xiyang1012 / local-crowd-counting Goto Github PK

Thanks for the great code!
I succeeded in executing demo.py and successfully estimated the number of people with my own image. Wonderful.

I would like to know how this model (QNRF_mae_86.61_mse_152.19.pth) was trained.

thanks.

我是一个入门级学徒，调试代码只能进行到使用大佬发出来的训练好的QNRF模型进行预测的程度。想更近一步学习大佬的代码。请问什么时候能放出源码？

你好，大佬，请问一个问题，UCF-CC-50，怎么没有像其他数据集一样的preapre-xxx文件呢？

大佬好，因为看到这篇是基于C3框架的论文，我就把C3中的train.py和train-config搬过来了，参数都是按照论文中设置Epoch=400，lr=1e-5,bz=1,
问题一：然后训练后发现一百多个epoch就MAE最低到214（Adam),200(SGD）,之后就一直在MAE200到300间浮动，怎么也降不下来，我改动了您SHHA.py中加了一个 den = den.transpose()
def read_image_and_gt(self, fname):
img = Image.open(os.path.join(self.img_path, fname))
if img.mode == 'L':
img = img.convert('RGB')

    den = pd.read_csv(os.path.join(self.gt_path,os.path.splitext(fname)[0] + '.csv'), sep=',',header=None).values
    den = den.astype(np.float32, copy=False)
    den = den.transpose()
    den = Image.fromarray(den)

    return img, den

因为发现Image.fromarray读进来后会使得HW相反，但是这样改动之后可以训练了，但是就会出现前面说的mae降不下来的问题，想问问具体原因在哪呢，Adam 优化的时候MAE最低214 val loss 4.16左右
问题二：最好的表现模型VGG16_LCM_Regression在定义 #SAM scale-aware-module模块中的Maxout函数似乎与原文中的以四个膨胀卷纪率1234卷积再concat得操作好像不太一样，我将您注释的代码concat的还原了一下好像size无法匹配，还是太菜了，没理解这块的原理，球球大佬解释一下呀！！
def Maxout(x1, x2, x3, x4):
mask_1 = torch.ge(x1, x2)
mask_1 = mask_1.float()
x = mask_1 * x1 + (1-mask_1) * x2

mask_2 = torch.ge(x, x3)
mask_2 = mask_2.float()
x = mask_2 * x + (1-mask_2) * x3

mask_3 = torch.ge(x, x4)
mask_3 = mask_3.float()
x = mask_3 * x + (1-mask_3) * x4
return x

想问下这里的troch.ge比较前后两个变量的大小输出0/1，没想明白具体是啥意思，大佬可以解释下吗
最后大佬可以加一波微信或者qq或者邮件私聊我吗，最近一直在想复现这篇论文问题具体出在那里了,实在是找不到问题，很着急，像热锅上的蚂蚁，拜托大佬可以赐教一波！！
我的微信：knightyxp qq:377525381 email:[email protected]

SHHB，test图片316张，但该脚本文件中是range(1, 401)；train图片400张，但该脚本文件中是range(1, 317)。请问这里是错误的吗？

Hi there! Congratulations on your publication. I'm getting the following error when trying to run test.py. Can you kindly guide me what am I doing wrong here?

RuntimeError: Error(s) in loading state_dict for CrowdCounter:
Missing key(s) in state_dict: "CCN.layer3.0.weight", "CCN.layer3.0.bias", "CCN.layer3.2.weight", "CCN.layer3.2.bias", "CCN.layer3.5.weight", "CCN.layer3.5.bias", "CCN.layer3.7.weight", "CCN.layer3.7.bias", "CCN.layer3.10.weight", "CCN.layer3.10.bias", "CCN.layer3.12.weight", "CCN.layer3.12.bias", "CCN.layer3.14.weight", "CCN.layer3.14.bias", "CCN.layer3.17.weight", "CCN.layer3.17.bias", "CCN.layer3.19.weight", "CCN.layer3.19.bias", "CCN.layer3.21.weight", "CCN.layer3.21.bias", "CCN.layer4.1.weight", "CCN.layer4.1.bias", "CCN.layer4.3.weight", "CCN.layer4.3.bias", "CCN.layer4.5.weight", "CCN.layer4.5.bias", "CCN.layer5.1.weight", "CCN.layer5.1.bias", "CCN.layer5.3.weight", "CCN.layer5.3.bias", "CCN.layer5.5.weight", "CCN.layer5.5.bias", "CCN.fuse_layer5.conv1x1_d1.weight", "CCN.fuse_layer5.conv1x1_d1.bias", "CCN.fuse_layer5.conv1x1_d2.weight", "CCN.fuse_layer5.conv1x1_d2.bias", "CCN.fuse_layer5.conv1x1_d3.weight", "CCN.fuse_layer5.conv1x1_d3.bias", "CCN.fuse_layer5.conv1x1_d4.weight", "CCN.fuse_layer5.conv1x1_d4.bias", "CCN.fuse_layer5.conv_d1.weight", "CCN.fuse_layer5.conv_d1.bias", "CCN.fuse_layer5.conv_d2.weight", "CCN.fuse_layer5.conv_d2.bias", "CCN.fuse_layer5.conv_d3.weight", "CCN.fuse_layer5.conv_d3.bias", "CCN.fuse_layer5.conv_d4.weight", "CCN.fuse_layer5.conv_d4.bias", "CCN.fuse_layer4.conv1x1_d1.weight", "CCN.fuse_layer4.conv1x1_d1.bias", "CCN.fuse_layer4.conv1x1_d2.weight", "CCN.fuse_layer4.conv1x1_d2.bias", "CCN.fuse_layer4.conv1x1_d3.weight", "CCN.fuse_layer4.conv1x1_d3.bias", "CCN.fuse_layer4.conv1x1_d4.weight", "CCN.fuse_layer4.conv1x1_d4.bias", "CCN.fuse_layer4.conv_d1.weight", "CCN.fuse_layer4.conv_d1.bias", "CCN.fuse_layer4.conv_d2.weight", "CCN.fuse_layer4.conv_d2.bias", "CCN.fuse_layer4.conv_d3.weight", "CCN.fuse_layer4.conv_d3.bias", "CCN.fuse_layer4.conv_d4.weight", "CCN.fuse_layer4.conv_d4.bias", "CCN.fuse_layer3.conv1x1_d1.weight", "CCN.fuse_layer3.conv1x1_d1.bias", "CCN.fuse_layer3.conv1x1_d2.weight", "CCN.fuse_layer3.conv1x1_d2.bias", "CCN.fuse_layer3.conv1x1_d3.weight", "CCN.fuse_layer3.conv1x1_d3.bias", "CCN.fuse_layer3.conv1x1_d4.weight", "CCN.fuse_layer3.conv1x1_d4.bias", "CCN.fuse_layer3.conv_d1.weight", "CCN.fuse_layer3.conv_d1.bias", "CCN.fuse_layer3.conv_d2.weight", "CCN.fuse_layer3.conv_d2.bias", "CCN.fuse_layer3.conv_d3.weight", "CCN.fuse_layer3.conv_d3.bias", "CCN.fuse_layer3.conv_d4.weight", "CCN.fuse_layer3.conv_d4.bias", "CCN.count_layer5.avgpool_layer.0.weight", "CCN.count_layer5.avgpool_layer.0.bias", "CCN.count_layer5.maxpool_layer.0.weight", "CCN.count_layer5.maxpool_layer.0.bias", "CCN.count_layer5.conv1x1.0.weight", "CCN.count_layer5.conv1x1.0.bias", "CCN.count_layer4.avgpool_layer.0.weight", "CCN.count_layer4.avgpool_layer.0.bias", "CCN.count_layer4.maxpool_layer.0.weight", "CCN.count_layer4.maxpool_layer.0.bias", "CCN.count_layer4.conv1x1.0.weight", "CCN.count_layer4.conv1x1.0.bias", "CCN.count_layer3.avgpool_layer.0.weight", "CCN.count_layer3.avgpool_layer.0.bias", "CCN.count_layer3.maxpool_layer.0.weight", "CCN.count_layer3.maxpool_layer.0.bias", "CCN.count_layer3.conv1x1.0.weight", "CCN.count_layer3.conv1x1.0.bias", "CCN.layer5_k.0.weight", "CCN.layer5_k.0.bias", "CCN.layer4_k.0.weight", "CCN.layer4_k.0.bias", "CCN.layer3_k.0.weight", "CCN.layer3_k.0.bias", "CCN.layer5_i.0.weight", "CCN.layer5_i.0.bias", "CCN.layer4_i.0.weight", "CCN.layer4_i.0.bias", "CCN.layer3_i.0.weight", "CCN.layer3_i.0.bias", "CCN.layer5_p.0.weight", "CCN.layer5_p.0.bias", "CCN.layer4_p.0.weight", "CCN.layer4_p.0.bias", "CCN.layer3_p.0.weight", "CCN.layer3_p.0.bias".
Unexpected key(s) in state_dict: "features.0.weight", "features.0.bias", "features.2.weight", "features.2.bias", "features.5.weight", "features.5.bias", "features.7.weight", "features.7.bias", "features.10.weight", "features.10.bias", "features.12.weight", "features.12.bias", "features.14.weight", "features.14.bias", "features.17.weight", "features.17.bias", "features.19.weight", "features.19.bias", "features.21.weight", "features.21.bias", "features.24.weight", "features.24.bias", "features.26.weight", "features.26.bias", "features.28.weight", "features.28.bias", "classifier.0.weight", "classifier.0.bias", "classifier.3.weight", "classifier.3.bias", "classifier.6.weight", "classifier.6.bias".

Hi, thanks for your excellent work and the open-source codes!
I'm very interested in the novel optimization criterion proposed in your paper and tried to reproduce it locally. But there is a confusing problem of performance degradation when I restrained the height and width of the test images to make sure that they are divisible by a particular number (like, 8, 16, or 32). This data preprocessing strategy is inspired by C^3 Framework and guarantees the output size of some down-sampling layers count meet the requirements of subsequent processing. More details of the test results are shown in the table below.

This issue is to sincerely ask you how such a simple operation can have such an obvious impact on the experimental results.

Local patch等于把Density map在Local 范围内加起来，那不就等价于直接把Density map降Resolution吗？

Thanks for your great work.I want to get density map of a single picture from other dataset, such as the photo took by myself.Can you provide the code? Thanks.

偶尔正常：
density_map: torch.Size([1, 1, 64, 64])
gt_data: torch.Size([1, 1, 64, 64])
match不上，程序就直接停掉了：
density_map: torch.Size([1, 1, 64, 64])
gt_data: torch.Size([1, 1, 64, 37])
报错信息：
RuntimeError: The size of tensor a (64) must match the size of tensor b (37) at non-singleton dimension 3

Hi,
I noticed that this model accept any image shape and produce different pred map shapes according to the input image shape,
So is this normal ?
Does there any recommended image size for the model ?, like the one it trained with?

"I am asking this question because I need specific input and output shapes to convert it to engine"

Hello, I find your work has been accepted by ECCV. Congratuations, but I want to ask when can you release your training code?