The inference outputs imagenet's label, not the label of my custom dataset.

[alphaProgrammingNow]

The training script is:

./distributed_train.sh 1 --data-dir /data/classification/ --train-split train --val-split val --class-map /data/classification/class_to_idx.txt --model efficientnet_b2

My dataset file directory is as follows:

and the content of class-map(i.e. class_to_idx.txt) is:

bird
cat
chicken
dog
monkey

And the training process is normal.

But when I use trained models to inference on two images, using the following script:

python inference.py --data-dir /data/classification/val_now --model efficientnet_b2 --checkpoint /home/pytorch-image-models/output/train/20230509-233141-efficientnet_b2-256/checkpoint-8.pth.tar

the output label is the label of imagenet, such as electric ray and crampfish, not my own dataset (which should be dog, cat and so on):

Model efficientnet_b2 created, param count: 9109994
Predict: [0/1] Time 2.594 (2.594)
--result
{
    "val_1.jpg":{
        "label":"electric ray, crampfish, numbfish, torpedo",
        "prob":0.2533043027
    },
    "val_2.jpg":{
        "label":"electric ray, crampfish, numbfish, torpedo",
        "prob":0.3099145889
    }
}

I don't understand why. Could someone help me?

[TorbenSDJohansen]

Hi @alphaProgrammingNow,

Did you try to also provide the argument --class-map when running inference? I.e., change your command to:

python inference.py --data-dir /data/classification/val_now --model efficientnet_b2 --checkpoint /home/pytorch-image-models/output/train/20230509-233141-efficientnet_b2-256/checkpoint-8.pth.tar --class-map /data/classification/class_to_idx.txt

[alphaProgrammingNow]

If I just add the --class-map /data/classification/class_to_idx.txt, it will report an error:

RuntimeError: Found 0 images in subfolders of /data/classification/val_now. Supported image extensions are .png, .jpg, .jpeg

It's ok because there is no subfolders in val_now:

So I create two subfolders in val_now, as follows:

and add the --class-map /datas/classification/class_to_idx.txt, the script is：

python inference.py --data-dir /data/classification/val_now --model efficientnet_b2 --checkpoint /home/pytorch-image-models/output/train/20230509-233141-efficientnet_b2-256/checkpoint-8.pth.tar --class-map /data/classification/class_to_idx.txt

The result is:

Model efficientnet_b2 created, param count: 9109994
Predict: [0/1] Time 8.049 (8.049)
--result
{
    "val_1.jpg":{
        "label":"electric ray, crampfish, numbfish, torpedo",
        "prob":0.3513013422
    },
    "val_2.jpg":{
        "label":"hammerhead, hammerhead shark",
        "prob":0.4176284969
    },
    "val_3.jpg":{
        "label":"hammerhead, hammerhead shark",
        "prob":0.2973769903
    },
    "val_4.jpg":{
        "label":"electric ray, crampfish, numbfish, torpedo",
        "prob":0.3641816974
    }
}

Also it outputs the label of imagenet, not my custom dataset.

[rwightman]

@alphaProgrammingNow so yeah, this has not been added to the inference script, the labelling (for imagenet) feature is relatively recent, some of the bits are there but nothing has been added to inference.py

A custom dataset info class (https://github.com/huggingface/pytorch-image-models/blob/main/timm/data/dataset_info.py#L35)
would need to be populated from a txt file passed as an arg to the inference script such that CustomDatasetInfo is used instead of ImageNetInfo when arg is set (https://github.com/huggingface/pytorch-image-models/blob/main/inference.py#L247-L261)

[rwightman]

I guess --class-map could be that arg if set, if --class-map set, load it, and populate CustomDatasetInfo instead of ImageNetInfo

[alphaProgrammingNow]

OK, I will try, thank you!