search

salesforce / pytorch-qrnn

PyTorch implementation of the Quasi-Recurrent Neural Network - up to 16 times faster than NVIDIA's cuDNN LSTM
BSD 3-Clause "New" or "Revised" License
1.26k stars 193 forks source link

Strings are encoded twice by both QRNN and Pynvrtc? #17

Open luungoc2005 opened 6 years ago

luungoc2005 commented 6 years ago

Usually the library runs fine (I often boot up fresh cloud GPU, Ubuntu 16.04 instances on many providers), but, on one occasion - instance by SnarkAI in particular, I saw this error:

'bytes' object has no attribute 'encode' caused by this line: https://github.com/salesforce/pytorch-qrnn/blob/daadb0f39a1811128df7eb03933f286aa5e319ed/torchqrnn/forget_mult.py#L102

which calls upon the constructor in pynvrtc

Upon further inspection, it seems that pynvrtc also performs encode() of its own:

https://github.com/NVIDIA/pynvrtc/blob/fffa9f6f4a7ee1d452346cbdf68b84b5246ccffb/pynvrtc/interface.py#L200

Which calls upon the encode_str function - which encodes the string into utf-8 bytes if using python 3

However I'm also running Python 3.6 on all machines...

Removing the encode() in QRNN seems to have made it work on that machine for me - but I have to wonder 1. why only that machine had the issue and 2. would removing the .encode() in QRNN be alright for all other cases?

Edit: Apparently this is caused by this recent PR merge: https://github.com/NVIDIA/pynvrtc/pull/2 Shouldn't QRNN be updated accordingly?

threefoldo commented 6 years ago

Thanks, removing those two 'encode' works for me. I using Python 3.5.2.