Pytorch RuntimeError: "host_softmax" not implemented for 'torch.cuda.LongTensor'

Question

I am using pytorch for training models. But I got an runtime error when it was computing the cross-entropy loss.

Traceback (most recent call last):
  File "deparser.py", line 402, in <module>
    d.train()
  File "deparser.py", line 331, in train
    total, correct, avgloss = self.train_util()
  File "deparser.py", line 362, in train_util
    loss = self.step(X_train, Y_train, correct, total)
  File "deparser.py", line 214, in step
    loss = nn.CrossEntropyLoss()(out.long(), y)
  File "/home/summer2018/TF/lib/python3.5/site-packages/torch/nn/modules/module.py", line 477, in __call__
    result = self.forward(*input, **kwargs)
  File "/home/summer2018/TF/lib/python3.5/site-packages/torch/nn/modules/loss.py", line 862, in forward
    ignore_index=self.ignore_index, reduction=self.reduction)
  File "/home/summer2018/TF/lib/python3.5/site-packages/torch/nn/functional.py", line 1550, in cross_entropy
    return nll_loss(log_softmax(input, 1), target, weight, None, ignore_index, None, reduction)
  File "/home/summer2018/TF/lib/python3.5/site-packages/torch/nn/functional.py", line 975, in log_softmax
    return input.log_softmax(dim)
RuntimeError: "host_softmax" not implemented for 'torch.cuda.LongTensor'

I think this is because the .cuda() function or conversion between torch.Float and torch.Long. But I have tried many ways to change the variable by .cpu()/.cuda() and .long()/.float(), but it still not work. This error message can't be found when searching it on google. Can anyone helps me? Thanks!!!

This is the code cause error:

def step(self, x, y, correct, total):
    self.optimizer.zero_grad()
    out = self.forward(*x)
    loss = nn.CrossEntropyLoss()(out.long(), y)
    loss.backward()
    self.optimizer.step()
    _, predicted = torch.max(out.data, 1)
    total += y.size(0)
    correct += int((predicted == y).sum().data)
    return loss.data

And this function step() is called by:

def train_util(self):
    total = 0
    correct = 0
    avgloss = 0
    for i in range(self.step_num_per_epoch):
        X_train, Y_train = self.trainloader()
        self.optimizer.zero_grad()
        if torch.cuda.is_available():
            self.cuda()
            for i in range(len(X_train)):
                X_train[i] = Variable(torch.from_numpy(X_train[i]))
                X_train[i].requires_grad = False
                X_train[i] = X_train[i].cuda()
            Y_train = torch.from_numpy(Y_train)
            Y_train.requires_grad = False
            Y_train = Y_train.cuda()
        loss = self.step(X_train, Y_train, correct, total)
        avgloss+=float(loss)*Y_train.size(0)
        self.optimizer.step()
        if i%100==99:
            print('STEP %d, Loss: %.4f, Acc: %.4f'%(i+1,loss,correct/total))

    return total, correct, avgloss/self.data_len

The input data X_train, Y_train = self.trainloader() are numpy arrays at begining.

This is a data sample:

>>> X_train, Y_train = d.trainloader()
>>> X_train[0].dtype
dtype('int64')
>>> X_train[1].dtype
dtype('int64')
>>> X_train[2].dtype
dtype('int64')
>>> Y_train.dtype
dtype('float32')
>>> X_train[0]
array([[   0,    6,    0, ...,    0,    0,    0],
       [   0, 1944, 8168, ...,    0,    0,    0],
       [   0,  815,  317, ...,    0,    0,    0],
       ...,
       [   0,    0,    0, ...,    0,    0,    0],
       [   0,   23,    6, ...,    0,    0,    0],
       [   0,    0,  297, ...,    0,    0,    0]])
>>> X_train[1]
array([ 6,  7,  8, 21,  2, 34,  3,  4, 19, 14, 15,  2, 13,  3, 11, 22,  4,
   13, 34, 10, 13,  3, 48, 18, 16, 19, 16, 17, 48,  3,  3, 13])
>>> X_train[2]
array([ 4,  5,  8, 36,  2, 33,  5,  3, 17, 16, 11,  0,  9,  3, 10, 20,  1,
   14, 33, 25, 19,  1, 46, 17, 14, 24, 15, 15, 51,  2,  1, 14])
>>> Y_train
array([[0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
        0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 1.],
       [0., 0., 0., 1., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
        0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.],
       [1., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
       ...,
       [0., 0., 0., 0., 0., 0., 1., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
        0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.]],
      dtype=float32)

Try all possible combinations:

case 1:
loss = nn.CrossEntropyLoss()(out, y)
I get:
RuntimeError: Expected object of type torch.cuda.LongTensor but found type torch.cuda.FloatTensor for argument #2 'target'

case 2:
loss = nn.CrossEntropyLoss()(out.long(), y)
as description above

case 3:
loss = nn.CrossEntropyLoss()(out.float(), y)
I get:
RuntimeError: Expected object of type torch.cuda.LongTensor but found type torch.cuda.FloatTensor for argument #2 'target'

case 4:
loss = nn.CrossEntropyLoss()(out, y.long())
I get:
RuntimeError: multi-target not supported at /pytorch/aten/src/THCUNN/generic/ClassNLLCriterion.cu:15

case 5:
loss = nn.CrossEntropyLoss()(out.long(), y.long())
I get:
RuntimeError: "host_softmax" not implemented for 'torch.cuda.LongTensor'

case 6:
loss = nn.CrossEntropyLoss()(out.float(), y.long())
I get:
RuntimeError: multi-target not supported at /pytorch/aten/src/THCUNN/generic/ClassNLLCriterion.cu:15

case 7:
loss = nn.CrossEntropyLoss()(out, y.float())
I get:
RuntimeError: Expected object of type torch.cuda.LongTensor but found type torch.cuda.FloatTensor for argument #2 'target'

case 8:
loss = nn.CrossEntropyLoss()(out.long(), y.float())
I get:
RuntimeError: "host_softmax" not implemented for 'torch.cuda.LongTensor'

case 9:
loss = nn.CrossEntropyLoss()(out.float(), y.float())
I get:
RuntimeError: Expected object of type torch.cuda.LongTensor but found type torch.cuda.FloatTensor for argument #2 'target'

Answer 1

I have figured out that if preds and loss aren't of type float, it will throw an error.

preds, feat = model.classify(images)
labels = labels.long().to("cuda:0")
loss =  torch.nn.CrossEntropyLoss(preds, labels)

RuntimeError: Expected floating point type for target with class probabilities, got Long

Or

RuntimeError: "host_softmax" not implemented for 'Long'

To solve this, you can explicitly convert it to a float while calculating loss. It worked for me.

loss =  torch.nn.CrossEntropyLoss(preds.float(), labels.float())

Answer 2

I have met this problem.

RuntimeError: "host_softmax" not implemented for 'Long'

My fault is when use nn.CrossEntropyLoss(), the order of loss_criterion(y_hat, y_label) parameters is wrong.

Answer 3

All right, this worked for me whenever I got error on similar lines (I am gonna keep it here for reference):

1. When calling your model on input ensure that your x is float.
2. However keep your target as long.
3. Call loss on model output and y without any typecasting.

from torch import nn
net = nn.Linear(input, out_classes)
loss_criterion = nn.CrossEntropyLoss()

net = net.to(device)
X = X.to(device).float()
y = y.to(device).long()

y_hat = net(X)
l = loss_criterion(y_hat, y)

Answer 4

In my case, it was because I had flipped targets and logits and since logits are obviously not torch.int64 it raised the error.

Answer 5

I know where the problem is.

y should be in torch.int64 dtype without one-hot encoding. And CrossEntropyLoss() will auto encoding it with one-hot (while out is the probability distribution of prediction like one-hot format).

It can run now!