How can I test the accuracy of the model for each class

[6sz]

I want to know how accurate is the model for each class, or how can I get the Confusion matrix.Looking forward to getting help

[mzr1996]

Hello, we haven't exposed the confusion matrix calculation API by now. But you can call it directly.
For example, get the classification score of every sample at first:

python tools/test.py your_config_file your_checkpoint --out result.pkl --out-items class_scores

And then, calculate the confusion matrix:

>>> import mmcv
>>> from mmcls.datasets import build_dataset
>>> from mmcls.core.evaluation import calculate_confusion_matrix
>>> cfg = mmcv.Config.fromfile("your_config_file")
>>> dataset = build_dataset(cfg.data.test)
>>> pred = mmcv.load("./result.pkl")['class_scores']
>>> matrix = calculate_confusion_matrix(pred, dataset.get_gt_labels())
>>> print(matrix)
tensor([[47.,  0.,  0.,  ...,  0.,  0.,  0.],
        [ 0., 46.,  0.,  ...,  0.,  0.,  0.],
        [ 0.,  0., 38.,  ...,  0.,  0.,  0.],
        ...,
        [ 0.,  0.,  0.,  ..., 36.,  0.,  0.],
        [ 0.,  0.,  0.,  ...,  0., 24.,  0.],
        [ 0.,  0.,  0.,  ...,  0.,  0., 26.]])
>>> import matplotlib.pyplot as plt
>>> plt.imshow(matrix[:20, :20])   # Visualize the first twenty classes.
>>> plt.show()

[SunTaoTao-666]

matrix = calculate_confusion_matrix(result['class_scores'], dataset.get_gt_labels())
result会报错，未定义

[mzr1996]

matrix = calculate_confusion_matrix(result['class_scores'], dataset.get_gt_labels()) result会报错，未定义

fixed

[6sz]

Wow, thank you very much

[mxmdpc]

There has an easy method to get each class ap.
1、You can change the file 'mean_ap.py' -> def mAP

2、return ap in the file of test.py,then print the ap array,like
print(f'\apple: {round(aps[0], 2)}')
print(f'banana : {round(aps[1], 2)}')

[vothaianh1997]

.