How to write the component without cloning and patching the whole DeepPavlov repo?

[sld]

Hi!

I'd like to write my own component in my private project.

It seems like I should clone the repo then patch it:

1. By writing write my own dataset readers, iterators,... and putting them into DeepPavlov folder.
2. Then I should put necessary imports into https://github.com/deepmipt/DeepPavlov/blob/master/deeppavlov/__init__.py.
3. In the end I should write my config.json and run python -m deeppavlov...

Can I write separate component independently without patching source code of DeepPavlov?

[seliverstov]

The only reason we write the imports in the __init__ file is to ensure that all standard components are registered through the decorator @register.

@register('ner')
class NER(TFModel):

It is not necessary to put your component import in this __init__ file, you only need to ensure that the appropriate decorator is called before the component is used through it code name in pipeline.

[sld]

I am running next command:
python -m deeppavlov.deep interact deeppavlov/configs/custom_config_component.json

How can I register custom component's dataset readers, iterators, model without putting to __init__.py of DeepPavlov repo?

It seems that I should write my own fork of deeppavlov.deep interact module.

[seliverstov]

deeppavlov.deep interact module is pretty simple. You can write two line of code to interact with your model directly in your own python script:

from deeppavlov.core.commands.infer import interact_model
interact_model(pipeline_config_path)

or train your model:

from deeppavlov.core.commands.train import train_model_from_config
train_model_from_config(pipeline_config_path)

[sshivaji]

This should be in the FAQ as its a very common need!