FEAT: Multi input, multi model and JSON output

[johnr14]

Hi, thanks for your work !

I changed a few things to make it work faster on my side, offering it back.

Multiple models

It will now allow --model=all and --model=vit,convnext that will help to evaluate model performance.

Multiple file

• It will allow multiple file input. Normal usage is not affected for a single file.
• It also support * globs that will be parsed inside python as well as normal shell expansion of file list.
• Will do all files with a specific model first to keep the model loaded, saving lots of time.
• TODO: Benchmark the speedup. Sorry I will not do it.

JSON export

Allow output to JSON with --json flag. Will not overwrite already existing files, but append to them. If --json is set it will not print results, less clogging when doing multiple models, use --print to force it to print. It logs the model used in the JSON output only.

JSON Summary

It's just a list of best tags as well as highest score. Also list what model have the winning score. Use with --json --summary

Updated README.md for more details, updated requirements.txt.

Possible enhancements

• Forgot to enable multi model selection, like --model=vit,convnext but I won't use it, in worst case, I'll just edit MODEL_REPO_MAP.
• JSON output to other directory ?
• Add more models or checkpoints ?
• Support URL and url.txt as input ?

[johnr14]

I hope summary is not bloating the application. It could be split in a second file in worst case.

[johnr14]

I have an other branch that fetch tag descriptions and save it to JSON's output. Also caches tag locally. Not sure if it fits here. tag-desc branch