TypeError: ZipperIterDataPipe instance doesn't have valid length

[SadGare]

Just copy the code in "LANGUAGE TRANSLATION WITH NN.TRANSFORMER AND TORCHTEXT" and try run it on my own PC, it pops up this error and it seems do not have solutions on the Internet now.

cc @pytorch/team-text-core @Nayef211

[hezhongyu]

I have the same question. It seems you cannot get the length of the variable "train_iter". Tracing it, I found the problem may occur in built-in function(len()) of torch.utils.data.datapipes.iter.combining.ZipperIterDataPipe, but I have no idea how to fix it.

[Krb686]

@Gare-Ng @hezhongyu I noticed the same issue today. If you just remove the " / len(train_dataloader)" and " / len(val_dataloader)" in the train_epoch and evaluate functions respectively, the model will run. I still haven't quite figured out if that's logically correct though. The model ran and spit out "A group of people stand in front of an igloo"

Edit:
From what I've read here, len(DataLoader) is supposed to return the number of batches, i.e. dataset size / batch size. The batch size is available in those functions. I hardcoded the dataset size as 29000, which appears to be the NUM_LINES for the 'train' set shown here.

Thus I have:
return losses / ( 29000 / BATCH_SIZE)

It seems to be working. Someone who actually knows what's really going on here can correct me though.

[lixiaojun2914]

The problem could be solved by:
return losses / len(list(train_dataloader))

[jiluojiluo]

yes, it works,thanks

The problem could be solved by: return losses / len(list(train_dataloader))

[Nayef211]

Closing as the issue seems to be resolved. Let me know if this is still a problem and I will reopen