Why is it that mAP is basically 7%-8% when I run your code,how can I achieve your best results?

[xiaonvxia]

I'm sorry to bother you.I have run your code several times, mAP only reaches 8%, and top-1 only reaches 20% at most. Otherwise,the results are different when each time I run the code. How can I achieve the effect in your paper and keep the results at the same?

[Yu-Wu]

Hi, I cannot reproduce this issue.

Can you please paste all the log shown in the screen?

[Yu-Wu]

Can you please try to use my original code, without any personal changes?

[xiaonvxia]

Thank you for your reply! I did run your source code without any changes，the final result of the last run is as follows: Extract 76 batch videos Mean AP: 8.2% CMC Scores: top-1 20.2% top-5 32.2% top-10 38.8% top-20 46.2% create dataloader for Test with batch_size 256 2%|██▎ | 1/48 [00:01<01:04, 1.37s/i 4%|████▋ | 2/48 [00:02<01:03, 1.37s 6%|███████ | 3/48 [00:04<01:02, 1.3 8%|█████████▎ | 4/48 [00:05<01:01, 10%|███████████▋ | 5/48 [00:07<01:00, 12%|██████████████ | 6/48 [00:08<01:0 15%|████████████████▎ | 7/48 [00:09<0 17%|██████████████████▋ | 8/48 [00:11 19%|█████████████████████ | 9/48 [00: 21%|███████████████████████▏ | 10/48 [ 23%|█████████████████████████▍ | 11/48 25%|███████████████████████████▊ | 12/ 27%|██████████████████████████████ | 1 29%|████████████████████████████████▍ 31%|██████████████████████████████████▋ 33%|█████████████████████████████████████ 35%|███████████████████████████████████████▎ 38%|█████████████████████████████████████████▋ 40%|███████████████████████████████████████████▉ 42%|██████████████████████████████████████████████▎ 44%|████████████████████████████████████████████████▌ 46%|██████████████████████████████████████████████████▉ 48%|█████████████████████████████████████████████████████▏ 50%|███████████████████████████████████████████████████████▌ 52%|█████████████████████████████████████████████████████████▊ 54%|████████████████████████████████████████████████████████████ 56%|██████████████████████████████████████████████████████████████▍ 58%|████████████████████████████████████████████████████████████████▊ 60%|███████████████████████████████████████████████████████████████████ 62%|█████████████████████████████████████████████████████████████████████▍ 65%|███████████████████████████████████████████████████████████████████████▋ 67%|████████████████████████████████████████████████████████████████████████ 69%|████████████████████████████████████████████████████████████████████████ 71%|████████████████████████████████████████████████████████████████████████ 73%|████████████████████████████████████████████████████████████████████████ 75%|████████████████████████████████████████████████████████████████████████ 77%|████████████████████████████████████████████████████████████████████████ 79%|████████████████████████████████████████████████████████████████████████ 81%|████████████████████████████████████████████████████████████████████████ 83%|████████████████████████████████████████████████████████████████████████ 85%|████████████████████████████████████████████████████████████████████████ 88%|████████████████████████████████████████████████████████████████████████ 90%|████████████████████████████████████████████████████████████████████████ 92%|████████████████████████████████████████████████████████████████████████ 94%|████████████████████████████████████████████████████████████████████████ 96%|████████████████████████████████████████████████████████████████████████ 98%|████████████████████████████████████████████████████████████████████████ 100%|████████████████████████████████████████████████████████████████████████ 100%|████████████████████████████████████████████████████████████████████████ ███████████████████████████████████████| 48/48 [01:06<00:00, 1.39s/it] Extract 48 batch videos create dataloader for Test with batch_size 256 33%|█████████████████████████████████████▋ 67%|████████████████████████████████████████████████████████████████████████ 100%|████████████████████████████████████████████████████████████████████████ 100%|████████████████████████████████████████████████████████████████████████ █████████████████████████████████████████| 3/3 [00:04<00:00, 1.40s/it] Extract 3 batch videos u_features (12185, 2048) l_features (751, 2048) Label predictions on all the unlabeled data: 2579 of 12185 is correct, accuracy = 0.212 selected pseudo-labeled data: 2513 of 10966 is correct, accuracy: 0.2292 new train data: 11717 Unselected Data:1219

[Yu-Wu]

Could you upload the full log.txt file?

[xiaonvxia]

Because the result of the code was so long, the terminal only kept the end of the log records, I copied all the logs that were displayed. log.txt

[Yu-Wu]

Hi, my code kept the log record in the log dir. Please find it and upload here.

Since the framework is in a progressive way, I should see the history log to check and find the reason.

[xiaonvxia]

Is the files in the logs folder?I'll upload here

[xiaonvxia]

This line in your run file sys.stdout = Logger(osp.join(args.logsdir, 'log'+ str(args.EF)+ time.strftime(".%m%d_%H:%M:%S") + '.txt')) Is this line of code used to keep a log of the results of the run?Because there was an error in this line of code when I ran the code, I logged it out. Now there is only .ckpt file in my logs folder

[Yu-Wu]

Yes, you're correct. This is exact the line that stores the log information to a file.

[Yu-Wu]

I cannot find the reason given the very few log information. Each training model is based on that of the previous one.

[xiaonvxia]

I'll run the program again, with this line of code added and reply to you.There is another problem: my code runs on CPU, but my computer has a single GPU. I don't know why the code doesn't run on GPU？

[Yu-Wu]

Please pay attention to the CUDA environment in your machines

[xiaonvxia]

I noticed your line of code model = nn.DataParallel(model).cuda() however, my computer is single GPU, does this affect the use of GPU?

[Yu-Wu]

model = nn.DataParallel(model).cuda() should be work even on a single GPU.

Could you please try the basic PyTorch examples and test whether it runs on GPU?

[xiaonvxia]

All right,I will have a try.Thank you for your patient explaination

[xiaonvxia]

When I was running the code today, I found that I had set the batch-size to 4, and the result I ran after I changed it to 16 reached the result in your paper.

[Yu-Wu]

Good to hear that you could reproduce our results.

The batch size is very important here, batch size =4 cannot lead to a stable Stochastic gradient descent training.

[xiaonvxia]

Indeed, I didn't expect that batchsize would have such an impact on the experimental results. So the bigger the batchsize, the better the result?

[xiaonvxia]

I use the following line to verify that pytorch is using GPU: torch.cuda.is_available() The returned result is true, but my GPU utilization rate is only 3%. Is it because most of the code operations are performed on the CPU, and GPU only undertakes a small amount of computing tasks?

[Yu-Wu]

Maybe not. The batch size is usually kept as 16 or 32 on the person re-id task. Too large batch size will definitely downgrade the performance, especially on the one-example training (which only has hundreds of data).

[Yu-Wu]

I use the following line to verify that pytorch is using GPU: torch.cuda.is_available() The returned result is true, but my GPU utilization rate is only 3%. Is it because most of the code operations are performed on the CPU, and GPU only undertakes a small amount of computing tasks?

In the training stage, the GPU utilization rate should be very high (close to 90%, maybe). In the test and pseudo labeling stage, there are lots of distance calculation. So at this stage, the most computation is done on CPU.

[xiaonvxia]

Thank you very much for your answer. I'll find out the reason for the low utilization rate on GPU