search

TuSimple / rl-multishot-reid

Multi-shot Pedestrian Re-identification via Sequential Decision Making (CVPR2018)
93 stars 26 forks source link

how to prepare the training samples for RL training #3

Open likelyzhao opened 6 years ago

likelyzhao commented 6 years ago

I have tried training the RL phase model with the baseline training samples(recs) , but it seems that the training sample preparing is different between RL training and baseline, could you share the data preprocessing scripts for RL training thank you

SMCamila commented 6 years ago

Please refer RL/batch_provider.py, we random select (id, cam) sequence pairs and then random select frames from these two sequences.

likelyzhao commented 6 years ago

I see it, so what's the format of input params of batch_provider? it seems to need a recordio file with labels and an lst file parsed from a txt file. I have no idea about the contents of these files, could you give me some more details? thank you

SMCamila commented 6 years ago

The first term input of batch_provider is a image record iterator, the second term is an lst file which is the same file to the lst file provided to the first term. I think this part is similar to the setting in baseline program. Tell me if it is still not clear. Thank you.

likelyzhao commented 6 years ago

what‘s in this txt file which used in valid_lst = np.loadtxt('%s/recs/%s.txt'%(args.data_dir, args.valid_set)).astype(int) ? I used the lst file provided to the first term but dose not work .