forked from huawei/mindspore2022
!22605 添加yolov5网络到master
Merge pull request !22605 from ziquan/yolov5-master
This commit is contained in:
Gitee
commit
cfd913dfd5
|
|
@ -1,361 +1,253 @@
|
||||||
# Contents
|
# Inference ProcessContents
|
||||||
|
|
||||||
- [YOLOv5 Description](#YOLOv5-description)
|
- [YOLOv5 Description](#YOLOv5-description)
|
||||||
- [Model Architecture](#model-architecture)
|
- [Model Architecture](#model-architecture)
|
||||||
- [Dataset](#dataset)
|
- [Dataset](#dataset)
|
||||||
- [Environment Requirements](#environment-requirements)
|
|
||||||
- [Quick Start](#quick-start)
|
- [Quick Start](#quick-start)
|
||||||
- [Script Description](#script-description)
|
- [Script Description](#script-description)
|
||||||
- [Script and Sample Code](#script-and-sample-code)
|
- [Script and Sample Code](#script-and-sample-code)
|
||||||
- [Script Parameters](#script-parameters)
|
- [Script Parameters](#script-parameters)
|
||||||
- [Training Process](#training-process)
|
- [Training Process](#training-process)
|
||||||
- [Training](#training)
|
- [Training](#training)
|
||||||
- [Testing Process](#testing-process)
|
- [Distributed Training](#distributed-training)
|
||||||
- [Evaluation](#testing)
|
|
||||||
- [Evaluation Process](#evaluation-process)
|
- [Evaluation Process](#evaluation-process)
|
||||||
- [Evaluation](#evaluation)
|
- [Evaluation](#evaluation)
|
||||||
- [Inference Process](#inference-process)
|
|
||||||
- [Export MindIR](#export-mindir)
|
|
||||||
- [Infer on Ascend310](#infer-on-ascend310)
|
|
||||||
- [result](#result)
|
|
||||||
- [Model Description](#model-description)
|
- [Model Description](#model-description)
|
||||||
- [Performance](#performance)
|
- [Performance](#performance)
|
||||||
- [Evaluation Performance](#evaluation-performance)
|
- [Evaluation Performance](#evaluation-performance)
|
||||||
- [Inference Performance](#inference-performance)
|
- [Inference Performance](#inference-performance)
|
||||||
- [310 Inference Performance](#310-inference-performance)
|
- [Transfer Learning](#transfer-learning)
|
||||||
|
- [Description of Random Situation](#description-of-random-situation)
|
||||||
- [ModelZoo Homepage](#modelzoo-homepage)
|
- [ModelZoo Homepage](#modelzoo-homepage)
|
||||||
|
|
||||||
# [YOLOv5 Description](#contents)
|
# [YOLOv5 Description](#contents)
|
||||||
|
|
||||||
YOLOv5 is a state-of-the-art detector which is faster (FPS) and more accurate (MS COCO AP50...95 and AP50) than all available alternative detectors.
|
Published in April 2020, YOLOv5 achieved state of the art performance on the COCO dataset for object detection. It is an important improvement of YoloV3, the implementation of a new architecture in the **Backbone** and the modifications in the **Neck** have improved the **mAP**(mean Average Precision) by **10%** and the number of **FPS**(Frame per Second) by **12%**.
|
||||||
YOLOv5 has verified a large number of features, and selected for use such of them for improving the accuracy of both the classifier and the detector.
|
|
||||||
These features can be used as best-practice for future studies and developments.
|
|
||||||
|
|
||||||
[Code](https://github.com/ultralytics/yolov5)
|
[code](https://github.com/ultralytics/yolov5)
|
||||||
|
|
||||||
# [Model Architecture](#contents)
|
# [Model Architecture](#contents)
|
||||||
|
|
||||||
YOLOv5 choose CSP with Focus backbone, SPP additional module, PANet path-aggregation neck, and YOLOv5 (anchor based) head as the architecture of YOLOv5.
|
The YOLOv5 network is mainly composed of CSP and Focus as a backbone, spatial pyramid pooling(SPP) additional module, PANet path-aggregation neck and YOLOv3 head. [CSP](https://arxiv.org/abs/1911.11929) is a novel backbone that can enhance the learning capability of CNN. The [spatial pyramid pooling](https://arxiv.org/abs/1406.4729) block is added over CSP to increase the receptive field and separate out the most significant context features. Instead of Feature pyramid networks (FPN) for object detection used in YOLOv3, the PANet is used as the method for parameter aggregation for different detector levels. To be more specifical, CSPDarknet53 contains 5 CSP modules which use the convolution **C** with kernel size k=3x3, stride s = 2x2; Within the PANet and SPP, **1x1, 5x5, 9x9, 13x13 max poolings are applied.
|
||||||
|
|
||||||
# [Dataset](#contents)
|
# [Dataset](#contents)
|
||||||
|
|
||||||
Dataset support: [MS COCO] or datasetd with the same format as MS COCO
|
Dataset used: [COCO2017](<https://cocodataset.org/#download>)
|
||||||
Annotation support: [MS COCO] or annotation as the same format as MS COCO
|
|
||||||
|
|
||||||
- The directory structure is as follows, the name of directory and file is user define:
|
Note that you can run the scripts with **COCO2017 **or any other datasets with the same format as MS COCO Annotation. But we do suggest user to use MS COCO dataset to experience our model.
|
||||||
|
|
||||||
```shell
|
|
||||||
├── dataset
|
|
||||||
├── YOLOv5
|
|
||||||
├── annotations
|
|
||||||
│ ├─ train.json
|
|
||||||
│ └─ val.json
|
|
||||||
├─ images
|
|
||||||
├─ train
|
|
||||||
│ └─images
|
|
||||||
│ ├─picture1.jpg
|
|
||||||
│ ├─ ...
|
|
||||||
│ └─picturen.jpg
|
|
||||||
└─ val
|
|
||||||
└─images
|
|
||||||
├─picture1.jpg
|
|
||||||
├─ ...
|
|
||||||
└─picturen.jpg
|
|
||||||
```
|
|
||||||
|
|
||||||
we suggest user to use MS COCO dataset to experience our model,
|
|
||||||
other datasets need to use the same format as MS COCO.
|
|
||||||
|
|
||||||
# [Environment Requirements](#contents)
|
|
||||||
|
|
||||||
- Hardware（Ascend）
|
|
||||||
- Prepare hardware environment with Ascend processor.
|
|
||||||
- Framework
|
|
||||||
- [MindSpore](https://www.mindspore.cn/install/en)
|
|
||||||
- For more information, please check the resources below：
|
|
||||||
- [MindSpore Tutorials](https://www.mindspore.cn/tutorials/en/master/index.html)
|
|
||||||
- [MindSpore Python API](https://www.mindspore.cn/docs/api/en/master/index.html)
|
|
||||||
|
|
||||||
# [Quick Start](#contents)
|
# [Quick Start](#contents)
|
||||||
|
|
||||||
After installing MindSpore via the official website, you can start training and evaluation as follows:
|
After installing MindSpore via the official website, you can start training and evaluation as follows:
|
||||||
|
|
||||||
``` shell
|
```bash
|
||||||
# The parameter of training_shape define image shape for network, default is [640, 640],
|
|
||||||
```
|
|
||||||
|
|
||||||
```shell
|
|
||||||
#run training example(1p) by python command
|
#run training example(1p) by python command
|
||||||
python train.py \
|
python train.py \
|
||||||
--data_dir=./dataset/xxx \
|
--data_dir=xxx/dataset \
|
||||||
--is_distributed=0 \
|
--is_distributed=0 \
|
||||||
--lr=0.01 \
|
--yolov5_version='yolov5s' \
|
||||||
--T_max=320 \
|
--lr=0.02 \
|
||||||
--max_epoch=320 \
|
--max_epoch=300 \
|
||||||
--warmup_epochs=4 \
|
--warmup_epochs=20 \
|
||||||
--training_shape=640 \
|
--per_batch_size=128 \
|
||||||
--lr_scheduler=cosine_annealing > log.txt 2>&1 &
|
--lr_scheduler=cosine_annealing > log.txt 2>&1 &
|
||||||
```
|
```
|
||||||
|
|
||||||
```shell
|
```bash
|
||||||
# standalone training example(1p) by shell script
|
|
||||||
bash run_standalone_train.sh dataset/xxx
|
|
||||||
```
|
|
||||||
|
|
||||||
```shell
|
|
||||||
# For Ascend device, distributed training example(8p) by shell script
|
# For Ascend device, distributed training example(8p) by shell script
|
||||||
bash run_distribute_train.sh dataset/xxx rank_table_8p.json
|
bash run_distribute_train.sh xxx/dateset/ xxx/cspdarknet.ckpt rank_table_8pcs.json
|
||||||
```
|
```
|
||||||
|
|
||||||
```python
|
```bash
|
||||||
# run evaluation by python command
|
# run evaluation by python command
|
||||||
python eval.py \
|
python eval.py \
|
||||||
--data_dir=./dataset/xxx \
|
--data_dir=xxx/dataset \
|
||||||
--pretrained=yolov5.ckpt \
|
--testing_shape=640 > log.txt 2>&1 &
|
||||||
--testing_shape=640 > log.txt 2>&1 &
|
|
||||||
```
|
```
|
||||||
|
|
||||||
```python
|
```bash
|
||||||
# run evaluation by shell script
|
# run evaluation by shell script
|
||||||
bash run_eval.sh dataset/xxx checkpoint/xxx.ckpt
|
bash run_eval.sh xxx/dataset xxx/yolov5.ckpt
|
||||||
```
|
```
|
||||||
|
|
||||||
# [Script Description](#contents)
|
# [Script Description](#contents)
|
||||||
|
|
||||||
## [Script and Sample Code](#contents)
|
## [Script and Sample Code](#contents)
|
||||||
|
|
||||||
```python
|
```bash
|
||||||
└─yolov5
|
├── model_zoo
|
||||||
├─README.md
|
├── README.md // descriptions about all the models
|
||||||
├─mindspore_hub_conf.md # config for mindspore hub
|
├── yolov5
|
||||||
├─ascend310_infer # application for 310 inference
|
├── README.md // descriptions about yolov5
|
||||||
├─scripts
|
├── scripts
|
||||||
├─run_standalone_train.sh # launch standalone training(1p) in ascend
|
│ ├──run_distribute_train.sh // launch distributed training(8p) in ascend
|
||||||
├─run_distribute_train.sh # launch distributed training(8p) in ascend
|
│ ├──run_eval.sh // shell script for evaluation
|
||||||
├─run_infer_310.sh # launch 310 inference in ascend
|
│ ├──rank_table_8pcs.json // the example of rank table settings for 8p training
|
||||||
└─run_eval.sh # launch evaluating in ascend
|
├── src
|
||||||
├─src
|
│ ├──config.py // parameter configuration
|
||||||
├─__init__.py # python init file
|
│ ├──backbone.py // backbone of network
|
||||||
├─config.py # parameter configuration
|
│ ├──distributed_sampler.py // iterator of dataset
|
||||||
├─yolov5_backbone.py # backbone of network
|
│ ├──initializer.py // initializer of parameters
|
||||||
├─distributed_sampler.py # iterator of dataset
|
│ ├──logger.py // log function
|
||||||
├─initializer.py # initializer of parameters
|
│ ├──loss.py // loss function
|
||||||
├─logger.py # log function
|
│ ├──lr_scheduler.py // generate learning rate
|
||||||
├─loss.py # loss function
|
│ ├──transforms.py // Preprocess data
|
||||||
├─lr_scheduler.py # generate learning rate
|
│ ├──util.py // util function
|
||||||
├─transforms.py # Preprocess data
|
│ ├──yolo.py // yolov5 network
|
||||||
├─util.py # util function
|
│ ├──yolo_dataset.py // create dataset for YOLOV5
|
||||||
├─yolo.py # yolov5 network
|
├── train.py // training script
|
||||||
├─yolo_dataset.py # create dataset for YOLOV5
|
├── eval.py // evaluation script
|
||||||
├─eval.py # evaluate val results
|
├── export.py // export script
|
||||||
├─export.py # convert mindspore model to air model
|
|
||||||
├─postprocess.py # postprocess script
|
|
||||||
└─train.py # train net
|
|
||||||
```
|
```
|
||||||
|
|
||||||
## [Script Parameters](#contents)
|
## [Script Parameters](#contents)
|
||||||
|
|
||||||
Major parameters train.py as follows:
|
```python
|
||||||
|
Major parameters in train.py are:
|
||||||
|
|
||||||
```shell
|
|
||||||
optional arguments:
|
optional arguments:
|
||||||
-h, --help show this help message and exit
|
|
||||||
--device_target device where the code will be implemented: "Ascend" | "GPU", default is "Ascend"
|
--device_target device where the code will be implemented: "Ascend", default is "Ascend"
|
||||||
--data_dir DATA_DIR Train dataset directory.
|
--data_dir Train dataset directory.
|
||||||
--per_batch_size PER_BATCH_SIZE
|
--per_batch_size Batch size for Training. Default: 8.
|
||||||
Batch size for Training. Default: 8.
|
--pretrained_backbone The ckpt file of CSPDarknet53. Default: "".
|
||||||
--pretrained_backbone PRETRAINED_BACKBONE
|
--resume_yolov5 The ckpt file of YOLOv5, which used to fine tune.Default: ""
|
||||||
The backbone file of yolov5. Default: "".
|
--lr_scheduler Learning rate scheduler, options: exponential,cosine_annealing.
|
||||||
--resume_yolov5 RESUME_YOLOV5
|
Default: cosine_annealing
|
||||||
The ckpt file of YOLOv5, which used to fine tune.
|
--lr Learning rate. Default: 0.02
|
||||||
Default: ""
|
--lr_epochs Epoch of changing of lr changing, split with ",". Default: '220,250'
|
||||||
--lr_scheduler LR_SCHEDULER
|
--lr_gamma Decrease lr by a factor of exponential lr_scheduler. Default: 0.1
|
||||||
Learning rate scheduler, options: exponential,
|
--eta_min Eta_min in cosine_annealing scheduler. Default: 0.
|
||||||
cosine_annealing. Default: exponential
|
--t_max T-max in cosine_annealing scheduler. Default: 320
|
||||||
--lr LR Learning rate. Default: 0.01
|
--max_epoch Max epoch num to train the model. Default: 320
|
||||||
--lr_epochs LR_EPOCHS
|
--warmup_epochs Warmup epochs. Default: 20
|
||||||
Epoch of changing of lr changing, split with ",".
|
--weight_decay Weight decay factor. Default: 0.0005
|
||||||
Default: 220,250
|
--momentum Momentum. Default: 0.9
|
||||||
--lr_gamma LR_GAMMA Decrease lr by a factor of exponential lr_scheduler.
|
--loss_scale Static loss scale. Default: 64
|
||||||
Default: 0.1
|
--label_smooth Whether to use label smooth in CE. Default:0
|
||||||
--eta_min ETA_MIN Eta_min in cosine_annealing scheduler. Default: 0
|
--label_smooth_factor Smooth strength of original one-hot. Default: 0.1
|
||||||
--T_max T_MAX T-max in cosine_annealing scheduler. Default: 320
|
--log_interval Logging interval steps. Default: 100
|
||||||
--max_epoch MAX_EPOCH
|
--ckpt_path Checkpoint save location. Default: outputs/
|
||||||
Max epoch num to train the model. Default: 320
|
--ckpt_interval Save checkpoint interval. Default: None
|
||||||
--warmup_epochs WARMUP_EPOCHS
|
--is_save_on_master Save ckpt on master or all rank, 1 for master, 0 for all ranks. Default: 1
|
||||||
Warmup epochs. Default: 0
|
--is_distributed Distribute train or not, 1 for yes, 0 for no. Default: 1
|
||||||
--weight_decay WEIGHT_DECAY
|
--rank Local rank of distributed. Default: 0
|
||||||
Weight decay factor. Default: 0.0005
|
--group_size World size of device. Default: 1
|
||||||
--momentum MOMENTUM Momentum. Default: 0.9
|
--need_profiler Whether use profiler. 0 for no, 1 for yes. Default: 0
|
||||||
--loss_scale LOSS_SCALE
|
--training_shape Fix training shape. Default: ""
|
||||||
Static loss scale. Default: 1024
|
--resize_rate Resize rate for multi-scale training. Default: 10
|
||||||
--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
|
|
||||||
```
|
```
|
||||||
|
|
||||||
## [Training Process](#contents)
|
## [Training Process](#contents)
|
||||||
|
|
||||||
### Training
|
### Training
|
||||||
|
|
||||||
|
For Ascend device, standalone training can be started like this:
|
||||||
|
|
||||||
```python
|
```python
|
||||||
|
#run training example(1p) by python command
|
||||||
python train.py \
|
python train.py \
|
||||||
--data_dir=/dataset/xxx \
|
--data_dir=xxx/dataset \
|
||||||
|
--yolov5_version='yolov5s' \
|
||||||
--is_distributed=0 \
|
--is_distributed=0 \
|
||||||
--lr=0.01 \
|
--lr=0.02 \
|
||||||
--T_max=320 \
|
--max_epoch=300 \
|
||||||
--max_epoch=320 \
|
--warmup_epochs=20 \
|
||||||
--warmup_epochs=4 \
|
--per_batch_size=128 \
|
||||||
--training_shape=640 \
|
|
||||||
--lr_scheduler=cosine_annealing > log.txt 2>&1 &
|
--lr_scheduler=cosine_annealing > log.txt 2>&1 &
|
||||||
```
|
```
|
||||||
|
|
||||||
The python command above will run in the background, you can view the results through the file log.txt.
|
The python command above will run in the background, you can view the results through the file `log.txt`.
|
||||||
|
|
||||||
After training, you'll get some checkpoint files under the outputs folder by default. The loss value will be achieved as follows:
|
After training, you'll get some checkpoint files under the **outputs** folder by default. The loss value will be achieved as follows:
|
||||||
|
|
||||||
```shell
|
```python
|
||||||
# grep "loss:" train/log.txt
|
# grep "loss:" log.txt
|
||||||
2021-05-13 20:50:25,617:INFO:epoch[0], iter[100], loss:loss:2648.764910, fps:61.59 imgs/sec, lr:1.7226087948074564e-05
|
2021-08-06 15:30:15,798:INFO:epoch[0], iter[600], loss:296.308071, fps:44.44 imgs/sec, lr:0.00010661844862625003
|
||||||
2021-05-13 20:50:39,821:INFO:epoch[0], iter[200], loss:loss:764.535622, fps:56.33 imgs/sec, lr:3.4281620173715055e-05
|
2021-08-06 15:31:21,119:INFO:epoch[0], iter[700], loss:276.071959, fps:48.99 imgs/sec, lr:0.00012435863027349114
|
||||||
2021-05-13 20:50:53,287:INFO:epoch[0], iter[300], loss:loss:494.950782, fps:59.47 imgs/sec, lr:5.1337152399355546e-05
|
2021-08-06 15:32:26,185:INFO:epoch[0], iter[800], loss:266.955208, fps:49.18 imgs/sec, lr:0.00014209879736881703
|
||||||
2021-05-13 20:51:06,138:INFO:epoch[0], iter[400], loss:loss:393.339678, fps:62.25 imgs/sec, lr:6.839268462499604e-05
|
2021-08-06 15:33:30,507:INFO:epoch[0], iter[900], loss:252.610914, fps:49.75 imgs/sec, lr:0.00015983897901605815
|
||||||
2021-05-13 20:51:17,985:INFO:epoch[0], iter[500], loss:loss:329.976604, fps:67.57 imgs/sec, lr:8.544822048861533e-05
|
2021-08-06 15:34:42,176:INFO:epoch[0], iter[1000], loss:243.106683, fps:44.65 imgs/sec, lr:0.00017757914611138403
|
||||||
2021-05-13 20:51:29,359:INFO:epoch[0], iter[600], loss:loss:294.734397, fps:70.37 imgs/sec, lr:0.00010250374907627702
|
2021-08-06 15:35:47,429:INFO:epoch[0], iter[1100], loss:240.498834, fps:49.04 imgs/sec, lr:0.00019531932775862515
|
||||||
2021-05-13 20:51:40,634:INFO:epoch[0], iter[700], loss:loss:281.497078, fps:70.98 imgs/sec, lr:0.00011955928493989632
|
2021-08-06 15:36:48,945:INFO:epoch[0], iter[1200], loss:245.711473, fps:52.02 imgs/sec, lr:0.00021305949485395104
|
||||||
2021-05-13 20:51:52,307:INFO:epoch[0], iter[800], loss:loss:264.300707, fps:68.54 imgs/sec, lr:0.0001366148208035156
|
2021-08-06 15:37:51,293:INFO:epoch[0], iter[1300], loss:231.388255, fps:51.33 imgs/sec, lr:0.00023079967650119215
|
||||||
2021-05-13 20:52:05,479:INFO:epoch[0], iter[900], loss:loss:261.971103, fps:60.76 imgs/sec, lr:0.0001536703493911773
|
2021-08-06 15:38:55,680:INFO:epoch[0], iter[1400], loss:238.904242, fps:49.70 imgs/sec, lr:0.00024853984359651804
|
||||||
2021-05-13 20:52:17,362:INFO:epoch[0], iter[1000], loss:loss:264.591175, fps:67.33 imgs/sec, lr:0.00017072587797883898
|
2021-08-06 15:39:57,419:INFO:epoch[0], iter[1500], loss:232.161600, fps:51.83 imgs/sec, lr:0.00026628002524375916
|
||||||
|
2021-08-06 15:41:03,808:INFO:epoch[0], iter[1600], loss:227.844698, fps:48.20 imgs/sec, lr:0.00028402020689100027
|
||||||
|
2021-08-06 15:42:06,155:INFO:epoch[0], iter[1700], loss:226.668858, fps:51.33 imgs/sec, lr:0.00030176035943441093
|
||||||
...
|
...
|
||||||
```
|
```
|
||||||
|
|
||||||
### Distributed Training
|
### Distributed Training
|
||||||
|
|
||||||
For Ascend device, distributed training example(8p) by shell script
|
For Ascend device, distributed training example(8p) by shell script：
|
||||||
|
|
||||||
```shell
|
```bash
|
||||||
bash run_distribute_train.sh dataset/coco2017 rank_table_8p.json
|
# For Ascend device, distributed training example(8p) by shell script
|
||||||
|
bash run_distribute_train.sh xxx/dateset/ xxx/cspdarknet.ckpt rank_table_8pcs.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:
|
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:
|
||||||
|
|
||||||
```shell
|
```bash
|
||||||
# distribute training result(8p)
|
# distribute training result(8p, dynamic shape)
|
||||||
...
|
...
|
||||||
2021-05-13 21:08:41,992:INFO:epoch[0], iter[600], loss:247.577421, fps:469.29 imgs/sec, lr:0.0001640283880988136
|
2021-08-05 16:01:34,116:INFO:epoch[0], iter[200], loss:415.453676, fps:580.07 imgs/sec, lr:0.0002742903889156878
|
||||||
2021-05-13 21:08:56,291:INFO:epoch[0], iter[700], loss:235.298894, fps:447.67 imgs/sec, lr:0.0001913209562189877
|
2021-08-05 16:01:57,588:INFO:epoch[0], iter[300], loss:273.358383, fps:545.96 imgs/sec, lr:0.00041075327317230403
|
||||||
2021-05-13 21:09:10,431:INFO:epoch[0], iter[800], loss:239.481037, fps:452.78 imgs/sec, lr:0.00021861353889107704
|
2021-08-05 16:02:26,247:INFO:epoch[0], iter[400], loss:244.621502, fps:446.64 imgs/sec, lr:0.0005472161574289203
|
||||||
2021-05-13 21:09:23,517:INFO:epoch[0], iter[900], loss:232.826709, fps:489.15 imgs/sec, lr:0.0002459061215631664
|
2021-08-05 16:02:55,532:INFO:epoch[0], iter[500], loss:234.524876, fps:437.10 imgs/sec, lr:0.000683679012581706
|
||||||
2021-05-13 21:09:36,407:INFO:epoch[0], iter[1000], loss:224.734599, fps:496.65 imgs/sec, lr:0.0002731987042352557
|
2021-08-05 16:03:25,046:INFO:epoch[0], iter[600], loss:235.185213, fps:434.08 imgs/sec, lr:0.0008201419259421527
|
||||||
2021-05-13 21:09:49,072:INFO:epoch[0], iter[1100], loss:232.334771, fps:505.34 imgs/sec, lr:0.0003004912578035146
|
2021-08-05 16:03:54,585:INFO:epoch[0], iter[700], loss:228.878598, fps:433.48 imgs/sec, lr:0.0009566047810949385
|
||||||
2021-05-13 21:10:03,597:INFO:epoch[0], iter[1200], loss:242.001476, fps:440.69 imgs/sec, lr:0.00032778384047560394
|
2021-08-05 16:04:23,932:INFO:epoch[0], iter[800], loss:219.259134, fps:436.29 imgs/sec, lr:0.0010930676944553852
|
||||||
2021-05-13 21:10:18,237:INFO:epoch[0], iter[1300], loss:225.391021, fps:437.20 imgs/sec, lr:0.0003550764231476933
|
2021-08-05 16:04:52,707:INFO:epoch[0], iter[900], loss:225.741833, fps:444.84 imgs/sec, lr:0.001229530549608171
|
||||||
2021-05-13 21:10:33,027:INFO:epoch[0], iter[1400], loss:228.738176, fps:432.76 imgs/sec, lr:0.0003823690058197826
|
2021-08-05 16:05:21,872:INFO:epoch[1], iter[1000], loss:218.811336, fps:438.91 imgs/sec, lr:0.0013659934047609568
|
||||||
2021-05-13 21:10:47,424:INFO:epoch[0], iter[1500], loss:225.712950, fps:444.54 imgs/sec, lr:0.0004096615593880415
|
2021-08-05 16:05:51,216:INFO:epoch[1], iter[1100], loss:219.491889, fps:436.50 imgs/sec, lr:0.0015024563763290644
|
||||||
2021-05-13 21:11:02,077:INFO:epoch[0], iter[1600], loss:221.249353, fps:436.77 imgs/sec, lr:0.00043695414206013083
|
2021-08-05 16:06:20,546:INFO:epoch[1], iter[1200], loss:219.895906, fps:436.57 imgs/sec, lr:0.0016389192314818501
|
||||||
2021-05-13 21:11:16,631:INFO:epoch[0], iter[1700], loss:222.449119, fps:439.89 imgs/sec, lr:0.00046424672473222017
|
2021-08-05 16:06:49,521:INFO:epoch[1], iter[1300], loss:218.516680, fps:441.79 imgs/sec, lr:0.001775382086634636
|
||||||
|
2021-08-05 16:07:18,303:INFO:epoch[1], iter[1400], loss:209.922935, fps:444.79 imgs/sec, lr:0.0019118449417874217
|
||||||
|
2021-08-05 16:07:47,702:INFO:epoch[1], iter[1500], loss:210.997816, fps:435.60 imgs/sec, lr:0.0020483077969402075
|
||||||
|
2021-08-05 16:08:16,482:INFO:epoch[1], iter[1600], loss:210.678421, fps:444.88 imgs/sec, lr:0.002184770768508315
|
||||||
|
2021-08-05 16:08:45,568:INFO:epoch[1], iter[1700], loss:203.285874, fps:440.07 imgs/sec, lr:0.0023212337400764227
|
||||||
|
2021-08-05 16:09:13,947:INFO:epoch[1], iter[1800], loss:203.014775, fps:451.11 imgs/sec, lr:0.0024576964788138866
|
||||||
|
2021-08-05 16:09:42,954:INFO:epoch[2], iter[1900], loss:194.683969, fps:441.28 imgs/sec, lr:0.0025941594503819942
|
||||||
...
|
...
|
||||||
```
|
```
|
||||||
|
|
||||||
## [Evaluation Process](#contents)
|
## [Evaluation Process](#contents)
|
||||||
|
|
||||||
### Valid
|
### Evaluation
|
||||||
|
|
||||||
|
Before running the command below, please check the checkpoint path used for evaluation. The file **yolov5.ckpt** used in the follow script is the last saved checkpoint file, but we renamed it to "yolov5.ckpt".
|
||||||
|
|
||||||
```python
|
```python
|
||||||
|
# run evaluation by python command
|
||||||
python eval.py \
|
python eval.py \
|
||||||
--data_dir=./dataset/coco2017 \
|
--data_dir=xxx/dataset \
|
||||||
--pretrained=yolov5.ckpt \
|
--pretrained=xxx/yolov5.ckpt \
|
||||||
--testing_shape=640 > log.txt 2>&1 &
|
--testing_shape=640 > log.txt 2>&1 &
|
||||||
OR
|
OR
|
||||||
bash run_eval.sh dataset/coco2017 checkpoint/yolov5.ckpt
|
# run evaluation by shell script
|
||||||
|
bash run_eval.sh xxx/dataset xxx/yolov5.ckpt
|
||||||
```
|
```
|
||||||
|
|
||||||
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:
|
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:
|
||||||
|
|
||||||
```shell
|
```python
|
||||||
# log.txt
|
# log.txt
|
||||||
=============coco eval reulst=========
|
=============coco eval reulst=========
|
||||||
Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.372
|
Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.369
|
||||||
Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.574
|
Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.573
|
||||||
Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.403
|
Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.395
|
||||||
Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.219
|
Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.218
|
||||||
Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.426
|
Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.418
|
||||||
Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.480
|
Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.482
|
||||||
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 1 ] = 0.302
|
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 1 ] = 0.298
|
||||||
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 10 ] = 0.504
|
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 10 ] = 0.501
|
||||||
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.560
|
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.557
|
||||||
Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.399
|
Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.395
|
||||||
Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.619
|
Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.619
|
||||||
Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.674
|
Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.677
|
||||||
```
|
2020-12-21 17:16:40,322:INFO:testing cost time 0.35h
|
||||||
|
|
||||||
## [Inference process](#contents)
|
|
||||||
|
|
||||||
### Export MindIR
|
|
||||||
|
|
||||||
```shell
|
|
||||||
python export.py --ckpt_file [CKPT_PATH] --file_format [EXPORT_FORMAT] --batch_size [BATCH_SIZE]
|
|
||||||
```
|
|
||||||
|
|
||||||
The ckpt_file parameter is required,
|
|
||||||
`EXPORT_FORMAT` should be in ["AIR", "MINDIR"].Current model only support CPU MODE.
|
|
||||||
`BATCH_SIZE` current batch_size can only be set to 1.
|
|
||||||
|
|
||||||
### Infer on Ascend310
|
|
||||||
|
|
||||||
Before performing inference, the mindir file must be exported by `export.py` script. We only provide an example of inference using MINDIR model.
|
|
||||||
Current batch_size can only be set to 1.
|
|
||||||
|
|
||||||
```shell
|
|
||||||
# Ascend310 inference
|
|
||||||
bash run_infer_310.sh [MINDIR_PATH] [DATA_PATH] [ANN_FILE] [DVPP] [DEVICE_ID]
|
|
||||||
```
|
|
||||||
|
|
||||||
- `ANN_FILE` annotations file path.
|
|
||||||
- `DVPP` is mandatory, and must choose from ["DVPP", "CPU"], it's case-insensitive. Current model only support CPU MODE.
|
|
||||||
- `DEVICE_ID` is optional, default value is 0.
|
|
||||||
|
|
||||||
### result
|
|
||||||
|
|
||||||
Inference result is saved in current path, you can find result like this in acc.log file.
|
|
||||||
|
|
||||||
```bash
|
|
||||||
# acc.log
|
|
||||||
=============coco 310 infer reulst=========
|
|
||||||
Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.369
|
|
||||||
Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.571
|
|
||||||
Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.398
|
|
||||||
Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.216
|
|
||||||
Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.421
|
|
||||||
Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.487
|
|
||||||
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 1 ] = 0.301
|
|
||||||
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 10 ] = 0.502
|
|
||||||
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.558
|
|
||||||
Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.388
|
|
||||||
Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.617
|
|
||||||
Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.677
|
|
||||||
```
|
```
|
||||||
|
|
||||||
# [Model Description](#contents)
|
# [Model Description](#contents)
|
||||||
|
|
@ -366,56 +258,41 @@ Inference result is saved in current path, you can find result like this in acc.
|
||||||
|
|
||||||
YOLOv5 on 118K images(The annotation and data format must be the same as coco2017)
|
YOLOv5 on 118K images(The annotation and data format must be the same as coco2017)
|
||||||
|
|
||||||
| Parameters | YOLOv5s |
|
| Parameters | YOLOv5 |
|
||||||
| -------------------------- | ----------------------------------------------------------- |
|
| -------------------------- | ------------------------------------------------------------ |
|
||||||
| Resource | Ascend 910; CPU 2.60GHz, 192cores; Memory, 755G |
|
| Resource | Ascend 910 ；CPU 2.60GHz，192cores; Memory, 755G |
|
||||||
| uploaded Date | 5/14/2021 (month/day/year) |
|
| uploaded Date | 7/12/2021 (month/day/year) |
|
||||||
| MindSpore Version | 1.0.0-alpha |
|
| MindSpore Version | 1.2.0 |
|
||||||
| Dataset | 118K images |
|
| Dataset | 118K images |
|
||||||
| Training Parameters | epoch=320, batch_size=8, lr=0.01, momentum=0.9 |
|
| Training Parameters | epoch=300, batch_size=8, lr=0.02,momentum=0.9,warmup_epoch=20 |
|
||||||
| Optimizer | Momentum |
|
| Optimizer | Momentum |
|
||||||
| Loss Function | Sigmoid Cross Entropy with logits, Giou Loss |
|
| Loss Function | Sigmoid Cross Entropy with logits, Giou Loss |
|
||||||
| outputs | heatmaps |
|
| outputs | boxes and label |
|
||||||
| Loss | 53 |
|
| Loss | 111.970097 |
|
||||||
| Speed | 1p 55 img/s 8p 440 img/s(shape=640) |
|
| Speed | 8p about 450 FPS |
|
||||||
| Total time | 24h(8pcs) |
|
| Total time | 8p 21h28min |
|
||||||
| Checkpoint for Fine tuning | 58M (.ckpt file) |
|
| Checkpoint for Fine tuning | 53.62M (.ckpt file) |
|
||||||
| Scripts | <https://gitee.com/mindspore/mindspore/tree/master/model_zoo/>|
|
| Scripts | https://gitee.com/mindspore/mindspore/tree/master/model_zoo/ |
|
||||||
|
|
||||||
### Inference Performance
|
### Inference Performance
|
||||||
|
|
||||||
YOLOv5 on 5K images(The annotation and data format must be the same as coco val2017 )
|
| Parameters | YOLOv5 |
|
||||||
|
| ------------------- | --------------------------- |
|
||||||
| Parameters | YOLOv5s |
|
| Resource | Ascend 910 ；CPU 2.60GHz，192cores; Memory, 755G |
|
||||||
| -------------------------- | ----------------------------------------------------------- |
|
| Uploaded Date | 7/12/2021 (month/day/year) |
|
||||||
| Resource | Ascend 910; CPU 2.60GHz, 192cores; Memory, 755G |
|
| MindSpore Version | 1.2.0 |
|
||||||
| uploaded Date | 5/14/2021 (month/day/year) |
|
| Dataset | 20K images |
|
||||||
| MindSpore Version | 1.2.0 |
|
| batch_size | 1 |
|
||||||
| Dataset | 5K images |
|
|
||||||
| batch_size | 1 |
|
|
||||||
| outputs | box position and sorces, and probability |
|
|
||||||
| Accuracy | map=36.8~37.2%(shape=640) |
|
|
||||||
| Model for inference | 58M (.ckpt file) |
|
|
||||||
|
|
||||||
### 310 Inference Performance
|
|
||||||
|
|
||||||
| Parameters | Ascend |
|
|
||||||
| ------------------- | ---------------------------------------- |
|
|
||||||
| Model Version | YOLOv5s |
|
|
||||||
| Resource | Ascend 310; CentOS 3.10 |
|
|
||||||
| Uploaded Date | 07/06/2021 (month/day/year) |
|
|
||||||
| MindSpore Version | 1.2.0 |
|
|
||||||
| Dataset | Coco2017 5K images |
|
|
||||||
| batch_size | 1 |
|
|
||||||
| outputs | box position and sorces, and probability |
|
| outputs | box position and sorces, and probability |
|
||||||
| Accuracy | Accuracy=0.71654 |
|
| Accuracy | mAP >= 36.7%(shape=640) |
|
||||||
| Model for inference | 58M(.ckpt file) |
|
| Model for inference | 56.67M (.ckpt file) |
|
||||||
|
|
||||||
|
### Transfer Learning
|
||||||
|
|
||||||
# [Description of Random Situation](#contents)
|
# [Description of Random Situation](#contents)
|
||||||
|
|
||||||
In dataset.py, we set the seed inside ```create_dataset``` function.
|
In dataset.py, we set the seed inside “create_dataset" function. We also use random seed in train.py.
|
||||||
In var_init.py, we set seed for weight initialization
|
|
||||||
|
|
||||||
# [ModelZoo Homepage](#contents)
|
# [ModelZoo Homepage](#contents)
|
||||||
|
|
||||||
Please check the official [homepage](https://gitee.com/mindspore/mindspore/tree/master/model_zoo).
|
Please check the official [homepage](https://gitee.com/mindspore/mindspore/tree/master/model_zoo).
|
||||||
|
|
|
||||||
|
|
@ -1,422 +0,0 @@
|
||||||
# 目录
|
|
||||||
|
|
||||||
- [YOLOv5说明](#yolov5说明)
|
|
||||||
- [模型架构](#模型架构)
|
|
||||||
- [数据集](#数据集)
|
|
||||||
- [环境要求](#环境要求)
|
|
||||||
- [快速入门](#快速入门)
|
|
||||||
- [脚本说明](#脚本说明)
|
|
||||||
- [脚本和示例代码](#脚本和示例代码)
|
|
||||||
- [脚本参数](#脚本参数)
|
|
||||||
- [训练过程](#训练过程)
|
|
||||||
- [训练](#训练)
|
|
||||||
- [测试过程](#测试过程)
|
|
||||||
- [测试](#测试)
|
|
||||||
- [评估过程](#评估过程)
|
|
||||||
- [评估](#评估)
|
|
||||||
- [推理过程](#推理过程)
|
|
||||||
- [导出MindIR](#导出mindir)
|
|
||||||
- [在Ascend310执行推理](#在ascend310执行推理)
|
|
||||||
- [结果](#结果)
|
|
||||||
- [模型说明](#模型说明)
|
|
||||||
- [性能](#性能)
|
|
||||||
- [评估性能](#评估性能)
|
|
||||||
- [推理性能](#推理性能)
|
|
||||||
- [310推理性能](#310推理性能)
|
|
||||||
- [ModelZoo主页](#modelzoo主页)
|
|
||||||
|
|
||||||
# [YOLOv5描述](#目录)
|
|
||||||
|
|
||||||
YOLOv5作为先进的检测器，它比所有可用的替代检测器更快（FPS）并且更准确（MS COCO AP50 ... 95和AP50）。
|
|
||||||
本文已经验证了大量的特征，并选择使用这些特征来提高分类和检测的精度。
|
|
||||||
这些特性可以作为未来研究和开发的最佳实践。
|
|
||||||
|
|
||||||
[代码](https://github.com/ultralytics/yolov5)
|
|
||||||
|
|
||||||
# [模型架构](#目录)
|
|
||||||
|
|
||||||
选择CSP Focus主干、SPP附加模块、PANet路径聚合网络和YOLOv5（基于锚点）头作为YOLOv5架构。
|
|
||||||
|
|
||||||
# [数据集](#目录)
|
|
||||||
|
|
||||||
支持的数据集：[MS COCO]或与MS COCO格式相同的数据集
|
|
||||||
支持的标注：[MS COCO]或与MS COCO相同格式的标注
|
|
||||||
|
|
||||||
- 目录结构如下，由用户定义目录和文件的名称：
|
|
||||||
|
|
||||||
```shell
|
|
||||||
├── dataset
|
|
||||||
├── YOLOv5
|
|
||||||
├── annotations
|
|
||||||
│ ├─ train.json
|
|
||||||
│ └─ val.json
|
|
||||||
├─ images
|
|
||||||
├─ train
|
|
||||||
│ └─images
|
|
||||||
│ ├─picture1.jpg
|
|
||||||
│ ├─ ...
|
|
||||||
│ └─picturen.jpg
|
|
||||||
└─ val
|
|
||||||
└─images
|
|
||||||
├─picture1.jpg
|
|
||||||
├─ ...
|
|
||||||
└─picturen.jpg
|
|
||||||
```
|
|
||||||
|
|
||||||
建议用户使用MS COCO数据集来体验模型，
|
|
||||||
其他数据集需要使用与MS COCO相同的格式。
|
|
||||||
|
|
||||||
# [环境要求](#目录)
|
|
||||||
|
|
||||||
- 硬件 Ascend
|
|
||||||
- 使用Ascend处理器准备硬件环境。
|
|
||||||
- 框架
|
|
||||||
- [MindSpore](https://www.mindspore.cn/install)
|
|
||||||
- 更多关于Mindspore的信息，请查看以下资源：
|
|
||||||
- [MindSpore教程](https://www.mindspore.cn/tutorials/zh-CN/master/index.html)
|
|
||||||
- [MindSpore API](https://www.mindspore.cn/docs/api/zh-CN/master/index.html)
|
|
||||||
|
|
||||||
# [快速入门](#目录)
|
|
||||||
|
|
||||||
通过官方网站安装MindSpore后，您可以按照如下步骤进行训练和评估：
|
|
||||||
|
|
||||||
``` shell
|
|
||||||
# training_shape参数定义网络图像形状，默认为[640, 640]。
|
|
||||||
```
|
|
||||||
|
|
||||||
```shell
|
|
||||||
# python命令执行训练示例（1卡）
|
|
||||||
python train.py \
|
|
||||||
--data_dir=./dataset/xxx \
|
|
||||||
--is_distributed=0 \
|
|
||||||
--lr=0.01 \
|
|
||||||
--T_max=320 \
|
|
||||||
--max_epoch=320 \
|
|
||||||
--warmup_epochs=4 \
|
|
||||||
--training_shape=640 \
|
|
||||||
--lr_scheduler=cosine_annealing > log.txt 2>&1 &
|
|
||||||
```
|
|
||||||
|
|
||||||
```shell
|
|
||||||
# shell脚本单机训练示例（1卡）
|
|
||||||
bash run_standalone_train.sh dataset/xxx
|
|
||||||
```
|
|
||||||
|
|
||||||
```shell
|
|
||||||
# 对于Ascend设备，使用shell脚本分布式训练示例（8卡）
|
|
||||||
bash run_distribute_train.sh dataset/xxx rank_table_8p.json
|
|
||||||
```
|
|
||||||
|
|
||||||
```python
|
|
||||||
# 使用python命令评估
|
|
||||||
python eval.py \
|
|
||||||
--data_dir=./dataset/xxx \
|
|
||||||
--pretrained=yolov5.ckpt \
|
|
||||||
--testing_shape=640 > log.txt 2>&1 &
|
|
||||||
```
|
|
||||||
|
|
||||||
```python
|
|
||||||
# shell脚本执行评估
|
|
||||||
bash run_eval.sh dataset/xxx checkpoint/xxx.ckpt
|
|
||||||
```
|
|
||||||
|
|
||||||
# [脚本说明](#目录)
|
|
||||||
|
|
||||||
## [脚本和示例代码](#目录)
|
|
||||||
|
|
||||||
```python
|
|
||||||
└─yolov5
|
|
||||||
├─README.md
|
|
||||||
├─mindspore_hub_conf.md # Mindspore Hub配置
|
|
||||||
├─ascend310_infer # 用于310推理
|
|
||||||
├─scripts
|
|
||||||
├─run_standalone_train.sh # 在Ascend中启动单机训练（1卡）
|
|
||||||
├─run_distribute_train.sh # 在Ascend中启动分布式训练（8卡）
|
|
||||||
├─run_infer_310.sh # 在Ascend中启动310推理
|
|
||||||
├─run_eval.sh # 在Ascend中启动评估
|
|
||||||
├─src
|
|
||||||
├─__init__.py # Python初始化文件
|
|
||||||
├─config.py # 参数配置
|
|
||||||
├─yolov5_backbone.py # 网络骨干
|
|
||||||
├─distributed_sampler.py # 数据集迭代器
|
|
||||||
├─initializer.py # 参数初始化器
|
|
||||||
├─logger.py # 日志函数
|
|
||||||
├─loss.py # 损失函数
|
|
||||||
├─lr_scheduler.py # 生成学习率
|
|
||||||
├─transforms.py # 预处理数据
|
|
||||||
├─util.py # 工具函数
|
|
||||||
├─yolo.py # YOLOv5网络
|
|
||||||
├─yolo_dataset.py # 为YOLOv5创建数据集
|
|
||||||
|
|
||||||
├─eval.py # 评估验证结果
|
|
||||||
├─export.py # 将MindSpore模型转换为AIR模型
|
|
||||||
├─preprocess.py # 310推理前处理脚本
|
|
||||||
├─postprocess.py # 310推理后处理脚本
|
|
||||||
├─train.py # 训练网络
|
|
||||||
```
|
|
||||||
|
|
||||||
## [脚本参数](#目录)
|
|
||||||
|
|
||||||
train.py中主要参数如下：
|
|
||||||
|
|
||||||
```shell
|
|
||||||
可选参数：
|
|
||||||
-h, --help 显示此帮助消息并退出
|
|
||||||
--device_target 实现代码的设备：“Ascend”（默认值）|“GPU”
|
|
||||||
--data_dir DATA_DIR 训练数据集目录
|
|
||||||
--per_batch_size PER_BATCH_SIZE
|
|
||||||
训练的批处理大小。 默认值：8。
|
|
||||||
--pretrained_backbone PRETRAINED_BACKBONE
|
|
||||||
YOLOv5主干文件。 默认值：""。
|
|
||||||
--resume_yolov5 RESUME_YOLOV5
|
|
||||||
YOLOv5的ckpt文件，用于微调。
|
|
||||||
默认值：""
|
|
||||||
--lr_scheduler LR_SCHEDULER
|
|
||||||
学习率调度器，取值选项：exponential，
|
|
||||||
cosine_annealing。 默认值：exponential
|
|
||||||
--lr LR 学习率。 默认值：0.01
|
|
||||||
--lr_epochs LR_EPOCHS
|
|
||||||
LR变化轮次，用“,”分隔。
|
|
||||||
默认值：220,250
|
|
||||||
--lr_gamma LR_GAMMA 将LR降低一个exponential lr_scheduler因子。
|
|
||||||
默认值：0.1
|
|
||||||
--eta_min ETA_MIN cosine_annealing调度器中的eta_min。 默认值：0
|
|
||||||
--T_max T_MAX cosine_annealing调度器中的T-max。 默认值：320
|
|
||||||
--max_epoch MAX_EPOCH
|
|
||||||
训练模型的最大轮次数。 默认值：320
|
|
||||||
--warmup_epochs WARMUP_EPOCHS
|
|
||||||
热身轮次。 默认值：0
|
|
||||||
--weight_decay WEIGHT_DECAY
|
|
||||||
权重衰减因子。 默认值：0.0005
|
|
||||||
--momentum MOMENTUM 动量。 默认值：0.9
|
|
||||||
--loss_scale LOSS_SCALE
|
|
||||||
静态损失尺度。 默认值：1024
|
|
||||||
--label_smooth LABEL_SMOOTH
|
|
||||||
CE中是否使用标签平滑。 默认值：0
|
|
||||||
--label_smooth_factor LABEL_SMOOTH_FACTOR
|
|
||||||
原one-hot的光滑强度。 默认值：0.1
|
|
||||||
--log_interval LOG_INTERVAL
|
|
||||||
日志记录间隔步数。 默认值：100
|
|
||||||
--ckpt_path CKPT_PATH
|
|
||||||
Checkpoint保存位置。 默认值：outputs/
|
|
||||||
--ckpt_interval CKPT_INTERVAL
|
|
||||||
保存checkpoint间隔。 默认值：None
|
|
||||||
--is_save_on_master IS_SAVE_ON_MASTER
|
|
||||||
在master或all rank上保存ckpt，1代表master，0代表
|
|
||||||
all ranks。 默认值：1
|
|
||||||
--is_distributed IS_DISTRIBUTED
|
|
||||||
是否分发训练，1代表是，0代表否。 默认值：
|
|
||||||
1
|
|
||||||
--rank RANK 分布式本地进程序号。 默认值：0
|
|
||||||
--group_size GROUP_SIZE
|
|
||||||
设备进程总数。 默认值：1
|
|
||||||
--need_profiler NEED_PROFILER
|
|
||||||
是否使用profiler。 0表示否，1表示是。 默认值：0
|
|
||||||
--training_shape TRAINING_SHAPE
|
|
||||||
恢复训练形状。 默认值：""
|
|
||||||
--resize_rate RESIZE_RATE
|
|
||||||
多尺度训练的缩放速率。 默认值：None
|
|
||||||
```
|
|
||||||
|
|
||||||
## [训练过程](#目录)
|
|
||||||
|
|
||||||
### 训练
|
|
||||||
|
|
||||||
```python
|
|
||||||
python train.py \
|
|
||||||
--data_dir=/dataset/xxx \
|
|
||||||
--is_distributed=0 \
|
|
||||||
--lr=0.01 \
|
|
||||||
--T_max=320 \
|
|
||||||
--max_epoch=320 \
|
|
||||||
--warmup_epochs=4 \
|
|
||||||
--training_shape=640 \
|
|
||||||
--lr_scheduler=cosine_annealing > log.txt 2>&1 &
|
|
||||||
```
|
|
||||||
|
|
||||||
上述python命令将在后台运行，您可以通过log.txt文件查看结果。
|
|
||||||
|
|
||||||
训练结束后，您可在默认输出文件夹下找到checkpoint文件。 得到如下损失值：
|
|
||||||
|
|
||||||
```shell
|
|
||||||
# grep "loss:" train/log.txt
|
|
||||||
2021-05-13 20:50:25,617:INFO:epoch[0], iter[100], loss:loss:2648.764910, fps:61.59 imgs/sec, lr:1.7226087948074564e-05
|
|
||||||
2021-05-13 20:50:39,821:INFO:epoch[0], iter[200], loss:loss:764.535622, fps:56.33 imgs/sec, lr:3.4281620173715055e-05
|
|
||||||
2021-05-13 20:50:53,287:INFO:epoch[0], iter[300], loss:loss:494.950782, fps:59.47 imgs/sec, lr:5.1337152399355546e-05
|
|
||||||
2021-05-13 20:51:06,138:INFO:epoch[0], iter[400], loss:loss:393.339678, fps:62.25 imgs/sec, lr:6.839268462499604e-05
|
|
||||||
2021-05-13 20:51:17,985:INFO:epoch[0], iter[500], loss:loss:329.976604, fps:67.57 imgs/sec, lr:8.544822048861533e-05
|
|
||||||
2021-05-13 20:51:29,359:INFO:epoch[0], iter[600], loss:loss:294.734397, fps:70.37 imgs/sec, lr:0.00010250374907627702
|
|
||||||
2021-05-13 20:51:40,634:INFO:epoch[0], iter[700], loss:loss:281.497078, fps:70.98 imgs/sec, lr:0.00011955928493989632
|
|
||||||
2021-05-13 20:51:52,307:INFO:epoch[0], iter[800], loss:loss:264.300707, fps:68.54 imgs/sec, lr:0.0001366148208035156
|
|
||||||
2021-05-13 20:52:05,479:INFO:epoch[0], iter[900], loss:loss:261.971103, fps:60.76 imgs/sec, lr:0.0001536703493911773
|
|
||||||
2021-05-13 20:52:17,362:INFO:epoch[0], iter[1000], loss:loss:264.591175, fps:67.33 imgs/sec, lr:0.00017072587797883898
|
|
||||||
...
|
|
||||||
```
|
|
||||||
|
|
||||||
### 分布式训练
|
|
||||||
|
|
||||||
对于Ascend设备，使用shell脚本分布式训练示例（8卡）
|
|
||||||
|
|
||||||
```shell
|
|
||||||
bash run_distribute_train.sh dataset/coco2017 rank_table_8p.json
|
|
||||||
```
|
|
||||||
|
|
||||||
上述shell脚本将在后台运行分布式训练。 您可以通过train_parallel[X]/log.txt文件查看结果。 得到如下损失值：
|
|
||||||
|
|
||||||
```shell
|
|
||||||
# 分布式训练示例（8卡）
|
|
||||||
...
|
|
||||||
2021-05-13 21:08:41,992:INFO:epoch[0], iter[600], loss:247.577421, fps:469.29 imgs/sec, lr:0.0001640283880988136
|
|
||||||
2021-05-13 21:08:56,291:INFO:epoch[0], iter[700], loss:235.298894, fps:447.67 imgs/sec, lr:0.0001913209562189877
|
|
||||||
2021-05-13 21:09:10,431:INFO:epoch[0], iter[800], loss:239.481037, fps:452.78 imgs/sec, lr:0.00021861353889107704
|
|
||||||
2021-05-13 21:09:23,517:INFO:epoch[0], iter[900], loss:232.826709, fps:489.15 imgs/sec, lr:0.0002459061215631664
|
|
||||||
2021-05-13 21:09:36,407:INFO:epoch[0], iter[1000], loss:224.734599, fps:496.65 imgs/sec, lr:0.0002731987042352557
|
|
||||||
2021-05-13 21:09:49,072:INFO:epoch[0], iter[1100], loss:232.334771, fps:505.34 imgs/sec, lr:0.0003004912578035146
|
|
||||||
2021-05-13 21:10:03,597:INFO:epoch[0], iter[1200], loss:242.001476, fps:440.69 imgs/sec, lr:0.00032778384047560394
|
|
||||||
2021-05-13 21:10:18,237:INFO:epoch[0], iter[1300], loss:225.391021, fps:437.20 imgs/sec, lr:0.0003550764231476933
|
|
||||||
2021-05-13 21:10:33,027:INFO:epoch[0], iter[1400], loss:228.738176, fps:432.76 imgs/sec, lr:0.0003823690058197826
|
|
||||||
2021-05-13 21:10:47,424:INFO:epoch[0], iter[1500], loss:225.712950, fps:444.54 imgs/sec, lr:0.0004096615593880415
|
|
||||||
2021-05-13 21:11:02,077:INFO:epoch[0], iter[1600], loss:221.249353, fps:436.77 imgs/sec, lr:0.00043695414206013083
|
|
||||||
2021-05-13 21:11:16,631:INFO:epoch[0], iter[1700], loss:222.449119, fps:439.89 imgs/sec, lr:0.00046424672473222017
|
|
||||||
...
|
|
||||||
```
|
|
||||||
|
|
||||||
## [评估过程](#目录)
|
|
||||||
|
|
||||||
### 验证
|
|
||||||
|
|
||||||
```python
|
|
||||||
python eval.py \
|
|
||||||
--data_dir=./dataset/coco2017 \
|
|
||||||
--pretrained=yolov5.ckpt \
|
|
||||||
--testing_shape=640 > log.txt 2>&1 &
|
|
||||||
OR
|
|
||||||
bash run_eval.sh dataset/coco2017 checkpoint/yolov5.ckpt
|
|
||||||
```
|
|
||||||
|
|
||||||
上述python命令将在后台运行。 您可以通过log.txt文件查看结果。 测试数据集的mAP如下：
|
|
||||||
|
|
||||||
```shell
|
|
||||||
# log.txt
|
|
||||||
=============coco eval reulst=========
|
|
||||||
Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.372
|
|
||||||
Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.574
|
|
||||||
Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.403
|
|
||||||
Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.219
|
|
||||||
Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.426
|
|
||||||
Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.480
|
|
||||||
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 1 ] = 0.302
|
|
||||||
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 10 ] = 0.504
|
|
||||||
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.560
|
|
||||||
Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.399
|
|
||||||
Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.619
|
|
||||||
Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.674
|
|
||||||
```
|
|
||||||
|
|
||||||
## [推理过程](#目录)
|
|
||||||
|
|
||||||
### 导出MindIR
|
|
||||||
|
|
||||||
```shell
|
|
||||||
python export.py --ckpt_file [CKPT_PATH] --file_format [EXPORT_FORMAT] --batch_size [BATCH_SIZE]
|
|
||||||
```
|
|
||||||
|
|
||||||
参数ckpt_file为必填项，
|
|
||||||
`EXPORT_FORMAT` 必须在 ["AIR", "MINDIR"]中选择。
|
|
||||||
`BATCH_SIZE` 目前仅支持batch_size为1的推理。
|
|
||||||
|
|
||||||
### 在Ascend310执行推理
|
|
||||||
|
|
||||||
在执行推理前，mindir文件必须通过`export.py`脚本导出。以下展示了使用mindir模型执行推理的示例。
|
|
||||||
|
|
||||||
```shell
|
|
||||||
# Ascend310 inference
|
|
||||||
bash run_infer_310.sh [MINDIR_PATH] [DATA_PATH] [ANN_FILE] [DVPP] [DEVICE_ID]
|
|
||||||
```
|
|
||||||
|
|
||||||
- `ANN_FILE` Annotations 文件路径。
|
|
||||||
- `DVPP` 为必填项，需要在["DVPP", "CPU"]选择，大小写均可。目前仅支持CPU算子推理。
|
|
||||||
- `DEVICE_ID` 可选，默认值为0。
|
|
||||||
|
|
||||||
### 结果
|
|
||||||
|
|
||||||
推理结果保存在脚本执行的当前路径，你可以在acc.log中看到以下精度计算结果。
|
|
||||||
|
|
||||||
```bash
|
|
||||||
=============coco 310 infer reulst=========
|
|
||||||
Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.369
|
|
||||||
Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.571
|
|
||||||
Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.398
|
|
||||||
Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.216
|
|
||||||
Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.421
|
|
||||||
Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.487
|
|
||||||
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 1 ] = 0.301
|
|
||||||
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 10 ] = 0.502
|
|
||||||
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.558
|
|
||||||
Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.388
|
|
||||||
Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.617
|
|
||||||
Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.677
|
|
||||||
```
|
|
||||||
|
|
||||||
# [模型说明](#目录)
|
|
||||||
|
|
||||||
## [性能](#目录)
|
|
||||||
|
|
||||||
### 评估性能
|
|
||||||
|
|
||||||
YOLOv5应用于118000张图像上（标注和数据格式必须与COCO 2017相同）
|
|
||||||
|
|
||||||
|参数| YOLOv5s |
|
|
||||||
| -------------------------- | ----------------------------------------------------------- |
|
|
||||||
| 资源 | Ascend 910；CPU 2.60GHz，192核；内存：755G |
|
|
||||||
|上传日期| 2021年05月14日 |
|
|
||||||
| MindSpore版本|1.0.0-alpha|
|
|
||||||
|数据集|118000张图像|
|
|
||||||
|训练参数|epoch=320, batch_size=8, lr=0.01, momentum=0.9|
|
|
||||||
| 优化器 | Momentum |
|
|
||||||
|损失函数|Sigmoid Cross Entropy with logits, Giou Loss|
|
|
||||||
|输出|heatmaps |
|
|
||||||
| 损失 | 53 |
|
|
||||||
|速度| 1卡：55 img/s；8卡：440 img/s（shape=640）|
|
|
||||||
| 总时长 | 24小时(8卡) |
|
|
||||||
| 微调检查点 | 58M （.ckpt文件） |
|
|
||||||
|脚本| <https://gitee.com/mindspore/mindspore/tree/master/model_zoo/> |
|
|
||||||
|
|
||||||
### 推理性能
|
|
||||||
|
|
||||||
YOLOv5应用于5000张图像上（标注和数据格式必须与COCO val 2017相同）
|
|
||||||
|
|
||||||
|参数| YOLOv5s |
|
|
||||||
| -------------------------- | ----------------------------------------------------------- |
|
|
||||||
| 资源 | Ascend 910；CPU 2.60GHz，192核；内存：755G |
|
|
||||||
|上传日期| 2021年05月14日 |
|
|
||||||
| MindSpore版本 | 1.2.0 |
|
|
||||||
|数据集|5000张图像|
|
|
||||||
|批处理大小|1|
|
|
||||||
|输出|边框位置和分数，以及概率|
|
|
||||||
|精度|map=36.8~37.2%（shape=640）|
|
|
||||||
|推理模型| 58M（.ckpt文件）|
|
|
||||||
|
|
||||||
### 310推理性能
|
|
||||||
|
|
||||||
YOLOv5应用于5000张图像上（标注和数据格式必须与COCO val 2017相同）
|
|
||||||
|
|
||||||
|参数| YOLOv5s |
|
|
||||||
| -------------------------- | ----------------------------------------------------------- |
|
|
||||||
| 资源 | Ascend 310；CPU 2.60GHz，192核；内存：755G |
|
|
||||||
|上传日期| 2021年06月28日 |
|
|
||||||
| MindSpore版本 | 1.2.0 |
|
|
||||||
|数据集|5000张图像|
|
|
||||||
|批处理大小|1|
|
|
||||||
|输出|边框位置和分数，以及概率|
|
|
||||||
|精度|map=36.9%（shape=640）|
|
|
||||||
|推理模型| 58M（.ckpt文件）|
|
|
||||||
|
|
||||||
# [随机情况说明](#目录)
|
|
||||||
|
|
||||||
在dataset.py中，我们设置了“create_dataset”函数内的种子。
|
|
||||||
在var_init.py中，我们设置了权重初始化的种子。
|
|
||||||
|
|
||||||
# [ModelZoo主页](#目录)
|
|
||||||
|
|
||||||
请浏览官网[主页](https://gitee.com/mindspore/mindspore/tree/master/model_zoo)。
|
|
||||||
|
|
@ -31,7 +31,7 @@ from mindspore import context
|
||||||
from mindspore.train.serialization import load_checkpoint, load_param_into_net
|
from mindspore.train.serialization import load_checkpoint, load_param_into_net
|
||||||
import mindspore as ms
|
import mindspore as ms
|
||||||
|
|
||||||
from src.yolo import YOLOV5s
|
from src.yolo import YOLOV5
|
||||||
from src.logger import get_logger
|
from src.logger import get_logger
|
||||||
from src.yolo_dataset import create_yolo_dataset
|
from src.yolo_dataset import create_yolo_dataset
|
||||||
from src.config import ConfigYOLOV5
|
from src.config import ConfigYOLOV5
|
||||||
|
|
@ -43,11 +43,13 @@ parser.add_argument('--device_target', type=str, default='Ascend',
|
||||||
help='device where the code will be implemented. (Default: Ascend)')
|
help='device where the code will be implemented. (Default: Ascend)')
|
||||||
|
|
||||||
# dataset related
|
# dataset related
|
||||||
parser.add_argument('--data_dir', type=str, default='', help='train data dir')
|
parser.add_argument('--data_dir', type=str, default='/data/coco', help='train data dir')
|
||||||
parser.add_argument('--per_batch_size', default=1, type=int, help='batch size for per gpu')
|
parser.add_argument('--per_batch_size', default=1, type=int, help='batch size for per gpu')
|
||||||
|
|
||||||
# network related
|
# network related
|
||||||
parser.add_argument('--pretrained', default='', type=str, help='model_path, local pretrained model to load')
|
parser.add_argument('--pretrained', default='', type=str, help='model_path, local pretrained model to load')
|
||||||
|
parser.add_argument('--yolov5_version', default='yolov5s', type=str,
|
||||||
|
help='The version of YOLOv5, options: yolov5s, yolov5m, yolov5l, yolov5x')
|
||||||
|
|
||||||
# logging related
|
# logging related
|
||||||
parser.add_argument('--log_path', type=str, default='outputs/', help='checkpoint save location')
|
parser.add_argument('--log_path', type=str, default='outputs/', help='checkpoint save location')
|
||||||
|
|
@ -59,11 +61,37 @@ parser.add_argument('--testing_shape', type=str, default='', help='shape for tes
|
||||||
parser.add_argument('--ignore_threshold', type=float, default=0.001, help='threshold to throw low quality boxes')
|
parser.add_argument('--ignore_threshold', type=float, default=0.001, help='threshold to throw low quality boxes')
|
||||||
parser.add_argument('--multi_label', type=ast.literal_eval, default=True, help='whether to use multi label')
|
parser.add_argument('--multi_label', type=ast.literal_eval, default=True, help='whether to use multi label')
|
||||||
parser.add_argument('--multi_label_thresh', type=float, default=0.1, help='threshhold to throw low quality boxes')
|
parser.add_argument('--multi_label_thresh', type=float, default=0.1, help='threshhold to throw low quality boxes')
|
||||||
|
parser.add_argument('--is_modelArts', type=int, default=0,
|
||||||
|
help='Trainning in modelArts or not, 1 for yes, 0 for no. Default: 0')
|
||||||
|
|
||||||
args, _ = parser.parse_known_args()
|
args, _ = parser.parse_known_args()
|
||||||
|
args.rank = 0
|
||||||
|
|
||||||
args.data_root = os.path.join(args.data_dir, 'val2017')
|
if args.is_modelArts:
|
||||||
args.ann_file = os.path.join(args.data_dir, 'annotations/instances_val2017.json')
|
args.data_root = os.path.join(args.data_dir, 'val2017')
|
||||||
|
args.ann_file = os.path.join(args.data_dir, 'annotations')
|
||||||
|
import moxing as mox
|
||||||
|
|
||||||
|
local_data_url = os.path.join('/cache/data', str(args.rank))
|
||||||
|
local_annFile = os.path.join('/cache/data', str(args.rank))
|
||||||
|
local_pretrained = os.path.join('/cache/data', str(args.rank))
|
||||||
|
|
||||||
|
temp_str = args.pretrained.split('/')[-1]
|
||||||
|
args.pretrained = args.pretrained[0:args.pretrained.rfind('/')]
|
||||||
|
|
||||||
|
mox.file.copy_parallel(args.data_root, local_data_url)
|
||||||
|
args.data_root = local_data_url
|
||||||
|
|
||||||
|
mox.file.copy_parallel(args.ann_file, local_annFile)
|
||||||
|
args.ann_file = os.path.join(local_data_url, 'instances_val2017.json')
|
||||||
|
|
||||||
|
mox.file.copy_parallel(args.pretrained, local_pretrained)
|
||||||
|
args.pretrained = os.path.join(local_data_url, temp_str)
|
||||||
|
else:
|
||||||
|
args.data_root = os.path.join(args.data_dir, 'val2017')
|
||||||
|
args.ann_file = os.path.join(
|
||||||
|
args.data_dir,
|
||||||
|
'annotations/instances_val2017.json')
|
||||||
|
|
||||||
|
|
||||||
class Redirct:
|
class Redirct:
|
||||||
|
|
@ -103,7 +131,7 @@ class DetectionEngine:
|
||||||
self.nms_thresh = args_detection.nms_thresh
|
self.nms_thresh = args_detection.nms_thresh
|
||||||
self.multi_label = args_detection.multi_label
|
self.multi_label = args_detection.multi_label
|
||||||
self.multi_label_thresh = args_detection.multi_label_thresh
|
self.multi_label_thresh = args_detection.multi_label_thresh
|
||||||
# self.coco_catids = self._coco.getCatIds()
|
self.coco_catids = self._coco.getCatIds()
|
||||||
self.coco_catIds = [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 27,
|
self.coco_catIds = [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 27,
|
||||||
28, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 46, 47, 48, 49, 50, 51, 52, 53,
|
28, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 46, 47, 48, 49, 50, 51, 52, 53,
|
||||||
54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 67, 70, 72, 73, 74, 75, 76, 77, 78, 79, 80,
|
54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 67, 70, 72, 73, 74, 75, 76, 77, 78, 79, 80,
|
||||||
|
|
@ -111,18 +139,15 @@ class DetectionEngine:
|
||||||
|
|
||||||
def do_nms_for_results(self):
|
def do_nms_for_results(self):
|
||||||
"""Get result boxes."""
|
"""Get result boxes."""
|
||||||
# np.save('/opt/disk1/hjc/yolov5_positive_policy/result.npy', self.results)
|
|
||||||
for img_id in self.results:
|
for img_id in self.results:
|
||||||
for clsi in self.results[img_id]:
|
for clsi in self.results[img_id]:
|
||||||
dets = self.results[img_id][clsi]
|
dets = self.results[img_id][clsi]
|
||||||
dets = np.array(dets)
|
dets = np.array(dets)
|
||||||
keep_index = self._diou_nms(dets, thresh=self.nms_thresh)
|
keep_index = self._diou_nms(dets, thresh=self.nms_thresh)
|
||||||
|
|
||||||
keep_box = [{'image_id': int(img_id),
|
keep_box = [{'image_id': int(img_id), 'category_id': int(clsi),
|
||||||
'category_id': int(clsi),
|
|
||||||
'bbox': list(dets[i][:4].astype(float)),
|
'bbox': list(dets[i][:4].astype(float)),
|
||||||
'score': dets[i][4].astype(float)}
|
'score': dets[i][4].astype(float)} for i in keep_index]
|
||||||
for i in keep_index]
|
|
||||||
self.det_boxes.extend(keep_box)
|
self.det_boxes.extend(keep_box)
|
||||||
|
|
||||||
def _nms(self, predicts, threshold):
|
def _nms(self, predicts, threshold):
|
||||||
|
|
@ -149,7 +174,8 @@ class DetectionEngine:
|
||||||
intersect_w = np.maximum(0.0, min_x2 - max_x1 + 1)
|
intersect_w = np.maximum(0.0, min_x2 - max_x1 + 1)
|
||||||
intersect_h = np.maximum(0.0, min_y2 - max_y1 + 1)
|
intersect_h = np.maximum(0.0, min_y2 - max_y1 + 1)
|
||||||
intersect_area = intersect_w * intersect_h
|
intersect_area = intersect_w * intersect_h
|
||||||
ovr = intersect_area / (areas[i] + areas[order[1:]] - intersect_area)
|
ovr = intersect_area / \
|
||||||
|
(areas[i] + areas[order[1:]] - intersect_area)
|
||||||
|
|
||||||
indexes = np.where(ovr <= threshold)[0]
|
indexes = np.where(ovr <= threshold)[0]
|
||||||
order = order[indexes + 1]
|
order = order[indexes + 1]
|
||||||
|
|
@ -289,8 +315,8 @@ class DetectionEngine:
|
||||||
flag[i, c] = True
|
flag[i, c] = True
|
||||||
confidence = cls_emb[flag] * conf
|
confidence = cls_emb[flag] * conf
|
||||||
|
|
||||||
for x_lefti, y_lefti, wi, hi, confi, clsi in zip(x_top_left, y_top_left, w, h, confidence,
|
for x_lefti, y_lefti, wi, hi, confi, clsi in zip(x_top_left, y_top_left,
|
||||||
cls_argmax):
|
w, h, confidence, cls_argmax):
|
||||||
if confi < self.ignore_threshold:
|
if confi < self.ignore_threshold:
|
||||||
continue
|
continue
|
||||||
if img_id not in self.results:
|
if img_id not in self.results:
|
||||||
|
|
@ -317,9 +343,8 @@ if __name__ == "__main__":
|
||||||
context.set_context(mode=context.GRAPH_MODE, device_target=args.device_target, device_id=device_id)
|
context.set_context(mode=context.GRAPH_MODE, device_target=args.device_target, device_id=device_id)
|
||||||
|
|
||||||
# logger
|
# logger
|
||||||
args.outputs_dir = os.path.join(args.log_path,
|
args.outputs_dir = os.path.join(args.log_path, datetime.datetime.now().strftime('%Y-%m-%d_time_%H_%M_%S'))
|
||||||
datetime.datetime.now().strftime('%Y-%m-%d_time_%H_%M_%S'))
|
rank_id = int(os.getenv('DEVICE_ID', '0'))
|
||||||
rank_id = int(os.environ.get('RANK_ID')) if os.environ.get('RANK_ID') else 0
|
|
||||||
args.logger = get_logger(args.outputs_dir, rank_id)
|
args.logger = get_logger(args.outputs_dir, rank_id)
|
||||||
|
|
||||||
context.reset_auto_parallel_context()
|
context.reset_auto_parallel_context()
|
||||||
|
|
@ -327,7 +352,8 @@ if __name__ == "__main__":
|
||||||
context.set_auto_parallel_context(parallel_mode=parallel_mode, gradients_mean=True, device_num=1)
|
context.set_auto_parallel_context(parallel_mode=parallel_mode, gradients_mean=True, device_num=1)
|
||||||
|
|
||||||
args.logger.info('Creating Network....')
|
args.logger.info('Creating Network....')
|
||||||
network = YOLOV5s(is_training=False)
|
dict_version = {'yolov5s': 0, 'yolov5m': 1, 'yolov5l': 2, 'yolov5x': 3}
|
||||||
|
network = YOLOV5(is_training=False, version=dict_version[args.yolov5_version])
|
||||||
|
|
||||||
args.logger.info(args.pretrained)
|
args.logger.info(args.pretrained)
|
||||||
if os.path.isfile(args.pretrained):
|
if os.path.isfile(args.pretrained):
|
||||||
|
|
@ -355,8 +381,7 @@ if __name__ == "__main__":
|
||||||
config.test_img_shape = convert_testing_shape(args.testing_shape)
|
config.test_img_shape = convert_testing_shape(args.testing_shape)
|
||||||
|
|
||||||
ds, data_size = create_yolo_dataset(data_root, ann_file, is_training=False, batch_size=args.per_batch_size,
|
ds, data_size = create_yolo_dataset(data_root, ann_file, is_training=False, batch_size=args.per_batch_size,
|
||||||
max_epoch=1, device_num=1, rank=rank_id, shuffle=False,
|
max_epoch=1, device_num=1, rank=rank_id, shuffle=False, config=config)
|
||||||
config=config)
|
|
||||||
|
|
||||||
args.logger.info('testing shape : {}'.format(config.test_img_shape))
|
args.logger.info('testing shape : {}'.format(config.test_img_shape))
|
||||||
args.logger.info('total {} images to eval'.format(data_size))
|
args.logger.info('total {} images to eval'.format(data_size))
|
||||||
|
|
|
||||||
|
|
@ -18,18 +18,21 @@ import numpy as np
|
||||||
import mindspore
|
import mindspore
|
||||||
from mindspore import context, Tensor
|
from mindspore import context, Tensor
|
||||||
from mindspore.train.serialization import export, load_checkpoint, load_param_into_net
|
from mindspore.train.serialization import export, load_checkpoint, load_param_into_net
|
||||||
from src.config import ConfigYOLOV5
|
|
||||||
from src.yolo import YOLOV5s_Infer
|
from src.yolo import YOLOV5s_Infer
|
||||||
|
|
||||||
parser = argparse.ArgumentParser(description='yolov5 export')
|
parser = argparse.ArgumentParser(description='yolov5 export')
|
||||||
parser.add_argument("--device_id", type=int, default=0, help="Device id")
|
parser.add_argument("--device_id", type=int, default=0, help="Device id")
|
||||||
parser.add_argument("--batch_size", type=int, default=1, help="batch size")
|
parser.add_argument("--batch_size", type=int, default=1, help="batch size")
|
||||||
|
parser.add_argument('--yolov5_version', default='yolov5s', type=str,
|
||||||
|
help='The version of YOLOv5, options: yolov5s, yolov5m, yolov5l, yolov5x')
|
||||||
|
|
||||||
parser.add_argument("--testing_shape", type=int, default=640, help="test shape")
|
parser.add_argument("--testing_shape", type=int, default=640, help="test shape")
|
||||||
parser.add_argument("--ckpt_file", type=str, required=True, help="Checkpoint file path.")
|
parser.add_argument("--ckpt_file", type=str, required=True, help="Checkpoint file path.")
|
||||||
parser.add_argument("--file_name", type=str, default="yolov5", help="output file name.")
|
parser.add_argument("--file_name", type=str, default="yolov5", help="output file name.")
|
||||||
parser.add_argument('--file_format', type=str, choices=["AIR", "MINDIR"], default='AIR', help='file format')
|
parser.add_argument('--file_format', type=str, choices=["AIR", "ONNX", "MINDIR"], default='MINDIR', help='file format')
|
||||||
parser.add_argument("--device_target", type=str, choices=["Ascend", "GPU", "CPU"], default="Ascend",
|
parser.add_argument("--device_target", type=str, choices=["Ascend", "GPU", "CPU"],
|
||||||
help="device target")
|
default="Ascend", help="device target")
|
||||||
args = parser.parse_args()
|
args = parser.parse_args()
|
||||||
|
|
||||||
context.set_context(mode=context.GRAPH_MODE, device_target=args.device_target)
|
context.set_context(mode=context.GRAPH_MODE, device_target=args.device_target)
|
||||||
|
|
@ -37,16 +40,18 @@ if args.device_target == "Ascend":
|
||||||
context.set_context(device_id=args.device_id)
|
context.set_context(device_id=args.device_id)
|
||||||
|
|
||||||
if __name__ == "__main__":
|
if __name__ == "__main__":
|
||||||
config = ConfigYOLOV5()
|
ts_shape = args.testing_shape // 2
|
||||||
if args.testing_shape:
|
|
||||||
config.test_img_shape = [int(args.testing_shape), int(args.testing_shape)]
|
|
||||||
ts_shape = config.test_img_shape[0]
|
|
||||||
|
|
||||||
network = YOLOV5s_Infer(config.test_img_shape)
|
dict_version = {'yolov5s': 0, 'yolov5m': 1, 'yolov5l': 2, 'yolov5x': 3}
|
||||||
|
args.file_name = args.file_name + '_' + args.yolov5_version
|
||||||
|
|
||||||
|
network = YOLOV5s_Infer(args.testing_shape, version=dict_version[args.yolov5_version])
|
||||||
|
network.set_train(False)
|
||||||
|
|
||||||
param_dict = load_checkpoint(args.ckpt_file)
|
param_dict = load_checkpoint(args.ckpt_file)
|
||||||
load_param_into_net(network, param_dict)
|
load_param_into_net(network, param_dict)
|
||||||
|
|
||||||
input_data = Tensor(np.zeros([args.batch_size, 12, int(ts_shape/2), int(ts_shape/2)]), mindspore.float32)
|
input_data = Tensor(np.zeros([args.batch_size, 12, ts_shape, ts_shape]), mindspore.float32)
|
||||||
|
|
||||||
export(network, input_data, file_name=args.file_name, file_format=args.file_format)
|
export(network, input_data, file_name=args.file_name, file_format=args.file_format)
|
||||||
|
print('==========success export===============')
|
||||||
|
|
|
||||||
|
|
@ -83,7 +83,6 @@ class DetectionEngine:
|
||||||
self.nms_thresh = args_detection.nms_thresh
|
self.nms_thresh = args_detection.nms_thresh
|
||||||
self.multi_label = args_detection.multi_label
|
self.multi_label = args_detection.multi_label
|
||||||
self.multi_label_thresh = args_detection.multi_label_thresh
|
self.multi_label_thresh = args_detection.multi_label_thresh
|
||||||
# self.coco_catids = self._coco.getCatIds()
|
|
||||||
self.coco_catIds = [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 27,
|
self.coco_catIds = [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 27,
|
||||||
28, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 46, 47, 48, 49, 50, 51, 52, 53,
|
28, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 46, 47, 48, 49, 50, 51, 52, 53,
|
||||||
54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 67, 70, 72, 73, 74, 75, 76, 77, 78, 79, 80,
|
54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 67, 70, 72, 73, 74, 75, 76, 77, 78, 79, 80,
|
||||||
|
|
@ -91,18 +90,15 @@ class DetectionEngine:
|
||||||
|
|
||||||
def do_nms_for_results(self):
|
def do_nms_for_results(self):
|
||||||
"""Get result boxes."""
|
"""Get result boxes."""
|
||||||
# np.save('/opt/disk1/hjc/yolov5_positive_policy/result.npy', self.results)
|
|
||||||
for image_id in self.results:
|
for image_id in self.results:
|
||||||
for clsi in self.results[image_id]:
|
for clsi in self.results[image_id]:
|
||||||
dets = self.results[image_id][clsi]
|
dets = self.results[image_id][clsi]
|
||||||
dets = np.array(dets)
|
dets = np.array(dets)
|
||||||
keep_index = self._diou_nms(dets, thresh=self.nms_thresh)
|
keep_index = self._diou_nms(dets, thresh=self.nms_thresh)
|
||||||
|
|
||||||
keep_box = [{'image_id': int(image_id),
|
keep_box = [{'image_id': int(image_id), 'category_id': int(clsi),
|
||||||
'category_id': int(clsi),
|
|
||||||
'bbox': list(dets[i][:4].astype(float)),
|
'bbox': list(dets[i][:4].astype(float)),
|
||||||
'score': dets[i][4].astype(float)}
|
'score': dets[i][4].astype(float)} for i in keep_index]
|
||||||
for i in keep_index]
|
|
||||||
self.det_boxes.extend(keep_box)
|
self.det_boxes.extend(keep_box)
|
||||||
|
|
||||||
def _nms(self, predicts, threshold):
|
def _nms(self, predicts, threshold):
|
||||||
|
|
|
||||||
|
|
@ -63,6 +63,7 @@ do
|
||||||
python train.py \
|
python train.py \
|
||||||
--data_dir=$DATASET_PATH \
|
--data_dir=$DATASET_PATH \
|
||||||
--is_distributed=1 \
|
--is_distributed=1 \
|
||||||
|
--yolov5_version='yolov5s' \
|
||||||
--lr=0.02 \
|
--lr=0.02 \
|
||||||
--T_max=300 \
|
--T_max=300 \
|
||||||
--max_epoch=300 \
|
--max_epoch=300 \
|
||||||
|
|
|
||||||
|
|
@ -61,6 +61,7 @@ env > env.log
|
||||||
echo "start inferring for device $DEVICE_ID"
|
echo "start inferring for device $DEVICE_ID"
|
||||||
python eval.py \
|
python eval.py \
|
||||||
--data_dir=$DATASET_PATH \
|
--data_dir=$DATASET_PATH \
|
||||||
|
--yolov5_version='yolov5s' \
|
||||||
--pretrained=$CHECKPOINT_PATH \
|
--pretrained=$CHECKPOINT_PATH \
|
||||||
--testing_shape=640 > log.txt 2>&1 &
|
--testing_shape=640 > log.txt 2>&1 &
|
||||||
cd ..
|
cd ..
|
||||||
|
|
|
||||||
|
|
@ -58,6 +58,7 @@ env > env.log
|
||||||
python train.py \
|
python train.py \
|
||||||
--data_dir=$DATASET_PATH \
|
--data_dir=$DATASET_PATH \
|
||||||
--is_distributed=0 \
|
--is_distributed=0 \
|
||||||
|
--yolov5_version='yolov5s' \
|
||||||
--lr=0.01 \
|
--lr=0.01 \
|
||||||
--T_max=320 \
|
--T_max=320 \
|
||||||
--max_epoch=320 \
|
--max_epoch=320 \
|
||||||
|
|
|
||||||
|
|
@ -0,0 +1,172 @@
|
||||||
|
# Copyright 2021 Huawei Technologies Co., Ltd
|
||||||
|
#
|
||||||
|
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||||
|
# you may not use this file except in compliance with the License.
|
||||||
|
# You may obtain a copy of the License at
|
||||||
|
#
|
||||||
|
# http://www.apache.org/licenses/LICENSE-2.0
|
||||||
|
#
|
||||||
|
# Unless required by applicable law or agreed to in writing, software
|
||||||
|
# distributed under the License is distributed on an "AS IS" BASIS,
|
||||||
|
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||||
|
# See the License for the specific language governing permissions and
|
||||||
|
# limitations under the License.
|
||||||
|
# ============================================================================
|
||||||
|
"""DarkNet model."""
|
||||||
|
import mindspore.nn as nn
|
||||||
|
from mindspore.ops import operations as P
|
||||||
|
|
||||||
|
|
||||||
|
class Bottleneck(nn.Cell):
|
||||||
|
# Standard bottleneck
|
||||||
|
# ch_in, ch_out, shortcut, groups, expansion
|
||||||
|
def __init__(self, c1, c2, shortcut=True, e=0.5):
|
||||||
|
super(Bottleneck, self).__init__()
|
||||||
|
c_ = int(c2 * e) # hidden channels
|
||||||
|
self.conv1 = Conv(c1, c_, 1, 1)
|
||||||
|
self.conv2 = Conv(c_, c2, 3, 1)
|
||||||
|
self.add = shortcut and c1 == c2
|
||||||
|
|
||||||
|
def construct(self, x):
|
||||||
|
c1 = self.conv1(x)
|
||||||
|
c2 = self.conv2(c1)
|
||||||
|
out = c2
|
||||||
|
if self.add:
|
||||||
|
out = x + out
|
||||||
|
return out
|
||||||
|
|
||||||
|
|
||||||
|
class BottleneckCSP(nn.Cell):
|
||||||
|
# CSP Bottleneck with 3 convolutions
|
||||||
|
def __init__(self, c1, c2, n=1, shortcut=True, e=0.5):
|
||||||
|
super(BottleneckCSP, self).__init__()
|
||||||
|
c_ = int(c2 * e) # hidden channels
|
||||||
|
self.conv1 = Conv(c1, c_, 1, 1)
|
||||||
|
self.conv2 = Conv(c1, c_, 1, 1)
|
||||||
|
self.conv3 = Conv(2 * c_, c2, 1) # act=FReLU(c2)
|
||||||
|
self.m = nn.SequentialCell(
|
||||||
|
[Bottleneck(c_, c_, shortcut, e=1.0) for _ in range(n)])
|
||||||
|
self.concat = P.Concat(axis=1)
|
||||||
|
|
||||||
|
def construct(self, x):
|
||||||
|
c1 = self.conv1(x)
|
||||||
|
c2 = self.m(c1)
|
||||||
|
c3 = self.conv2(x)
|
||||||
|
c4 = self.concat((c2, c3))
|
||||||
|
c5 = self.conv3(c4)
|
||||||
|
|
||||||
|
return c5
|
||||||
|
|
||||||
|
|
||||||
|
class SPP(nn.Cell):
|
||||||
|
# Spatial pyramid pooling layer used in YOLOv3-SPP
|
||||||
|
def __init__(self, c1, c2, k=(5, 9, 13)):
|
||||||
|
super(SPP, self).__init__()
|
||||||
|
c_ = c1 // 2 # hidden channels
|
||||||
|
self.conv1 = Conv(c1, c_, 1, 1)
|
||||||
|
self.conv2 = Conv(c_ * (len(k) + 1), c2, 1, 1)
|
||||||
|
|
||||||
|
self.maxpool1 = nn.MaxPool2d(kernel_size=5, stride=1, pad_mode='same')
|
||||||
|
self.maxpool2 = nn.MaxPool2d(kernel_size=9, stride=1, pad_mode='same')
|
||||||
|
self.maxpool3 = nn.MaxPool2d(kernel_size=13, stride=1, pad_mode='same')
|
||||||
|
self.concat = P.Concat(axis=1)
|
||||||
|
|
||||||
|
def construct(self, x):
|
||||||
|
c1 = self.conv1(x)
|
||||||
|
m1 = self.maxpool1(c1)
|
||||||
|
m2 = self.maxpool2(c1)
|
||||||
|
m3 = self.maxpool3(c1)
|
||||||
|
c4 = self.concat((c1, m1, m2, m3))
|
||||||
|
c5 = self.conv2(c4)
|
||||||
|
return c5
|
||||||
|
|
||||||
|
|
||||||
|
class Focus(nn.Cell):
|
||||||
|
# Focus wh information into c-space
|
||||||
|
def __init__(self, c1, c2, k=1, s=1, p=None, act=True):
|
||||||
|
super(Focus, self).__init__()
|
||||||
|
self.conv = Conv(c1 * 4, c2, k, s, p, act)
|
||||||
|
|
||||||
|
def construct(self, x):
|
||||||
|
c1 = self.conv(x)
|
||||||
|
return c1
|
||||||
|
|
||||||
|
|
||||||
|
class SiLU(nn.Cell):
|
||||||
|
def __init__(self):
|
||||||
|
super(SiLU, self).__init__()
|
||||||
|
self.sigmoid = P.Sigmoid()
|
||||||
|
|
||||||
|
def construct(self, x):
|
||||||
|
return x * self.sigmoid(x)
|
||||||
|
|
||||||
|
|
||||||
|
def auto_pad(k, p=None): # kernel, padding
|
||||||
|
# Pad to 'same'
|
||||||
|
if p is None:
|
||||||
|
p = k // 2 if isinstance(k, int) else [x // 2 for x in k] # auto-pad
|
||||||
|
return p
|
||||||
|
|
||||||
|
|
||||||
|
class Conv(nn.Cell):
|
||||||
|
# Standard convolution
|
||||||
|
def __init__(self, c1, c2, k=1, s=1, p=None,
|
||||||
|
dilation=1,
|
||||||
|
alpha=0.1,
|
||||||
|
momentum=0.97,
|
||||||
|
eps=1e-3,
|
||||||
|
pad_mode="same",
|
||||||
|
act=True): # ch_in, ch_out, kernel, stride, padding
|
||||||
|
super(Conv, self).__init__()
|
||||||
|
self.padding = auto_pad(k, p)
|
||||||
|
self.pad_mode = None
|
||||||
|
if self.padding == 0:
|
||||||
|
self.pad_mode = 'same'
|
||||||
|
elif self.padding == 1:
|
||||||
|
self.pad_mode = 'pad'
|
||||||
|
self.conv = nn.Conv2d(
|
||||||
|
c1,
|
||||||
|
c2,
|
||||||
|
k,
|
||||||
|
s,
|
||||||
|
padding=self.padding,
|
||||||
|
pad_mode=self.pad_mode,
|
||||||
|
has_bias=False)
|
||||||
|
self.bn = nn.BatchNorm2d(c2, momentum=momentum, eps=eps)
|
||||||
|
self.act = SiLU() if act is True else (
|
||||||
|
act if isinstance(act, nn.Cell) else P.Identity())
|
||||||
|
|
||||||
|
def construct(self, x):
|
||||||
|
return self.act(self.bn(self.conv(x)))
|
||||||
|
|
||||||
|
|
||||||
|
class YOLOv5Backbone(nn.Cell):
|
||||||
|
def __init__(self, shape):
|
||||||
|
super(YOLOv5Backbone, self).__init__()
|
||||||
|
self.focus = Focus(shape[0], shape[1], k=3, s=1)
|
||||||
|
self.conv1 = Conv(shape[1], shape[2], k=3, s=2)
|
||||||
|
self.CSP1 = BottleneckCSP(shape[2], shape[2], n=1 * shape[6])
|
||||||
|
self.conv2 = Conv(shape[2], shape[3], k=3, s=2)
|
||||||
|
self.CSP2 = BottleneckCSP(shape[3], shape[3], n=3 * shape[6])
|
||||||
|
self.conv3 = Conv(shape[3], shape[4], k=3, s=2)
|
||||||
|
self.CSP3 = BottleneckCSP(shape[4], shape[4], n=3 * shape[6])
|
||||||
|
self.conv4 = Conv(shape[4], shape[5], k=3, s=2)
|
||||||
|
self.spp = SPP(shape[5], shape[5], k=[5, 9, 13])
|
||||||
|
self.CSP4 = BottleneckCSP(shape[5], shape[5], n=1 * shape[6], shortcut=False)
|
||||||
|
|
||||||
|
def construct(self, x):
|
||||||
|
"""construct method"""
|
||||||
|
c1 = self.focus(x)
|
||||||
|
c2 = self.conv1(c1)
|
||||||
|
c3 = self.CSP1(c2)
|
||||||
|
c4 = self.conv2(c3)
|
||||||
|
# out
|
||||||
|
c5 = self.CSP2(c4)
|
||||||
|
c6 = self.conv3(c5)
|
||||||
|
# out
|
||||||
|
c7 = self.CSP3(c6)
|
||||||
|
c8 = self.conv4(c7)
|
||||||
|
c9 = self.spp(c8)
|
||||||
|
# out
|
||||||
|
c10 = self.CSP4(c9)
|
||||||
|
return c5, c7, c10
|
||||||
|
|
@ -30,9 +30,22 @@ class ConfigYOLOV5:
|
||||||
jitter = 0.3
|
jitter = 0.3
|
||||||
|
|
||||||
resize_rate = 10
|
resize_rate = 10
|
||||||
multi_scale = [[320, 320], [352, 352], [384, 384], [416, 416], [448, 448],
|
multi_scale = [
|
||||||
[480, 480], [512, 512], [544, 544], [576, 576], [608, 608],
|
[320, 320],
|
||||||
[640, 640], [672, 672], [704, 704], [736, 736], [768, 768]]
|
[352, 352],
|
||||||
|
[384, 384],
|
||||||
|
[416, 416],
|
||||||
|
[448, 448],
|
||||||
|
[480, 480],
|
||||||
|
[512, 512],
|
||||||
|
[544, 544],
|
||||||
|
[576, 576],
|
||||||
|
[608, 608],
|
||||||
|
[640, 640],
|
||||||
|
[672, 672],
|
||||||
|
[704, 704],
|
||||||
|
[736, 736],
|
||||||
|
[768, 768]]
|
||||||
num_classes = 80
|
num_classes = 80
|
||||||
max_box = 150
|
max_box = 150
|
||||||
|
|
||||||
|
|
@ -51,5 +64,10 @@ class ConfigYOLOV5:
|
||||||
(459, 401)]
|
(459, 401)]
|
||||||
out_channel = 3 * (num_classes + 5)
|
out_channel = 3 * (num_classes + 5)
|
||||||
|
|
||||||
|
input_shape = [[3, 32, 64, 128, 256, 512, 1],
|
||||||
|
[3, 48, 96, 192, 384, 768, 2],
|
||||||
|
[3, 64, 128, 256, 512, 1024, 3],
|
||||||
|
[3, 80, 160, 320, 640, 1280, 4]]
|
||||||
|
|
||||||
# test_param
|
# test_param
|
||||||
test_img_shape = [640, 640]
|
test_img_shape = [640, 640]
|
||||||
|
|
|
||||||
|
|
@ -20,6 +20,7 @@ import numpy as np
|
||||||
|
|
||||||
class DistributedSampler:
|
class DistributedSampler:
|
||||||
"""Distributed sampler."""
|
"""Distributed sampler."""
|
||||||
|
|
||||||
def __init__(self, dataset_size, num_replicas=None, rank=None, shuffle=True):
|
def __init__(self, dataset_size, num_replicas=None, rank=None, shuffle=True):
|
||||||
if num_replicas is None:
|
if num_replicas is None:
|
||||||
print("***********Setting world_size to 1 since it is not passed in ******************")
|
print("***********Setting world_size to 1 since it is not passed in ******************")
|
||||||
|
|
@ -39,7 +40,8 @@ class DistributedSampler:
|
||||||
# deterministically shuffle based on epoch
|
# deterministically shuffle based on epoch
|
||||||
if self.shuffle:
|
if self.shuffle:
|
||||||
indices = np.random.RandomState(seed=self.epoch).permutation(self.dataset_size)
|
indices = np.random.RandomState(seed=self.epoch).permutation(self.dataset_size)
|
||||||
# np.array type. number from 0 to len(dataset_size)-1, used as index of dataset
|
# np.array type. number from 0 to len(dataset_size)-1, used as
|
||||||
|
# index of dataset
|
||||||
indices = indices.tolist()
|
indices = indices.tolist()
|
||||||
self.epoch += 1
|
self.epoch += 1
|
||||||
# change to list type
|
# change to list type
|
||||||
|
|
|
||||||
|
|
@ -20,7 +20,7 @@ from mindspore.common import initializer as init
|
||||||
from mindspore.common.initializer import Initializer as MeInitializer
|
from mindspore.common.initializer import Initializer as MeInitializer
|
||||||
from mindspore.train.serialization import load_checkpoint, load_param_into_net
|
from mindspore.train.serialization import load_checkpoint, load_param_into_net
|
||||||
import mindspore.nn as nn
|
import mindspore.nn as nn
|
||||||
from .util import load_backbone
|
|
||||||
|
|
||||||
def calculate_gain(nonlinearity, param=None):
|
def calculate_gain(nonlinearity, param=None):
|
||||||
r"""Return the recommended gain value for the given nonlinearity function.
|
r"""Return the recommended gain value for the given nonlinearity function.
|
||||||
|
|
@ -58,7 +58,8 @@ def calculate_gain(nonlinearity, param=None):
|
||||||
# True/False are instances of int, hence check above
|
# True/False are instances of int, hence check above
|
||||||
negative_slope = param
|
negative_slope = param
|
||||||
else:
|
else:
|
||||||
raise ValueError("negative_slope {} not a valid number".format(param))
|
raise ValueError(
|
||||||
|
"negative_slope {} not a valid number".format(param))
|
||||||
return math.sqrt(2.0 / (1 + negative_slope ** 2))
|
return math.sqrt(2.0 / (1 + negative_slope ** 2))
|
||||||
|
|
||||||
raise ValueError("Unsupported nonlinearity {}".format(nonlinearity))
|
raise ValueError("Unsupported nonlinearity {}".format(nonlinearity))
|
||||||
|
|
@ -118,7 +119,8 @@ def kaiming_uniform_(arr, a=0, mode='fan_in', nonlinearity='leaky_relu'):
|
||||||
fan = _calculate_correct_fan(arr, mode)
|
fan = _calculate_correct_fan(arr, mode)
|
||||||
gain = calculate_gain(nonlinearity, a)
|
gain = calculate_gain(nonlinearity, a)
|
||||||
std = gain / math.sqrt(fan)
|
std = gain / math.sqrt(fan)
|
||||||
bound = math.sqrt(3.0) * std # Calculate uniform bounds from standard deviation
|
# Calculate uniform bounds from standard deviation
|
||||||
|
bound = math.sqrt(3.0) * std
|
||||||
return np.random.uniform(-bound, bound, arr.shape)
|
return np.random.uniform(-bound, bound, arr.shape)
|
||||||
|
|
||||||
|
|
||||||
|
|
@ -141,6 +143,7 @@ def _calculate_fan_in_and_fan_out(arr):
|
||||||
|
|
||||||
class KaimingUniform(MeInitializer):
|
class KaimingUniform(MeInitializer):
|
||||||
"""Kaiming uniform initializer."""
|
"""Kaiming uniform initializer."""
|
||||||
|
|
||||||
def __init__(self, a=0, mode='fan_in', nonlinearity='leaky_relu'):
|
def __init__(self, a=0, mode='fan_in', nonlinearity='leaky_relu'):
|
||||||
super(KaimingUniform, self).__init__()
|
super(KaimingUniform, self).__init__()
|
||||||
self.a = a
|
self.a = a
|
||||||
|
|
@ -156,34 +159,23 @@ def default_recurisive_init(custom_cell):
|
||||||
"""Initialize parameter."""
|
"""Initialize parameter."""
|
||||||
for _, cell in custom_cell.cells_and_names():
|
for _, cell in custom_cell.cells_and_names():
|
||||||
if isinstance(cell, nn.Conv2d):
|
if isinstance(cell, nn.Conv2d):
|
||||||
cell.weight.set_data(init.initializer(KaimingUniform(a=math.sqrt(5)),
|
cell.weight.set_data(init.initializer(KaimingUniform(a=math.sqrt(5)), cell.weight.shape, cell.weight.dtype))
|
||||||
cell.weight.shape,
|
|
||||||
cell.weight.dtype))
|
|
||||||
if cell.bias is not None:
|
if cell.bias is not None:
|
||||||
fan_in, _ = _calculate_fan_in_and_fan_out(cell.weight)
|
fan_in, _ = _calculate_fan_in_and_fan_out(cell.weight)
|
||||||
bound = 1 / math.sqrt(fan_in)
|
bound = 1 / math.sqrt(fan_in)
|
||||||
cell.bias.set_data(init.initializer(init.Uniform(bound),
|
cell.bias.set_data(init.initializer(init.Uniform(bound), cell.bias.shape, cell.bias.dtype))
|
||||||
cell.bias.shape,
|
|
||||||
cell.bias.dtype))
|
|
||||||
elif isinstance(cell, nn.Dense):
|
elif isinstance(cell, nn.Dense):
|
||||||
cell.weight.set_data(init.initializer(KaimingUniform(a=math.sqrt(5)),
|
cell.weight.set_data(init.initializer(KaimingUniform(a=math.sqrt(5)), cell.weight.shape, cell.weight.dtype))
|
||||||
cell.weight.shape,
|
|
||||||
cell.weight.dtype))
|
|
||||||
if cell.bias is not None:
|
if cell.bias is not None:
|
||||||
fan_in, _ = _calculate_fan_in_and_fan_out(cell.weight)
|
fan_in, _ = _calculate_fan_in_and_fan_out(cell.weight)
|
||||||
bound = 1 / math.sqrt(fan_in)
|
bound = 1 / math.sqrt(fan_in)
|
||||||
cell.bias.set_data(init.initializer(init.Uniform(bound),
|
cell.bias.set_data(init.initializer(init.Uniform(bound), cell.bias.shape, cell.bias.dtype))
|
||||||
cell.bias.shape,
|
|
||||||
cell.bias.dtype))
|
|
||||||
elif isinstance(cell, (nn.BatchNorm2d, nn.BatchNorm1d)):
|
elif isinstance(cell, (nn.BatchNorm2d, nn.BatchNorm1d)):
|
||||||
pass
|
pass
|
||||||
|
|
||||||
|
|
||||||
def load_yolov5_params(args, network):
|
def load_yolov5_params(args, network):
|
||||||
"""Load yolov5 backbone parameter from checkpoint."""
|
"""Load yolov5 backbone parameter from checkpoint."""
|
||||||
if args.pretrained_backbone:
|
|
||||||
network = load_backbone(network, args.pretrained_backbone, args)
|
|
||||||
args.logger.info('load pre-trained backbone {} into network'.format(args.pretrained_backbone))
|
|
||||||
|
|
||||||
if args.resume_yolov5:
|
if args.resume_yolov5:
|
||||||
param_dict = load_checkpoint(args.resume_yolov5)
|
param_dict = load_checkpoint(args.resume_yolov5)
|
||||||
param_dict_new = {}
|
param_dict_new = {}
|
||||||
|
|
@ -200,3 +192,20 @@ def load_yolov5_params(args, network):
|
||||||
args.logger.info('resume finished')
|
args.logger.info('resume finished')
|
||||||
load_param_into_net(network, param_dict_new)
|
load_param_into_net(network, param_dict_new)
|
||||||
args.logger.info('load_model {} success'.format(args.resume_yolov5))
|
args.logger.info('load_model {} success'.format(args.resume_yolov5))
|
||||||
|
|
||||||
|
if args.pretrained_backbone:
|
||||||
|
param_dict = load_checkpoint(args.pretrained_backbone)
|
||||||
|
param_dict_new = {}
|
||||||
|
for key, values in param_dict.items():
|
||||||
|
if key.startswith('moments.'):
|
||||||
|
continue
|
||||||
|
elif key.startswith('yolo_network.'):
|
||||||
|
param_dict_new[key[13:]] = values
|
||||||
|
args.logger.info('in resume {}'.format(key))
|
||||||
|
else:
|
||||||
|
param_dict_new[key] = values
|
||||||
|
args.logger.info('in resume {}'.format(key))
|
||||||
|
|
||||||
|
args.logger.info('pretrained finished')
|
||||||
|
load_param_into_net(network, param_dict_new)
|
||||||
|
args.logger.info('load_model {} success'.format(args.pretrained_backbone))
|
||||||
|
|
|
||||||
|
|
@ -27,6 +27,7 @@ class LOGGER(logging.Logger):
|
||||||
logger_name: String. Logger name.
|
logger_name: String. Logger name.
|
||||||
rank: Integer. Rank id.
|
rank: Integer. Rank id.
|
||||||
"""
|
"""
|
||||||
|
|
||||||
def __init__(self, logger_name, rank=0):
|
def __init__(self, logger_name, rank=0):
|
||||||
super(LOGGER, self).__init__(logger_name)
|
super(LOGGER, self).__init__(logger_name)
|
||||||
self.rank = rank
|
self.rank = rank
|
||||||
|
|
@ -65,11 +66,11 @@ class LOGGER(logging.Logger):
|
||||||
if self.isEnabledFor(logging.INFO) and self.rank == 0:
|
if self.isEnabledFor(logging.INFO) and self.rank == 0:
|
||||||
line_width = 2
|
line_width = 2
|
||||||
important_msg = '\n'
|
important_msg = '\n'
|
||||||
important_msg += ('*'*70 + '\n')*line_width
|
important_msg += ('*' * 70 + '\n') * line_width
|
||||||
important_msg += ('*'*line_width + '\n')*2
|
important_msg += ('*' * line_width + '\n') * 2
|
||||||
important_msg += '*'*line_width + ' '*8 + msg + '\n'
|
important_msg += '*' * line_width + ' ' * 8 + msg + '\n'
|
||||||
important_msg += ('*'*line_width + '\n')*2
|
important_msg += ('*' * line_width + '\n') * 2
|
||||||
important_msg += ('*'*70 + '\n')*line_width
|
important_msg += ('*' * 70 + '\n') * line_width
|
||||||
self.info(important_msg, *args, **kwargs)
|
self.info(important_msg, *args, **kwargs)
|
||||||
|
|
||||||
|
|
||||||
|
|
|
||||||
|
|
@ -16,8 +16,10 @@
|
||||||
from mindspore.ops import operations as P
|
from mindspore.ops import operations as P
|
||||||
import mindspore.nn as nn
|
import mindspore.nn as nn
|
||||||
|
|
||||||
|
|
||||||
class ConfidenceLoss(nn.Cell):
|
class ConfidenceLoss(nn.Cell):
|
||||||
"""Loss for confidence."""
|
"""Loss for confidence."""
|
||||||
|
|
||||||
def __init__(self):
|
def __init__(self):
|
||||||
super(ConfidenceLoss, self).__init__()
|
super(ConfidenceLoss, self).__init__()
|
||||||
self.cross_entropy = P.SigmoidCrossEntropyWithLogits()
|
self.cross_entropy = P.SigmoidCrossEntropyWithLogits()
|
||||||
|
|
@ -32,6 +34,7 @@ class ConfidenceLoss(nn.Cell):
|
||||||
|
|
||||||
class ClassLoss(nn.Cell):
|
class ClassLoss(nn.Cell):
|
||||||
"""Loss for classification."""
|
"""Loss for classification."""
|
||||||
|
|
||||||
def __init__(self):
|
def __init__(self):
|
||||||
super(ClassLoss, self).__init__()
|
super(ClassLoss, self).__init__()
|
||||||
self.cross_entropy = P.SigmoidCrossEntropyWithLogits()
|
self.cross_entropy = P.SigmoidCrossEntropyWithLogits()
|
||||||
|
|
|
||||||
|
|
@ -76,7 +76,7 @@ def warmup_cosine_annealing_lr(lr, steps_per_epoch, warmup_epochs, max_epoch, T_
|
||||||
if i < warmup_steps:
|
if i < warmup_steps:
|
||||||
lr = linear_warmup_lr(i + 1, warmup_steps, base_lr, warmup_init_lr)
|
lr = linear_warmup_lr(i + 1, warmup_steps, base_lr, warmup_init_lr)
|
||||||
else:
|
else:
|
||||||
lr = eta_min + (base_lr - eta_min) * (1. + math.cos(math.pi*last_epoch / T_max)) / 2
|
lr = eta_min + (base_lr - eta_min) * (1. + math.cos(math.pi * last_epoch / T_max)) / 2
|
||||||
lr_each_step.append(lr)
|
lr_each_step.append(lr)
|
||||||
|
|
||||||
return np.array(lr_each_step).astype(np.float32)
|
return np.array(lr_each_step).astype(np.float32)
|
||||||
|
|
@ -92,7 +92,7 @@ def warmup_cosine_annealing_lr_V2(lr, steps_per_epoch, warmup_epochs, max_epoch,
|
||||||
last_lr = 0
|
last_lr = 0
|
||||||
last_epoch_V1 = 0
|
last_epoch_V1 = 0
|
||||||
|
|
||||||
T_max_V2 = int(max_epoch*1/3)
|
T_max_V2 = int(max_epoch * 1 / 3)
|
||||||
|
|
||||||
lr_each_step = []
|
lr_each_step = []
|
||||||
for i in range(total_steps):
|
for i in range(total_steps):
|
||||||
|
|
@ -100,13 +100,13 @@ def warmup_cosine_annealing_lr_V2(lr, steps_per_epoch, warmup_epochs, max_epoch,
|
||||||
if i < warmup_steps:
|
if i < warmup_steps:
|
||||||
lr = linear_warmup_lr(i + 1, warmup_steps, base_lr, warmup_init_lr)
|
lr = linear_warmup_lr(i + 1, warmup_steps, base_lr, warmup_init_lr)
|
||||||
else:
|
else:
|
||||||
if i < total_steps*2/3:
|
if i < total_steps * 2 / 3:
|
||||||
lr = eta_min + (base_lr - eta_min) * (1. + math.cos(math.pi*last_epoch / T_max)) / 2
|
lr = eta_min + (base_lr - eta_min) * (1. + math.cos(math.pi * last_epoch / T_max)) / 2
|
||||||
last_lr = lr
|
last_lr = lr
|
||||||
last_epoch_V1 = last_epoch
|
last_epoch_V1 = last_epoch
|
||||||
else:
|
else:
|
||||||
base_lr = last_lr
|
base_lr = last_lr
|
||||||
last_epoch = last_epoch-last_epoch_V1
|
last_epoch = last_epoch - last_epoch_V1
|
||||||
lr = eta_min + (base_lr - eta_min) * (1. + math.cos(math.pi * last_epoch / T_max_V2)) / 2
|
lr = eta_min + (base_lr - eta_min) * (1. + math.cos(math.pi * last_epoch / T_max_V2)) / 2
|
||||||
|
|
||||||
lr_each_step.append(lr)
|
lr_each_step.append(lr)
|
||||||
|
|
@ -118,8 +118,8 @@ def warmup_cosine_annealing_lr_sample(lr, steps_per_epoch, warmup_epochs, max_ep
|
||||||
start_sample_epoch = 60
|
start_sample_epoch = 60
|
||||||
step_sample = 2
|
step_sample = 2
|
||||||
tobe_sampled_epoch = 60
|
tobe_sampled_epoch = 60
|
||||||
end_sampled_epoch = start_sample_epoch + step_sample*tobe_sampled_epoch
|
end_sampled_epoch = start_sample_epoch + step_sample * tobe_sampled_epoch
|
||||||
max_sampled_epoch = max_epoch+tobe_sampled_epoch
|
max_sampled_epoch = max_epoch + tobe_sampled_epoch
|
||||||
T_max = max_sampled_epoch
|
T_max = max_sampled_epoch
|
||||||
|
|
||||||
base_lr = lr
|
base_lr = lr
|
||||||
|
|
@ -137,7 +137,7 @@ def warmup_cosine_annealing_lr_sample(lr, steps_per_epoch, warmup_epochs, max_ep
|
||||||
if i < warmup_steps:
|
if i < warmup_steps:
|
||||||
lr = linear_warmup_lr(i + 1, warmup_steps, base_lr, warmup_init_lr)
|
lr = linear_warmup_lr(i + 1, warmup_steps, base_lr, warmup_init_lr)
|
||||||
else:
|
else:
|
||||||
lr = eta_min + (base_lr - eta_min) * (1. + math.cos(math.pi*last_epoch / T_max)) / 2
|
lr = eta_min + (base_lr - eta_min) * (1. + math.cos(math.pi * last_epoch / T_max)) / 2
|
||||||
lr_each_step.append(lr)
|
lr_each_step.append(lr)
|
||||||
|
|
||||||
assert total_steps == len(lr_each_step)
|
assert total_steps == len(lr_each_step)
|
||||||
|
|
@ -147,34 +147,17 @@ def warmup_cosine_annealing_lr_sample(lr, steps_per_epoch, warmup_epochs, max_ep
|
||||||
def get_lr(args):
|
def get_lr(args):
|
||||||
"""generate learning rate."""
|
"""generate learning rate."""
|
||||||
if args.lr_scheduler == 'exponential':
|
if args.lr_scheduler == 'exponential':
|
||||||
lr = warmup_step_lr(args.lr,
|
lr = warmup_step_lr(args.lr, args.lr_epochs, args.steps_per_epoch, args.warmup_epochs, args.max_epoch,
|
||||||
args.lr_epochs,
|
gamma=args.lr_gamma)
|
||||||
args.steps_per_epoch,
|
|
||||||
args.warmup_epochs,
|
|
||||||
args.max_epoch,
|
|
||||||
gamma=args.lr_gamma,
|
|
||||||
)
|
|
||||||
elif args.lr_scheduler == 'cosine_annealing':
|
elif args.lr_scheduler == 'cosine_annealing':
|
||||||
lr = warmup_cosine_annealing_lr(args.lr,
|
lr = warmup_cosine_annealing_lr(args.lr, args.steps_per_epoch, args.warmup_epochs,
|
||||||
args.steps_per_epoch,
|
args.max_epoch, args.T_max, args.eta_min)
|
||||||
args.warmup_epochs,
|
|
||||||
args.max_epoch,
|
|
||||||
args.T_max,
|
|
||||||
args.eta_min)
|
|
||||||
elif args.lr_scheduler == 'cosine_annealing_V2':
|
elif args.lr_scheduler == 'cosine_annealing_V2':
|
||||||
lr = warmup_cosine_annealing_lr_V2(args.lr,
|
lr = warmup_cosine_annealing_lr_V2(args.lr, args.steps_per_epoch, args.warmup_epochs,
|
||||||
args.steps_per_epoch,
|
args.max_epoch, args.T_max, args.eta_min)
|
||||||
args.warmup_epochs,
|
|
||||||
args.max_epoch,
|
|
||||||
args.T_max,
|
|
||||||
args.eta_min)
|
|
||||||
elif args.lr_scheduler == 'cosine_annealing_sample':
|
elif args.lr_scheduler == 'cosine_annealing_sample':
|
||||||
lr = warmup_cosine_annealing_lr_sample(args.lr,
|
lr = warmup_cosine_annealing_lr_sample(args.lr, args.steps_per_epoch, args.warmup_epochs,
|
||||||
args.steps_per_epoch,
|
args.max_epoch, args.T_max, args.eta_min)
|
||||||
args.warmup_epochs,
|
|
||||||
args.max_epoch,
|
|
||||||
args.T_max,
|
|
||||||
args.eta_min)
|
|
||||||
else:
|
else:
|
||||||
raise NotImplementedError(args.lr_scheduler)
|
raise NotImplementedError(args.lr_scheduler)
|
||||||
return lr
|
return lr
|
||||||
|
|
|
||||||
|
|
@ -22,6 +22,7 @@ from PIL import Image
|
||||||
import cv2
|
import cv2
|
||||||
import mindspore.dataset.vision.py_transforms as PV
|
import mindspore.dataset.vision.py_transforms as PV
|
||||||
|
|
||||||
|
|
||||||
def _rand(a=0., b=1.):
|
def _rand(a=0., b=1.):
|
||||||
return np.random.rand() * (b - a) + a
|
return np.random.rand() * (b - a) + a
|
||||||
|
|
||||||
|
|
@ -64,7 +65,7 @@ def statistic_normalize_img(img, statistic_norm):
|
||||||
# img: RGB
|
# img: RGB
|
||||||
if isinstance(img, Image.Image):
|
if isinstance(img, Image.Image):
|
||||||
img = np.array(img)
|
img = np.array(img)
|
||||||
img = img/255.
|
img = img / 255.
|
||||||
mean = np.array([0.485, 0.456, 0.406])
|
mean = np.array([0.485, 0.456, 0.406])
|
||||||
std = np.array([0.229, 0.224, 0.225])
|
std = np.array([0.229, 0.224, 0.225])
|
||||||
if statistic_norm:
|
if statistic_norm:
|
||||||
|
|
@ -137,16 +138,15 @@ def _preprocess_true_boxes(true_boxes, anchors, in_shape, num_classes, max_boxes
|
||||||
"""
|
"""
|
||||||
Introduction
|
Introduction
|
||||||
------------
|
------------
|
||||||
对训练数据的ground truth box进行预处理
|
preprocessing ground truth box
|
||||||
Parameters
|
Parameters
|
||||||
----------
|
----------
|
||||||
true_boxes: ground truth box 形状为[boxes, 5], x_min, y_min, x_max, y_max, class_id
|
true_boxes: ground truth box shape as [boxes, 5], x_min, y_min, x_max, y_max, class_id
|
||||||
"""
|
"""
|
||||||
anchors = np.array(anchors)
|
anchors = np.array(anchors)
|
||||||
num_layers = anchors.shape[0] // 3
|
num_layers = anchors.shape[0] // 3
|
||||||
anchor_mask = [[6, 7, 8], [3, 4, 5], [0, 1, 2]]
|
anchor_mask = [[6, 7, 8], [3, 4, 5], [0, 1, 2]]
|
||||||
true_boxes = np.array(true_boxes, dtype='float32')
|
true_boxes = np.array(true_boxes, dtype='float32')
|
||||||
# input_shape = np.array([in_shape, in_shape], dtype='int32')
|
|
||||||
input_shape = np.array(in_shape, dtype='int32')
|
input_shape = np.array(in_shape, dtype='int32')
|
||||||
boxes_xy = (true_boxes[..., 0:2] + true_boxes[..., 2:4]) // 2.
|
boxes_xy = (true_boxes[..., 0:2] + true_boxes[..., 2:4]) // 2.
|
||||||
# trans to box center point
|
# trans to box center point
|
||||||
|
|
@ -160,17 +160,14 @@ def _preprocess_true_boxes(true_boxes, anchors, in_shape, num_classes, max_boxes
|
||||||
y_true = [np.zeros((grid_shapes[l][0], grid_shapes[l][1], len(anchor_mask[l]),
|
y_true = [np.zeros((grid_shapes[l][0], grid_shapes[l][1], len(anchor_mask[l]),
|
||||||
5 + num_classes), dtype='float32') for l in range(num_layers)]
|
5 + num_classes), dtype='float32') for l in range(num_layers)]
|
||||||
# y_true [gridy, gridx]
|
# y_true [gridy, gridx]
|
||||||
# 这里扩充维度是为了后面应用广播计算每个图中所有box的anchor互相之间的iou
|
|
||||||
anchors = np.expand_dims(anchors, 0)
|
anchors = np.expand_dims(anchors, 0)
|
||||||
anchors_max = anchors / 2.
|
anchors_max = anchors / 2.
|
||||||
anchors_min = -anchors_max
|
anchors_min = -anchors_max
|
||||||
# 因为之前对box做了padding, 因此需要去除全0行
|
|
||||||
valid_mask = boxes_wh[..., 0] > 0
|
valid_mask = boxes_wh[..., 0] > 0
|
||||||
wh = boxes_wh[valid_mask]
|
wh = boxes_wh[valid_mask]
|
||||||
if wh.size != 0:
|
if wh.size != 0:
|
||||||
# 为了应用广播扩充维度
|
|
||||||
wh = np.expand_dims(wh, -2)
|
wh = np.expand_dims(wh, -2)
|
||||||
# wh 的shape为[box_num, 1, 2]
|
# wh shape[box_num, 1, 2]
|
||||||
boxes_max = wh / 2.
|
boxes_max = wh / 2.
|
||||||
boxes_min = -boxes_max
|
boxes_min = -boxes_max
|
||||||
intersect_min = np.maximum(boxes_min, anchors_min)
|
intersect_min = np.maximum(boxes_min, anchors_min)
|
||||||
|
|
@ -180,10 +177,8 @@ def _preprocess_true_boxes(true_boxes, anchors, in_shape, num_classes, max_boxes
|
||||||
box_area = wh[..., 0] * wh[..., 1]
|
box_area = wh[..., 0] * wh[..., 1]
|
||||||
anchor_area = anchors[..., 0] * anchors[..., 1]
|
anchor_area = anchors[..., 0] * anchors[..., 1]
|
||||||
iou = intersect_area / (box_area + anchor_area - intersect_area)
|
iou = intersect_area / (box_area + anchor_area - intersect_area)
|
||||||
#topk iou
|
|
||||||
# 找出和ground truth box的iou最大的anchor box,
|
|
||||||
# 然后将对应不同比例的负责该ground turth box 的位置置为ground truth box坐标
|
|
||||||
|
|
||||||
|
# topk iou
|
||||||
topk = 4
|
topk = 4
|
||||||
topk_flag = iou.argsort()
|
topk_flag = iou.argsort()
|
||||||
topk_flag = topk_flag >= topk_flag.shape[1] - topk
|
topk_flag = topk_flag >= topk_flag.shape[1] - topk
|
||||||
|
|
@ -210,7 +205,7 @@ def _preprocess_true_boxes(true_boxes, anchors, in_shape, num_classes, max_boxes
|
||||||
y_true[l][j, i, k, 5 + c] = 1 - label_smooth_factor
|
y_true[l][j, i, k, 5 + c] = 1 - label_smooth_factor
|
||||||
else:
|
else:
|
||||||
y_true[l][j, i, k, 5 + c] = 1.
|
y_true[l][j, i, k, 5 + c] = 1.
|
||||||
#best anchor for gt
|
# best anchor for gt
|
||||||
best_anchor = np.argmax(iou, axis=-1)
|
best_anchor = np.argmax(iou, axis=-1)
|
||||||
for t, n in enumerate(best_anchor):
|
for t, n in enumerate(best_anchor):
|
||||||
for l in range(num_layers):
|
for l in range(num_layers):
|
||||||
|
|
@ -352,9 +347,7 @@ def _choose_candidate_by_constraints(max_trial, input_w, input_h, image_w, image
|
||||||
(None, 1),
|
(None, 1),
|
||||||
)
|
)
|
||||||
else:
|
else:
|
||||||
constraints = (
|
constraints = ((None, None),)
|
||||||
(None, None),
|
|
||||||
)
|
|
||||||
# add default candidate
|
# add default candidate
|
||||||
candidates = [(0, 0, input_w, input_h)]
|
candidates = [(0, 0, input_w, input_h)]
|
||||||
for constraint in constraints:
|
for constraint in constraints:
|
||||||
|
|
@ -411,7 +404,8 @@ def _correct_bbox_by_candidates(candidates, input_w, input_h, image_w,
|
||||||
if allow_outside_center:
|
if allow_outside_center:
|
||||||
pass
|
pass
|
||||||
else:
|
else:
|
||||||
t_box = t_box[np.logical_and((t_box[:, 0] + t_box[:, 2])/2. >= 0., (t_box[:, 1] + t_box[:, 3])/2. >= 0.)]
|
t_box = t_box[
|
||||||
|
np.logical_and((t_box[:, 0] + t_box[:, 2]) / 2. >= 0., (t_box[:, 1] + t_box[:, 3]) / 2. >= 0.)]
|
||||||
t_box = t_box[np.logical_and((t_box[:, 0] + t_box[:, 2]) / 2. <= input_w,
|
t_box = t_box[np.logical_and((t_box[:, 0] + t_box[:, 2]) / 2. <= input_w,
|
||||||
(t_box[:, 1] + t_box[:, 3]) / 2. <= input_h)]
|
(t_box[:, 1] + t_box[:, 3]) / 2. <= input_h)]
|
||||||
|
|
||||||
|
|
@ -455,24 +449,12 @@ def _data_aug(image, box, jitter, hue, sat, val, image_input_size, max_boxes,
|
||||||
flip = _rand() < .5
|
flip = _rand() < .5
|
||||||
box_data = np.zeros((max_boxes, 5))
|
box_data = np.zeros((max_boxes, 5))
|
||||||
|
|
||||||
candidates = _choose_candidate_by_constraints(use_constraints=False,
|
candidates = _choose_candidate_by_constraints(use_constraints=False, max_trial=max_trial, input_w=input_w,
|
||||||
max_trial=max_trial,
|
input_h=input_h, image_w=image_w, image_h=image_h,
|
||||||
input_w=input_w,
|
jitter=jitter, box=box)
|
||||||
input_h=input_h,
|
box_data, candidate = _correct_bbox_by_candidates(candidates=candidates, input_w=input_w, input_h=input_h,
|
||||||
image_w=image_w,
|
image_w=image_w, image_h=image_h, flip=flip, box=box,
|
||||||
image_h=image_h,
|
box_data=box_data, allow_outside_center=True, max_boxes=max_boxes)
|
||||||
jitter=jitter,
|
|
||||||
box=box)
|
|
||||||
box_data, candidate = _correct_bbox_by_candidates(candidates=candidates,
|
|
||||||
input_w=input_w,
|
|
||||||
input_h=input_h,
|
|
||||||
image_w=image_w,
|
|
||||||
image_h=image_h,
|
|
||||||
flip=flip,
|
|
||||||
box=box,
|
|
||||||
box_data=box_data,
|
|
||||||
allow_outside_center=True,
|
|
||||||
max_boxes=max_boxes)
|
|
||||||
dx, dy, nw, nh = candidate
|
dx, dy, nw, nh = candidate
|
||||||
interp = get_interp_method(interp=10)
|
interp = get_interp_method(interp=10)
|
||||||
image = image.resize((nw, nh), pil_image_reshape(interp))
|
image = image.resize((nw, nh), pil_image_reshape(interp))
|
||||||
|
|
@ -514,6 +496,7 @@ def reshape_fn(image, img_id, config):
|
||||||
|
|
||||||
class MultiScaleTrans:
|
class MultiScaleTrans:
|
||||||
"""Multi scale transform."""
|
"""Multi scale transform."""
|
||||||
|
|
||||||
def __init__(self, config, device_num):
|
def __init__(self, config, device_num):
|
||||||
self.config = config
|
self.config = config
|
||||||
self.seed = 0
|
self.seed = 0
|
||||||
|
|
|
||||||
|
|
@ -13,8 +13,6 @@
|
||||||
# limitations under the License.
|
# limitations under the License.
|
||||||
# ============================================================================
|
# ============================================================================
|
||||||
"""Util class or function."""
|
"""Util class or function."""
|
||||||
from mindspore.train.serialization import load_checkpoint
|
|
||||||
import mindspore.nn as nn
|
|
||||||
import mindspore.common.dtype as mstype
|
import mindspore.common.dtype as mstype
|
||||||
|
|
||||||
from .yolo import YoloLossBlock
|
from .yolo import YoloLossBlock
|
||||||
|
|
@ -54,65 +52,6 @@ class AverageMeter:
|
||||||
return fmtstr.format(**self.__dict__)
|
return fmtstr.format(**self.__dict__)
|
||||||
|
|
||||||
|
|
||||||
def load_backbone(net, ckpt_path, args):
|
|
||||||
"""Load cspdarknet53 backbone checkpoint."""
|
|
||||||
param_dict = load_checkpoint(ckpt_path)
|
|
||||||
yolo_backbone_prefix = 'feature_map.backbone'
|
|
||||||
darknet_backbone_prefix = 'backbone'
|
|
||||||
find_param = []
|
|
||||||
not_found_param = []
|
|
||||||
net.init_parameters_data()
|
|
||||||
for name, cell in net.cells_and_names():
|
|
||||||
if name.startswith(yolo_backbone_prefix):
|
|
||||||
name = name.replace(yolo_backbone_prefix, darknet_backbone_prefix)
|
|
||||||
if isinstance(cell, (nn.Conv2d, nn.Dense)):
|
|
||||||
darknet_weight = '{}.weight'.format(name)
|
|
||||||
darknet_bias = '{}.bias'.format(name)
|
|
||||||
if darknet_weight in param_dict:
|
|
||||||
cell.weight.set_data(param_dict[darknet_weight].data)
|
|
||||||
find_param.append(darknet_weight)
|
|
||||||
else:
|
|
||||||
not_found_param.append(darknet_weight)
|
|
||||||
if darknet_bias in param_dict:
|
|
||||||
cell.bias.set_data(param_dict[darknet_bias].data)
|
|
||||||
find_param.append(darknet_bias)
|
|
||||||
else:
|
|
||||||
not_found_param.append(darknet_bias)
|
|
||||||
elif isinstance(cell, (nn.BatchNorm2d, nn.BatchNorm1d)):
|
|
||||||
darknet_moving_mean = '{}.moving_mean'.format(name)
|
|
||||||
darknet_moving_variance = '{}.moving_variance'.format(name)
|
|
||||||
darknet_gamma = '{}.gamma'.format(name)
|
|
||||||
darknet_beta = '{}.beta'.format(name)
|
|
||||||
if darknet_moving_mean in param_dict:
|
|
||||||
cell.moving_mean.set_data(param_dict[darknet_moving_mean].data)
|
|
||||||
find_param.append(darknet_moving_mean)
|
|
||||||
else:
|
|
||||||
not_found_param.append(darknet_moving_mean)
|
|
||||||
if darknet_moving_variance in param_dict:
|
|
||||||
cell.moving_variance.set_data(param_dict[darknet_moving_variance].data)
|
|
||||||
find_param.append(darknet_moving_variance)
|
|
||||||
else:
|
|
||||||
not_found_param.append(darknet_moving_variance)
|
|
||||||
if darknet_gamma in param_dict:
|
|
||||||
cell.gamma.set_data(param_dict[darknet_gamma].data)
|
|
||||||
find_param.append(darknet_gamma)
|
|
||||||
else:
|
|
||||||
not_found_param.append(darknet_gamma)
|
|
||||||
if darknet_beta in param_dict:
|
|
||||||
cell.beta.set_data(param_dict[darknet_beta].data)
|
|
||||||
find_param.append(darknet_beta)
|
|
||||||
else:
|
|
||||||
not_found_param.append(darknet_beta)
|
|
||||||
|
|
||||||
args.logger.info('================found_param {}========='.format(len(find_param)))
|
|
||||||
args.logger.info(find_param)
|
|
||||||
args.logger.info('================not_found_param {}========='.format(len(not_found_param)))
|
|
||||||
args.logger.info(not_found_param)
|
|
||||||
args.logger.info('=====load {} successfully ====='.format(ckpt_path))
|
|
||||||
|
|
||||||
return net
|
|
||||||
|
|
||||||
|
|
||||||
def default_wd_filter(x):
|
def default_wd_filter(x):
|
||||||
"""default weight decay filter."""
|
"""default weight decay filter."""
|
||||||
parameter_name = x.name
|
parameter_name = x.name
|
||||||
|
|
@ -120,10 +59,12 @@ def default_wd_filter(x):
|
||||||
# all bias not using weight decay
|
# all bias not using weight decay
|
||||||
return False
|
return False
|
||||||
if parameter_name.endswith('.gamma'):
|
if parameter_name.endswith('.gamma'):
|
||||||
# bn weight bias not using weight decay, be carefully for now x not include BN
|
# bn weight bias not using weight decay, be carefully for now x not
|
||||||
|
# include BN
|
||||||
return False
|
return False
|
||||||
if parameter_name.endswith('.beta'):
|
if parameter_name.endswith('.beta'):
|
||||||
# bn weight bias not using weight decay, be carefully for now x not include BN
|
# bn weight bias not using weight decay, be carefully for now x not
|
||||||
|
# include BN
|
||||||
return False
|
return False
|
||||||
|
|
||||||
return True
|
return True
|
||||||
|
|
@ -139,19 +80,23 @@ def get_param_groups(network):
|
||||||
# all bias not using weight decay
|
# all bias not using weight decay
|
||||||
no_decay_params.append(x)
|
no_decay_params.append(x)
|
||||||
elif parameter_name.endswith('.gamma'):
|
elif parameter_name.endswith('.gamma'):
|
||||||
# bn weight bias not using weight decay, be carefully for now x not include BN
|
# bn weight bias not using weight decay, be carefully for now x not
|
||||||
|
# include BN
|
||||||
no_decay_params.append(x)
|
no_decay_params.append(x)
|
||||||
elif parameter_name.endswith('.beta'):
|
elif parameter_name.endswith('.beta'):
|
||||||
# bn weight bias not using weight decay, be carefully for now x not include BN
|
# bn weight bias not using weight decay, be carefully for now x not
|
||||||
|
# include BN
|
||||||
no_decay_params.append(x)
|
no_decay_params.append(x)
|
||||||
else:
|
else:
|
||||||
decay_params.append(x)
|
decay_params.append(x)
|
||||||
|
|
||||||
return [{'params': no_decay_params, 'weight_decay': 0.0}, {'params': decay_params}]
|
return [{'params': no_decay_params, 'weight_decay': 0.0},
|
||||||
|
{'params': decay_params}]
|
||||||
|
|
||||||
|
|
||||||
class ShapeRecord:
|
class ShapeRecord:
|
||||||
"""Log image shape."""
|
"""Log image shape."""
|
||||||
|
|
||||||
def __init__(self):
|
def __init__(self):
|
||||||
self.shape_record = {
|
self.shape_record = {
|
||||||
416: 0,
|
416: 0,
|
||||||
|
|
@ -178,7 +123,7 @@ class ShapeRecord:
|
||||||
def show(self, logger):
|
def show(self, logger):
|
||||||
for key in self.shape_record:
|
for key in self.shape_record:
|
||||||
rate = self.shape_record[key] / float(self.shape_record['total'])
|
rate = self.shape_record[key] / float(self.shape_record['total'])
|
||||||
logger.info('shape {}: {:.2f}%'.format(key, rate*100))
|
logger.info('shape {}: {:.2f}%'.format(key, rate * 100))
|
||||||
|
|
||||||
|
|
||||||
def keep_loss_fp32(network):
|
def keep_loss_fp32(network):
|
||||||
|
|
|
||||||
|
|
@ -24,29 +24,28 @@ from mindspore.ops import operations as P
|
||||||
from mindspore.ops import functional as F
|
from mindspore.ops import functional as F
|
||||||
from mindspore.ops import composite as C
|
from mindspore.ops import composite as C
|
||||||
|
|
||||||
from src.yolov5_backbone import YOLOv5Backbone, Conv, C3
|
from src.backbone import YOLOv5Backbone, Conv, BottleneckCSP
|
||||||
from src.config import ConfigYOLOV5
|
from src.config import ConfigYOLOV5
|
||||||
from src.loss import ConfidenceLoss, ClassLoss
|
from src.loss import ConfidenceLoss, ClassLoss
|
||||||
|
|
||||||
|
|
||||||
class YOLOv5(nn.Cell):
|
class YOLO(nn.Cell):
|
||||||
def __init__(self, backbone, out_channel):
|
def __init__(self, backbone, shape):
|
||||||
super(YOLOv5, self).__init__()
|
super(YOLO, self).__init__()
|
||||||
self.out_channel = out_channel
|
|
||||||
self.backbone = backbone
|
self.backbone = backbone
|
||||||
|
self.config = ConfigYOLOV5()
|
||||||
|
|
||||||
self.conv1 = Conv(512, 256, k=1, s=1) # 10
|
self.conv1 = Conv(shape[5], shape[4], k=1, s=1)
|
||||||
self.C31 = C3(512, 256, n=1, shortcut=False) # 11
|
self.CSP5 = BottleneckCSP(shape[5], shape[4], n=1*shape[6], shortcut=False)
|
||||||
self.conv2 = Conv(256, 128, k=1, s=1)
|
self.conv2 = Conv(shape[4], shape[3], k=1, s=1)
|
||||||
self.C32 = C3(256, 128, n=1, shortcut=False) # 13
|
self.CSP6 = BottleneckCSP(shape[4], shape[3], n=1*shape[6], shortcut=False)
|
||||||
self.conv3 = Conv(128, 128, k=3, s=2)
|
self.conv3 = Conv(shape[3], shape[3], k=3, s=2)
|
||||||
self.C33 = C3(256, 256, n=1, shortcut=False) # 15
|
self.CSP7 = BottleneckCSP(shape[4], shape[4], n=1*shape[6], shortcut=False)
|
||||||
self.conv4 = Conv(256, 256, k=3, s=2)
|
self.conv4 = Conv(shape[4], shape[4], k=3, s=2)
|
||||||
self.C34 = C3(512, 512, n=1, shortcut=False) # 17
|
self.CSP8 = BottleneckCSP(shape[5], shape[5], n=1*shape[6], shortcut=False)
|
||||||
|
self.back_block1 = YoloBlock(shape[3], self.config.out_channel)
|
||||||
self.backblock1 = YoloBlock(128, 255)
|
self.back_block2 = YoloBlock(shape[4], self.config.out_channel)
|
||||||
self.backblock2 = YoloBlock(256, 255)
|
self.back_block3 = YoloBlock(shape[5], self.config.out_channel)
|
||||||
self.backblock3 = YoloBlock(512, 255)
|
|
||||||
|
|
||||||
self.concat = P.Concat(axis=1)
|
self.concat = P.Concat(axis=1)
|
||||||
|
|
||||||
|
|
@ -57,29 +56,32 @@ class YOLOv5(nn.Cell):
|
||||||
feature_map2 is (batch_size, backbone_shape[3], h/16, w/16)
|
feature_map2 is (batch_size, backbone_shape[3], h/16, w/16)
|
||||||
feature_map3 is (batch_size, backbone_shape[4], h/32, w/32)
|
feature_map3 is (batch_size, backbone_shape[4], h/32, w/32)
|
||||||
"""
|
"""
|
||||||
img_hight = P.Shape()(x)[2] * 2
|
img_height = P.Shape()(x)[2] * 2
|
||||||
img_width = P.Shape()(x)[3] * 2
|
img_width = P.Shape()(x)[3] * 2
|
||||||
|
|
||||||
backbone4, backbone6, backbone9 = self.backbone(x)
|
feature_map1, feature_map2, feature_map3 = self.backbone(x)
|
||||||
|
|
||||||
cv1 = self.conv1(backbone9) # 10
|
c1 = self.conv1(feature_map3)
|
||||||
ups1 = P.ResizeNearestNeighbor((img_hight / 16, img_width / 16))(cv1)
|
ups1 = P.ResizeNearestNeighbor((img_height // 16, img_width // 16))(c1)
|
||||||
concat1 = self.concat((ups1, backbone6))
|
c2 = self.concat((ups1, feature_map2))
|
||||||
bcsp1 = self.C31(concat1) # 13
|
c3 = self.CSP5(c2)
|
||||||
cv2 = self.conv2(bcsp1)
|
c4 = self.conv2(c3)
|
||||||
ups2 = P.ResizeNearestNeighbor((img_hight / 8, img_width / 8))(cv2) # 15
|
ups2 = P.ResizeNearestNeighbor((img_height // 8, img_width // 8))(c4)
|
||||||
concat2 = self.concat((ups2, backbone4))
|
c5 = self.concat((ups2, feature_map1))
|
||||||
bcsp2 = self.C32(concat2) # 17
|
# out
|
||||||
cv3 = self.conv3(bcsp2)
|
c6 = self.CSP6(c5)
|
||||||
|
c7 = self.conv3(c6)
|
||||||
|
|
||||||
concat3 = self.concat((cv3, cv2))
|
c8 = self.concat((c7, c4))
|
||||||
bcsp3 = self.C33(concat3) # 20
|
# out
|
||||||
cv4 = self.conv4(bcsp3)
|
c9 = self.CSP7(c8)
|
||||||
concat4 = self.concat((cv4, cv1))
|
c10 = self.conv4(c9)
|
||||||
bcsp4 = self.C34(concat4) # 23
|
c11 = self.concat((c10, c1))
|
||||||
small_object_output = self.backblock1(bcsp2) # h/8, w/8
|
# out
|
||||||
medium_object_output = self.backblock2(bcsp3) # h/16, w/16
|
c12 = self.CSP8(c11)
|
||||||
big_object_output = self.backblock3(bcsp4) # h/32, w/32
|
small_object_output = self.back_block1(c6)
|
||||||
|
medium_object_output = self.back_block2(c9)
|
||||||
|
big_object_output = self.back_block3(c12)
|
||||||
return small_object_output, medium_object_output, big_object_output
|
return small_object_output, medium_object_output, big_object_output
|
||||||
|
|
||||||
|
|
||||||
|
|
@ -98,19 +100,17 @@ class YoloBlock(nn.Cell):
|
||||||
YoloBlock(12, 255)
|
YoloBlock(12, 255)
|
||||||
|
|
||||||
"""
|
"""
|
||||||
|
|
||||||
def __init__(self, in_channels, out_channels):
|
def __init__(self, in_channels, out_channels):
|
||||||
super(YoloBlock, self).__init__()
|
super(YoloBlock, self).__init__()
|
||||||
|
|
||||||
self.cv = nn.Conv2d(in_channels, out_channels, kernel_size=1, stride=1, has_bias=True)
|
self.conv = nn.Conv2d(in_channels, out_channels, kernel_size=1, stride=1, has_bias=True)
|
||||||
|
|
||||||
def construct(self, x):
|
def construct(self, x):
|
||||||
"""construct method"""
|
"""construct method"""
|
||||||
|
|
||||||
out = self.cv(x)
|
out = self.conv(x)
|
||||||
return out
|
return out
|
||||||
|
|
||||||
|
|
||||||
class DetectionBlock(nn.Cell):
|
class DetectionBlock(nn.Cell):
|
||||||
"""
|
"""
|
||||||
YOLOv5 detection Network. It will finally output the detection result.
|
YOLOv5 detection Network. It will finally output the detection result.
|
||||||
|
|
@ -146,13 +146,14 @@ class DetectionBlock(nn.Cell):
|
||||||
raise KeyError("Invalid scale value for DetectionBlock")
|
raise KeyError("Invalid scale value for DetectionBlock")
|
||||||
self.anchors = Tensor([self.config.anchor_scales[i] for i in idx], ms.float32)
|
self.anchors = Tensor([self.config.anchor_scales[i] for i in idx], ms.float32)
|
||||||
self.num_anchors_per_scale = 3
|
self.num_anchors_per_scale = 3
|
||||||
self.num_attrib = 4 + 1 + self.config.num_classes
|
self.num_attrib = 4+1+self.config.num_classes
|
||||||
self.lambda_coord = 1
|
self.lambda_coord = 1
|
||||||
|
|
||||||
self.sigmoid = nn.Sigmoid()
|
self.sigmoid = nn.Sigmoid()
|
||||||
self.reshape = P.Reshape()
|
self.reshape = P.Reshape()
|
||||||
self.tile = P.Tile()
|
self.tile = P.Tile()
|
||||||
self.concat = P.Concat(axis=-1)
|
self.concat = P.Concat(axis=-1)
|
||||||
|
self.pow = P.Pow()
|
||||||
self.conf_training = is_training
|
self.conf_training = is_training
|
||||||
|
|
||||||
def construct(self, x, input_shape):
|
def construct(self, x, input_shape):
|
||||||
|
|
@ -190,6 +191,7 @@ class DetectionBlock(nn.Cell):
|
||||||
P.Cast()(F.tuple_to_array((grid_size[1], grid_size[0])), ms.float32)
|
P.Cast()(F.tuple_to_array((grid_size[1], grid_size[0])), ms.float32)
|
||||||
# box_wh is w->h
|
# box_wh is w->h
|
||||||
box_wh = P.Exp()(box_wh) * self.anchors / input_shape
|
box_wh = P.Exp()(box_wh) * self.anchors / input_shape
|
||||||
|
|
||||||
box_confidence = self.sigmoid(box_confidence)
|
box_confidence = self.sigmoid(box_confidence)
|
||||||
box_probs = self.sigmoid(box_probs)
|
box_probs = self.sigmoid(box_probs)
|
||||||
|
|
||||||
|
|
@ -200,7 +202,6 @@ class DetectionBlock(nn.Cell):
|
||||||
|
|
||||||
class Iou(nn.Cell):
|
class Iou(nn.Cell):
|
||||||
"""Calculate the iou of boxes"""
|
"""Calculate the iou of boxes"""
|
||||||
|
|
||||||
def __init__(self):
|
def __init__(self):
|
||||||
super(Iou, self).__init__()
|
super(Iou, self).__init__()
|
||||||
self.min = P.Minimum()
|
self.min = P.Minimum()
|
||||||
|
|
@ -214,8 +215,8 @@ class Iou(nn.Cell):
|
||||||
"""
|
"""
|
||||||
box1_xy = box1[:, :, :, :, :, :2]
|
box1_xy = box1[:, :, :, :, :, :2]
|
||||||
box1_wh = box1[:, :, :, :, :, 2:4]
|
box1_wh = box1[:, :, :, :, :, 2:4]
|
||||||
box1_mins = box1_xy - box1_wh / F.scalar_to_array(2.0) # topLeft
|
box1_mins = box1_xy - box1_wh / F.scalar_to_array(2.0) # topLeft
|
||||||
box1_maxs = box1_xy + box1_wh / F.scalar_to_array(2.0) # rightDown
|
box1_maxs = box1_xy + box1_wh / F.scalar_to_array(2.0) # rightDown
|
||||||
|
|
||||||
box2_xy = box2[:, :, :, :, :, :2]
|
box2_xy = box2[:, :, :, :, :, :2]
|
||||||
box2_wh = box2[:, :, :, :, :, 2:4]
|
box2_wh = box2[:, :, :, :, :, 2:4]
|
||||||
|
|
@ -239,7 +240,6 @@ class YoloLossBlock(nn.Cell):
|
||||||
"""
|
"""
|
||||||
Loss block cell of YOLOV5 network.
|
Loss block cell of YOLOV5 network.
|
||||||
"""
|
"""
|
||||||
|
|
||||||
def __init__(self, scale, config=ConfigYOLOV5()):
|
def __init__(self, scale, config=ConfigYOLOV5()):
|
||||||
super(YoloLossBlock, self).__init__()
|
super(YoloLossBlock, self).__init__()
|
||||||
self.config = config
|
self.config = config
|
||||||
|
|
@ -261,7 +261,7 @@ class YoloLossBlock(nn.Cell):
|
||||||
self.class_loss = ClassLoss()
|
self.class_loss = ClassLoss()
|
||||||
|
|
||||||
self.reduce_sum = P.ReduceSum()
|
self.reduce_sum = P.ReduceSum()
|
||||||
self.giou = Giou()
|
self.g_iou = GIou()
|
||||||
|
|
||||||
def construct(self, prediction, pred_xy, pred_wh, y_true, gt_box, input_shape):
|
def construct(self, prediction, pred_xy, pred_wh, y_true, gt_box, input_shape):
|
||||||
"""
|
"""
|
||||||
|
|
@ -315,15 +315,15 @@ class YoloLossBlock(nn.Cell):
|
||||||
pred_boxes_me = P.Reshape()(pred_boxes_me, (-1, 4))
|
pred_boxes_me = P.Reshape()(pred_boxes_me, (-1, 4))
|
||||||
true_boxes_me = xywh2x1y1x2y2(true_boxes)
|
true_boxes_me = xywh2x1y1x2y2(true_boxes)
|
||||||
true_boxes_me = P.Reshape()(true_boxes_me, (-1, 4))
|
true_boxes_me = P.Reshape()(true_boxes_me, (-1, 4))
|
||||||
ciou = self.giou(pred_boxes_me, true_boxes_me)
|
c_iou = self.g_iou(pred_boxes_me, true_boxes_me)
|
||||||
ciou_loss = object_mask_me * box_loss_scale_me * (1 - ciou)
|
c_iou_loss = object_mask_me * box_loss_scale_me * (1 - c_iou)
|
||||||
ciou_loss_me = self.reduce_sum(ciou_loss, ())
|
c_iou_loss_me = self.reduce_sum(c_iou_loss, ())
|
||||||
loss = ciou_loss_me * 4 + confidence_loss + class_loss
|
loss = c_iou_loss_me * 4 + confidence_loss + class_loss
|
||||||
batch_size = P.Shape()(prediction)[0]
|
batch_size = P.Shape()(prediction)[0]
|
||||||
return loss / batch_size
|
return loss / batch_size
|
||||||
|
|
||||||
|
|
||||||
class YOLOV5s(nn.Cell):
|
class YOLOV5(nn.Cell):
|
||||||
"""
|
"""
|
||||||
YOLOV5 network.
|
YOLOV5 network.
|
||||||
|
|
||||||
|
|
@ -337,13 +337,13 @@ class YOLOV5s(nn.Cell):
|
||||||
YOLOV5s(True)
|
YOLOV5s(True)
|
||||||
"""
|
"""
|
||||||
|
|
||||||
def __init__(self, is_training):
|
def __init__(self, is_training, version=0):
|
||||||
super(YOLOV5s, self).__init__()
|
super(YOLOV5, self).__init__()
|
||||||
self.config = ConfigYOLOV5()
|
self.config = ConfigYOLOV5()
|
||||||
|
|
||||||
# YOLOv5 network
|
# YOLOv5 network
|
||||||
self.feature_map = YOLOv5(backbone=YOLOv5Backbone(),
|
self.shape = self.config.input_shape[version]
|
||||||
out_channel=self.config.out_channel)
|
self.feature_map = YOLO(backbone=YOLOv5Backbone(shape=self.shape), shape=self.shape)
|
||||||
|
|
||||||
# prediction on the default anchor boxes
|
# prediction on the default anchor boxes
|
||||||
self.detect_1 = DetectionBlock('l', is_training=is_training)
|
self.detect_1 = DetectionBlock('l', is_training=is_training)
|
||||||
|
|
@ -364,18 +364,17 @@ class YOLOV5s_Infer(nn.Cell):
|
||||||
YOLOV5 Infer.
|
YOLOV5 Infer.
|
||||||
"""
|
"""
|
||||||
|
|
||||||
def __init__(self, inputshape):
|
def __init__(self, input_shape, version=0):
|
||||||
super(YOLOV5s_Infer, self).__init__()
|
super(YOLOV5s_Infer, self).__init__()
|
||||||
self.network = YOLOV5s(is_training=False)
|
self.network = YOLOV5(is_training=False, version=version)
|
||||||
self.inputshape = inputshape
|
self.input_shape = input_shape
|
||||||
|
|
||||||
def construct(self, x):
|
def construct(self, x):
|
||||||
return self.network(x, self.inputshape)
|
return self.network(x, self.input_shape)
|
||||||
|
|
||||||
|
|
||||||
class YoloWithLossCell(nn.Cell):
|
class YoloWithLossCell(nn.Cell):
|
||||||
"""YOLOV5 loss."""
|
"""YOLOV5 loss."""
|
||||||
|
|
||||||
def __init__(self, network):
|
def __init__(self, network):
|
||||||
super(YoloWithLossCell, self).__init__()
|
super(YoloWithLossCell, self).__init__()
|
||||||
self.yolo_network = network
|
self.yolo_network = network
|
||||||
|
|
@ -398,7 +397,6 @@ class YoloWithLossCell(nn.Cell):
|
||||||
|
|
||||||
class TrainingWrapper(nn.Cell):
|
class TrainingWrapper(nn.Cell):
|
||||||
"""Training wrapper."""
|
"""Training wrapper."""
|
||||||
|
|
||||||
def __init__(self, network, optimizer, sens=1.0):
|
def __init__(self, network, optimizer, sens=1.0):
|
||||||
super(TrainingWrapper, self).__init__(auto_prefix=False)
|
super(TrainingWrapper, self).__init__(auto_prefix=False)
|
||||||
self.network = network
|
self.network = network
|
||||||
|
|
@ -427,15 +425,13 @@ class TrainingWrapper(nn.Cell):
|
||||||
grads = self.grad(self.network, weights)(*args, sens)
|
grads = self.grad(self.network, weights)(*args, sens)
|
||||||
if self.reducer_flag:
|
if self.reducer_flag:
|
||||||
grads = self.grad_reducer(grads)
|
grads = self.grad_reducer(grads)
|
||||||
self.optimizer(grads)
|
return F.depend(loss, self.optimizer(grads))
|
||||||
return loss
|
|
||||||
|
|
||||||
|
|
||||||
class Giou(nn.Cell):
|
class GIou(nn.Cell):
|
||||||
"""Calculating giou"""
|
"""Calculating giou"""
|
||||||
|
|
||||||
def __init__(self):
|
def __init__(self):
|
||||||
super(Giou, self).__init__()
|
super(GIou, self).__init__()
|
||||||
self.cast = P.Cast()
|
self.cast = P.Cast()
|
||||||
self.reshape = P.Reshape()
|
self.reshape = P.Reshape()
|
||||||
self.min = P.Minimum()
|
self.min = P.Minimum()
|
||||||
|
|
|
||||||
|
|
@ -19,135 +19,126 @@ import argparse
|
||||||
import datetime
|
import datetime
|
||||||
import mindspore as ms
|
import mindspore as ms
|
||||||
from mindspore.context import ParallelMode
|
from mindspore.context import ParallelMode
|
||||||
from mindspore.nn.optim.momentum import Momentum
|
from mindspore.nn import Momentum
|
||||||
from mindspore import Tensor
|
from mindspore import Tensor
|
||||||
from mindspore import context
|
from mindspore import context
|
||||||
from mindspore.communication.management import init, get_rank, get_group_size
|
from mindspore.communication.management import init, get_rank, get_group_size
|
||||||
from mindspore.train.callback import ModelCheckpoint, RunContext
|
from mindspore.train.callback import ModelCheckpoint, RunContext
|
||||||
from mindspore.train.callback import _InternalCallbackParam, CheckpointConfig
|
from mindspore.train.callback import _InternalCallbackParam, CheckpointConfig
|
||||||
|
|
||||||
from src.yolo import YOLOV5s, YoloWithLossCell, TrainingWrapper
|
from src.yolo import YOLOV5, YoloWithLossCell, TrainingWrapper
|
||||||
from src.logger import get_logger
|
from src.logger import get_logger
|
||||||
from src.util import AverageMeter, get_param_groups
|
from src.util import AverageMeter, get_param_groups
|
||||||
from src.lr_scheduler import get_lr
|
from src.lr_scheduler import get_lr
|
||||||
from src.yolo_dataset import create_yolo_dataset
|
from src.yolo_dataset import create_yolo_dataset
|
||||||
from src.initializer import default_recurisive_init, load_yolov5_params
|
from src.initializer import default_recurisive_init, load_yolov5_params
|
||||||
from src.config import ConfigYOLOV5
|
from src.config import ConfigYOLOV5
|
||||||
|
|
||||||
ms.set_seed(1)
|
ms.set_seed(1)
|
||||||
|
|
||||||
|
parser = argparse.ArgumentParser('mindspore coco training')
|
||||||
|
|
||||||
def parse_args(cloud_args=None):
|
# device related
|
||||||
"""Parse train arguments."""
|
parser.add_argument('--device_target', type=str, default='Ascend', help='device where the code will be implemented.')
|
||||||
parser = argparse.ArgumentParser('mindspore coco training')
|
|
||||||
|
|
||||||
# device related
|
# dataset related
|
||||||
parser.add_argument('--device_target', type=str, default='Ascend',
|
parser.add_argument('--data_dir', default='/data/coco', type=str, help='Train dataset directory.')
|
||||||
help='device where the code will be implemented.')
|
parser.add_argument('--per_batch_size', default=32, type=int, help='Batch size for Training. Default: 8')
|
||||||
|
|
||||||
# dataset related
|
# network related
|
||||||
parser.add_argument('--data_dir', type=str, help='Train dataset directory.')
|
parser.add_argument('--yolov5_version', default='yolov5s', type=str,
|
||||||
parser.add_argument('--per_batch_size', default=8, type=int, help='Batch size for Training. Default: 8')
|
help='The version of YOLOv5, options: yolov5s, yolov5m, yolov5l, yolov5x')
|
||||||
|
parser.add_argument('--pretrained_backbone', default='', type=str, help='The pretrained file of yolov5. Default: "".')
|
||||||
|
parser.add_argument('--resume_yolov5', default='', type=str,
|
||||||
|
help='The ckpt file of YOLOv5, which used to fine tune. Default: ""')
|
||||||
|
|
||||||
# network related
|
# optimizer and lr related
|
||||||
parser.add_argument('--pretrained_backbone', default='', type=str,
|
parser.add_argument('--lr_scheduler', default='cosine_annealing', type=str,
|
||||||
help='The backbone file of YOLOv5. Default: "".')
|
help='Learning rate scheduler, options: exponential, cosine_annealing. Default: exponential')
|
||||||
parser.add_argument('--resume_yolov5', default='', type=str,
|
parser.add_argument('--lr', default=0.013, type=float, help='Learning rate. Default: 0.01')
|
||||||
help='The ckpt file of YOLOv5, which used to fine tune. Default: ""')
|
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. 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=300, help='T-max in cosine_annealing scheduler. Default: 320')
|
||||||
|
parser.add_argument('--max_epoch', type=int, default=300, help='Max epoch num to train the model. Default: 320')
|
||||||
|
parser.add_argument('--warmup_epochs', default=20, 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')
|
||||||
|
|
||||||
# optimizer and lr related
|
# loss related
|
||||||
parser.add_argument('--lr_scheduler', default='cosine_annealing', type=str,
|
parser.add_argument('--loss_scale', type=int, default=1024, help='Static loss scale. Default: 1024')
|
||||||
help='Learning rate scheduler, options: exponential, cosine_annealing. Default: exponential')
|
parser.add_argument('--label_smooth', type=int, default=0, help='Whether to use label smooth in CE. Default:0')
|
||||||
parser.add_argument('--lr', default=0.013, type=float, help='Learning rate. Default: 0.01')
|
parser.add_argument('--label_smooth_factor', type=float, default=0.1,
|
||||||
parser.add_argument('--lr_epochs', type=str, default='220,250',
|
help='Smooth strength of original one-hot. Default: 0.1')
|
||||||
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. 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=300, help='T-max in cosine_annealing scheduler. Default: 320')
|
|
||||||
parser.add_argument('--max_epoch', type=int, default=300, help='Max epoch num to train the model. Default: 320')
|
|
||||||
parser.add_argument('--warmup_epochs', default=20, 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
|
# logging related
|
||||||
parser.add_argument('--loss_scale', type=int, default=1024, help='Static loss scale. Default: 1024')
|
parser.add_argument('--log_interval', type=int, default=100, help='Logging interval steps. Default: 100')
|
||||||
parser.add_argument('--label_smooth', type=int, default=0, help='Whether to use label smooth in CE. Default:0')
|
parser.add_argument('--ckpt_path', type=str, default='outputs/', help='Checkpoint save location. Default: outputs/')
|
||||||
parser.add_argument('--label_smooth_factor', type=float, default=0.1,
|
parser.add_argument('--ckpt_interval', type=int, default=None, help='Save checkpoint interval. Default: None')
|
||||||
help='Smooth strength of original one-hot. Default: 0.1')
|
|
||||||
|
|
||||||
# logging related
|
parser.add_argument('--is_save_on_master', type=int, default=1,
|
||||||
parser.add_argument('--log_interval', type=int, default=100, help='Logging interval steps. Default: 100')
|
help='Save ckpt on master or all rank, 1 for master, 0 for all ranks. Default: 1')
|
||||||
parser.add_argument('--ckpt_path', type=str, default='outputs/', help='Checkpoint save location. Default: outputs/')
|
|
||||||
parser.add_argument('--ckpt_interval', type=int, default=10, help='Save checkpoint interval. Default: 10')
|
|
||||||
|
|
||||||
parser.add_argument('--is_save_on_master', type=int, default=1,
|
# distributed related
|
||||||
help='Save ckpt on master or all rank, 1 for master, 0 for all ranks. Default: 1')
|
parser.add_argument('--is_distributed', type=int, default=0,
|
||||||
|
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')
|
||||||
|
|
||||||
# distributed related
|
# roma obs
|
||||||
parser.add_argument('--is_distributed', type=int, default=1,
|
parser.add_argument('--train_url', type=str, default="", help='train url')
|
||||||
help='Distribute train or not, 1 for yes, 0 for no. Default: 1')
|
# profiler init
|
||||||
parser.add_argument('--rank', type=int, default=0, help='Local rank of distributed. Default: 0')
|
parser.add_argument('--need_profiler', type=int, default=0,
|
||||||
parser.add_argument('--group_size', type=int, default=1, help='World size of device. Default: 1')
|
help='Whether use profiler. 0 for no, 1 for yes. Default: 0')
|
||||||
|
|
||||||
# roma obs
|
# reset default config
|
||||||
parser.add_argument('--train_url', type=str, default="", help='train url')
|
parser.add_argument('--training_shape', type=str, default="", help='Fix training shape. Default: ""')
|
||||||
# profiler init
|
parser.add_argument('--resize_rate', type=int, default=10, help='Resize rate for multi-scale training. Default: None')
|
||||||
parser.add_argument('--need_profiler', type=int, default=0,
|
parser.add_argument('--is_modelArts', type=int, default=0,
|
||||||
help='Whether use profiler. 0 for no, 1 for yes. Default: 0')
|
help='Trainning in modelArts or not, 1 for yes, 0 for no. Default: 0')
|
||||||
|
|
||||||
# reset default config
|
args, _ = parser.parse_known_args()
|
||||||
parser.add_argument('--training_shape', type=str, default="", help='Fix training shape. Default: ""')
|
|
||||||
parser.add_argument('--resize_rate', type=int, default=10,
|
|
||||||
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:
|
||||||
args = merge_args(args, cloud_args)
|
args.T_max = args.max_epoch
|
||||||
if args.lr_scheduler == 'cosine_annealing' and args.max_epoch > args.T_max:
|
|
||||||
args.T_max = args.max_epoch
|
|
||||||
|
|
||||||
args.lr_epochs = list(map(int, args.lr_epochs.split(',')))
|
args.lr_epochs = list(map(int, args.lr_epochs.split(',')))
|
||||||
|
|
||||||
|
if args.is_modelArts:
|
||||||
args.data_root = os.path.join(args.data_dir, 'train2017')
|
args.data_root = os.path.join(args.data_dir, 'train2017')
|
||||||
args.annFile = os.path.join(args.data_dir, 'annotations/instances_train2017.json')
|
args.annFile = os.path.join(args.data_dir, 'annotations')
|
||||||
|
outputs_dir = os.path.join('/cache', args.ckpt_path)
|
||||||
|
else:
|
||||||
|
args.data_root = os.path.join(args.data_dir, 'train2017')
|
||||||
|
args.annFile = os.path.join(
|
||||||
|
args.data_dir, 'annotations/instances_train2017.json')
|
||||||
|
outputs_dir = args.ckpt_path
|
||||||
|
|
||||||
devid = int(os.getenv('DEVICE_ID', '0'))
|
deviced = int(os.getenv('DEVICE_ID', '0'))
|
||||||
context.set_context(mode=context.GRAPH_MODE, enable_auto_mixed_precision=True,
|
context.set_context(mode=context.GRAPH_MODE, enable_auto_mixed_precision=True, device_target=args.device_target,
|
||||||
device_target=args.device_target, save_graphs=False, device_id=devid)
|
save_graphs=False, device_id=deviced)
|
||||||
# init distributed
|
# init distributed
|
||||||
if args.is_distributed:
|
if args.is_distributed:
|
||||||
if args.device_target == "Ascend":
|
if args.device_target == "Ascend":
|
||||||
init()
|
init()
|
||||||
else:
|
|
||||||
init("nccl")
|
|
||||||
args.rank = get_rank()
|
|
||||||
args.group_size = get_group_size()
|
|
||||||
|
|
||||||
# select for master rank save ckpt or all rank save, compatible for model parallel
|
|
||||||
args.rank_save_ckpt_flag = 0
|
|
||||||
if args.is_save_on_master:
|
|
||||||
if args.rank == 0:
|
|
||||||
args.rank_save_ckpt_flag = 1
|
|
||||||
else:
|
else:
|
||||||
|
init("nccl")
|
||||||
|
args.rank = get_rank()
|
||||||
|
args.group_size = get_group_size()
|
||||||
|
|
||||||
|
args.rank_save_ckpt_flag = 0
|
||||||
|
if args.is_save_on_master:
|
||||||
|
if args.rank == 0:
|
||||||
args.rank_save_ckpt_flag = 1
|
args.rank_save_ckpt_flag = 1
|
||||||
|
else:
|
||||||
|
args.rank_save_ckpt_flag = 1
|
||||||
|
|
||||||
# logger
|
# logger
|
||||||
args.outputs_dir = os.path.join(args.ckpt_path,
|
args.outputs_dir = os.path.join(outputs_dir, datetime.datetime.now().strftime('%Y-%m-%d_time_%H_%M_%S'))
|
||||||
datetime.datetime.now().strftime('%Y-%m-%d_time_%H_%M_%S'))
|
args.logger = get_logger(args.outputs_dir, args.rank)
|
||||||
args.logger = get_logger(args.outputs_dir, args.rank)
|
args.logger.save_args(args)
|
||||||
args.logger.save_args(args)
|
|
||||||
|
|
||||||
return args
|
|
||||||
|
|
||||||
def merge_args(args, cloud_args):
|
|
||||||
args_dict = vars(args)
|
|
||||||
if isinstance(cloud_args, dict):
|
|
||||||
for key in cloud_args.keys():
|
|
||||||
val = cloud_args[key]
|
|
||||||
if key in args_dict and val:
|
|
||||||
arg_type = type(args_dict[key])
|
|
||||||
if arg_type is not type(None):
|
|
||||||
val = arg_type(val)
|
|
||||||
args_dict[key] = val
|
|
||||||
return args
|
|
||||||
|
|
||||||
|
|
||||||
def convert_training_shape(args_training_shape):
|
def convert_training_shape(args_training_shape):
|
||||||
|
|
@ -155,111 +146,114 @@ def convert_training_shape(args_training_shape):
|
||||||
return training_shape
|
return training_shape
|
||||||
|
|
||||||
|
|
||||||
def train(cloud_args=None):
|
loss_meter = AverageMeter('loss')
|
||||||
args = parse_args(cloud_args)
|
|
||||||
loss_meter = AverageMeter('loss')
|
|
||||||
|
|
||||||
context.reset_auto_parallel_context()
|
if args.is_modelArts:
|
||||||
parallel_mode = ParallelMode.STAND_ALONE
|
import moxing as mox
|
||||||
degree = 1
|
local_data_url = os.path.join('/cache/data', str(args.rank))
|
||||||
if args.is_distributed:
|
local_annFile = os.path.join('/cache/data', str(args.rank))
|
||||||
parallel_mode = ParallelMode.DATA_PARALLEL
|
mox.file.copy_parallel(args.data_root, local_data_url)
|
||||||
degree = get_group_size()
|
args.data_root = local_data_url
|
||||||
context.set_auto_parallel_context(parallel_mode=parallel_mode, gradients_mean=True, device_num=degree)
|
|
||||||
|
|
||||||
network = YOLOV5s(is_training=True)
|
mox.file.copy_parallel(args.annFile, local_annFile)
|
||||||
# default is kaiming-normal
|
args.annFile = os.path.join(local_data_url, 'instances_train2017.json')
|
||||||
default_recurisive_init(network)
|
|
||||||
load_yolov5_params(args, network)
|
|
||||||
|
|
||||||
network = YoloWithLossCell(network)
|
context.reset_auto_parallel_context()
|
||||||
config = ConfigYOLOV5()
|
parallel_mode = ParallelMode.STAND_ALONE
|
||||||
|
degree = 1
|
||||||
|
if args.is_distributed:
|
||||||
|
parallel_mode = ParallelMode.DATA_PARALLEL
|
||||||
|
degree = get_group_size()
|
||||||
|
context.set_auto_parallel_context(parallel_mode=parallel_mode, gradients_mean=True, device_num=degree)
|
||||||
|
|
||||||
config.label_smooth = args.label_smooth
|
dict_version = {'yolov5s': 0, 'yolov5m': 1, 'yolov5l': 2, 'yolov5x': 3}
|
||||||
config.label_smooth_factor = args.label_smooth_factor
|
network = YOLOV5(is_training=True, version=dict_version[args.yolov5_version])
|
||||||
|
# default is kaiming-normal
|
||||||
|
default_recurisive_init(network)
|
||||||
|
load_yolov5_params(args, network)
|
||||||
|
|
||||||
if args.training_shape:
|
network = YoloWithLossCell(network)
|
||||||
config.multi_scale = [convert_training_shape(args.training_shape)]
|
config = ConfigYOLOV5()
|
||||||
if args.resize_rate:
|
|
||||||
config.resize_rate = args.resize_rate
|
|
||||||
|
|
||||||
ds, data_size = create_yolo_dataset(image_dir=args.data_root, anno_path=args.annFile, is_training=True,
|
config.label_smooth = args.label_smooth
|
||||||
batch_size=args.per_batch_size, max_epoch=args.max_epoch,
|
config.label_smooth_factor = args.label_smooth_factor
|
||||||
device_num=args.group_size, rank=args.rank, config=config)
|
|
||||||
args.logger.info('Finish loading dataset')
|
|
||||||
|
|
||||||
args.steps_per_epoch = int(data_size / args.per_batch_size / args.group_size)
|
if args.training_shape:
|
||||||
|
config.multi_scale = [convert_training_shape(args.training_shape)]
|
||||||
|
if args.resize_rate:
|
||||||
|
config.resize_rate = args.resize_rate
|
||||||
|
|
||||||
if not args.ckpt_interval:
|
ds, data_size = create_yolo_dataset(image_dir=args.data_root, anno_path=args.annFile, is_training=True,
|
||||||
args.ckpt_interval = args.steps_per_epoch
|
batch_size=args.per_batch_size, max_epoch=args.max_epoch,
|
||||||
|
device_num=args.group_size, rank=args.rank, config=config)
|
||||||
|
|
||||||
lr = get_lr(args)
|
args.logger.info('Finish loading dataset')
|
||||||
|
|
||||||
opt = Momentum(params=get_param_groups(network),
|
args.steps_per_epoch = int(data_size / args.per_batch_size / args.group_size)
|
||||||
learning_rate=Tensor(lr),
|
|
||||||
momentum=args.momentum,
|
|
||||||
weight_decay=args.weight_decay,
|
|
||||||
loss_scale=args.loss_scale)
|
|
||||||
|
|
||||||
network = TrainingWrapper(network, opt, args.loss_scale // 2)
|
if not args.ckpt_interval:
|
||||||
network.set_train()
|
args.ckpt_interval = args.steps_per_epoch
|
||||||
|
|
||||||
|
lr = get_lr(args)
|
||||||
|
|
||||||
|
opt = Momentum(params=get_param_groups(network), momentum=args.momentum, learning_rate=Tensor(lr),
|
||||||
|
weight_decay=args.weight_decay, loss_scale=args.loss_scale)
|
||||||
|
|
||||||
|
network = TrainingWrapper(network, opt, args.loss_scale // 2)
|
||||||
|
network.set_train()
|
||||||
|
|
||||||
|
if args.rank_save_ckpt_flag:
|
||||||
|
# checkpoint save
|
||||||
|
ckpt_max_num = args.max_epoch * args.steps_per_epoch // args.ckpt_interval
|
||||||
|
ckpt_config = CheckpointConfig(save_checkpoint_steps=args.ckpt_interval, keep_checkpoint_max=1)
|
||||||
|
save_ckpt_path = os.path.join(args.outputs_dir, 'ckpt_' + str(args.rank) + '/')
|
||||||
|
ckpt_cb = ModelCheckpoint(config=ckpt_config, directory=save_ckpt_path, prefix='{}'.format(args.rank))
|
||||||
|
cb_params = _InternalCallbackParam()
|
||||||
|
cb_params.train_network = network
|
||||||
|
cb_params.epoch_num = ckpt_max_num
|
||||||
|
cb_params.cur_epoch_num = 1
|
||||||
|
run_context = RunContext(cb_params)
|
||||||
|
ckpt_cb.begin(run_context)
|
||||||
|
|
||||||
|
old_progress = -1
|
||||||
|
t_end = time.time()
|
||||||
|
data_loader = ds.create_dict_iterator(output_numpy=True, num_epochs=1)
|
||||||
|
|
||||||
|
for i, data in enumerate(data_loader):
|
||||||
|
images = data["image"]
|
||||||
|
input_shape = images.shape[2:4]
|
||||||
|
images = Tensor.from_numpy(images)
|
||||||
|
batch_y_true_0 = Tensor.from_numpy(data['bbox1'])
|
||||||
|
batch_y_true_1 = Tensor.from_numpy(data['bbox2'])
|
||||||
|
batch_y_true_2 = Tensor.from_numpy(data['bbox3'])
|
||||||
|
batch_gt_box0 = Tensor.from_numpy(data['gt_box1'])
|
||||||
|
batch_gt_box1 = Tensor.from_numpy(data['gt_box2'])
|
||||||
|
batch_gt_box2 = Tensor.from_numpy(data['gt_box3'])
|
||||||
|
input_shape = Tensor(tuple(input_shape[::-1]), ms.float32)
|
||||||
|
loss = network(images, batch_y_true_0, batch_y_true_1, batch_y_true_2, batch_gt_box0, batch_gt_box1,
|
||||||
|
batch_gt_box2, input_shape)
|
||||||
|
loss_meter.update(loss.asnumpy())
|
||||||
|
|
||||||
if args.rank_save_ckpt_flag:
|
if args.rank_save_ckpt_flag:
|
||||||
# checkpoint save
|
# ckpt progress
|
||||||
ckpt_max_num = args.max_epoch * args.steps_per_epoch // args.ckpt_interval
|
cb_params.cur_step_num = i + 1 # current step number
|
||||||
ckpt_config = CheckpointConfig(save_checkpoint_steps=args.ckpt_interval,
|
cb_params.batch_num = i + 2
|
||||||
keep_checkpoint_max=ckpt_max_num)
|
ckpt_cb.step_end(run_context)
|
||||||
save_ckpt_path = os.path.join(args.outputs_dir, 'ckpt_' + str(args.rank) + '/')
|
|
||||||
ckpt_cb = ModelCheckpoint(config=ckpt_config,
|
|
||||||
directory=save_ckpt_path,
|
|
||||||
prefix='{}'.format(args.rank))
|
|
||||||
cb_params = _InternalCallbackParam()
|
|
||||||
cb_params.train_network = network
|
|
||||||
cb_params.epoch_num = ckpt_max_num
|
|
||||||
cb_params.cur_epoch_num = 1
|
|
||||||
run_context = RunContext(cb_params)
|
|
||||||
ckpt_cb.begin(run_context)
|
|
||||||
|
|
||||||
old_progress = -1
|
if i % args.log_interval == 0:
|
||||||
t_end = time.time()
|
time_used = time.time() - t_end
|
||||||
data_loader = ds.create_dict_iterator(output_numpy=True, num_epochs=1)
|
epoch = int(i / args.steps_per_epoch)
|
||||||
|
fps = args.per_batch_size * (i - old_progress) * args.group_size / time_used
|
||||||
|
if args.rank == 0:
|
||||||
|
args.logger.info('epoch[{}], iter[{}], {}, fps:{:.2f} imgs/sec, '
|
||||||
|
'lr:{}'.format(epoch, i, loss_meter, fps, lr[i]))
|
||||||
|
t_end = time.time()
|
||||||
|
loss_meter.reset()
|
||||||
|
old_progress = i
|
||||||
|
|
||||||
for i, data in enumerate(data_loader):
|
if (i + 1) % args.steps_per_epoch == 0 and args.rank_save_ckpt_flag:
|
||||||
images = data["image"]
|
cb_params.cur_epoch_num += 1
|
||||||
input_shape = images.shape[2:4]
|
|
||||||
images = Tensor.from_numpy(images)
|
|
||||||
batch_y_true_0 = Tensor.from_numpy(data['bbox1'])
|
|
||||||
batch_y_true_1 = Tensor.from_numpy(data['bbox2'])
|
|
||||||
batch_y_true_2 = Tensor.from_numpy(data['bbox3'])
|
|
||||||
batch_gt_box0 = Tensor.from_numpy(data['gt_box1'])
|
|
||||||
batch_gt_box1 = Tensor.from_numpy(data['gt_box2'])
|
|
||||||
batch_gt_box2 = Tensor.from_numpy(data['gt_box3'])
|
|
||||||
input_shape = Tensor(tuple(input_shape[::-1]), ms.float32)
|
|
||||||
loss = network(images, batch_y_true_0, batch_y_true_1, batch_y_true_2, batch_gt_box0, batch_gt_box1,
|
|
||||||
batch_gt_box2, input_shape)
|
|
||||||
loss_meter.update(loss.asnumpy())
|
|
||||||
|
|
||||||
if args.rank_save_ckpt_flag:
|
if args.is_modelArts:
|
||||||
# ckpt progress
|
mox.file.copy_parallel(src_url='/cache/outputs/', dst_url='obs://hit-cyf/yolov5_npu/outputs/')
|
||||||
cb_params.cur_step_num = i + 1 # current step number
|
args.logger.info('==========end training===============')
|
||||||
cb_params.batch_num = i + 2
|
|
||||||
ckpt_cb.step_end(run_context)
|
|
||||||
|
|
||||||
if i % args.log_interval == 0:
|
|
||||||
time_used = time.time() - t_end
|
|
||||||
epoch = int(i / args.steps_per_epoch)
|
|
||||||
fps = args.per_batch_size * (i - old_progress) * args.group_size / time_used
|
|
||||||
if args.rank == 0:
|
|
||||||
args.logger.info(
|
|
||||||
'epoch[{}], iter[{}], {}, fps:{:.2f} imgs/sec, lr:{}'.format(epoch, i, loss_meter, fps, lr[i]))
|
|
||||||
t_end = time.time()
|
|
||||||
loss_meter.reset()
|
|
||||||
old_progress = i
|
|
||||||
|
|
||||||
if (i + 1) % args.steps_per_epoch == 0 and args.rank_save_ckpt_flag:
|
|
||||||
cb_params.cur_epoch_num += 1
|
|
||||||
|
|
||||||
args.logger.info('==========end training===============')
|
|
||||||
|
|
||||||
if __name__ == "__main__":
|
|
||||||
train()
|
|
||||||
|
|
|
||||||
Loading…
Reference in New Issue