I’m trying to run the predict_rec script using the multilingual recognition model file—can I use the ViTSTR backbone instead of MobileNetV3?

[yongchoooon]

I’m executing the file with the command provided in the guidelines. When I checked tools/infer/pytorchocr_utility.py, I saw that whenever model_type is set to 'rec', the backbone is hard-coded to MobileNetV3. Is there a way to switch the backbone to ViTSTR? I noticed in pytorchocr/modeling/backbones/__init__.py that ViTSTR is included in the support_dict. If it’s possible, could you let me know how to adjust the network_config or other parts of the code to make this change?

For reference, I’m using paddlepaddle-gpu==2.4.2