Experience sharing request

[G-Naughty]

Hello, I'm a student who is learning object detection. I downloaded your ROItransformer model and ran it. Migrating the code from mmdetection0.6 to mmdetection2.0 is a perfect work. I want to migrate a model(s2anet which is also published by Wuhan university) from mmdetection1.0 to mmdetection2.0 ,do you know how to migrate it? Would you like to share your experience about migrating model?

[nijkah]

Actually, I already have a plan to include s2anet to this repository.
Anyway, this is my personal advice for your questions.

First of all, you have to specify a goal whether you want to mitigate the model for a long time or just test or reproduce the model.
For former one, you have to struggle with very actively updating mmdetection. I feel exhausted to follow so many changes of mmdetection. So, I recommend you to mitigate it as possible as you can execute.

Apart from that, one of the most important thing that I experienced is that you have to consider mmcv version first.
Since, mmcv is ongoing project, its structures and modularity is still changing actively.
If you find appropriate version of mmcv, it will reduce large parts of mitigation.

[G-Naughty]

Actually, I already have a plan to include s2anet to this repository.
Anyway, this is my personal advice for your questions.

First of all, you have to specify a goal whether you want to mitigate the model for a long time or just test or reproduce the model.
For former one, you have to struggle with very actively updating mmdetection. I feel exhausted to follow so many changes of mmdetection. So, I recommend you to mitigate it as possible as you can execute.

Apart from that, one of the most important thing that I experienced is that you have to consider mmcv version first.
Since, mmcv is ongoing project, its structures and modularity is still changing actively.
If you find appropriate version of mmcv, it will reduce large parts of mitigation.

Thank you for your reply. I'd like to mitigate the model for a long time and want to do research and Study on the basis of this model However as a beginners of Remote sensing target detection，It's hard for me to mitigate model. Would you like to mitigate s2anet in the near future?

[nijkah]

Thank you for your reply. I'd like to mitigate the model for a long time and want to do research and Study on the basis of this model However as a beginners of Remote sensing target detection，It's hard for me to mitigate model. Would you like to mitigate s2anet in the near future?

I'll try it till this year.

[nijkah]

@G-Naughty Hi, I run s2anet recently. And I found that it is not difficult to run it.
Any difficulty you are struggling?

By the way, I will integrate s2anet into this project.

[G-Naughty]

@nijkah Congratulations on successfully upgrading the model, and thank you very much for integrating s2anet into this project.
Actually , as a beginner of object detection, I have many problem of mitigating the model. First, I even can't fully understand the code of s2anet( by reading paper I think the most contribution of the paper is AlignConv, however till now I still can't really know how to write it). Another thing is I don't have the experience of mitigating the model, and have no idea about f mitigating model.(I usually just use model ,but know my computer can't support mmdetection1).

[G-Naughty]

@nijkah I look forward to your mitigating S2ANET

[nijkah]

@G-Naughty So sorry. I am concentrating on the other project which ends in February.
But it does not mean I gave up to mitigate s2anet.
I'll try my best on holidays for this project. Thanks for your interests.

[G-Naughty]

@nijkah thanks for your reply. First of all, I wish you a smooth project research. and I will still keep a watchful eye on your project, looking forward to your mitigating.

[nijkah]

@G-Naughty

https://github.com/csuhan/s2anet

Fortunately, s2anet which was SOTA model on DOTA dataset releases its update for mmdetection v2.
So it will be better to use that repo.
Sorry for late update. :(

[G-Naughty]

It's true that s2anet has update, however it just in mmdetection2's style, Its core is still mdetection1. So I try to update it into mmdetection2 and now it can run . The problem is the model that I updated isn't work. it's misconvergence . Have you encountered this problem before? And how to solve it.
{"mode": "train", "epoch": 1, "iter": 50, "lr": 0.00025, "memory": 5439, "data_time": 0.04951, "loss_cls": 1.61558, "loss_bbox": 3.7731, "loss": 5.38868, "grad_norm": 1.64646, "time": 0.50323}
{"mode": "train", "epoch": 1, "iter": 100, "lr": 0.0005, "memory": 5439, "data_time": 0.00484, "loss_cls": 1.66116, "loss_bbox": 2.77849, "loss": 4.43964, "grad_norm": 12.58261, "time": 0.45653}
{"mode": "train", "epoch": 1, "iter": 150, "lr": 0.00075, "memory": 5439, "data_time": 0.00498, "loss_cls": 1.56718, "loss_bbox": 2.25654, "loss": 3.82372, "grad_norm": 41.11334, "time": 0.46665}
{"mode": "train", "epoch": 1, "iter": 200, "lr": 0.001, "memory": 5524, "data_time": 0.00485, "loss_cls": 1.37763, "loss_bbox": 2.39505, "loss": 3.77268, "grad_norm": 32.25086, "time": 0.46999}
{"mode": "train", "epoch": 1, "iter": 250, "lr": 0.00125, "memory": 5524, "data_time": 0.00475, "loss_cls": 1.24957, "loss_bbox": 1.72512, "loss": 2.97469, "grad_norm": 21.88243, "time": 0.46359}
{"mode": "train", "epoch": 1, "iter": 300, "lr": 0.0015, "memory": 5563, "data_time": 0.00478, "loss_cls": 1.24225, "loss_bbox": 2.53604, "loss": 3.77829, "grad_norm": 19.60273, "time": 0.47814}
{"mode": "train", "epoch": 1, "iter": 350, "lr": 0.00175, "memory": 5563, "data_time": 0.0047, "loss_cls": 1.14714, "loss_bbox": 1.61101, "loss": 2.75814, "grad_norm": 13.90954, "time": 0.47372}
{"mode": "train", "epoch": 1, "iter": 400, "lr": 0.002, "memory": 5563, "data_time": 0.00467, "loss_cls": 1.12006, "loss_bbox": 1.55968, "loss": 2.67974, "grad_norm": 5.11082, "time": 0.47158}
{"mode": "train", "epoch": 1, "iter": 450, "lr": 0.00225, "memory": 5607, "data_time": 0.00466, "loss_cls": 1.12379, "loss_bbox": 1.4593, "loss": 2.58309, "grad_norm": 4.62821, "time": 0.47467}
{"mode": "train", "epoch": 1, "iter": 500, "lr": 0.0025, "memory": 5607, "data_time": 0.00468, "loss_cls": 1.10848, "loss_bbox": 1.60575, "loss": 2.71423, "grad_norm": 5.75681, "time": 0.47896}
{"mode": "train", "epoch": 1, "iter": 550, "lr": 0.0025, "memory": 5607, "data_time": 0.00475, "loss_cls": 1.14408, "loss_bbox": 1.58183, "loss": 2.72591, "grad_norm": 4.51427, "time": 0.47594}
{"mode": "train", "epoch": 1, "iter": 600, "lr": 0.0025, "memory": 5607, "data_time": 0.00475, "loss_cls": 1.13299, "loss_bbox": 1.51068, "loss": 2.64367, "grad_norm": 4.22345, "time": 0.47734}
{"mode": "train", "epoch": 1, "iter": 650, "lr": 0.0025, "memory": 5607, "data_time": 0.00478, "loss_cls": 1.14294, "loss_bbox": 1.52628, "loss": 2.66923, "grad_norm": 3.63465, "time": 0.47958}
{"mode": "train", "epoch": 1, "iter": 700, "lr": 0.0025, "memory": 5607, "data_time": 0.00469, "loss_cls": 1.01544, "loss_bbox": 1.30221, "loss": 2.31765, "grad_norm": 3.81383, "time": 0.47745}
{"mode": "train", "epoch": 1, "iter": 750, "lr": 0.0025, "memory": 5607, "data_time": 0.00479, "loss_cls": 1.16248, "loss_bbox": 1.53069, "loss": 2.69318, "grad_norm": 4.64493, "time": 0.47873}
{"mode": "train", "epoch": 1, "iter": 800, "lr": 0.0025, "memory": 5607, "data_time": 0.00469, "loss_cls": 1.09673, "loss_bbox": 1.45115, "loss": 2.54788, "grad_norm": 3.74426, "time": 0.48132}
{"mode": "train", "epoch": 1, "iter": 850, "lr": 0.0025, "memory": 5607, "data_time": 0.00475, "loss_cls": 1.08604, "loss_bbox": 1.42378, "loss": 2.50982, "grad_norm": 3.37275, "time": 0.47997}
{"mode": "train", "epoch": 1, "iter": 900, "lr": 0.0025, "memory": 5607, "data_time": 0.00466, "loss_cls": 1.06428, "loss_bbox": 1.40702, "loss": 2.4713, "grad_norm": 3.62493, "time": 0.48248}
{"mode": "train", "epoch": 12, "iter": 5200, "lr": 3e-05, "memory": 5745, "data_time": 0.00526, "loss_cls": 0.79183, "loss_bbox": 0.79386, "loss": 1.58569, "grad_norm": 6.08125, "time": 0.50276}
{"mode": "train", "epoch": 12, "iter": 5250, "lr": 3e-05, "memory": 5745, "data_time": 0.00546, "loss_cls": 0.82034, "loss_bbox": 0.85067, "loss": 1.67101, "grad_norm": 6.63815, "time": 0.50692}