Add --onnx to wd14 tagger

[Isotr0py]

Related issue: #854

• Add --onnx for wd14 tagger to run onnx model
• Make tensorflow optional for wd14 tagger

Though the onnx model in SmilingWolf/wd-v1-4-convnext-tagger-v2 can only use batch_size=1, it still has a faster tagging speed than keras.

Tested on i9-13900HX CPU windows environment:

• onnx: about 3.37it/s
• tensorflow: about 1.42s/it

[kohya-ss]

Thank you for this! This is really useful!

[kohya-ss]

I have tried to change the batch size of the model on the fly according to this issue:
onnx/onnx#2182

but no luck 😅 The result for all images was the same.

[Isotr0py]

Yes, changing the batch size of the model on the fly will cause wrong results, because we may break the model graph on the fly.
For other tagger, change the batch size of the model on the fly will fail because they have different calculation graph.

It seems that the dynamic shape onnx model exported from original tensorflow model using onnx/tensorflow-onnx can work well:

python -m tf2onnx.convert --saved-model ./wd-v1-4-convnext-tagger-v2/ --output convnext-tagger-v2.onnx --opset 13 --verbose --inputs input_1:0[-1,448,448,3]