Using a field representing real numbers with the iterator

[pytorch]

Using a field representing real numbers with the iterator

[nelson-liu]

Thanks for the issue.

Torchtext needs to convert the string number to an int or float somewhere down the line and it currently doesn't do this. A quick fix would be to manually add a pipeline to the postprocessing argument that converts everything in the TARGETS field to int. With a slightly modified version of your code:

Edit: just noticed that your example uses doubles. changed my code accordingly

(tab separated file)

$ cat test.txt
1.1   test string
1.2   test string2
1.3   test string3

The following works on my machine in the meantime while we patch this:

In [1]: import torch

In [2]: from torchtext import data

In [3]: TEXT = data.Field(batch_first=True)

In [4]: TARGETS = data.Field(sequential=False, tensor_type=torch.DoubleTensor, batch_first=True, use_vocab=False, postprocessing=data.Pipeline(lambda x: float(x)))

In [5]: fields = [('targets', TARGETS), ('text', TEXT)]

In [6]: dataset = data.TabularDataset(path="test.txt", format="tsv", fields=fields)

In [7]: TEXT.build_vocab(dataset)

In [8]: train_iter = data.Iterator(dataset, batch_size=1, sort_key=lambda x: len(x.text), shuffle=True)

In [9]: batch = next(iter(train_iter))

In [10]: batch.targets
Out[10]: 
Variable containing:
 1.3000
[torch.cuda.DoubleTensor of size 1 (GPU 0)]

Hope that helps.

[greed2411]

for me the above one, didn't work.
if anyone is still wondering,
change postprocessing=data.Pipeline(lambda x: float(x)) to preprocessing= lambda x: float(x)
that made it work for me (pytorch 0.4 and torchtext 0.2.3)

[giovannipcarvalho]

@greed2411 you don't even need the lambda. Field(use_vocab=False, preprocessing=float) is enough.

edit: It seems to work for RawField but not Field. 😕
edit2: ah, forgot to set sequential=False.

[finiteautomata]

data.LabelField(dtype = torch.float, use_vocab=False, preprocessing=float) does the trick as data.LabelField already sets use_sequential=False (and also removes <unk> token)

[ashudeep]

Found use_vocab argument 😞

[ashudeep]

Even after setting use_vocab=False. I get

RuntimeError: already counted a million dimensions in a given sequence. Most likely your items are also sequences and there's no way to infer how many dimension should the tensor have

It is the same error that one gets when you try to do torch.DoubleTensor('1.2'). Is there something I am doing wrong?

[ashudeep]

Thanks for the solution @nelson-liu

[nelson-liu]

could you leave this open for now --- there is a bug behind this that would be nice to track (the fact that we do not actually convert values with use_vocab=False to numbers). Thanks!

[ashudeep]

Sure, I agree.

[jekbradbury]

Yeah, I was originally imagining that values would be provided as Python numerical types -- but that isn't really consistent with the nature of the library as loading mostly text values. Certainly if it sees strings it should convert them!

[ImtiazKhanDS]

If both my fields like target and source are sequences then also we get the same error , any idea on how to resolve this?