Как оценить точность и потерю обученной модели хорошо или нет в кафе?

Question

Я тренировал свои данные, используя Caffe. Ниже приведены параметры в файле solver.prototxt.

test_iter: 100
test_interval: 1000
base_lr: 0.00001
lr_policy: "step"
gamma: 0.1
stepsize: 20000
display: 20
max_iter: 13000
momentum: 0.9
weight_decay: 0.0005
snapshot: 10000
solver_mode: GPU

Точность теста составляет около 0,0185, а потери при тестировании - около 2,85 при первом тесте. На iter 13000 точность испытания составляет 0,92, а потери при испытаниях - 0,52. Как я оцениваю, является ли точность хорошей или это переоценка данных? Могу ли я использовать какой-либо метод для оценки?
Ниже приведены точность и потеря при каждом тесте.

I1031 11:36:47.063542 18188 solver.cpp:404]     Test net output #0: accuracy = 0.0184
I1031 11:36:47.063623 18188 solver.cpp:404]     Test net output #1: loss = 2.85812 (* 1 = 2.85812 loss)   

I1031 11:38:49.832749 18188 solver.cpp:337] Iteration 1000, Testing net (#0)
I1031 11:38:55.510437 18188 solver.cpp:404]     Test net output #0: accuracy = 0.764
I1031 11:38:55.510509 18188 solver.cpp:404]     Test net output #1: loss = 1.1725 (* 1 = 1.1725 loss)  

I1031 11:40:59.547670 18188 solver.cpp:337] Iteration 2000, Testing net (#0)
I1031 11:41:04.451655 18188 solver.cpp:404]     Test net output #0: accuracy = 0.6372
I1031 11:41:04.451717 18188 solver.cpp:404]     Test net output #1: loss = 1.44847 (* 1 = 1.44847 loss)  

I1031 11:43:05.340741 18188 solver.cpp:337] Iteration 3000, Testing net (#0)
I1031 11:43:09.262504 18188 solver.cpp:404]     Test net output #0: accuracy = 0.8844
I1031 11:43:09.262568 18188 solver.cpp:404]     Test net output #1: loss = 0.576498 (* 1 = 0.576498 loss)

I1031 11:45:10.821233 18188 solver.cpp:337] Iteration 4000, Testing net (#0)
I1031 11:45:14.686005 18188 solver.cpp:404]     Test net output #0: accuracy = 0.5484
I1031 11:45:14.686064 18188 solver.cpp:404]     Test net output #1: loss = 1.91799 (* 1 = 1.91799 loss)

I1031 11:47:17.107151 18188 solver.cpp:337] Iteration 5000, Testing net (#0)
I1031 11:47:21.160307 18188 solver.cpp:404]     Test net output #0: accuracy = 0.8908
I1031 11:47:21.160372 18188 solver.cpp:404]     Test net output #1: loss = 0.54212 (* 1 = 0.54212 loss)

I1031 11:49:23.325654 18188 solver.cpp:337] Iteration 6000, Testing net (#0)
I1031 11:49:27.229637 18188 solver.cpp:404]     Test net output #0: accuracy = 0.7384
I1031 11:49:27.229677 18188 solver.cpp:404]     Test net output #1: loss = 1.1355 (* 1 = 1.1355 loss)

I1031 11:51:29.619175 18188 solver.cpp:337] Iteration 7000, Testing net (#0)
I1031 11:51:33.568794 18188 solver.cpp:404]     Test net output #0: accuracy = 0.6264
I1031 11:51:33.568837 18188 solver.cpp:404]     Test net output #1: loss = 1.10095 (* 1 = 1.10095 loss)

I1031 11:53:36.075034 18188 solver.cpp:337] Iteration 8000, Testing net (#0)
I1031 11:53:39.969431 18188 solver.cpp:404]     Test net output #0: accuracy = 0.9156
I1031 11:53:39.969481 18188 solver.cpp:404]     Test net output #1: loss = 0.522906 (* 1 = 0.522906 loss)

I1031 11:55:42.594107 18188 solver.cpp:337] Iteration 9000, Testing net (#0)
I1031 11:55:46.473902 18188 solver.cpp:404]     Test net output #0: accuracy = 0.5228
I1031 11:55:46.473961 18188 solver.cpp:404]     Test net output #1: loss = 1.63102 (* 1 = 1.63102 loss)

I1031 11:57:55.669351 18188 solver.cpp:337] Iteration 10000, Testing net (#0)
I1031 11:57:59.571413 18188 solver.cpp:404]     Test net output #0: accuracy = 0.8472
I1031 11:57:59.571485 18188 solver.cpp:404]     Test net output #1: loss = 0.638568 (* 1 = 0.638568 loss

I1031 12:00:01.984112 18188 solver.cpp:337] Iteration 11000, Testing net (#0)
I1031 12:00:05.870985 18188 solver.cpp:404]     Test net output #0: accuracy = 0.7476
I1031 12:00:05.871040 18188 solver.cpp:404]     Test net output #1: loss = 1.19568 (* 1 = 1.19568 loss)

I1031 12:02:08.464495 18188 solver.cpp:337] Iteration 12000, Testing net (#0)
I1031 12:02:12.397044 18188 solver.cpp:404]     Test net output #0: accuracy = 0.596
I1031 12:02:12.397104 18188 solver.cpp:404]     Test net output #1: loss = 1.32557 (* 1 = 1.32557 loss)

I1031 12:04:24.876971 18188 solver.cpp:337] Iteration 13000, Testing net (#0)
I1031 12:04:28.693732 18188 solver.cpp:404]     Test net output #0: accuracy = 0.9284
I1031 12:04:28.693771 18188 solver.cpp:404]     Test net output #1: loss = 0.521483 (* 1 = 0.521483 loss)

Пожалуйста, дайте мне знать, если вам нужна дополнительная информация

Answer 1

Подгонка относится к ситуации, когда повышение точности обучения не приводит к увеличению точности тестирования. Поэтому я предлагаю вам продолжить тренировку до тех пор, пока это не произойдет, или вы будете удовлетворены точностью теста.