yxlijun / pyramidbox.pytorch Goto Github PK

在你的code中，conv6的形式是：
conv6 = nn.Conv2d(512, 1024, kernel_size=3, padding=6, dilation=6)
但这个形式其实是SSD中的。
而在PyramidBox中，底层是基于S3FD中的VGG16。跟SSD没关系。所以conv6的形式应该是：
conv6 = nn.Conv2d(512, 1024, kernel_size=3, padding=1)

您好，我在看完论文后，感觉您写的代码和论文中cpm那张图有点不一样，想请问一下是我理解的不到位，还是这部分代码并没有根据论文进行编写。十分感谢！！

I met an error.
Traceback (most recent call last):
File "demo.py", line 111, in
detect(net, path, args.thresh)
File "demo.py", line 70, in detect
y = net(x)
File "/lib/python3.7/site-packages/torch/nn/modules/module.py", line 727, in _call_impl
result = self.forward(*input, **kwargs)
TypeError: forward() missing 1 required positional argument: 'x'

how do I fix it?

Trying to run wider_test.py on validation part:

python wider_test.py --model ../weights/pyramidbox_5000.pth

but getting an error (on 0004th image?):

Detect 0001 th image costs 4.4119
Detect 0002 th image costs 2.4946
Detect 0003 th image costs 3.0682
THCudaCheck FAIL file=/opt/conda/conda-bld/pytorch_1518244421288/work/torch/lib/THC/generic/THCStorage.cu line=58 error=2 : out of memory
Traceback (most recent call last):
  File "wider_test.py", line 238, in <module>
    det0 = detect_face(net, img, shrink)
  File "wider_test.py", line 61, in detect_face
    y = net(x)
  File "/root/miniconda3/envs/pytorch0.3.1/lib/python3.6/site-packages/torch/nn/modules/module.py", line 357, in __call__
    result = self.forward(*input, **kwargs)
  File "../pyramidbox.py", line 153, in forward
    ssh_conv3_norm = self.cpm[0](self.L2Norm3_3(lfpn0_on_conv3))
  File "/root/miniconda3/envs/pytorch0.3.1/lib/python3.6/site-packages/torch/nn/modules/module.py", line 357, in __call__
    result = self.forward(*input, **kwargs)
  File "../pyramidbox.py", line 57, in forward
    out_residual = self.branch1(x)
  File "/root/miniconda3/envs/pytorch0.3.1/lib/python3.6/site-packages/torch/nn/modules/module.py", line 357, in __call__
    result = self.forward(*input, **kwargs)
  File "../pyramidbox.py", line 36, in forward
    return self.bn1(x)
  File "/root/miniconda3/envs/pytorch0.3.1/lib/python3.6/site-packages/torch/nn/modules/module.py", line 357, in __call__
    result = self.forward(*input, **kwargs)
  File "/root/miniconda3/envs/pytorch0.3.1/lib/python3.6/site-packages/torch/nn/modules/batchnorm.py", line 37, in forward
    self.training, self.momentum, self.eps)
  File "/root/miniconda3/envs/pytorch0.3.1/lib/python3.6/site-packages/torch/nn/functional.py", line 1013, in batch_norm
    return f(input, weight, bias)
RuntimeError: cuda runtime error (2) : out of memory at /opt/conda/conda-bld/pytorch_1518244421288/work/torch/lib/THC/generic/THCStorage.cu:58

My pytorch=0.3.1
I'm using GeForce GTX 1080 Ti (11178Mb).

@yxlijun you also work with 11Gb Titan1080. Can you help to manage this memory problem?

@yxlijun HI
看了您的code，受益匪浅
起初看paper时，因为公式较多，很多地方难以理解．结合您的code来看，很多都理解了．非常感谢！

整体感觉：
从code来看，对于损失函数这里，其实还是常规的．重点在于head标签信息的设置，其本质上是通过face的标签信息得到的：

1. 关于head坐标信息，体现在data/widerface.py中，head的坐标就是在face的基础上扩大了一倍．这个感觉就对应的是**公式(5)**的内容
2. 关于head类别标签（0/1）信息，和face标签保持一致．

在正负样本匹配时：
从code来看，不论是face还是head，共用一组anchor（priors）．
比如一个face，尺度128．那么这个face对应的合理的anchor就会出现在第4个预测分支上．这个face对应head的尺度就是256，那么这个head对应的合理的anchor就会出现在第5个预测分支上．
有个疑问：因为head是由这个face衍生来的，那么需不需要把这个2个anchor标记一下，表示其来自于同一个face呢？？从**公式(1),(4)**来看，face和head的anchor总感觉有一种匹配关系．但是从code来看，没有这样的标记，针对face和head的损失都是独立进行的．

另外有另一个版本的PyramidBox复现代码，他针对face和head设计了2组anchors．但总感觉head_anchors设计的有问题．不知道您参考过没？

when i run demo.py，it will happen a mistake which cuda out of memory，but my computer is 2080ti，i don't know how to solve it,Do you have such a mistake?

Hi everyone,

Can anyone provide a download link for pretrained model (vgg16_reducedfc) other than pan.baidu link preferably Google Drive. I have been struggling to download the pth file and I would really appreciate the help. Thanks

请问在哪里设置max_im_shrink

您好，谢谢您提供的代码！我在测试时遇到问题。detect到第1241张图片后traceback，在调用wider_test.py下的bbox_vote函数时，没有进入while循环（报错：UnboundLocalError:local variable 'dets' referenced before assignment），您可以对我的遇到的错误给一些意见吗？谢谢！

Excuse me
I get errors when I run the demo.py ，I can't find the module detection.
the erro: ModuleNotFoundError: No module named 'detection' from detection import Detect

I also want to ask a question whether you used the body to train the module.

and if you can speak chinese ,you can return in chinese to me . thanks.

首先很感谢你提供的项目，很出色的工作。我按照步骤修改了一些小bug,跑通了代码并取得了十分接近原论文的结果。由于论文中许多地方没有给出细节，我在学习代码的过程中有一些看不懂的地方，不知是否能给予初学者一些帮助。
1 prior_box.py文件里 forward函数

这里 s_kw = self.min_sizes[k] / self.imw ，用anchor_size / image_size 有什么意思？

2 pyramidbox.py文件里 multibox函数，

为啥第0层的confidence是8个通道的，而别的层都是6通道？

3 augmentations.py文件里preprocess函数

这里else 后面的操作是干啥意思，算法思路是啥啊，augmentations.py也没有注释，简直要把我看晕了……能否讲一下主要**

4 您给出的代码和原论文给出的代码都是用的SSD_loss, 请问论文里3.4节 PyramidBox loss.那一段里给出的几个公式，


在代码中有体现吗？我没有找到相关的代码

再次感谢你的帮助和不吝赐教！！这对像我一样的小白帮助很大！

只有将max_im_shrink设为1的时候才不会溢出，1080Ti的环境，感觉是什么东西一下子占了很多显存，虽然只allocate不到2G，但free的显存很少。

您好：现在用您的代码跑了个基本结果，fddb数据集基本对上，在测试wideerface时总是out of memory,我的显卡为k80,请问您的显卡显存时多大？谢谢。

File ".\demo.py", line 74, in detect
y = net(x)
File "C:\Users\xxxxx\Miniconda3\lib\site-packages\torch\nn\modules\module.py", line 493, in call
result = self.forward(*input, **kwargs)
File "C:\Users\xxxxx\Desktop\local\Pyramidbox.pytorch\pyramidbox.py", line 252, in forward
priors)
File "C:\Users\xxxxx\Desktop\local\Pyramidbox.pytorch\layers\functions\detection.py", line 65, in forward
boxes_, scores, self.nms_thresh, self.top_k)
ValueError: not enough values to unpack (expected 2, got 0)
PS C:\Users\shuxin\Desktop\local\Pyramidbox.pytorch>

运行train.py时。报错提示
RuntimeError: cuda runtime error (2) : out of memory at /pytorch/torch/lib/THC/generic/THCStorage.cu:58
使用的训练显卡是 TITAN XP 12196Mb
请问如何解决

您好，非常感谢您的代码，我现在使用的是pytorch1.0稳定版，我按照您的设置进行训练时，尝试了不同的学习率，发现1e-3就可以开始收敛，您问什么把学习率设置那么小，是pytorch版本问题，还是什么trick。

您好！请问您提供的pyramidbox.pth是通过您给的代码训练出来的吗？我用您的代码训练了模型，在测试的时候每张图片得到的结果都是如下图左所示（最后的图是用您提供的模型测试出的）。我是否是在训练的某一步骤出了错呢？谢谢！

在文件tools/wider_test.py中，函数bbox_vote中
if merge_index.shape[0] <= 1:
continue
不是应该if merge_index.shape[0] < 1: continue
吗？

Thanks a lot yxlijun.
Your work has helped me a lot with my project.
I tried to download trained model for running demo.py
But it seems I can't get to baidu drive.
Uploading to ex)google drive will be very helpful.

I am currently working on implementation of pyramidbox using pre trained model in google colab, I have tried to follow all the instruction mentioned in the repo for the same. While implementing the command(python tools/wider_test.py) I am getting the following error:
Traceback (most recent call last):
File "tools/wider_test.py", line 22, in
from data.config import cfg
ModuleNotFoundError: No module named 'data'
Please help me out!!