# compute average precision 计算平均精度 # reference: https://en.wikipedia.org/wiki/Evaluation_measures_(information_retrieval)#Average_precision num_rel = orig_cmc.sum() # 所有元素求和 tmp_cmc = orig_cmc.cumsum() # 累加和,不改变数据形状 tmp_cmc = [x / (i+1.) for i, x in enumerate(tmp_cmc)] # enumerate() 函数,返回数据下标和数据(i,x) # 计算top_i的cmc ..x / (i+1.) tmp_cmc = np.asarray(tmp_cmc) * orig_cmc # np.asarray(tmp_cmc),数据类型转换为数组, 只保留正确的匹配 , 错误的值为0 AP = tmp_cmc.sum() / num_rel # 平均精度 = 正确匹配的元素求和 除以原来所有元素的总和 all_AP.append(AP)

你好，有个问题请教一下。代码里的 AP的计算公式对应的是 P=TP/TP+NP吧，感觉怎么像是在计算精确度，不是用PR曲线的面积来求解的。能麻烦您讲解一下吗？

在ResNet.py文件中，有如下注释：

我阅读过代码后，觉得这个传入的x维度应该是[batch_size,seq_len,channels,height,width]
这样的话，下面这句代码才合理：
x = x.view(bt, x.size(2), x.size(3), x.size(4))
否则，没有理由将x resize成bt，batch_size*channels，很莫名其妙

Hi, when i run the main_video_person_reid.py at the Epoch 50 steps run test. I got an error --------> AttributeError: 'tuple' object has no attribute 'view' at the code features = features.view(n, -1).
Can you tell me how to fix this error.
Thank you

elif self.sample == 'dense': """ Sample all frames in a video into a list of clips, each clip contains seq_len frames, batch_size needs to be set to 1. 将视频中的所有帧采样到一系列的clips，每个clip包含seq_len个帧,批次大小需要设置为1 This sampling strategy is used in test phase. 在测试阶段采用密集采样策略. """

你好，我想问一下 dense的取样方式 是不是类似于 把 整个tracklet 划分成 多个 seq_length长的clip ？

你好，我想问一下，在main文件中设定50个epoch进行一次test， 是相当于使用了整个测试集的数据了吗？ 在训练完成后，进行测试的时候，测试的数据集用还是用原来的query_idx 吗？ 谢谢

Hi, I run demo of jiyanggao/Video-Person-ReID in Mars dataset, I got confused how to split a eval_set from train_set , I want to test the model on the eval set to save time. Thx

Hi, I learn from you project, and I can see the steps is:
1-----> I download the Mars data,
2-----> run the main_video_person_reid.py with pre_trained model----resnet-50-kinetic.pth
3-----> get the trained model to run test.py
of course change some path in the code.

Am I right?

你好，我自己制作了一个测试集，但是在测试的时候，imgs 的shape 为（26 , 4 , 3 , 224 , 112)，而visdom.images要求输入为（channel， height，weight） 请问可以怎么转换？ 谢谢。

Hi I am trying to understand different methodologies and architecture to train video reid datasets. My question is that can you please tell me what are the best hyperparameters combination, i.e. learning rate, margin for triplet loss, batch size train, batch size test, sequence length for video reid dataset training etc.

Also will you please tell me that in one of your questions asked in the TCLNet repository, you asked that while training for the mars dataset, your result was not similar to those in the research paper. The author said that try to use all frames while testing instead of 4 (the default value) will you please tell me what does it actually meant by this? Which value should be replaced instead of 4 in test_frames of argument parser?

I have also attached a screenshot for the reference

  # Compute pairwise distance, replace by the official when merged
        dist = torch.pow(inputs, 2).sum(dim=1, keepdim=True).expand(n, n)
        dist = dist + dist.t()
        dist.addmm_(1, -2, inputs, inputs.t())
        dist = dist.clamp(min=1e-12).sqrt()  # for numerical stability

你好，在损失函数的实现部分，TripletLoss(nn.Module)里面的上述代码我不太理解，能麻烦你解读一下吗？ 非常非常感谢。