coincheung / triplet-reid-pytorch Goto Github PK

Hi,
I used your code and changed the stride to 2 for faster computation.
I ran the training three times and got the following results in different runs:
Run1: mAP 74.2, cmc1: 86.1
Run2: mAP 64.7 cmc1: 80.5
Run3: mAP 75.2 cmc1: 87.5

As you see, in the second run the map and cmc1 dropped by large amount. I am not sure what is the reason for this.
Do you think this is because of the random erasing being used?
Have you also observed a large variation in results across different trainings?

Regards

I use the batch sampler in this repo, but the data fetch speed is quite slow, which is ~250s for 50 batches. The batch size is 256, P=32, K=8, use 4 workers and 4 GPUs. Have you encountered with this problem?

Hi CoinCheung,

Thank you for you code. I have tested it on the market1501 dataset, the result is lower than the readme result:

mAP is: 0.525827546826, cmc is: [0.7185273]

Can you tell the what's wrong with it?

for it, (img, lb_id) in enumerate(dl):
    print('\r=======>  processing iter {} / {}'.format(it, all_iter_nums),
            end = '', flush = True)
    label_ids.append(lb_id)
    #label_cams.append(lb_cam)
    embds = []
    for im in img:
        im = im.cuda()
        embd = model(im).detach().cpu().numpy() #通过网络生成特征向量
        embds.append(embd)
    embed = sum(embds) / len(embds)
    embeddings.append(embed)

请问代码为什么每次要把一个ｂａｔｃｈ的特征求和除去长度

Hi ! Thank you for your tricks, I would like to know what do you mean by saying "adjust the stride of the last stage of resnet from 2 to 1." Please let me know thank you