the detection_eval cannot increase any more

[justttry]

I train the model on icdar2015. The loss reaches 1.1 and can not decline any more. The detection_eval is only 8.9 and it is low on the training set. My training dataset and testing dataset are the same, they are all the whole icdar2015 dataset. Is any problem?

Below is the print content.

I0330 13:37:36.891244 10562 solver.cpp:259] Train net output #0: mbox_loss = 0.91194 ( 1 = 0.91194 loss) I0330 13:37:36.891259 10562 sgd_solver.cpp:138] Iteration 63500, lr = 1e-05 I0330 13:44:41.892575 10562 solver.cpp:243] Iteration 63600, loss = 1.0592 I0330 13:44:41.933722 10562 solver.cpp:259] Train net output #0: mbox_loss = 1.26222 ( 1 = 1.26222 loss) I0330 13:44:41.933746 10562 sgd_solver.cpp:138] Iteration 63600, lr = 1e-05 I0330 13:51:47.677556 10562 solver.cpp:243] Iteration 63700, loss = 1.08502 I0330 13:51:47.677860 10562 solver.cpp:259] Train net output #0: mbox_loss = 0.743075 ( 1 = 0.743075 loss) I0330 13:51:47.677875 10562 sgd_solver.cpp:138] Iteration 63700, lr = 1e-05 I0330 13:58:52.916873 10562 solver.cpp:243] Iteration 63800, loss = 1.04121 I0330 13:58:52.917120 10562 solver.cpp:259] Train net output #0: mbox_loss = 0.70995 ( 1 = 0.70995 loss) I0330 13:58:52.917130 10562 sgd_solver.cpp:138] Iteration 63800, lr = 1e-05 I0330 14:06:09.572183 10562 solver.cpp:243] Iteration 63900, loss = 1.0946 I0330 14:06:09.573016 10562 solver.cpp:259] Train net output #0: mbox_loss = 1.46517 ( 1 = 1.46517 loss) I0330 14:06:09.573030 10562 sgd_solver.cpp:138] Iteration 63900, lr = 1e-05 I0330 14:13:15.293570 10562 solver.cpp:593] Snapshotting to binary proto file models/VGGNet_icdar2015/text/text_polygon_pre cise_fix_order_384x384/VGG_text_text_polygon_precise_fix_order_384x384_iter_64000.caffemodel I0330 14:13:16.178711 10562 sgd_solver.cpp:307] Snapshotting solver state to binary proto file models/VGGNet_icdar2015/text /text_polygon_precise_fix_order_384x384/VGG_text_text_polygon_precise_fix_order_384x384_iter_64000.solverstate I0330 14:13:16.435995 10562 solver.cpp:433] Iteration 64000, Testing net (#0) I0330 14:13:16.436141 10562 net.cpp:693] Ignoring source layer mbox_loss I0330 14:13:25.917605 10562 blocking_queue.cpp:50] Data layer prefetch queue empty I0330 14:13:33.016362 10562 solver.cpp:543] Test net output #0: detection_eval = 0.895085 I0330 14:13:34.952975 10562 solver.cpp:243] Iteration 64000, loss = 1.04576 I0330 14:13:34.953037 10562 solver.cpp:259] Train net output #0: mbox_loss = 1.17127 ( 1 = 1.17127 loss) I0330 14:13:34.953048 10562 sgd_solver.cpp:138] Iteration 64000, lr = 1e-05 I0330 14:20:57.065918 10562 solver.cpp:243] Iteration 64100, loss = 1.10268 I0330 14:20:57.066187 10562 solver.cpp:259] Train net output #0: mbox_loss = 0.987974 ( 1 = 0.987974 loss) I0330 14:20:57.066200 10562 sgd_solver.cpp:138] Iteration 64100, lr = 1e-05 I0330 14:28:14.230305 10562 solver.cpp:243] Iteration 64200, loss = 1.08085 I0330 14:28:14.230576 10562 solver.cpp:259] Train net output #0: mbox_loss = 0.83759 ( 1 = 0.83759 loss) I0330 14:28:14.230587 10562 sgd_solver.cpp:138] Iteration 64200, lr = 1e-05 I0330 14:35:39.323906 10562 solver.cpp:243] Iteration 64300, loss = 1.04889 I0330 14:35:39.324132 10562 solver.cpp:259] Train net output #0: mbox_loss = 1.06259 ( 1 = 1.06259 loss) I0330 14:35:39.324146 10562 sgd_solver.cpp:138] Iteration 64300, lr = 1e-05 I0330 14:43:40.534621 10562 solver.cpp:243] Iteration 64400, loss = 1.10161 I0330 14:43:40.534920 10562 solver.cpp:259] Train net output #0: mbox_loss = 1.12481 ( 1 = 1.12481 loss) I0330 14:43:40.534938 10562 sgd_solver.cpp:138] Iteration 64400, lr = 1e-05 I0330 14:51:23.179654 10562 solver.cpp:243] Iteration 64500, loss = 1.04438 I0330 14:51:23.179877 10562 solver.cpp:259] Train net output #0: mbox_loss = 0.88916 ( 1 = 0.88916 loss) I0330 14:51:23.179889 10562 sgd_solver.cpp:138] Iteration 64500, lr = 1e-05 I0330 14:59:18.520834 10562 solver.cpp:243] Iteration 64600, loss = 1.06642 I0330 14:59:18.521076 10562 solver.cpp:259] Train net output #0: mbox_loss = 1.14982 ( 1 = 1.14982 loss) I0330 14:59:18.521085 10562 sgd_solver.cpp:138] Iteration 64600, lr = 1e-05 I0330 15:08:24.289408 10562 solver.cpp:243] Iteration 64700, loss = 1.03248 I0330 15:08:24.289634 10562 solver.cpp:259] Train net output #0: mbox_loss = 1.22763 (* 1 = 1.22763 loss) I0330 15:08:24.289647 10562 sgd_solver.cpp:138] Iteration 64700, lr = 1e-05 I0330 15:17:31.931952 10562 blocking_queue.cpp:50] Data layer prefetch queue empty

[MhLiao]

@justttry The det_eval (MAP) is 0.89, which is of very high performance.

[justttry]

thanks for your reply! @MhLiao