Error : "host_softmax" not implemented for 'Long'

I am trying to train a model for image segmentation. I have 3 different class to segment whihc is denoted by [0,1,2] in the ground truth image. These are the output from different steps:

Model Output : torch.Size([5, 3, 120, 160]) #batch,channel,height,width
Argmax Output : torch.Size([5, 120, 160]) #check maximum along the channel axis
Ground Truth : torch.Size([5, 120, 160])

I am using below code snippet to find the loss using CrossEntropy

loss_seg = criterion_seg(x, label)
#x = output from argmax
#label = ground truth

While doing so I am getting below error

RuntimeError                              Traceback (most recent call last)
<ipython-input-40-e1fd37d7abc6> in <module>
      9     print("epoch",epoch)
---> 10     train_loss, model = train_model(model, "seg", seg_train_loader, criterion_blob, criterion_seg, optimizer, avDev)
     11     loss_details[epoch] = train_loss
     12     print("train loss",train_loss)

<ipython-input-29-f9e95c6cbb49> in train_model(model, mode, train_loader, criterion_blob, criterion_seg, optimizer, device)
     36 #             print(label.size())
     37             optimizer.zero_grad()
---> 38             loss_seg = criterion_seg(torch.tensor(x, dtype=torch.long, device=device), label)

~/anaconda3/lib/python3.7/site-packages/torch/nn/modules/ in __call__(self, *input, **kwargs)
    539             result = self._slow_forward(*input, **kwargs)
    540         else:
--> 541             result = self.forward(*input, **kwargs)
    542         for hook in self._forward_hooks.values():
    543             hook_result = hook(self, input, result)

~/anaconda3/lib/python3.7/site-packages/torch/nn/modules/ in forward(self, input, target)
    914     def forward(self, input, target):
    915         return F.cross_entropy(input, target, weight=self.weight,
--> 916                                ignore_index=self.ignore_index, reduction=self.reduction)

~/anaconda3/lib/python3.7/site-packages/torch/nn/ in cross_entropy(input, target, weight, size_average, ignore_index, reduce, reduction)
   2007     if size_average is not None or reduce is not None:
   2008         reduction = _Reduction.legacy_get_string(size_average, reduce)
-> 2009     return nll_loss(log_softmax(input, 1), target, weight, None, ignore_index, None, reduction)

~/anaconda3/lib/python3.7/site-packages/torch/nn/ in log_softmax(input, dim, _stacklevel, dtype)
   1315         dim = _get_softmax_dim('log_softmax', input.dim(), _stacklevel)
   1316     if dtype is None:
-> 1317         ret = input.log_softmax(dim)
   1318     else:
   1319         ret = input.log_softmax(dim, dtype=dtype)

RuntimeError: "host_softmax" not implemented for 'Long'

Kindly help me to get rid of this error.

Hi Abishek!

I assume that criterion_seg is something you wrote yourself
that basically forwards its arguments to F.cross_entropy().

You most likely should pass Model Output in place of x as
the first argument of criterion_seg, and don’t use argmax()
at all.

This is (most likely) telling you that your are passing the Long
result of argmax() to F.cross_entropy() which is expecting
Float as its “predictions” input. (cross_entropy()'s
target – your label – should, however, be a LongTensor
containing integer class labels ranging over [0, 1, 2]).

Without seeing your code for criterion_seg() I don’t really
know, but I would have expected you to get a dimensions
mismatch error before the type error.

Based on what I think you are doing, x should be your model
output, a FloatTensor of shape [5, 3, 120, 160], where
nBatch = 5, and nClass = 3, containing raw-score logits
that run from -inf to inf. label should be your ground truth,
a LongTensor of shape [5, 120, 160], containing integer
class labels in [0, nClass - 1], inclusive.

The point is you don’t want to convert a set of nClass logits
(or probabilities) to a single integer class label (by using
argmax()) and then pass it to cross_entropy(), because
that is not what cross_entropy() is expecting.

Good luck.

K. Frank


Hi Frank,

Thank you again for the solution. Just passing the output from the model into the Cross_Entropy() worked.

Thank you :slight_smile: