Spacy: English models returning incorrect results

[mukulseth]

Trying to run a simple example from the website --

import spacy
spacy.require_gpu()
nlp = spacy.load("en_core_web_sm")
doc = nlp("Apple is looking at buying U.K. startup for $1 billion")
for token in doc:
    print(token.text, token.pos_, token.dep_)

Outputs --

Apple VERB dep
is VERB dep
looking VERB dep
at VERB dep
buying VERB dep
U.K. VERB dep
startup VERB dep
for VERB dep
$ VERB dep
1 VERB dep
billion VERB ROOT

I've tried with the md and lg models as well -- here are the results for those resp. --

en_core_web_md --

Apple PROPN dep
is PROPN dep
looking PROPN dep
at PROPN dep
buying PROPN dep
U.K. PROPN dep
startup PROPN dep
for PROPN dep
$ PROPN dep
1 PROPN dep
billion PROPN ROOT

en_core_web_lg --

Apple PROPN dep
is AUX dep
looking ADJ dep
at PROPN dep
buying PROPN dep
U.K. PROPN dep
startup PROPN dep
for PROPN dep
$ PROPN dep
1 PROPN dep
billion PROPN ROOT

Not sure if I missed something. Using CUDA 10.2, and cupy 7.1.1. The result is the same whether I use the GPU or not though.

Info about spaCy

• spaCy version: 2.2.3
• Platform: Windows-10-10.0.18362-SP0
• Python version: 3.7.1

[svlandeg]

That definitely doesn't look right!

This reminds me of an issue in thinc we found when using Windows + GPU, and which was fixed recently: https://github.com/explosion/thinc/pull/149

Are you 100% certain that these regressions happen also without GPU? I haven't been able to reproduce this.

If it's only on GPU, we can consider this issue fixed.

[mukulseth]

Yep. I think that was it. I tried again with the GPU disabled, and the output was as expected. Guess thinc was to blame.

Will try to get that fix.

Thanks!

[svlandeg]

Ok, happy to hear it was the old bug and not a new one! :-)

[svlandeg]

With respect to getting the fix: thinc has been entirely revamped since that fix, which means that you can't just upgrade thinc - spaCy will break. We're currently working on a new version of spaCy to work with the new thinc, but that's not ready yet.

In short, you're probably best of to just adjust the _murmur3.cu file on your system for now, cf. https://github.com/explosion/thinc/pull/149/files

[mukulseth]

Worked! Thank you :)

[lock[bot]]

This thread has been automatically locked since there has not been any recent activity after it was closed. Please open a new issue for related bugs.