CLI command for assembling (but not training) a model

[kinghuang]

We've recently upgraded to spaCy 3, thanks to the commercial support! We do a "two-phase" approach to building our language model, where the ner component is trained and versioned in one project, and then assembled with other components in another project.

The final assembly is being done by running spacy train … with a training config that has every trainable pipeline component frozen. The goal of the assembly isn't to train any of the components, but to assemble a bunch of components together into a complete NLP model/package. However, using the train command requires train and dev data to be passed in and causes a training pipeline to run, even though there's no actual training being performed.

It would be useful if there was a command akin to assemble, that does all the package assembly that the train command does (e.g., collecting components and generating a nlp module), without invoking any training.

Your Environment

• Operating System: Debian 10 (python container)
• Python Version Used: 3.8.
• spaCy Version Used: 3.0.5
• Environment Information: N/A

[polm]

Thanks for the suggestion, that's an interesting use case. Can you clarify why you assemble the model in a separate step rather than loading all the components, even if frozen, in the initial training step?

As far as the changes necessary for that, if you look at the implementation of the train command, it looks like you could just skip running the last line and replace it with saving the model.

[polm]

Note that this description of building a pipeline in code may also be a useful reference - sticking trained components into a pipeline and serializing them is actually quite simple.

[kinghuang]

Separation of concerns. The two major parts of the final model are the NER component and a bunch of custom components. We are building and versioning the trained NER component as one asset, which is then combined with the components to produce a final versioned model.

I have looked at the underlying code, and that's exactly what I mean. I'm basically looking for a CLI command that's the same as train_cli(), but just outputs the assembled model instead of calling train().

[kinghuang]

We used to have custom scripts for assembling the model, but the training config file is way more flexible. I just don't want to actually train in some cases.

[svlandeg]

Hi King,

I actually had this written down as a task on my TODO list ;-)

I think some type of "assemble" command would make a lot of sense to ensure you can keep using the config-based format even for steps where you're not actually training. So yes I think this could be a useful feature in many use-cases! (or you can do the "train for 1 epoch trick ;-))

[kinghuang]

Haha, it took me a little while to realize what was happening there. That “train for 1 epoch trick” is pretty handy workaround right now! 😄

[svlandeg]

PR with the proposal for implementation: #7783
This will be available from 3.0.6 onwards!