Are you sure you want to delete this task? Once this task is deleted, it cannot be recovered.
You can not select more than 25 topics
Topics must start with a chinese character,a letter or number, can include dashes ('-') and can be up to 35 characters long.
loyolh
3abacf6f40
|
2 years ago | |
|---|---|---|
| ascend310_infer | 2 years ago | |
| configs | 2 years ago | |
| model_utils | 2 years ago | |
| scripts | 2 years ago | |
| src | 2 years ago | |
| README.md | 2 years ago | |
| eval.py | 2 years ago | |
| export.py | 2 years ago | |
| postprocess.py | 2 years ago | |
| requirements.txt | 2 years ago | |
| train.py | 2 years ago | |
README.md
Contents
- Contents
Model Description
ReID Strong Baseline proposes a novel neck structure named as batch normalization neck (BNNeck).
BNNeck adds a batch normalization layer after global pooling layer to separate metric and classification
losses into two different feature spaces because we observe they are inconsistent in one embedding space.
Extended experiments show that BNNeck can boost the baseline.
Paper: Luo H., Jiang W., et al. “A Strong Baseline and Batch Normalization Neck for Deep Person Re-identification”. CVPRW2019, Oral.
Model Architecture
Model uses ResNet50 as backbone. BNNeck adds a BN layer after features and before classifier FC layers.
The BN and FC layers are initialized through Kaiming initialization.
Dataset
Market1501
Market1501 dataset is used to train and test model.
Market1501 contains 32,668 images of 1,501 labeled persons of six camera views.
There are 751 identities in the training set and 750 identities in the testing set.
In the original study on this proposed dataset, the author also uses mAP as the evaluation criteria to test the algorithms.
Data structure:
Market-1501-v15.09.15
├── bounding_box_test [19733 entries]
├── bounding_box_train [12937 entries]
├── gt_bbox [25260 entries]
├── gt_query [6736 entries]
├── query [3369 entries]
└── readme.txt
DukeMTMC-reID
DukeMTMC-reID is a new large-scale person ReID dataset and collects 36,411 images from 1,404 identities of eight camera views.
The training set has 16,522 images from 702 identities, and the testing set has 19,889 images from other 702 identities
Data structure:
DukeMTMC-reID
├── bounding_box_test [17661 entries]
├── bounding_box_train [16522 entries]
├── CITATION.txt
├── LICENSE_DukeMTMC-reID.txt
├── LICENSE_DukeMTMC.txt
├── query [2228 entries]
└── README.md
Environment Requirements
- Hardware（Ascend/GPU）
- Prepare hardware environment with Ascend or GPU processor.
- Framework
- For more information, please check the resources below：
Quick Start
Model uses pre-trained backbone ResNet50 trained on ImageNet2012. Link
After dataset preparation, you can start training and evaluation as follows:
(Note that you must specify dataset path in configs/market1501_config.yml)
Running on Ascend
# run distributed training example
bash scripts/run_distribute_train_ascend.sh ./configs/market1501_config.yml 8 /path/to/dataset/ /path/to/output/ /path/to/pretrained_resnet50.ckpt
# run evaluation example
bash scripts/run_eval_ascend.sh ./configs/market1501_config.yml /your/path/checkpoint_file /path/to/dataset/
Running on GPU
# run training example
bash scripts/run_standalone_train_gpu.sh ./configs/market1501_config.yml 0 /path/to/dataset/ /path/to/output/ /path/to/pretrained_resnet50.pth
# run distributed training example
bash scripts/run_distribute_train_gpu.sh ./configs/market1501_config.yml 8 /path/to/dataset/ /path/to/output/ /path/to/pretrained_resnet50.pth
# run evaluation example
bash scripts/run_eval_gpu.sh ./configs/market1501_config.yml /your/path/checkpoint_file /path/to/dataset/
Script Description
Script and Sample Code
ReIDStrongBaseline
├── ascend310_infer # application for 310 inference
├── configs
│ ├── dukemtmc_config.yml # Training/evaluation config on DukeMTMC dataset
│ └── market1501_config.yml # Training/evaluation config on Market1501 dataset
├── model_utils
│ ├── __init__.py
│ ├── config.py # Config parser
│ ├── device_adapter.py # Device adapter for ModelArts
│ ├── local_adapter.py # Environment variables parser
│ └── moxing_adapter.py # Moxing adapter for ModelArts
├── scripts
│ ├── run_distribute_train_Ascend.sh # Start multi Ascend training
│ ├── run_distribute_train_gpu.sh # Start multi GPU training
│ ├── run_eval_Ascend.sh # Start single Ascend evaluation
│ ├── run_eval_gpu.sh # Start single GPU evaluation
│ ├── run_infer_310.sh # Start 310 inference
│ └── run_standalone_train_gpu.sh # Start single GPU training
├── src
│ ├── callbacks.py # Logging to file callbacks
│ ├── center_loss.py # Center Loss definition
│ ├── dataset.py # Data preprocessing
│ ├── datasets
│ │ ├── __init__.py
│ │ ├── bases.py # Base dataset loader
│ │ ├── dukemtmcreid.py # DukeMTMC dataset loader
│ │ └── market1501.py # Market1501 dataset loader
│ ├── __init__.py
│ ├── loss.py # Losses definition
│ ├── lr_schedule.py # Learning rate scheduler
│ ├── metric_utils.py # Evaluation functions
│ ├── model
│ │ ├── __init__.py
│ │ ├── cell_wrapper.py # Model wrappers
│ │ ├── resnet.py # ResNet 50 network structure
│ │ └── strong_reid.py # ReID Strong Baseline network structure
│ ├── sampler.py # ReID sampler definition
│ └── triplet_loss.py # Triplet Loss definition
├── eval.py # Evaluate the network
├── export.py # Export the network
├── postprogress.py # post process for 310 inference
├── train.py # Train the network
├── requirements.txt # Required libraries
└── README.md
Script Parameters
Training Script Parameters
usage: train.py --config_path CONFIG_PATH [--distribute DISTRIBUTE] [--device_target DEVICE]
[--max_epoch N] [--start_decay_epoch EPOCH]
[--ids_per_batch IDS_NUM] [--ims_per_id IMS_NUM]
[--is_save_on_master SAVE_FLAG] [--is_print_on_master PRINT_FLAG]
[--pre_trained MODEL_PATH] [--pre_trained_backbone BACKBONE_PATH] [--data_dir DATA_DIR]
[--ckpt_path SAVE_CHECKPOINT_PATH] [--train_log_path SAVE_LOGS_PATH]
[--keep_checkpoint_max CKPT_NUM] [--ckpt_interval CKPT_STEP]
[--log_interval LOG_STEP]
options:
--config_path path to .yml config file
--distribute pre_training by several devices: "true"(training by more than 1 device) | "false", default is "false"
--device_target target device ("Ascend" | "GPU" | "CPU")
--max_epoch epoch size: N, default is 120
--start_decay_epoch epoch to decay, default is '40,70'
--ids_per_batch number of person in batch, default is 16 (8 for distributed)
--ims_per_id number of images for person in epoch
--is_save_on_master save checkpoint only from main thread for distributed
--is_print_on_master print loss logs only from main thread for distributed
--pre_trained path to pretrained model checkpoint
--pre_trained_backbone path to resnet50 backbone model checkpoint
--data_dir path to dateset images
--ckpt_path path to save checkpoint files: PATH
--train_log_path path to save loss logs
--keep_checkpoint_max number for saving checkpoint files: N, default is 2
--ckpt_interval number of epochs between checkpoints, default 1
--log_interval logging batch interval, default 32
Running Options
default_config.yaml:
weight_decay optimizer weight decay
global_loss_margin margin for global MarginRankingLoss
local_loss_margin margin for local MarginRankingLoss
center_loss_weight weight of center loss
crossentropy_loss_weight weight of identity loss
Network Parameters
Parameters for dataset and network (Training/Evaluation):
image_size size of input image
image_mean image mean value for normalization
image_std image std value for normalization
padding image padding before random crop
Parameters for learning rate:
lr_init initial network learning rate
lr_cri center loss learning rate
warmup_factor warmup scale
warmup_epoch warmup number of epochs
Training Process
- Set options in
configs/market1501_config.yamlorconfigs/dukemtmc_config.yaml,
including paths, learning rate and network hyperparameters.
Usage
on Ascend
-
Run
run_distribute_train_Ascend.shfor distributed training of ReID Strong Baseline model.bash scripts/run_distribute_train_Ascend.sh CONFIG_PATH DEVICE_NUM DATA_DIR OUTPUT_PATH PRETRAINED_RESNET50
on GPU
-
Run
run_standalone_train_gpu.shfor non-distributed training of model.bash scripts/run_standalone_train_gpu.sh CONFIG_PATH DEVICE_ID DATA_DIR OUTPUT_PATH PRETRAINED_RESNET50 -
Run
run_distribute_train_gpu.shfor distributed training of ReID Strong Baseline model.bash scripts/run_distribute_train_gpu.sh CONFIG_PATH DEVICE_NUM DATA_DIR OUTPUT_PATH PRETRAINED_RESNET50
Evaluation Process
- Set options in
market1501_config.yaml.
Usage
on Ascend
-
Run
bash scripts/run_eval_Ascend.shfor evaluation of ReID Strong Baseline model.bash scripts/run_eval_Ascend.sh CONFIG_PATH CKPT_PATH DATA_DIR
on GPU
-
Run
bash scripts/run_eval_gpu.shfor evaluation of ReID Strong Baseline model.bash scripts/run_eval_gpu.sh CONFIG_PATH CKPT_PATH DATA_DIR
Inference Process
Export MindIR
python export.py --config_path [CONFIG_PATH] --ckpt_file [CKPT_PATH] --file_name [FILE_NAME] --file_format [FILE_FORMAT]
options:
--config_path path to .yml config file
--ckpt_file checkpoint file
--file_name output file name
--file_format output file format, choices in ['MINDIR']
The ckpt_file and config_path parameters are required,
FILE_FORMAT should be in "MINDIR"
result
Inference result will be shown in the terminal
Model Description
Performance
Market1501 Training Performance
| Parameters | Ascend | GPU |
|---|---|---|
| Resource | 8x Ascend 910 32G | 8x Tesla V100-PCIE 32G |
| uploaded Date | 04/21/2022 (month/day/year) | 03/11/2022 (month/day/year) |
| MindSpore Version | 1.3.0 | 1.5.0 |
| Dataset | Market1501 | Market1501 |
| Training Parameters | max_epoch=120, ids_per_batch=8, start_decay_epoch=151, lr_init=0.0014, lr_cri=1.0, decay_epochs='40,70' | max_epoch=120, ids_per_batch=8, start_decay_epoch=151, lr_init=0.0014, lr_cri=1.0, decay_epochs='40,70' |
| Optimizer | Adam, SGD | Adam, SGD |
| Loss Function | Triplet, Smooth Identity, Center | Triplet, Smooth Identity, Center |
| Speed | 536ms/step (8pcs) | 182ms/step (8pcs) |
| Loss | 0.28 | 0.24 |
| Params (M) | 24.8 | 25.1 |
| Checkpoint for inference | 305Mb (.ckpt file) | 319Mb (.ckpt file) |
| Scripts | ReID Strong Baseline scripts | ReID Strong Baseline scripts |
Market1501 Evaluation Performance
| Parameters | Ascend | GPU |
|---|---|---|
| Resource | 1x Ascend 910 32G | 1x Tesla V100-PCIE 32G |
| Uploaded Date | 04/21/2022 (month/day/year) | 03/11/2022 (month/day/year) |
| MindSpore Version | 1.3.0 | 1.5.0 |
| Dataset | Market1501 | Market1501 |
| batch_size | 32 | 32 |
| outputs | mAP, Rank-1 | mAP, Rank-1 |
| Accuracy | mAP: 86.85%, rank-1: 94.36% | mAP: 86.99%, rank-1: 94.48% |
DukeMTMC-reID Training Performance
| Parameters | Ascend | GPU |
|---|---|---|
| Resource | 8x Ascend 910 32G | 8x Tesla V100-PCIE 32G |
| uploaded Date | 04/21/2022 (month/day/year) | 03/11/2022 (month/day/year) |
| MindSpore Version | 1.3.0 | 1.5.0 |
| Dataset | DukeMTMC-reID | DukeMTMC-reID |
| Training Parameters | max_epoch=120, ids_per_batch=8, start_decay_epoch=151, lr_init=0.0014, lr_cri=1.0, decay_epochs='40,70' | max_epoch=120, ids_per_batch=8, start_decay_epoch=151, lr_init=0.0014, lr_cri=1.0, decay_epochs='40,70' |
| Optimizer | Adam, SGD | Adam, SGD |
| Loss Function | Triplet, Smooth Identity, Center | Triplet, Smooth Identity, Center |
| Speed | 524ms/step (8pcs) | 180ms/step (8pcs) |
| Loss | 0.27 | 0.24 |
| Params (M) | 24.8 | 25.1 |
| Checkpoint for inference | 302Mb (.ckpt file) | 319Mb (.ckpt file) |
| Scripts | ReID Strong Baseline scripts | ReID Strong Baseline scripts |
DukeMTMC-reID Evaluation Performance
| Parameters | Ascend | GPU |
|---|---|---|
| Resource | 1x Ascend 910 32G | 1x Tesla V100-PCIE 32G |
| Uploaded Date | 04/21/2022 (month/day/year) | 03/11/2022 (month/day/year) |
| MindSpore Version | 1.3.0 | 1.5.0 |
| Dataset | DukeMTMC-reID | DukeMTMC-reID |
| batch_size | 32 | 32 |
| outputs | mAP, Rank-1 | mAP, Rank-1 |
| Accuracy | mAP: 76.58%, rank-1: 87.43% | mAP: 76.68%, rank-1: 87.34% |
Description of Random Situation
There are six random situations:
- Shuffle of the persons in the dataset.
- Select random images for persons.
- Random flip images.
- Random crop images.
- Random erasing images.
- Initialization of some model weights.
Some seeds have already been set in train.py and sampler.py to avoid the randomness of dataset shuffle and weight initialization.
ModelZoo Homepage
Please check the official homepage.
本研究提出了一个简单而有力的深层人再识别(ReID)baseline。在文献中，一些有效的训练技巧简要地出现在几篇论文或源代码中。本研究收集并评估了这些有效的训练技巧。同时我们也提出了一种新的瓶颈结构，称为批归一化瓶颈(BNNeck)。BNNeck在global pooling layer之后增加了一个batch normalization neck，将度量和分类损失分离到两个不同的特征空间中。
Python Shell C++ Text other
Contributors (1)