Open junqiangwu opened 4 years ago
Using test data ccpd_base
, the acc is around 0.50
total 250 correct 124 error 0 precision 0.496 six 0.836 avg_time 0.10912780570983886
total 300 correct 147 error 0 precision 0.49 six 0.84 avg_time 0.10871603012084961
total 350 correct 180 error 0 precision 0.5142857142857142 six 0.8514285714285714 avg_time 0.10944753987448556
total 400 correct 205 error 0 precision 0.5125 six 0.8575 avg_time 0.1089352834224701
total 450 correct 232 error 0 precision 0.5155555555555555 six 0.86 avg_time 0.10881534046596951
total 500 correct 255 error 0 precision 0.51 six 0.862 avg_time 0.10848832941055297
total 550 correct 280 error 0 precision 0.509090909090909 six 0.8563636363636363 avg_time 0.10873743317343972
total 600 correct 303 error 0 precision 0.505 six 0.8466666666666667 avg_time 0.10865229447682699
total 650 correct 336 error 0 precision 0.5169230769230769 six 0.8507692307692307 avg_time 0.10865203710702749
total 700 correct 362 error 0 precision 0.5171428571428571 six 0.8542857142857143 avg_time 0.10874608993530273
total 750 correct 391 error 0 precision 0.5213333333333333 six 0.8546666666666667 avg_time 0.1086995906829834
total 800 correct 412 error 0 precision 0.515 six 0.85625 avg_time 0.10871987730264664
fh02.pth that is provided is trained only for 5 epochs, in the paper its given that 300 epochs trained network will give the best results. So I think we need to train for 300 epochs atleast and try to test the accuracy.
fh02.pth that is provided is trained only for 5 epochs, in the paper its given that 300 epochs trained network will give the best results. So I think we need to train for 300 epochs atleast and try to test the accuracy.
i will try
结果根本复现不出来啊
Using the fh02.pth that you provided, and test data is
ccpd_rotate
, but the test_acc is very bad. loger show acc is around 0.187? Is this model not the final version?