Add pretrained ckpt loading (via local path or url) to support finetuning

[SamitHuang]

…ning

Thank you for your contribution to the MindOCR repo.
Before submitting this PR, please make sure:

• You have read the Contributing Guidelines on pull requests
• Your code builds clean without any errors or warnings
• You are using approved terminology
• You have added unit tests

Motivation

(Write your motivation for proposed changes here.)

Test Plan

(How should this PR be tested? Do you require special setup to run the test or repro the fixed bug?)

Related Issues and PRs

(Is this PR part of a group of changes? Link the other relevant PRs and Issues here. Use https://help.github.com/en/articles/closing-issues-using-keywords for help on GitHub syntax)

[hadipash]

I think pretrained parameter should be versatile, i.e. able to accept bool, str, url, etc. For now, we have many different parameters that don't match each other and look confusing. For example, some of them: pretrained, ckpt_path, checkpoint_path, and now pretrained_ckpt.

I think we should stick to one pretrained parameter, because eventually it doesn't matter whether it's pretrained on ImageNet, SynthText, or local weights, the purpose of it remains the same.

[SamitHuang]

I see. But we cannot map the customized architecture defined in yaml to a pretrained model if we use pretrained and accept bool input.
Indeed, it can be confusing for pretrained (for backbone only), pretrained_ckpt, and ckpt_load_path (for eval only) in one yaml file. I think renaming pretrained_ckpt to
initialize_from will help alleviate it?.

[hadipash]

For the custom architecture scenario, when a user sets pretrained to True we can issue a warning saying that the pretrained model was not found and asking a user to specify a path to the pretrained model instead. Then, continue model initialization without loading pretrained weights.

Moreover, it can be handy if we want to load custom weights to, for example, backbone. We can reuse pretrained parameter to specify a path to the weights instead of setting it to True.