Result on HRSC2016 about s2anet

[ming71]

The multiple versions of my RetinaNet also reached mAP of about 80% on HRSC2016, which are consistent with your results. Thank you again for your experiment and nice work! 😆

[ming71]

S2ANet reaches the mAP higher than 89%, while RetinaNet obtains only 56%.

[csuhan]

HRSC2016 is very sensitive to the hyperparameters, e.g., lr, schedules, warmup_iters.
My experiments on HRSC2016 with Retinanet get poor preformance as well. But I think it will be better by carefully adjusting the hyperparameters.
You can validate the AP with different lr (e.g., 0.05 for 4gpus), schedules (e.g., 24 epochs), less warmup_iters and so on.

[ming71]

HRSC2016 is very sensitive to the hyperparameters, e.g., lr, schedules, warmup_iters.
My experiments on HRSC2016 with Retinanet get poor preformance as well. But I think it will be better by carefully adjusting the hyperparameters.
You can validate the AP with different lr (e.g., 0.05 for 4gpus), schedules (e.g., 24 epochs), less warmup_iters and so on.

But it works well on DOTA... it's amazing that with the same config file, it reached mAP higher than 70%

[Fly-dream12]

Thanks for your code, i can not reproduce the same results on HRSC2016 as reported in your paper, what may be wrong?
I have changed the lr to 0.001, and the other settings are the same.
@ csuhan @ ming71

[Fly-dream12]

When I keep the learning rate as 0.01, the loss becomes too large as this:
2020-10-26 10:45:04,367 - INFO - Epoch [7][200/219] lr: 0.01000, eta: 0:24:59, time: 0.249, data_time: 0.002, memory: 2323, loss_fam_cls: 0.8456, loss_fam_bbox: 0.9042, loss_odm_cls: 0.3194, loss_odm_bbox: 58701399372.6484, loss: 58701399374.7175

[ming71]

I don‘t know. I achieved the mAP of 89+% with the original settings for s2anet.

[Fly-dream12]

But the result in the original paper is 90.17. @ming71

[ming71]

But the result in the original paper is 90.17. @ming71

Note that I just run about 20 epochs to achieve the performance, not the whole scheduler.

[Fly-dream12]

I have run 36 epochs, what is the lr in your config 0.01? when i trained the model under 0.01 learning rate, the loss become too large to converge. @ming71

[ming71]

I have run 36 epochs, what is the lr in your config 0.01? when i trained the model under 0.01 learning rate, the loss become too large to converge. @ming71

I trained via dist_train.sh with 4 2080 Ti, everything is OK, maybe it's your problem.

[Fly-dream12]

I trained with single gpu @ming71

[csuhan]

Refer to https://github.com/csuhan/s2anet/blob/master/docs/GETTING_STARTED.md#train-a-model
4GPU * 2img/GPU = 0.01lr
1GPU * 2img/GPU = 0.0025lr

[Sp2-Hybrid]

When I keep the learning rate as 0.01, the loss becomes too large as this:
2020-10-26 10:45:04,367 - INFO - Epoch [7][200/219] lr: 0.01000, eta: 0:24:59, time: 0.249, data_time: 0.002, memory: 2323, loss_fam_cls: 0.8456, loss_fam_bbox: 0.9042, loss_odm_cls: 0.3194, loss_odm_bbox: 58701399372.6484, loss: 58701399374.7175

Hello，how did you solve this problem? I met the same problem like you, when lr was equal to 0.01, loss was too large to train, I reduced lr even to 0.00001, but eventually it became large, even Nan.

[csuhan]

Hi, @ming71 , I know the reason why the mAP of RetinaNet in my codebase is so low. First, 3x is too short for RetinaNet. Then horizontal anchors make the angle prediction hard to converge.

Therefore, I changes some settings:
(1) longer training schedule: 6x RetinaNet reaches 73% mAP.
(2) more anchor angles: when set anchor_angles=[0., PI/3, PI/6, PI/2], the mAP is about 81%.