Using fine-tuned models on HMDB but got unexpected classes output

[hiankun]

Are models such as resnext-101-kinetics-hmdb51_split1.pth, resnext-101-64f-kinetics-hmdb51_split1.pth, and resnet-101-kinetics-hmdb51_split1.pth fine-tuned on HMDB data sets? If it's the case, should they output action classes within the 51 classes of HMDB?

I downloaded the three pre-triained models mentioned above and ran commands as the follows (using the first model as the example):

python main.py --input input --video_root ../test_videos/ --output output.json --model_name resnext --model ../pretrained_models/resnext-101-kinetics-hmdb51_split1.pth --model_depth 101 --resnet_shortcut B --batch_size 8

but the output (generated video) always show actions named belly dancing, bending metal, braiding hair,... and so on. All the actions are not from the 51 classes of HMDB.

Did I do something wrong when I was using the pre-trained models? Or my understanding on the fine-tuned models is not correct?

[hiankun]

It turns out that I misused the class label file (class_names_list)...