Yet-another-efficientdet-pytorch: 训练自己的数据

[zhangxue123]

首先感谢作者提供的代码，我在用最新版本的代码采用D2和D0训练我自己的数据时，tensorboard查看loss, loss最低只能到0.54，500多个epoch也未能收敛，测试时在验证集和训练集精度都基本为0。我的训练集为5000个样本（检测9类目标，标注框有13893个）完全制作为coco格式喂给模型，检查过数据是正确的，请问您知道大概是哪里有问题呢，导致无法检测？

[zylo117]

you should check for overfitting.
And 0.54 is not low enough.
When training coco, D0's val cls loss is only 0.25 and val reg loss is only 0.025.

[Cli98]

首先感谢作者提供的代码，我在用最新版本的代码采用D2和D0训练我自己的数据时，tensorboard查看loss, loss最低只能到0.54，500多个epoch也未能收敛，测试时在验证集和训练集精度都基本为0。我的训练集为5000个样本（检测9类目标，标注框有13893个）完全制作为coco格式喂给模型，检查过数据是正确的，请问您知道大概是哪里有问题呢，导致无法检测？

@zhangxue123 This is quite common in this repo... As far as I know, no customer dataset get satisfying performance by now.

[zylo117]

What about shape dataset? It trains well.
This model is very sensitive to batchsize. Not smaller, not larger, it has to be perfect.

[zhangxue123]

What about shape dataset? It trains well.
This model is very sensitive to batchsize. Not smaller, not larger, it has to be perfect.

I checked the data, no problem, I think the model is probably only good for the COCO dataset because of the neural network search.

[JimLee1996]

One of the critical reasons is that no other vital augmentations is implemented except random flip. LMAO.

[zylo117]

Seriously?
No augmentation is applied on official impl except random filp. They have AA, but none is applied by default. If you want it, you can implement it yourself.

[Cli98]

What about shape dataset? It trains well.
This model is very sensitive to batchsize. Not smaller, not larger, it has to be perfect.

I checked the data, no problem, I think the model is probably only good for the COCO dataset because of the neural network search.

@zhangxue123 Try my code if you want to modify you anchors.
Here is the link: https://github.com/Cli98/anchor_computation_tool

[pierrelzw]

This model is very sensitive to batchsize. Not smaller, not larger, it has to be perfect.

Is that mean it's better to use batchsize=8 for D0-D7? I also train custom data with this repo but the result AP is quite bad. With D3 and D6, i got [email protected]=0.4, while i got [email protected]=0.85 with CenterNet_res101 using the same dataset.

I didn't change any code but those for logging. My best train and val loss=0.1 with cls_loss =0.082 and reg_loss=0.017.

After that the model seems overfit: the val_loss is increased to 0.13 and train loss is decreased to 0.05 or less.

Do you have more suggestions for training the custom data?

[zylo117]

effdet is much deeper and has more activations, so it is more easily overfitting. especially when training larger network like d6/d7, AA should be applied like the paper says.
I've trained lots of custom datasets with D0 myself, no extra augmentation is applied, and the results are great. A little bit better than yolov3. I personally recommend D0 to D3, maybe including D4, the others are too painful to train and consume too much memory.

[mad-fogs]

首先感谢作者提供的代码，我在用最新版本的代码采用D2和D0训练我自己的数据时，tensorboard查看loss, loss最低只能到0.54，500多个epoch也未能收敛，测试时在验证集和训练集精度都基本为0。我的训练集为5000个样本（检测9类目标，标注框有13893个）完全制作为coco格式喂给模型，检查过数据是正确的，请问您知道大概是哪里有问题呢，导致无法检测？

@zhangxue123 This is quite common in this repo... As far as I know, no customer dataset get satisfying performance by now.

Well, i dont think it is wise choice to post words of this kind.
I got prefect performance on my private datasets, at least.

[zylo117]

首先感谢作者提供的代码，我在用最新版本的代码采用D2和D0训练我自己的数据时，tensorboard查看loss, loss最低只能到0.54，500多个epoch也未能收敛，测试时在验证集和训练集精度都基本为0。我的训练集为5000个样本（检测9类目标，标注框有13893个）完全制作为coco格式喂给模型，检查过数据是正确的，请问您知道大概是哪里有问题呢，导致无法检测？

@zhangxue123 This is quite common in this repo... As far as I know, no customer dataset get satisfying performance by now.

Well, i dont think it is wise choice to post words of this kind.
I got prefect performance on my private datasets, at least.

Same here, I actually trained a lot of custom datasets already, all of them are much better than maskrcnn.

[newjokker]

首先感谢作者提供的代码，我在用最新版本的代码采用D2和D0训练我自己的数据时，tensorboard查看loss, loss最低只能到0.54，500多个epoch也未能收敛，测试时在验证集和训练集精度都基本为0。我的训练集为5000个样本（检测9类目标，标注框有13893个）完全制作为coco格式喂给模型，检查过数据是正确的，请问您知道大概是哪里有问题呢，导致无法检测？

我也遇到了同样的问题，不知道是哪一步出现了错误

[zylo117]

首先感谢作者提供的代码，我在用最新版本的代码采用D2和D0训练我自己的数据时，tensorboard查看loss, loss最低只能到0.54，500多个epoch也未能收敛，测试时在验证集和训练集精度都基本为0。我的训练集为5000个样本（检测9类目标，标注框有13893个）完全制作为coco格式喂给模型，检查过数据是正确的，请问您知道大概是哪里有问题呢，导致无法检测？

我也遇到了同样的问题，不知道是哪一步出现了错误

underfitting