Interesting usage of SPACY_CONFIG_OVERRIDES

[polkaYK]

Hi!
Today I discovered a pretty interesting behavior of variable interpolation. I don't think this is a bug, but I think it's worth documenting somewhere.

Recently, I have played with variable interpolation in spacy project and read about the SPACY_CONFIG_OVERRIDES variable. It turned out that you can utilize it to override values in project.yml file. I say "turned out", because I'm not sure this was listed in docs, but I may be wrong. Given you have something like this in your project.yml file:

vars:
  value: null

commands:
  - name: test
    script: 
      - echo ${vars.value}

and you run:

SPACY_CONFIG_OVERRIDES="--vars.value HELLO_THERE" python -m spacy project run test

you get:

==================================== test ====================================
Running command: echo HELLO_THERE
HELLO_THERE

I think this behavior can be leveraged in some situations. So when I tried to run a whole pipeline in such manner, I was surprised when train command crashed with the message:

✘ Error parsing config overrides
vars -> value   not a section value that can be overwritten

I dug into the code and realized that when you run a command that contains python -m spacy train subcommand, interpolation is performed two times: the first is when project.yml is filled with provided values, the second is when a training config is filled with provided values during the execution of the train subcommand.
I found a workaround: turns out you can add conflicting sections to config.cfg in the following manner:

...
[initialize]
vectors = ${paths.vectors}
init_tok2vec = ${paths.init_tok2vec}
vocab_data = null
lookups = null
before_init = null
after_init = null

[initialize.components]


[initialize.tokenizer]

[vars]

And if you have nested values, you need to provide all subsections.
project.yml

vars:
  some_values:
    value: null

config.cfg

...
[initialize]
vectors = ${paths.vectors}
init_tok2vec = ${paths.init_tok2vec}
vocab_data = null
lookups = null
before_init = null
after_init = null

[initialize.components]


[initialize.tokenizer]

[vars]

[vars.some_values]

That's how you make spacy train ignore values provided for project.yml

Maybe all the aforementioned info is not how spacy project is supposed to work in the first place, so I'm not creating the issue. On the other hand, you may want to consider changing some things about it.