6 ms·

SpaCy v2.2

130 points by syllogism 7y ago

wyldfire 7y ago
Keep up the great work, Matthew and Ines (et al)! This doesn't mention the transformer models [1] that were recently announced, so even though it's not 2.2 related, I'll just point it out because it's recent spaCy news :) I had thought I'd seen somewhere (spaCy IRL?) that this kinda stuff was not supposed to show up until v3. [1] https://explosion.ai/blog/spacy-pytorch-transformers https://explosion.ai/blog/spacy-pytorch-transformers
- najarvg 7y ago
  SpaCy PyTorch transformers are pip installed and require Python 3.6+ and SpaCy 2.1.7+ to work - https://github.com/explosion/spacy-pytorch-transformers https://github.com/explosion/spacy-pytorch-transformers So probably not directly related to core SpaCy itself.
- syllogism 7y ago
  We're definitely planning to do more there -- we took a bit longer than anticipated getting v2.2 out, due to various packaging and automation hiccups. Huggingface are doing a great job keeping transformers moving quickly, so we already have some catching up to do. We need to update the wrapper so it works with the new version. I also want to check out Tensorflow 2, and build a Thinc wrapper for it.
benj111 7y ago
Is it just me, or does the Upper case C lead to a hard C pronunciation, which leads to an unfortunate insult, in UK English at least. https://www.urbandictionary.com/define.php?term=spacky https://www.urbandictionary.com/define.php?term=spacky Although I understood it to be derived from: https://www.urbandictionary.com/define.php?term=spacker https://www.urbandictionary.com/define.php?term=spacker
- syllogism 7y ago
  The intended pronounciation is /speɪˈsiː/, i.e. spay-SEE. The capitalisation refers to the fact that it's a C extension, implemented in Cython.
ashish01 7y ago
Thanks for your hard work. Really top quality software. I hope and wish they are making enough money to keep working on it.
- syllogism 7y ago
  Definitely a fair concern, and thanks for the kind words :). We haven't had to fund-raise, so our company performance has actually been a bit under-the-radar. But I can assure you everything's been going well, and most people would be surprised at the numbers. We've been profitable since day one, but this year we've been making enough to hire three more full-time developers: https://explosion.ai/about https://explosion.ai/about . Our revenue all comes from our annotation tool Prodigy: https://prodi.gy https://prodi.gy . Prodigy's very popular, we have a few thousand users in total. There aren't many alternative tools that are scriptable like Prodigy is, and annotation is something that has to be done on almost every project. Even if you don't need to annotate training data (e.g. if you're training on click data or something), you'll usually want to do small annotation tasks to do error analysis and understand your performance.
wgerard 7y ago
Awesome news! Really excited to try out the new CLI features for training.
sontek 7y ago
Nice! I'm excited! I've been working on a heavy NLP project with spanish and been having some issues and so this will be nice to test out and see if it helps! I love spaCy and how its improving all the time

130 points by syllogism 7y ago

Keep up the great work, Matthew and Ines (et al)! This doesn't mention the transformer models [1] that were recently announced, so even though it's not 2.2 related, I'll just point it out because it's recent spaCy news :) I had thought I'd seen somewhere (spaCy IRL?) that this kinda stuff was not supposed to show up until v3. [1] https://explosion.ai/blog/spacy-pytorch-transformers https://explosion.ai/blog/spacy-pytorch-transformers

najarvg 7y ago
SpaCy PyTorch transformers are pip installed and require Python 3.6+ and SpaCy 2.1.7+ to work - https://github.com/explosion/spacy-pytorch-transformers https://github.com/explosion/spacy-pytorch-transformers So probably not directly related to core SpaCy itself.
syllogism 7y ago
We're definitely planning to do more there -- we took a bit longer than anticipated getting v2.2 out, due to various packaging and automation hiccups. Huggingface are doing a great job keeping transformers moving quickly, so we already have some catching up to do. We need to update the wrapper so it works with the new version. I also want to check out Tensorflow 2, and build a Thinc wrapper for it.

benj111 7y ago

Is it just me, or does the Upper case C lead to a hard C pronunciation, which leads to an unfortunate insult, in UK English at least. https://www.urbandictionary.com/define.php?term=spacky https://www.urbandictionary.com/define.php?term=spacky Although I understood it to be derived from: https://www.urbandictionary.com/define.php?term=spacker https://www.urbandictionary.com/define.php?term=spacker

syllogism 7y ago
The intended pronounciation is /speɪˈsiː/, i.e. spay-SEE. The capitalisation refers to the fact that it's a C extension, implemented in Cython.

ashish01 7y ago

Thanks for your hard work. Really top quality software. I hope and wish they are making enough money to keep working on it.

syllogism 7y ago
Definitely a fair concern, and thanks for the kind words :). We haven't had to fund-raise, so our company performance has actually been a bit under-the-radar. But I can assure you everything's been going well, and most people would be surprised at the numbers. We've been profitable since day one, but this year we've been making enough to hire three more full-time developers: https://explosion.ai/about https://explosion.ai/about . Our revenue all comes from our annotation tool Prodigy: https://prodi.gy https://prodi.gy . Prodigy's very popular, we have a few thousand users in total. There aren't many alternative tools that are scriptable like Prodigy is, and annotation is something that has to be done on almost every project. Even if you don't need to annotate training data (e.g. if you're training on click data or something), you'll usually want to do small annotation tasks to do error analysis and understand your performance.

wgerard 7y ago

Awesome news! Really excited to try out the new CLI features for training.

sontek 7y ago

Nice! I'm excited! I've been working on a heavy NLP project with spanish and been having some issues and so this will be nice to test out and see if it helps! I love spaCy and how its improving all the time