我直接训练最好得到54，但是我用给出的权重（51.9，阈值保持默认为0.5）跑出来是53.13，但是阈值改成0.45是可以到52的，请问51.9是在当前代码默认setting下训练出来的还是有所调整呢？ 如果是在当前代码下跑出来的，请问能给一下你们的SHTechA的数据集和组织格式吗？因为我只改了loader。

Hi,

I could reproduce MAE results on ShanghaiTech part A using supplied arguments for train script. However, with same parameters, for ShanghaiTech part B I can not reproduce results, getting a MAE of 11.18 instead of article results which get a 6.25 .

I have created an script to build train files out of ShanghaiTech datasets and I can confirm that this part is ok because I can reproduce MAE with generated files (for part A).

For partB, I have tried with same supplied parameters and also changing learning rate but MAE does not go below 11.

Any more tweaks to be done in dataset loading? Any different parameter values?

Many thanks for your great article and code,

Ricard

Hi all!

I'd love to use this model in our ONNX flows but wasn't able to convert it to ONNX. Is there any known way of converting this model to ONNX?

Code I am using:

import os
import sys
import torch

sys.path.append(os.path.abspath(f"{os.getcwd()}/model"))

# Available after the above append
# it's in the model folder
from model.models.p2pnet import P2PNet
from model.models.backbone import Backbone_VGG

def main():
    onnx_model_name = sys.argv[1] or "model"
    onnx_model_name = f"{onnx_model_name}.onnx"

    print("Loading Model")
    # Create the model
    model_backbone = Backbone_VGG("vgg16_bn", True)
    model = P2PNet(model_backbone, 2, 2)

    # Load Weights
    checkpoint = torch.load("./model/weights/SHTechA.pth", map_location=torch.device('cpu'))
    model.load_state_dict(checkpoint["model"])
    model.eval() # Put in inference mode
    
    # Create dummy input
    dummy_input = torch.randn(1, 3, 640, 640)
    # dummy_input1 = torch.randn(1, 3, 1024, 1024)
    # dummy_input = (dummy_input0, dummy_input1)

    # Export as ONNX
    print(f"Exporting as ONNX: {onnx_model_name}")
    torch.onnx._export(
        model,
        dummy_input,
        onnx_model_name, # Output name
        opset_version=13, # ONNX Opset Version
        export_params=True, # Store the trained parameters in the model file
        do_constant_folding=True, # Execute constant folding for optimization
        input_names = ['input'],   # the model's input names 
        # output_names = ['pred_logits', 'pred_points'], # the model's output names (see forward in the architecture)
        output_names = ['pred_logits', 'pred_points'], # the model's output names (see forward in the architecture)
        dynamic_axes={
            # Input is an image [batch_size, channels, width, height]
            # all of it can be variable so we need to add it in dynamic_axes
            'input': {
                0: 'batch_size',
                1: 'channels',
                2: 'width',
                3: 'height'
            }, 
            'pred_logits': [0, 1, 2],
            'pred_points': [0, 1, 2],
        } 
    )


if __name__ == "__main__":
    main()

Error I receive:

[ONNXRuntimeError] : 1 : FAIL : Non-zero status code returned while running Add node. Name:'Add_88' Status Message: Add_88: right operand cannot broadcast on dim 1 LeftShape: {1,40960,2}, RightShape: {1,25600,2}

https://github.com/TencentYoutuResearch/CrowdCounting-P2PNet/blob/5c91a81ca062b1c7fd3db3ad1c55b1c21f0a7455/train.py#L187 As I run code above, an error "out of memory" pops while I run the script on single gpu RTX6000 with 24G memory.

Hello,

thanks for uploading the code!

When trying to run a test using supplied pretrained Shanghai Dataset Part A, an error is raised about VGG_BN pretrained weights:

No such file or directory: '/apdcephfs/private_changanwang/checkpoints/vgg16_bn-6c64b313.pth

Could you share it?

Thanks

    您好，我也是一样的情况，在出现了Namespace(backbone='vgg16_bn', gpu_id=0, line=2, output_dir='./logs/', row=2, weight_path='./weights/SHTechA.pth')之后，后面就没有反应了。不知道有没有顺利解决这个问题的

Originally posted by @fatakWang in https://github.com/TencentYoutuResearch/CrowdCounting-P2PNet/issues/16#issuecomment-1364650983

In models/p2pnet.py:

Line 278: losses['loss_point'] = loss_bbox.sum() / num_points Line 335: weight_dict = {'loss_ce': 1, 'loss_points': args.point_loss_coef}

one is "loss_points" and another is "loss_point" (an "s" is different) ! This would cause that the point loss is actually not used, which is fully wrong!

While reading the paper, I came to know that the authors used 13 convs from VGG16 network to extract deep features. Since VGG nets have been around for a quite a long time, why didn't you choose some more efficient and accurate networks? Or is there something I don't know of? Thank you.

hello ~thanks for this great job. I get some trouble when resizing images in run_tset.py. I modify new_width and new_height like this:

# load the images
img_raw = Image.open(data_path+img_path).convert('RGB')
# round the size
width, height = img_raw.size
img_raw = img_raw.resize((int(width*0.5), int(height*0.5)), Image.ANTIALIAS)

there is the error message： Namespace(backbone='vgg16_bn', gpu_id=0, line=2, output_dir='./logs/', row=2, weight_path='./weights/SHTechA.pth') File "run_test.py", line 140, in main(args) File "run_test.py", line 107, in main result = self.forward(*input, **kwargs) File "D:\working_space_HJ\CrowdCounting-P2PNet-main\models\p2pnet.py", line 215, in forward features_fpn = self.fpn([features[1], features[2], features[3]]) File "C:\ProgramData\Anaconda3\envs\crowd\lib\site-packages\torch\nn\modules\module.py", line 550, in call result = self.forward(*input, **kwargs) File "D:\working_space_HJ\CrowdCounting-P2PNet-main\models\p2pnet.py", line 183, in forward P4_x = P5_upsampled_x + P4_x RuntimeError: The size of tensor a (134) must match the size of tensor b (135) at non-singleton dimension 2

please help me and I'm really appreciate it if anyone could answer these questions. Thanks very much. ^ ^

Can anyone reproduce the accuracy of UCF-QNRF? I hava conducted acout ten experiments on UCF-QNRF, however, the MAE and RMSE is 100 and 180. I can not achieve the reported MAE:85 and RMSE:154.5, even I use the settings in this paper.