p04896132
/
mindspore
forked from mindspore-Ecosystem/mindspore
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

!5133 modify readme.md for yolov3_darknet53 in r0.7 

Merge pull request !5133 from yangyongjie/r0.7
This commit is contained in:
mindspore-ci-bot 2020-08-26 09:27:25 +08:00 committed by Gitee
parent 9e20e17590 32e7eb9148
commit 6017521219
2 changed files with 288 additions and 84 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

304
model_zoo/official/cv/yolov3_darknet53/README.md
View File

@ -1,35 +1,118 @@
# YOLOV3-DarkNet53 Example
# Contents
## Description
- [YOLOv3-DarkNet53 Description](#yolov3-darknet53-description)
- [Model Architecture](#model-architecture)
- [Dataset](#dataset)
- [Environment Requirements](#environment-requirements)
- [Quick Start](#quick-start)
- [Script Description](#script-description)
- [Script and Sample Code](#script-and-sample-code)
- [Script Parameters](#script-parameters)
- [Training Process](#training-process)
- [Training](#training)
- [Distributed Training](#distributed-training)
- [Evaluation Process](#evaluation-process)
- [Evaluation](#evaluation)
- [Model Description](#model-description)
- [Performance](#performance)
- [Evaluation Performance](#evaluation-performance)
- [Inference Performance](#evaluation-performance)
- [Description of Random Situation](#description-of-random-situation)
- [ModelZoo Homepage](#modelzoo-homepage)
This is an example of training YOLOV3-DarkNet53 with COCO2014 dataset in MindSpore.
## Requirements
# [YOLOv3-DarkNet53 Description](#contents)
- Install [MindSpore](https://www.mindspore.cn/install/en).
You only look once (YOLO) is a state-of-the-art, real-time object detection system. YOLOv3 is extremely fast and accurate.
- Download the dataset COCO2014.
Prior detection systems repurpose classifiers or localizers to perform detection. They apply the model to an image at multiple locations and scales. High scoring regions of the image are considered detections.
YOLOv3 use a totally different approach. It apply a single neural network to the full image. This network divides the image into regions and predicts bounding boxes and probabilities for each region. These bounding boxes are weighted by the predicted probabilities.
> Unzip the COCO2014 dataset to any path you want, the folder should include train and eval dataset as follows:
YOLOv3 uses a few tricks to improve training and increase performance, including: multi-scale predictions, a better backbone classifier, and more. The full details are in the paper!
[Paper](https://pjreddie.com/media/files/papers/YOLOv3.pdf): YOLOv3: An Incremental Improvement. Joseph Redmon, Ali Farhadi,
University of Washington
# [Model Architecture](#contents)
YOLOv3 use DarkNet53 for performing feature extraction, which is a hybrid approach between the network used in YOLOv2, Darknet-19, and that newfangled residual network stuff. DarkNet53 uses successive 3 × 3 and 1 × 1 convolutional layers and has some shortcut connections as well and is significantly larger. It has 53 convolutional layers.
# [Dataset](#contents)
Dataset used: [COCO2014](https://cocodataset.org/#download)
- Dataset size: 19G, 123,287 images, 80 object categories.
- Train：13G, 82,783 images
- Val：6GM, 40,504 images
- Annotations: 241M, Train/Val annotations
- Data format：zip files
- Note：Data will be processed in yolo_dataset.py, and unzip files before uses it.
# [Environment Requirements](#contents)
- Hardware（Ascend）
- Prepare hardware environment with Ascend processor. If you want to try Ascend , please send the [application form](https://obs-9be7.obs.cn-east-2.myhuaweicloud.com/file/other/Ascend%20Model%20Zoo%E4%BD%93%E9%AA%8C%E8%B5%84%E6%BA%90%E7%94%B3%E8%AF%B7%E8%A1%A8.docx) to ascend@huawei.com. Once approved, you can get the resources.
- Framework
- [MindSpore](http://10.90.67.50/mindspore/archive/20200506/OpenSource/me_vm_x86/)
- For more information, please check the resources below：
- [MindSpore tutorials](https://www.mindspore.cn/tutorial/zh-CN/master/index.html)
- [MindSpore API](https://www.mindspore.cn/api/zh-CN/master/index.html)
# [Quick Start](#contents)
After installing MindSpore via the official website, you can start training and evaluation in Ascend as follows:
```
.
└─dataset
├─train2014
├─val2014
└─annotations
# The darknet53_backbone.ckpt in the follow script is got from darknet53 training like paper.
# The parameter of pretrained_backbone is not necessary.
# The parameter of training_shape define image shape for network, default is "".
# It means use 10 kinds of shape as input shape, or it can be set some kind of shape.
# run training example(1p) by python command.
python train.py \
--data_dir=./dataset/coco2014 \
--pretrained_backbone=darknet53_backbone.ckpt \
--is_distributed=0 \
--lr=0.1 \
--T_max=320 \
--max_epoch=320 \
--warmup_epochs=4 \
--train_shape=416 \
--lr_scheduler=cosine_annealing > log.txt 2>&1 &
# standalone training example(1p) by shell script
sh run_standalone_train.sh dataset/coco2014 darknet53_backbone.ckpt
# distributed training example(8p) by shell script
sh run_distribute_train.sh dataset/coco2014 darknet53_backbone.ckpt rank_table_8p.json
# run evaluation by python command
python eval.py \
--data_dir=./dataset/coco2014 \
--pretrained=yolov3.ckpt \
--testing_shape=416 > log.txt 2>&1 &
# run evaluation by shell script
sh run_eval.sh dataset/coco2014/ checkpoint/0-319_102400.ckpt
```
## Structure
```shell
# [Script Description](#contents)
## [Script and Sample Code](#contents)
```
.
└─yolov3_darknet53
├─README.md
├─scripts
├─run_standalone_train.sh # launch standalone training(1p)
├─run_distribute_train.sh # launch distributed training(8p)
└─run_eval.sh # launch evaluating
├─run_standalone_train.sh # launch standalone training(1p) in ascend
├─run_distribute_train.sh # launch distributed training(8p) in ascend
└─run_eval.sh # launch evaluating in ascend
├─src
├─__init__.py # python init file
├─config.py # parameter configuration
@ -47,35 +130,109 @@ This is an example of training YOLOV3-DarkNet53 with COCO2014 dataset in MindSpo
└─train.py # train net
```
## Running the example
### Train
#### Usage
## [Script Parameters](#contents)
```
# distributed training
sh run_distribute_train.sh [DATASET_PATH] [PRETRAINED_BACKBONE] [RANK_TABLE_FILE]
# standalone training
sh run_standalone_train.sh [DATASET_PATH] [PRETRAINED_BACKBONE]
Major parameters in train.py as follow.
optional arguments:
-h, --help show this help message and exit
--data_dir DATA_DIR Train dataset directory.
--per_batch_size PER_BATCH_SIZE
Batch size for Training. Default: 32.
--pretrained_backbone PRETRAINED_BACKBONE
The ckpt file of DarkNet53. Default: "".
--resume_yolov3 RESUME_YOLOV3
The ckpt file of YOLOv3, which used to fine tune.
Default: ""
--lr_scheduler LR_SCHEDULER
Learning rate scheduler, options: exponential,
cosine_annealing. Default: exponential
--lr LR Learning rate. Default: 0.001
--lr_epochs LR_EPOCHS
Epoch of changing of lr changing, split with ",".
Default: 220,250
--lr_gamma LR_GAMMA Decrease lr by a factor of exponential lr_scheduler.
Default: 0.1
--eta_min ETA_MIN Eta_min in cosine_annealing scheduler. Default: 0
--T_max T_MAX T-max in cosine_annealing scheduler. Default: 320
--max_epoch MAX_EPOCH
Max epoch num to train the model. Default: 320
--warmup_epochs WARMUP_EPOCHS
Warmup epochs. Default: 0
--weight_decay WEIGHT_DECAY
Weight decay factor. Default: 0.0005
--momentum MOMENTUM Momentum. Default: 0.9
--loss_scale LOSS_SCALE
Static loss scale. Default: 1024
--label_smooth LABEL_SMOOTH
Whether to use label smooth in CE. Default:0
--label_smooth_factor LABEL_SMOOTH_FACTOR
Smooth strength of original one-hot. Default: 0.1
--log_interval LOG_INTERVAL
Logging interval steps. Default: 100
--ckpt_path CKPT_PATH
Checkpoint save location. Default: outputs/
--ckpt_interval CKPT_INTERVAL
Save checkpoint interval. Default: None
--is_save_on_master IS_SAVE_ON_MASTER
Save ckpt on master or all rank, 1 for master, 0 for
all ranks. Default: 1
--is_distributed IS_DISTRIBUTED
Distribute train or not, 1 for yes, 0 for no. Default:
1
--rank RANK Local rank of distributed. Default: 0
--group_size GROUP_SIZE
World size of device. Default: 1
--need_profiler NEED_PROFILER
Whether use profiler. 0 for no, 1 for yes. Default: 0
--training_shape TRAINING_SHAPE
Fix training shape. Default: ""
--resize_rate RESIZE_RATE
Resize rate for multi-scale training. Default: None
```
#### Launch
```bash
# distributed training example(8p)
sh run_distribute_train.sh dataset/coco2014 backbone/backbone.ckpt rank_table_8p.json
# standalone training example(1p)
sh run_standalone_train.sh dataset/coco2014 backbone/backbone.ckpt
## [Training Process](#contents)
### Training
```
python train.py \
--data_dir=./dataset/coco2014 \
--pretrained_backbone=darknet53_backbone.ckpt \
--is_distributed=0 \
--lr=0.1 \
--T_max=320 \
--max_epoch=320 \
--warmup_epochs=4 \
--training_shape=416 \
--lr_scheduler=cosine_annealing > log.txt 2>&1 &
```
> About rank_table.json, you can refer to the [distributed training tutorial](https://www.mindspore.cn/tutorial/en/master/advanced_use/distributed_training.html).
The python command above will run in the background, you can view the results through the file `log.txt`.
#### Result
After training, you'll get some checkpoint files under the outputs folder by default. The loss value will be achieved as follows:
Training result will be stored in the scripts path, whose folder name begins with "train" or "train_parallel". You can find checkpoint file together with result like the followings in log.txt.
```
# grep "loss:" train/log.txt
2020-08-20 14:14:43,640:INFO:epoch[0], iter[0], loss:7809.262695, 0.15 imgs/sec, lr:9.746589057613164e-06
2020-08-20 14:15:05,142:INFO:epoch[0], iter[100], loss:2778.349033, 133.92 imgs/sec, lr:0.0009844054002314806
2020-08-20 14:15:31,796:INFO:epoch[0], iter[200], loss:535.517361, 130.54 imgs/sec, lr:0.0019590642768889666
...
```
The model checkpoint will be saved in outputs directory.
### Distributed Training
```
sh run_distribute_train.sh dataset/coco2014 darknet53_backbone.ckpt rank_table_8p.json
```
The above shell script will run distribute training in the background. You can view the results through the file `train_parallel[X]/log.txt`. The loss value will be achieved as follows:
```
# distribute training result(8p)
@ -92,31 +249,26 @@ epoch[319], iter[102300], loss:35.430038, 423.49 imgs/sec, lr:2.409552052995423e
...
```
### Infer
#### Usage
## [Evaluation Process](#contents)
### Evaluation
Before running the command below.
```
# infer
sh run_eval.sh [DATASET_PATH] [CHECKPOINT_PATH]
```
#### Launch
```bash
# infer with checkpoint
python eval.py \
--data_dir=./dataset/coco2014 \
--pretrained=yolov3.ckpt \
--testing_shape=416 > log.txt 2>&1 &
OR
sh run_eval.sh dataset/coco2014/ checkpoint/0-319_102400.ckpt
```
> checkpoint can be produced in training process.
#### Result
Inference result will be stored in the scripts path, whose folder name is "eval". Under this, you can find result like the followings in log.txt.
The above python command will run in the background. You can view the results through the file "log.txt". The mAP of the test dataset will be as follows:
```
# log.txt
=============coco eval reulst=========
Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.311
Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.528
@ -131,3 +283,51 @@ Inference result will be stored in the scripts path, whose folder name is "eval"
Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.442
Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.551
```
# [Model Description](#contents)
## [Performance](#contents)
### Evaluation Performance
| Parameters | YOLO |
| -------------------------- | ----------------------------------------------------------- |
| Model Version | YOLOv3 |
| Resource | Ascend 910; CPU 2.60GHz, 192cores; Memory, 755G |
| uploaded Date | 06/31/2020 (month/day/year) |
| MindSpore Version | 0.5.0-alpha |
| Dataset | COCO2014 |
| Training Parameters | epoch=320, batch_size=32, lr=0.001, momentum=0.9 |
| Optimizer | Momentum |
| Loss Function | Sigmoid Cross Entropy with logits |
| outputs | boxes and label |
| Loss | 34 |
| Speed | 1pc: 350 ms/step; |
| Total time | 8pc: 25 hours |
| Parameters (M) | 62.1 |
| Checkpoint for Fine tuning | 474M (.ckpt file) |
| Scripts | https://gitee.com/mindspore/mindspore/tree/master/model_zoo/official/cv/yolov3_darknet53 |
### Inference Performance
| Parameters | YOLO |
| ------------------- | --------------------------- |
| Model Version | YOLOv3 |
| Resource | Ascend 910 |
| Uploaded Date | 06/31/2020 (month/day/year) |
| MindSpore Version | 0.5.0-alpha |
| Dataset | COCO2014, 40,504 images |
| batch_size | 1 |
| outputs | mAP |
| Accuracy | 8pcs: 31.1% |
| Model for inference | 474M (.ckpt file) |
# [Description of Random Situation](#contents)
There are random seeds in distributed_sampler.py, transforms.py, yolo_dataset.py files.
# [ModelZoo Homepage](#contents)
Please check the official [homepage](https://gitee.com/mindspore/mindspore/tree/master/model_zoo).

68
model_zoo/official/cv/yolov3_darknet53/train.py
View File

@ -63,54 +63,58 @@ def parse_args():
parser = argparse.ArgumentParser('mindspore coco training')
# dataset related
parser.add_argument('--data_dir', type=str, default='', help='train data dir')
parser.add_argument('--per_batch_size', default=32, type=int, help='batch size for per gpu')
parser.add_argument('--data_dir', type=str, help='Train dataset directory.')
parser.add_argument('--per_batch_size', default=32, type=int, help='Batch size for Training. Default: 32.')
# network related
parser.add_argument('--pretrained_backbone', default='', type=str, help='model_path, local pretrained backbone'
' model to load')
parser.add_argument('--resume_yolov3', default='', type=str, help='path of pretrained yolov3')
parser.add_argument('--pretrained_backbone', default='', type=str,
help='The ckpt file of DarkNet53. Default: "".')
parser.add_argument('--resume_yolov3', default='', type=str,
help='The ckpt file of YOLOv3, which used to fine tune. Default: ""')
# optimizer and lr related
parser.add_argument('--lr_scheduler', default='exponential', type=str,
help='lr-scheduler, option type: exponential, cosine_annealing')
parser.add_argument('--lr', default=0.001, type=float, help='learning rate of the training')
parser.add_argument('--lr_epochs', type=str, default='220,250', help='epoch of lr changing')
help='Learning rate scheduler, options: exponential, cosine_annealing. Default: exponential')
parser.add_argument('--lr', default=0.001, type=float, help='Learning rate. Default: 0.001')
parser.add_argument('--lr_epochs', type=str, default='220,250',
help='Epoch of changing of lr changing, split with ",". Default: 220,250')
parser.add_argument('--lr_gamma', type=float, default=0.1,
help='decrease lr by a factor of exponential lr_scheduler')
parser.add_argument('--eta_min', type=float, default=0., help='eta_min in cosine_annealing scheduler')
parser.add_argument('--T_max', type=int, default=320, help='T-max in cosine_annealing scheduler')
parser.add_argument('--max_epoch', type=int, default=320, help='max epoch num to train the model')
parser.add_argument('--warmup_epochs', default=0, type=float, help='warmup epoch')
parser.add_argument('--weight_decay', type=float, default=0.0005, help='weight decay')
parser.add_argument('--momentum', type=float, default=0.9, help='momentum')
help='Decrease lr by a factor of exponential lr_scheduler. Default: 0.1')
parser.add_argument('--eta_min', type=float, default=0., help='Eta_min in cosine_annealing scheduler. Default: 0')
parser.add_argument('--T_max', type=int, default=320, help='T-max in cosine_annealing scheduler. Default: 320')
parser.add_argument('--max_epoch', type=int, default=320, help='Max epoch num to train the model. Default: 320')
parser.add_argument('--warmup_epochs', default=0, type=float, help='Warmup epochs. Default: 0')
parser.add_argument('--weight_decay', type=float, default=0.0005, help='Weight decay factor. Default: 0.0005')
parser.add_argument('--momentum', type=float, default=0.9, help='Momentum. Default: 0.9')
# loss related
parser.add_argument('--loss_scale', type=int, default=1024, help='static loss scale')
parser.add_argument('--label_smooth', type=int, default=0, help='whether to use label smooth in CE')
parser.add_argument('--label_smooth_factor', type=float, default=0.1, help='smooth strength of original one-hot')
parser.add_argument('--loss_scale', type=int, default=1024, help='Static loss scale. Default: 1024')
parser.add_argument('--label_smooth', type=int, default=0, help='Whether to use label smooth in CE. Default:0')
parser.add_argument('--label_smooth_factor', type=float, default=0.1,
help='Smooth strength of original one-hot. Default: 0.1')
# logging related
parser.add_argument('--log_interval', type=int, default=100, help='logging interval')
parser.add_argument('--ckpt_path', type=str, default='outputs/', help='checkpoint save location')
parser.add_argument('--ckpt_interval', type=int, default=None, help='ckpt_interval')
parser.add_argument('--log_interval', type=int, default=100, help='Logging interval steps. Default: 100')
parser.add_argument('--ckpt_path', type=str, default='outputs/', help='Checkpoint save location. Default: outputs/')
parser.add_argument('--ckpt_interval', type=int, default=None, help='Save checkpoint interval. Default: None')
parser.add_argument('--is_save_on_master', type=int, default=1, help='save ckpt on master or all rank')
parser.add_argument('--is_save_on_master', type=int, default=1,
help='Save ckpt on master or all rank, 1 for master, 0 for all ranks. Default: 1')
# distributed related
parser.add_argument('--is_distributed', type=int, default=1, help='if multi device')
parser.add_argument('--rank', type=int, default=0, help='local rank of distributed')
parser.add_argument('--group_size', type=int, default=1, help='world size of distributed')
# roma obs
parser.add_argument('--train_url', type=str, default="", help='train url')
parser.add_argument('--is_distributed', type=int, default=1,
help='Distribute train or not, 1 for yes, 0 for no. Default: 1')
parser.add_argument('--rank', type=int, default=0, help='Local rank of distributed. Default: 0')
parser.add_argument('--group_size', type=int, default=1, help='World size of device. Default: 1')
# profiler init
parser.add_argument('--need_profiler', type=int, default=0, help='whether use profiler')
parser.add_argument('--need_profiler', type=int, default=0,
help='Whether use profiler. 0 for no, 1 for yes. Default: 0')
# reset default config
parser.add_argument('--training_shape', type=str, default="", help='fix training shape')
parser.add_argument('--resize_rate', type=int, default=None, help='resize rate for multi-scale training')
parser.add_argument('--training_shape', type=str, default="", help='Fix training shape. Default: ""')
parser.add_argument('--resize_rate', type=int, default=None,
help='Resize rate for multi-scale training. Default: None')
args, _ = parser.parse_known_args()
if args.lr_scheduler == 'cosine_annealing' and args.max_epoch > args.T_max:
@ -153,7 +157,7 @@ def train():
args.logger.save_args(args)
if args.need_profiler:
from mindinsight.profiler.profiling import Profiler
from mindspore.profiler.profiling import Profiler
profiler = Profiler(output_path=args.outputs_dir, is_detail=True, is_show_op_path=True)
loss_meter = AverageMeter('loss')