Thanks for your work to implement BMN on PyTorch!

May I ask, what is the model performance that you get?

On my side, the AUC is 64.5% and AR@100 is 73.5. Please find the details below.

I did not change any parameters before running the code. Minor typos were fixed.

+ python main.py --module Evaluation                                                                                                                                        
[INIT] Loaded annotations from validation subset.                                                                                                     
        Number of ground truth instances: 7293                                                                                                                              
        Number of proposals: 472628                                                                                                                   
        Fixed threshold for tiou score: [0.5  0.55 0.6  0.65 0.7  0.75 0.8  0.85 0.9  0.95]                                                                                 
[RESULTS] Performance on ActivityNet proposal task.                                                                                                   
        Area Under the AR vs AN curve: 64.50481283422461%                                                                                                                   
/home/xum/anaconda3/envs/GTAL/lib/python3.7/site-packages/matplotlib/figure.py:98: MatplotlibDeprecationWarning:                                      
Adding an axes using the same arguments as a previous axes currently reuses the earlier instance.  In a future version,                                                     a new instance will always be created and returned.  Meanwhile, this warning can be suppressed, and the future behavior                               ensured, by passing a unique label to each axes instance.                                                                                                                   
  "Adding an axes using the same arguments as a previous axes "                                                                                       
AR@1 is          0.30287947346770877                                                                                                                                        
AR@5 is          0.43831070889894425                                                                                                                  
AR@10 is         0.5157136980666392                                                                                                                                         
AR@100 is        0.7352392705333882

Since the training used fixed length of the feature's temporal dimension (~100), is that can be used to unfixed the length of the temporal dimension? and how? thanks a lot.

Thank you for the great work!

I tried your training program, and printed the loss after each batch. The loss seems correct in the beginning of the training, then "nan" values appear as shown in the following. Wonder if you guessed the possible reasons. ... BMN training loss(epoch 0): tem_loss: 1.184, pem class_loss: 0.458, pem reg_loss: 0.024, total_loss: 1.883 BMN training loss(epoch 0): tem_loss: 1.184, pem class_loss: 0.458, pem reg_loss: 0.024, total_loss: 1.884 BMN training loss(epoch 0): tem_loss: 1.184, pem class_loss: nan, pem reg_loss: 0.024, total_loss: nan BMN training loss(epoch 0): tem_loss: nan, pem class_loss: nan, pem reg_loss: 0.024, total_loss: nan

Thanks!

i download the BMN_best.pth.tar from the BaiduYun(pan.baidu.com) link u provided, for 2 times, both cannot be decompressed.

>tar -xvf BMN_best.pth.tar
tar: Error opening archive: Unrecognized archive format

Could u pls. help to check if the compressed file corrupted?

Hi, I am a graduate student who is interseted in TAD and hope more communication. First, thank you for the implementation of BMN. I find BMN model's structure is not the same as introduced in the paper, so I checked the source code by Tianwei Lin. I think the output dimension of x_1d_p and input dimension of x_3d_p may be changed to 128(hidden_dim_2d) as follows: # Proposal Evaluation Module self.x_1d_p = nn.Sequential( nn.Conv1d(self.hidden_dim_1d, self.hidden_dim_2d, kernel_size=3, padding=1), nn.ReLU(inplace=True) ) self.x_3d_p = nn.Sequential( nn.Conv3d(self.hidden_dim_2d, self.hidden_dim_3d, kernel_size=(self.num_sample, 1, 1),stride=(self.num_sample, 1, 1)), nn.ReLU(inplace=True) ) Lin's source code in this link: https://github.com/PaddlePaddle/models/blob/release/1.8/PaddleCV/video/models/bmn/bmn_net.py#L141-L149

Best wishes!

Maybe Chinese characters need to be handled...

$ sh bmn.sh
  File "main.py", line 136
SyntaxError: Non-ASCII character '\xe9' in file main.py on line 136, but no encoding declared; see http://python.org/dev/peps/pep-0263/ for details
  File "main.py", line 136
SyntaxError: Non-ASCII character '\xe9' in file main.py on line 136, but no encoding declared; see http://python.org/dev/peps/pep-0263/ for details

Thanks a lot for this code! I simply run the script to train BMN model. However I got PEM loss=NaN , and I find that label confidence is all 0. What 's wrong there?

In default BMN, the temporal scale is 100. if I want to generate 200*200 proposal map by feature with 200-length, need I change some hyperparameters ?

For instance, need I change the positive threshold of cls_pem_loss or three thresholds of reg_pem_loss ? Is there anything else that needs to be modified

谢谢你的工作，但是在你最新的代码中，opt文件中的"video_info_backup.csv"文件和"anet_anno_action_backup.json"文件好像没有提供，使用的还是之前的label文件？ 另外一个问题是：在训练过程中，你使用了best_loss变量，按照我的理解，这应该是一个局部变量，每一个epoch都会初始化为一个极大的值，而每一次保存都是会保存当前epoch的参数？请你确认一下。

Is each video sampled at different intervals How does the author turn unequal video into equal length? rescaled the feature length of all videos to same length 100,

Hi, thanks for the code. I'd like to try this on my own datasets, but when I was extracting features by using TSN, I can only get RGB.jpg and Flow_x_y.jpg. How can I get csv features which is used in your code? (like csv_mean_100/v_oOnKQgQZOZ0.csv) Or can I use jpg to do training and testing?

Hi！ Thanks for your great work! I'd like to combine your algorithm with temporal action classification,but I have no idea to do it.Would you please give me some advice?

Good afternoon! Thank you very much for your work! It's great that the model is not only on paddlepaddle.

Could you please share model training speed? I'm using a non-standard environment and would like to compare to the "original".

1. What is the meaning of featureframe,is it randomly extract from video then extract feature and reshape to (400,100)?
2. i have generated featurefrom video,and i extract per frame from 200ms,the length of feature is unfixed because of unfixed video,how can i make its length to be fixed ?
3. my dataset is hoi video,so the orginal label is (subject,object,relation),but in this project ,the label is action ,how can i present it?
4. the means of temeporal_scale?

在网络模型的输入中没有发现类别label 信息的载入？太奇怪了，这样自己做一个数据集和修改都不知道从何处 下手。 现在我能把数据集做成 json 文件格式的样子，能够包含起始位置，类型信息等。可以指没看见label类别信息的映射~帮忙指点一下，按照我的理解。比如activitynet 有个200类别，最后肯定有一个分类模块。并且输入输出结果映射指导类别上

Thanks for sharing your code. I'd like to change the number of temporal proposals of the model, but I couldn't find this parameter. Would you please tell me the specific location?