training on my own dataset

[liuliu66]

@unsky Hi, unsky. I really appreciate the code for FPN. But when I used it I got very bad mAP for my single object detection task on my own dataset. I have changed the parameters in train.prototxt and train_mergercnn.prototxt with changing 21 to 2 (object + background) and 84 to 8. The code could run well and loss also looks fine. So could you please tell me are there other places I need to change parameters to for training on my dataset?

[unsky]

you have bug in test code,
i think you havenot restore the rois.
i do this in

FPN/lib/fast_rcnn/test.py

Line 292 in 2f3e5c3

if cfg.TRAIN.BBOX_NORMALIZE_TARGETS_PRECOMPUTED:

rgb's faster rcnn do this in:
https://github.com/rbgirshick/py-faster-rcnn/blob/4e199d792f625cf712ca9b9a16278bafe0806201/lib/fast_rcnn/train.py#L65

[liuliu66]

@unsky Solved! Amazing result! Thanks for your help!

[TongYoungG]

@liuliu66 Hello, I got a problem same as you. Can you show me the detail that you solved the problem? I still don't know how to fix it. Thanks!

[liuliu66]

@gk616828144 Hello, my problem is duo to run a wrong sh file so I just check the test.sh and change the model path to my trained model path and the problem is solved. I think if you could run the code on your dataset and got seemly correct loss you could get correct caffemodel file. You can check test.sh and test_mergercnn.sh carefully to solve it.

[lixiangamy]

I have changed the parameters in train.prototxt with changing 21 to 3 (object + background) and 84 to 12.
cls_score_concat ：Top shape:4 3（12）？
error：Creating layer loss_cls
F0118 13:55:03.960583 15560 loss_layer.cpp:19] Check failed: bottom[0]->num() == bottom[1]->num() (4 vs. 1) The data and label should have the same number.

[unsky]

@lixiangamy use fpn-caffe!

[lixiangamy]

What are the differences between fpn-caffe and caffe-fast-rcnn?I used caffe-fast-rcnn

[unsky]

Duplicate of #21 #21

[TongYoungG]

@unsky hello,how to restore the rois? I still don't what to do.When I test imgs,the rois is all the same. I don't know where the error is.

[foralliance]

@unsky
虽然@liuliu66的问题是由于file文件导致的，但您指出 "i think you havenot restore the rois."这该如何理解呢？

[unsky]

FPN属于faster rcnn框架发展而来，保留了faster rcnn对于rois的设置，faster rcnn 有一个trick就是对第二阶段的rcnn的rois做了白化处理，这种策略可以加速收敛，当然也需要我们在测试阶段进行一次反白化。

[foralliance]

@unsky HI
在FPN的test.py中，在im_detect函数中，对最终的结果确实多了一个normalize操作，这个操作在py-faster-rcnn中是没有的．
在py-faster-rcnn中的trian.py中，也多了一个normalize操作，这个操作在FPN中是没有的．

这2者之间有关系吗？一个是训练的，一个是测试的．而且，您所谓的反白话又是什么意思？？

[mantou22]

作者您好， 请问训练集像素大小为100100， 目标大小为4040 ，请问 训练时如何设置anchor大小？？
谢谢作者啦

[mantou22]

您好，请问为什么我训练自己的模型会怎么慢速度？
前面几百次迭代，每20次只需要 34S 左右， 到了2000次左右时， 每20次需要 5-6分钟 为什么会那么慢？？？？？ 我的显卡是 GPU 1080Ti 的 12G运行内存