TensorIterator expected type torch.cuda.DoubleTensor but got torch.cuda.FloatTensor

fastai
1

Hi everyone, i’m having the error from the title in a text Multilabel Classification problems. My data is already tokenized, so I use

data_clas = TextClasDataBunch.from_ids(dir_path,train_ids,train_lbls,val_ids,val_lbls, 258, bs = 16)

to create my DataBunch. I then create a RNNLearner with :

learn = RNNLearner.classifier(data_clas, bptt = bptt, max_len = bptt*20  ,qrnn=_p.qrnn, drop_mult=0.3)

I make sure i have the loss i want with :

learn.loss_func = torch.nn.functional.binary_cross_entropy_with_logits

Now i want to find the best learning rate, so I do

learn.lr_find()
learn.recorder.plot()

But i get the following error :

RuntimeError: TensorIterator expected type torch.cuda.DoubleTensor but got torch.cuda.FloatTensor[8, 5]

I have 5 labels so that’s where the 5 is coming from. The weird thing is that i do almost the same exact thing with another dataset, the only difference being using TextClasDataBunch.from_id_files instead of TextClasDataBunch.from_id and it works perfectly there.

Any tips to solve that error would be welcomed :slight_smile:

PS : I’m not on the latest commit of fastai, i’m just before the refactoring of NLP.

2

That’s super weird. I don’t get why a DoubleTensor would be expected anywhere… Can you share the type of your labels passed as trn_lbl and the type you get when using from_id_fnames?

3

train_lbls and val_lbls are both float array obtained from :

np.load(...).astype(float)

This is a matrix with 5 columns. I changed the type, because if you have integers, you get long tensor which are not valid to compute BCEloss with logits.

In my other case when i use from_id_files, the labels are stored in an .npy files and are already converted to float.

4

I managed to fix my issue. I changed

np.load(...).astype(float)

by

np.load(...).astype(np.float32)

And it worked. Maybe calling TextClasDataBunch.from_ids should convert the labels to float 32 automatically to avoid that kind of problem.

5

From 1.0.21 and onward, this is dealt with at the very base of the fastai datasets, so you shouldn’t have that problemn anymore.

1 Like
6

Great!!!

7

def get_ip(img,pts): return ImagePoints(FlowField(img.size , pts), scale = True ,y_first = False)

def format_ip(pts): return torch.from_numpy(np.array(np.split(pts,9)))

points = np.genfromtxt(‘CAT_01/00000100_002.jpg.cat’, delimiter= " ",usecols= cols);points

array([251., 477., 463., 489., 360., 672., 104., 350., 141., 110., 280., 269., 451., 282., 634., 144., 612., 394.])

pts = get_ip(img, format_ip(points));pts

RuntimeError: expected device cpu and dtype Double but got device cpu and dtype Float

Any function that turns double into float??