.nll_loss(input, target, weight, size_average, ignore_index, reduce)
   1050 elif dim == 4:
-> 1051 return torch._C._nn.nll_loss2d(input, target, weight, size_average, ignore_index, reduce)
   1052 else:
   1053 raise ValueError('Expected 2 or 4 dimensions (got {})'.format(dim))

RuntimeError: Assertion `cur_target >= 0 && cur_target < n_classes' failed. at /opt/conda/conda-bld/pytorch_1513368888240/work/torch/lib/THNN/generic/SpatialClassNLLCriterion.c:111

```

I am taking number of classes as 2. So , class\_values should be 0 and 1

Does it mean output is generating some different value ?

If that is the case , how can I find / fix it ?  
Can I use `Sigmoid` here to squash the output value in the range from 0 to 1 ?

Thanks in advance.

---

<div class="post-metadata">

**Author:** ![ptrblck](https://discuss.pytorch.org/user_avatar/discuss.pytorch.org/ptrblck/32/1823_2.png) [@ptrblck](https://discuss.pytorch.org/u/ptrblck)\
**Post date:** [April 19, 2018, 11:01am UTC](https://discuss.pytorch.org/t/inconsistent-tensor-sizes-at-the-last-layer-of-segnet/16388/12 "2018-04-19T11:01:08Z")

</div>

Could you print the shape of `outputs` and `targets`?  
`outputs` should have the dimensions `[batch_size, n_classes, w, h]`  
while targets this one `[batch_size, w, h]`.  
You are indexing it with `targets[:, 0]`, so maybe this yields to the error.

---

<div class="post-metadata">

**Author:** ![Shounak\_Kundu](https://discuss.pytorch.org/letter_avatar_proxy/v4/letter/s/a88e4f/32.png) [@Shounak\_Kundu](https://discuss.pytorch.org/u/Shounak_Kundu)\
**Post date:** [April 19, 2018, 1:42pm UTC](https://discuss.pytorch.org/t/inconsistent-tensor-sizes-at-the-last-layer-of-segnet/16388/13 "2018-04-19T13:42:46Z")

</div>

I got the root cause of the problem.

my target tensor is containing values like 15 , 40 etc.

x,y=train\_dataset. **getitem** (0)

y contains values like 15,40 etc.

So , can you please help me in converting image tensor to 0,1 ?

Thanks in advance.

---

<div class="post-metadata">

**Author:** ![ptrblck](https://discuss.pytorch.org/user_avatar/discuss.pytorch.org/ptrblck/32/1823_2.png) [@ptrblck](https://discuss.pytorch.org/u/ptrblck)\
**Post date:** [April 19, 2018, 1:45pm UTC](https://discuss.pytorch.org/t/inconsistent-tensor-sizes-at-the-last-layer-of-segnet/16388/14 "2018-04-19T13:45:00Z")

</div>

The more important question is, why does your target contain these values?  
If your goal is a segmentation, each target values should represent a class.  
Could you explain a bit more about the target, i.e. how did you load it etc.?

---

<div class="post-metadata">

**Author:** ![Shounak\_Kundu](https://discuss.pytorch.org/letter_avatar_proxy/v4/letter/s/a88e4f/32.png) [@Shounak\_Kundu](https://discuss.pytorch.org/u/Shounak_Kundu)\
**Post date:** [April 19, 2018, 2:16pm UTC](https://discuss.pytorch.org/t/inconsistent-tensor-sizes-at-the-last-layer-of-segnet/16388/15 "2018-04-19T14:16:49Z")

</div>

Hi @ptrblck ,

thanks a lot for your reply.

Here is my dataloader code :

```auto
import os
import collections
import torch
import torchvision
import numpy as np
import scipy.misc as m
import matplotlib.pyplot as plt
import yaml
from PIL import Image
from torch.utils import data

class ImageLoader(data.Dataset):
    
    
    def make_dataset(self,dir, set):
        images = []
        if set == 'train':
            fname = os.path.join(dir, 'train_test_split.yaml')
        elif set == 'test':
            fname = os.path.join(dir, 'train_test_split.yaml')

        # read the content of the file
        with open(fname,'r') as f:
            doc = yaml.load(f)

        imagesNum = doc[set]
        imageFolderPath = dir + os.sep + 'images' + os.sep
        file_list = []
        
        for x in imagesNum:
            item = imageFolderPath + os.sep + str(x)+'_image.png'

            width, height = Image.open(open(item,'rb')).size
            #print("Width ==> ",width," Height ==> ",height)
            
            file_list.append(item)
        
        self.files[set]=file_list
            
        return
    
    def __init__ (self, root, split="train", img_size=None,input_transform=None, target_transform=None):
        self.root = root
        self.split = split
        self.img_size = [1296, 966]
        
        
        self.n_classes = 2
        self.files = collections.defaultdict(list)
        
        self.input_transform = input_transform
        self.target_transform = target_transform
        
        self.make_dataset(root,split)

    def __len__ (self):
        return len(self.files[self.split])

    def __getitem__ (self, index):
        img_name = self.files[self.split][index]
                
        img_path = img_name
        img_name = os.path.split(img_path)[1]
        #print("Img name ==> ",img_name)
        img_num = img_name.split("_")[0]
        #print("Img num ==> ",img_num)
        
        lbl_path = self.root + os.sep + 'annotations' + os.sep + img_num +"_" + "annotation.png"

        with open(img_path, 'rb') as f:
            image = Image.open(f).convert('RGB')
            
        
        with open(lbl_path, 'rb') as f:
            label = Image.open(f).convert('P')
            
        if self.input_transform is not None:
            image = self.input_transform(image)
        if self.target_transform is not None:
            label = self.target_transform(label) 

        return image, label

```

and I am transforming the images before passing them to model :

```auto

import numpy as np
import torch

from PIL import Image

def colormap(n):
    cmap=np.zeros([n, 3]).astype(np.uint8)

    for i in np.arange(n):
        r, g, b = np.zeros(3)

        for j in np.arange(8):
            r = r + (1<<(7-j))*((i&(1<<(3*j))) >> (3*j))
            g = g + (1<<(7-j))*((i&(1<<(3*j+1))) >> (3*j+1))
            b = b + (1<<(7-j))*((i&(1<<(3*j+2))) >> (3*j+2))

        cmap[i,:] = np.array([r, g, b])

    return cmap

class Relabel:

    def __init__ (self, olabel, nlabel):
        self.olabel = olabel
        self.nlabel = nlabel

    def __call__ (self, tensor):
        assert isinstance(tensor, torch.LongTensor), 'tensor needs to be LongTensor'
        tensor[tensor == self.olabel] = self.nlabel
        return tensor

class ToLabel:

    def __call__ (self, image):
        return torch.from_numpy(np.array(image)).long().unsqueeze(0)

class Colorize:

    def __init__ (self, n=22):
        self.cmap = colormap(256)
        self.cmap[n] = self.cmap[-1]
        self.cmap = torch.from_numpy(self.cmap[:n])

    def __call__ (self, gray_image):
        size = gray_image.size()
        color_image = torch.ByteTensor(3, size[1], size[2]).fill_(0)

        for label in range(1, len(self.cmap)):
            mask = gray_image[0] == label

            color_image[0][mask] = self.cmap[label][0]
            color_image[1][mask] = self.cmap[label][1]
            color_image[2][mask] = self.cmap[label][2]

        return color_image

```

and in the following way I am calling this transformation functions

```auto
color_transform = Colorize()
image_transform = ToPILImage()

input_transform = Compose([
    CenterCrop(256),
    ToTensor(),
    Normalize([.485, .456, .406], [.229, .224, .225]),
])
target_transform = Compose([
    CenterCrop(256),
    ToLabel(),
    Relabel(255, 2),
])

```

I am suspecting the erroneous labelling is occurring due to `target_transform`. But , I am not able to figure out exact place .

Could you please help ?

Thanks a lot in advance

---

<div class="post-metadata">

**Author:** ![ptrblck](https://discuss.pytorch.org/user_avatar/discuss.pytorch.org/ptrblck/32/1823_2.png) [@ptrblck](https://discuss.pytorch.org/u/ptrblck)\
**Post date:** [April 19, 2018, 2:41pm UTC](https://discuss.pytorch.org/t/inconsistent-tensor-sizes-at-the-last-layer-of-segnet/16388/16 "2018-04-19T14:41:48Z")

</div>

The `Relabel` class should set all pixel values of 255 to 2?  
If so, change it to `Relabel(255, 1)`, since you have two classes.

However, this does not solve the issue of having target values of 15 and 40.  
Could you add the following line after loading the label:

```auto
print(np.unique(np.array(label)))

```

---

<div class="post-metadata">

**Author:** ![Shounak\_Kundu](https://discuss.pytorch.org/letter_avatar_proxy/v4/letter/s/a88e4f/32.png) [@Shounak\_Kundu](https://discuss.pytorch.org/u/Shounak_Kundu)\
**Post date:** [April 19, 2018, 2:44pm UTC](https://discuss.pytorch.org/t/inconsistent-tensor-sizes-at-the-last-layer-of-segnet/16388/17 "2018-04-19T14:44:48Z")

</div>

I can not thank you enough @ptrblck

I will make all these changes and let you know.

=========================

**Update**  
Hi @ptrblck ,  
This is how I modified the code

```auto
target_transform = Compose([
    CenterCrop(256),
    #Resize(136),
    ToLabel(),
    Relabel(255, 1),
])

```

and I added a print statement here to print labels :

```auto

for epoch in range(1, num_epochs+1):
    epoch_loss = []
    iteration=1
    for step, (images, labels) in enumerate(trainLoader):
        print("Iter:"+str(iteration))
        print("Labels ==> ",np.unique(np.array(labels)))
        iteration=iteration+1
        inputs = Variable(images)
        targets = Variable(labels)
        
        outputs = model(inputs)
        optimizer.zero_grad()
        print("outputs size ==> ",outputs.size())
        
        print("outputs[:,0] size ==> ",outputs[:,0].size())
        print("targets[:, 0] size ==> ",targets[:, 0].size())

        #loss = criterion(outputs, targets[:, 0])
        
        #print(targets[:,0])
        loss = criterion(outputs,targets[:,0])
        
        loss.backward()
        optimizer.step()
        epoch_loss.append(loss.data[0])
        
        average = sum(epoch_loss) / len(epoch_loss)
        
        print("loss: "+str(average)+" epoch: "+str(epoch)+", step: "+str(step))

```

And this is the output I am getting ,

```auto
Iter:1
Labels ==> [0 40]
x torch.Size([1, 3, 256, 256])
conv1 torch.Size([1, 64, 256, 256])
layer1 torch.Size([1, 256, 256, 256])
layer2 torch.Size([1, 512, 256, 256])
layer3 torch.Size([1, 1024, 256, 256])

```

Could you please help ?

---

<div class="post-metadata">

**Author:** ![ptrblck](https://discuss.pytorch.org/user_avatar/discuss.pytorch.org/ptrblck/32/1823_2.png) [@ptrblck](https://discuss.pytorch.org/u/ptrblck)\
**Post date:** [April 19, 2018, 3:08pm UTC](https://discuss.pytorch.org/t/inconsistent-tensor-sizes-at-the-last-layer-of-segnet/16388/18 "2018-04-19T15:08:49Z")

</div>

Could you post the image source or upload one target image?  
Somehow you have two values in the first target image: `[0, 40]`.  
You could try to call `Relable(40, 1)`, but maybe the next image has a different value, so that’s not the solution.

If it’s possible, I would like to load one image and see, what kind of values are stored there.

Also, the output does not match the code exactly. I’m still wondering about the shape of `targets`.

---

<div class="post-metadata">

**Author:** ![Shounak\_Kundu](https://discuss.pytorch.org/letter_avatar_proxy/v4/letter/s/a88e4f/32.png) [@Shounak\_Kundu](https://discuss.pytorch.org/u/Shounak_Kundu)\
**Post date:** [April 19, 2018, 3:12pm UTC](https://discuss.pytorch.org/t/inconsistent-tensor-sizes-at-the-last-layer-of-segnet/16388/19 "2018-04-19T15:12:57Z")

</div>

![001_image](https://discuss.pytorch.org/uploads/default/original/2X/8/8dd69bfb3d581ad5f15c469a0310152081b7c8c2.jpg)

 ![001_annotation](https://discuss.pytorch.org/uploads/default/original/2X/8/805562b4f7c4ee97fc2ec84e41b131292cf82449.png)

---

<div class="post-metadata">

**Author:** ![ptrblck](https://discuss.pytorch.org/user_avatar/discuss.pytorch.org/ptrblck/32/1823_2.png) [@ptrblck](https://discuss.pytorch.org/u/ptrblck)\
**Post date:** [April 19, 2018, 3:36pm UTC](https://discuss.pytorch.org/t/inconsistent-tensor-sizes-at-the-last-layer-of-segnet/16388/20 "2018-04-19T15:36:21Z")

</div>

Thanks for the image.  
The target image still has 20 different target values.  
Using a histogram and a quick visual inspection of the array, it seems that the only valid values should be:

- 0 for background
- 15 for class0
- 40 for class1

All other values seem to be created by some interpolation technique.  
Did you resampled/interpolated the target image using something like bilinear interpolation?  
If so, you should use the nearest interpolation technique to avoid undefined classes.

However, you could just remove the interp values and go with the images you have now.  
Try the following code to clean your targets:

```auto
im_arr = np.array(label)
# Remove interp values
im_arr[(im_arr!=0) & (im_arr!=15) & (im_arr!=40)] = 0
# Set to classes
im_arr[im_arr==15] = 1
im_arr[im_arr==40] = 2
print(np.unique(im_arr))

```

Using this code you will get small borders, where the interpolation was previously.  
If you have the raw data, you could re-create the dataset with the nearest interpolation, if not, this could be good enough.

Also, change the number of classes to 3, since you have a background class and two plant classes. 😉

[Next page](https://discuss.pytorch.org/t/inconsistent-tensor-sizes-at-the-last-layer-of-segnet/16388.md?page=2)