- Python 3.6
- GPU Memory >= 6G
- Numpy
- Pytorch 0.3+
First, run python prepare.py to prepare the dataset. Preparation creates views in each identity. The output
folder should likes
0743--
|--view_0
|--view_1
|--view_3--
|--1479_c3s3_080744_06.jpg
|--1479_c3s3_080694_04To train the network, use the command
python train.py
`--data` '/home/paul/datasets/market1501/multiviews'
`--epochs` 100
`--b 4` (batch size)
`--lr` 0.0001 (Learning rate)
`--momentum` 0.9
`--lr-decay-freq` 30
`--lr-decay` 0.1Most of these parameters are set to default, so you can only run python train.py
To resume or use the the pretrained weight resnet from ImageNet, add
--resume --pretrained
To test a trained model, use:
python test.py
python evaluate_gpu (for a gpu evaluation)
python evaluate.py (run on CPU)
| Datasets | Rank 1 | Rank5 | Rank10 | Rank20 | mAP |
|---|---|---|---|---|---|
| Market-1501 | - | - | - | - | |
| DukeMTMC-ReID | |||||
| CUHK03 |