feat: add CAN model

[zhangjunlongtech]

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

• MindOCR does not currently support handwritten mathematical formula recognition. I hope to solve this problem by contributing a CAN model.

• HMER(Handwritten Mathematical Expression Recognition) mostly uses the encoder-decoder mechanism. However, when identifying long or complex formulas, it cannot guarantee the accuracy of the region of interest of the Attention module. CAN(Counting-Aware Network) utilizes Multi-Scale Counting Module to improve the accuracy of formula recognition, by introducing counting vectors that can provide global information and spatial position codes that can provide position information.

• The CAN model consists of the backbone and head modules. For details about the backbone module, see rec_densenet.py. For details about the head module, see rec_can_head.py.

Test Plan

• The model unit test file test_can_model has been provided. This file sets model parameters and test case parameters.
• The test file tests the BaseModel class in mindocr/models/base_model.py. The shape of the tensor in the test result meets the expectation.

• To verify the output of backbone, the build_backbone method in mindocr/models/backbones/builder.py is tested based on rec_densenet in CAN model. The tested tensor shape meets the requirements of the paper.

• Tested build_head method in mindocr/models/heads/builder.py based on backbone's output.

Related Issues and PRs

Related Issues: Model CAN Model Support

[SamitHuang]

why set pynative mode in model definition file?

[SamitHuang]

it can change the global context when imported