I got worse performance in triplet_loss.py

[Niculuse]

I ran the triplet_loss.py on market1501 as told in benchmark, but I got worse result that every item is about 5% lower than presented in bemchmark. BTW, I didn't change code. I do not understand why.Can you help me? Thanks!

[zydou]

Please provide as much information as possible. For example:

• your platform?
• your pytorch version?
• open-reid version?

[Niculuse]

@zydou sorry! I ran it in Ubuntu system with python 2.7.6, pytorch version 0.2.0_3 and open-reid version 0.2.0.

[zydou]

Try to use python3.

[xizero00]

Hi @zydou I changed my python version from 2.7 to 3.5 But I still cannot achieve the best results in the benchmark table The triplet loss on the market-1501 has 67.9 mAP, but I got the 60.5 mAP. Here is my results on the test set

Mean AP: 60.5%
CMC Scores    allshots      cuhk03  market1501
  top-1          33.2%       62.9%       78.2%
  top-5          50.4%       88.0%       91.2%
  top-10         59.1%       93.3%       94.4%

I trained the model for market-1501 dataset using the following parameters:

batchsize=64
lr=0.00005
epochs=600

I have one GPU, so I modified the batchsize from 256 to 64, changed the learning rate from 0.0002 to 0.00005 and changed the epochs from 150 to 600 according to tips-and-tricks I used the triplet loss and resnet-50 network. What's wrong with my training? Would you please help me?

[zydou]

Larger batch size is better for hard mining, and gives better performance. Please refer to #28 and #43 for more information.

[xizero00]

Thank you for your help.

[Cysu]

@zydou thanks a lot for your detailed explanations!