I want to know how to deal with this kind of error.Thank you very much!
the error is:
RuntimeError: arguments are located on different GPUs at /opt/conda/conda-bld/pytorch_1532581333611/work/aten/src/THC/generated/../generic/THCTensorMathPointwise.cu:314

The score seems too strange to me.
some of them are 18-19 and others 8-6.
and the threshold has been kept to 6.
What is the score range?

@louis-she HI

在paper中，在计算loss时，有这样的描述：

The two terms are normalized by Ncls and Nreg , and weighted by a balancing parameter λ . 
In our implementation, the cls term is normalized by the number of positive and negative anchors, and the reg term is normalized by the number of positive anchors.  Because of the imbalance between the number of positive and negative anchors, λ is used to balance these two loss terms.

换句话说，理论上，分类损失除以（正样本+负样本）个数，回归损失除以（正样本）个数，然后给分类损失乘以一个权重（4）．
比如正样本个数是100，那么分类损失就是除以400，然后乘以4；回归损失就是除以100．

但在code中，在计算2个损失时，

loss_class = 4 * F.cross_entropy(total_effective_pred, total_targets)
loss_reg = F.smooth_l1_loss(total_t, total_gt)

此时，相当于分类损失和回归损失都除以了（正样本+负样本）个数，都是400．

那么再给分类损失乘以权重4是不是有问题啊？？

@louis-she HI

感觉anchor的设置有点问题．感觉您是完全按照paper中的fig1(3)设置的．
比如：
image的size为640．
在5*5的特征图上，anchor尺度为512，stride为128．根据您的代码，第1个anchor的坐标形式为(256,256,512,512)．中心点的位置很奇怪．难道不应该是(64,64,512,512)这样的anchor吗？？

@louis-she HI

结合caffe源码的#12和paper，感觉：

对于本文的scale compensation anchor matching strategy，其是指在降低匹配阈值（0.5到0.35）后，如果image中某个face匹配的anchor数小于6个，才进行2-stage匹配，而且每个face的匹配上限是6．
但在您的代码中，考虑的是image中所有face匹配的anchor数（上限100），并不是单个face匹配的anchor数

can the code trained on multiple gpus?
when training on a single gpu, it just fine.
while training on multiple ones, errors like as below happen
"RuntimeError: arguments are located on different GPUs at /opt/conda/conda-bld/pytorch_153│| No running processes found |
3672544752/work/aten/src/THC/generated/../generic/THCTensorMathPointwise.cu:314"

thanks~

您好，我想请问一下您复现出S3FD代码，在测试时间的速度是多少？谢谢

Dear @louis-she,
Thank you for your nice work.
As Shifeng Zhang has mentioned in his paper: Our anchors are 1:1 aspect ratio (_i.e.,_ square anchor), the aspect ratios of all anchor boxes are 1:1. However, if one wants to change the aspect ratios of anchor boxes for example, to 1:1.4 (width:height) the network structure must be changed or just the training code & decoding strategy (i.e., utils.py or anchor.py) must be changed? Would you please explain how one can change the aspect ratios of anchor boxes?

Hi,sir ,could this object be run on python3.5.2?

hi Louis,
I think there are some attributes in config.py such as RANDOM_CROP, RANDOM_FLIP ,MIN_CROPPED_RATIO, etc.
Could you please share the config.py file you used for training?
Many thanks,
Yiming

hi，I don't seem to see you has done ‘’maxout‘’ in the code. It is different from the paper？

I can't find it in your code.

@louis-she HI
在利用anchors坐标和网络预测的偏移量得到最终box坐标时，都会引入一个参数variances．但在您的代码中，并未看到此参数．

x = (predictions[:, 0] * anchors[:, 2] + anchors[:, 0])
y = (predictions[:, 1] * anchors[:, 3] + anchors[:, 1])
w = (torch.exp(predictions[:, 2]) * anchors[:, 2])
h = (torch.exp(predictions[:, 3]) * anchors[:, 3])

bounding_boxes = torch.stack((x, y, w, h), dim=1).cpu().data.numpy()  #(200,4)
bounding_boxes = change_coordinate_inv(bounding_boxes)

In the paper,
3.2 Scale compensation anchor matching strategy
stage two is used to compensate the faces with few anchors.
However, in anchor.py, mark_anchors seems to compensate all the faces

Does this use nms during testing? I do not find nms part.

Then I run the command:
python3 detector.py --image ./image/test.jpg --model ./epoch_204.pth.tar
Then I got this error:
Traceback (most recent call last):
File "detector.py", line 11, in
from model import Net
File "E:\code\S3FD\sfd.pytorch-master\model.py", line 5, in
from torchvision.models.vgg import VGG, cfg, make_layers, vgg16
ImportError: cannot import name 'cfg'
And I still don't figure out what goes wrong

@louis-she HI
根据caffe源码的#7，因为eval_tools的原因，需要屏蔽掉tiny faces．原作者指出：When training，we do not use these extremely tiny faces (i.g., width or height＜6 pixels).
但在您的代码中，并未体现这一点！！

谢谢！

Dear @louis-she & @pengbo0054,
In

The format of txt ground truth is as follows: 
File name
Number of bounding box
x1, y1, w, h, blur, expression, illumination, invalid, occlusion, pos

I mean that, I guess a mistake occur in this part of code. Maybe you should change the dataset.py as follows:

        # scale coordinate
        height, width = image.shape[:2]
        width_scale, height_scale = 640.0 / width, 640.0 / height
        coordinates = np.array(list(map(lambda x: [
            x[0] * width_scale,  # Change this part
            x[1] * height_scale,  # Change this part
            x[2] * width_scale,  # Change this part
            x[3] * height_scale  # Change this part
], coordinates)))

Am I correct?

Just as another note. I suggest that you add a general ListDataset class instead of specific data sets (e.g., Wider-Face, Pascal VOC, etc.). For example the ListDataset class can use data set annotations as follows:

Load image/labels/boxes from a list file (e.g., *.txt file).
The list file is like:
a.jpg xmin ymin xmax ymax label xmin ymin xmax ymax label ...

Thank you

Hi, thank you for the great work.
I was trying to test your code, but when I run:
python3 detector.py --image ./images/test.jpg --model ./epoch_41.pth.tar
I got the following error.

predicted bounding boxes of faces:
Traceback (most recent call last):
  File "detector.py", line 160, in <module>
    main(args)
  File "detector.py", line 141, in main
    bboxes = Detector(args.model).infer(args.image)
  File "detector.py", line 24, in __init__
    self.model.load_state_dict(checkpoint['state_dict'], strict=True)
  File "/usr/local/lib/python3.6/dist-packages/torch/nn/modules/module.py", line 721, in load_state_dict
    self.__class__.__name__, "\n\t".join(error_msgs)))
RuntimeError: Error(s) in loading state_dict for Net:
	Missing key(s) in state_dict: "predict3_3_reg.weight", "predict3_3_reg.bias", "predict4_3_reg.weight", "predict4_3_reg.bias", "predict5_3_reg.weight", "predict5_3_reg.bias", "predict_fc7_reg.weight", "predict_fc7_reg.bias", "predict6_2_reg.weight", "predict6_2_reg.bias", "predict7_2_reg.weight", "predict7_2_reg.bias", "predict3_3_cls.weight", "predict3_3_cls.bias", "predict4_3_cls.weight", "predict4_3_cls.bias", "predict5_3_cls.weight", "predict5_3_cls.bias", "predict_fc7_cls.weight", "predict_fc7_cls.bias", "predict6_2_cls.weight", "predict6_2_cls.bias", "predict7_2_cls.weight", "predict7_2_cls.bias". 
	Unexpected key(s) in state_dict: "predict3_3.weight", "predict3_3.bias", "predict4_3.weight", "predict4_3.bias", "predict5_3.weight", "predict5_3.bias", "predict_fc7.weight", "predict_fc7.bias", "predict6_2.weight", "predict6_2.bias", "predict7_2.weight", "predict7_2.bias". 

It seems to be something wrong in the provided model or the way to read the model, but I am new to pytorch and couldn't figure out the problem. I have pytorch 0.4 and python 3.6 as required.

Thank you.

It cost 100 ms on one GPU by Tesla P100 . How can i do to speed up the infer?