Hi, I’m finetuning on the pretrained efficientnet model. A weird thing happens. Here’s the log. For training it’s fine, but in the testing, it always shows 10%, (10 classes). I have checked the input and output in the testing. input is different, but output is all the same vectors. I have also check training, but input and output are all different, not like testing. I don’t understand why training and testing use same model, but behave differently?
Epoch: [1][0/3] Time 26.747 (26.747) Data 26.546 (26.546) Loss 2.1273 (2.1273) Prec@1 22.656 (22.656) f1 5.775 (5.775)
Test: [0/1] Time 15.901 (15.901) Loss 2.3027 (2.3027) Prec@1 10.000 (10.000) f1 4.235 (4.235)
- Prec@1 10.000
Epoch: [2][0/3] Time 28.770 (28.770) Data 28.631 (28.631) Loss 2.0814 (2.0814) Prec@1 26.562 (26.562) f1 5.616 (5.616)
Test: [0/1] Time 15.997 (15.997) Loss 2.3027 (2.3027) Prec@1 10.000 (10.000) f1 4.235 (4.235) - Prec@1 10.000
Epoch: [3][0/3] Time 28.335 (28.335) Data 28.197 (28.197) Loss 2.0738 (2.0738) Prec@1 25.000 (25.000) f1 5.587 (5.587)
Test: [0/1] Time 19.465 (19.465) Loss 2.3028 (2.3028) Prec@1 10.000 (10.000) f1 6.000 (6.000) - Prec@1 10.000
Epoch: [4][0/3] Time 25.278 (25.278) Data 25.145 (25.145) Loss 2.1045 (2.1045) Prec@1 18.750 (18.750) f1 5.035 (5.035)
Test: [0/1] Time 18.216 (18.216) Loss 2.3029 (2.3029) Prec@1 10.000 (10.000) f1 6.000 (6.000) - Prec@1 10.000
Epoch: [5][0/3] Time 26.717 (26.717) Data 26.583 (26.583) Loss 1.9554 (1.9554) Prec@1 26.562 (26.562) f1 4.563 (4.563)
Test: [0/1] Time 15.682 (15.682) Loss 2.3031 (2.3031) Prec@1 10.000 (10.000) f1 6.000 (6.000) - Prec@1 10.000