[Feature] Implement of RAM with a gradio interface

[Coobiw]

Thanks for your contribution and we appreciate it a lot. The following instructions would make your pull request more healthy and more easily get feedback. If you do not understand some items, don't worry, just make the pull request and seek help from maintainers.

Motivation

After implementing CLIPZeroShot in mmpretrain, zero-shot image classification is well done by it. However, if there is an image with multi-objects, CLIP cannot do well on these multi-classification tasks. Recently, image tagging is a hot topic. Tag2Text and RAM(Recognize Anything Model) can recognize multi-objects in one image well. What's more coincidental is the implement of RAM depends on implement of CLIP. So it's natural for this feature.

Modification

• convert the checkpoint of RAM(especially, SwinTransformer) to mmpretrain style
• implement RAM based on mmpretrain components and register it in MODELS interface
• implement RAM inference in two modes (normal and openset(users can define the category for themselves))
• the openset inference relies on mmpretrain CLIP
• wrapped to a gradio interface for users to experence

BC-breaking (Optional)

Does the modification introduce changes that break the backward compatibility of the downstream repositories?
If so, please describe how it breaks the compatibility and how the downstream projects should modify their code to keep compatibility with this PR.

Use cases (Optional)

1. Convert RAM weights to mmpretrain style(Optional)

python tools/model_converters/ram2mmpretrain.py /xxx/ram_swin_large_14m.pth /xxx/ram_swin_large_14m_mmpretrain.pth

2. Weights Preparation
   You need to prepare RAM weights and CLIP weights. You can get them in ram_swin_large_14m_mmpretrain and clip-vit-b-p16_converted.
   The step of converting CLIP weights is in my previous PR(Implement of Zero-Shot CLIP Classifier #1737).
3. Gradio Installation

pip install gradio==3.44.0

4. Launch Gradio WebUI

cd mmpretrain
python -m mmpretrain.models.multimodal.ram.gradio_demo /xxx/ram_swin_large_14m_mmpretrain.pth /xxx/clip-vit-b-p16_converted.pth

5. Demos

• If you choose normal, you don't need to set the threshold and tag_list, just uploading the image, because these two are not used.

• If you choose openset, the threshold must be set. The tag_list is optional(default category includes something rare and unseen).

Checklist

Before PR:

• Pre-commit or other linting tools are used to fix the potential lint issues.
• Bug fixes are fully covered by unit tests, the case that causes the bug should be added in the unit tests.
• The modification is covered by complete unit tests. If not, please add more unit test to ensure the correctness.
• The documentation has been modified accordingly, like docstring or example tutorials.

After PR:

• If the modification has potential influence on downstream or other related projects, this PR should be tested with those projects, like MMDet or MMSeg.
• CLA has been signed and all committers have signed the CLA in this PR.

[codecov]

Codecov Report

Attention: 1 lines in your changes are missing coverage. Please review.

Files	Coverage Δ
configs/_base_/datasets/imagenet_bs128_mbv3.py	100.00% <ø> (ø)
configs/_base_/datasets/imagenet_bs32.py	100.00% <ø> (ø)
...onfigs/_base_/datasets/imagenet_bs32_pil_resize.py	100.00% <ø> (ø)
configs/_base_/datasets/imagenet_bs64_hivit_224.py	100.00% <100.00%> (ø)
configs/_base_/datasets/imagenet_bs64_swin_224.py	100.00% <ø> (ø)
configs/_base_/datasets/imagenet_bs64_swin_384.py	100.00% <ø> (ø)
configs/_base_/models/hivit/tiny_224.py	100.00% <100.00%> (ø)
...gs/_base_/schedules/imagenet_bs1024_adamw_hivit.py	100.00% <100.00%> (ø)
configs/dinov2/vit-base-p14_dinov2-pre_headless.py	100.00% <100.00%> (ø)
configs/sam/vit-base-p16_sam_headless.py	100.00% <100.00%> (ø)
... and 1 more

... and 189 files with indirect coverage changes

📢 Thoughts on this report? Let us know!.