How can I know how much is classification confidence in tlt’s classification?
There is no confidence showing. But there is similar log for each class when run “tlt-infer classification”.
For example,
2019-09-06 03:01:17.786559: I tensorflow/stream_executor/dso_loader.cc:152] successfully opened CUDA library libcublas.so.10.0 locally
Current predictions: [[0.01346104 0.06986671 0.01300252 0.02348311 0.03231168 0.05912479
0.27588853 0.00847786 0.02060444 0.00560192 0.00739526 0.00758662
0.01668995 0.01754443 0.29776257 0.02832825 0.00525219 0.01012738
0.07145583 0.01603499]]
Class label = 14
Class name = person
I do inference of multiple images in a folder. The results were written into a file for each image in classification. How can I know confidence?
There is no confidence showing. As above mentioned, the predictions are showing for each class.
Any way to find confidence in batch classification?
Why I am asking for confidence is that I have a lot of false positives in classification. One way is to put all those false positives to default folder and retrain again. But quite a lot of false positives and may be not enough high confidence images can be considered as normal.
For detection networks, there are confidence.
For classification network, there is not confidence during tlt-infer.
I will sync with internal team about your request.