Docker available for TransformsReader Class

[guillim]

I worked on making Docker available for TransformsReader Class as FARMReader Class is already implemented.

I had to make a few changes in search.py, in order to add some env variable enabling to choose between the two Classes to instantiate the Reader.
In config.py, i created three Variables:

• READER_TOKENIZER,
• READER_USE_TRANSFORMERS
• and GPU_NUMBER.
  They simply follow the Class requirements, nothing fancy.

I created another conf file called docker-compose-custom.yml reflecting how you can use TransformerReader by changing some env variables.

A couple of small unrelated changes i did :

• Dockerfile : Better to copy the /data folder as well, since the container needs it when using InMemoryDocumentStore to load the data. (espcially useful if you have downloaded your data and you run the ipyhton command to make this happen)
• transformers.py : I am not 100% sure on this one, but I think there is a typo for use_gpu values indications

[tholor]

Hey @guillim ,

Thanks for this great addition! Very helpful.

Two comments:

1. The fix in the doc string in transformers.py is absolutely right. Thx!
2. Why do you suggest a new docker-compose file? I could imagine that it's more user friendly to have a single one here but to add more comments in the "environment" section to clarify the usage of a TransformersReader. What do you think?

[guillim]

Hello,

Answering point 2. here :
Creating another docker-compose file is a way to ease people who want to build from Dockerfile, with custom ElasticSearch Store. But it is absolutely deletable if you think it confuses people. I am going to delete it so that you can merge the rest.

[tholor]

Hey @guillim ,

Sorry for the delay here. I did some testing and added some slight changes to your branch. I hope that's okay for you, if not feel free to rollback and I create a new PR with my own branch.

Suggested Changes:

• default of READER_USE_TRANSFORMERS should be False (Might otherwise break existing deployments using a FARMReader)
• merged your docker-compose-custom into the existing docker-compose (comments for TransformersReader + build + plain elasticsearch)
• made COPY of "data" folder optional in the Dockerfile as this is not present in the normal haystack repo and therefore breaks when building in environments without it

What do you think? If that's all good with you, we'll be ready to merge this.

[guillim]

Yeah that's perfect ! You're actually write, if we can make everything "backward-compatible" that's better. totally agreed