Skip to content
Snippets Groups Projects
Commit 3941f823 authored by Vladimir Nechaev's avatar Vladimir Nechaev Committed by mo-hai
Browse files

Implemented training and evaluation of Retinanet-Resnet152 on a GPU

parent 208cd92f
Branches
No related tags found
No related merge requests found
Show whitespace changes
Inline Side-by-side
Showing
with 919 additions and 102 deletions
research/cv/retinanet_resnet152/README.md 0 → 100644
+ 507
− 0
View file @ 3941f823
# 1. 内容
<!-- TOC -->
[查看中文](./README_CN.md)
- [Retinanet Description](#retinanet-description)
- [Model Architecture](#model-architecture)
- [Dataset](#dataset)
- [Environment Requirements](#environment-requirements)
- [Script Description](#script-description)
- [Script and Sample Code](#ccript-and-sample-code)
- [Script Parameters](#script-parameters)
- [Training Process](#training-process)
- [Usage](#usage1)
- [Run](#run1)
- [Result](#result1)
- [Evaluation Process](#evaluation-process)
- [Usage](#usage2)
- [Run](#run2)
- [Result](#result2)
- [Model Export](#model-export)
- [Usage](#usage3)
- [Run](#run3)
- [Inference Process](#inference-process)
- [Usage](#usage4)
- [Run](#run4)
- [Result](#result4)
- [Model Description](#model-description)
- [Performance](#performance)
- [Training Performance](#training-performance)
- [Evaluation Performance](#evaluation-performance)
- [Description of Random State](#description-of-random-situation)
- [ModelZoo Homepage](#modelzoo-homepage)
<!-- /TOC -->
## [Retinanet Description](#content)
RetinaNet was proposed in "Focal Loss for Dense Object Detection" 2018 paper by Facebook AI Research. The biggest contribution of this paper is to propose Focal Loss to solve the problem of category imbalance, thus creating RetinaNet (One Stage target detection algorithm), a target detection network whose accuracy exceeds the classic Two Stage Faster-RCNN.
[Paper](https://arxiv.org/pdf/1708.02002.pdf)
Lin T Y , Goyal P , Girshick R , et al. Focal Loss for Dense Object Detection[C]// 2017 IEEE International Conference on Computer Vision (ICCV). IEEE, 2017:2999-3007.
## [Model Architecture](#content)
The overall network architecture of Retinanet is [here](https://arxiv.org/pdf/1708.02002.pdf)
## [Dataset](#content)
Dataset used (refer to paper): [COCO2017](https://cocodataset.org/#download)
- Dataset size: 19.3G, 123287 pcs 80 classes of colored images
- [train](http://images.cocodataset.org/zips/train2017.zip): 19.3G, 118287 images
- [val](http://images.cocodataset.org/zips/val2017.zip): 814.3M, 5000 images
- Data format: RGB
>Note：The data will be processed with src/dataset.py.
## [Environment Requirements](#content)
- Hardware (Ascend/GPU)
- Prepare hardware environment with Ascend or GPU.
- Framework
- [MindSpore](https://www.mindspore.cn/install)
- For more information about MindSpore, please check the resources below：
- [MindSpore 教程](https://www.mindspore.cn/tutorials/zh-CN/master/index.html)
- [MindSpore Python API](https://www.mindspore.cn/docs/api/zh-CN/master/index.html)
- Training in ModelArts (If you want to run on ModelArts, you can refer to the following [modelarts documents](https://support.huaweicloud.com/modelarts/))
```python
# Train with 8 cards on ModelArts
# (1) choose a or b
# a. in default_config.yaml set "enable_modelarts=True"
# in default_config.yaml set "distribute=True"
# in default_config.yaml set "coco_root='/cache/data'"
# in default_config.yaml set "epoch_size=500"
# (optional)in default_config.yaml set "checkpoint_path='s3://dir_to_your_pretrained/'"
# in default_config.yaml set other parameters
# b. Set on the web page "enable_modelarts=True"
# Set on the web page "distribute=True"
# Set on the web page "coco_root=/cache/data"
# Set on the web page "epoch_size=500"
# (optional)Set on the web page "checkpoint_path='s3://dir_to_your_pretrained/'"
# Set on the web page other parameters
# (2) Prepare model code
# (3) If you choose to fine-tune your model, upload your pretrained model to the S3 bucket
# (4) choose a or b (recommended choice a)
# a. First, compress the dataset into a ".zip" file.
# Second, upload your compressed dataset to the S3 bucket (you can also upload the uncompressed dataset, but that may be slow.)
# b. Upload the original dataset to the S3 bucket.
# (Dataset conversion occurs during the training process, which takes more time. It will be re-transformed every time you train.)
# (5) Set your code path on the web page to "/path/retinanet"
# (6) Set the startup file on the web page as "train.py"
# (7) Set "training dataset", "training output file path", "job log path", etc. on the webpage
# (8) Create a training job
#
# Train with 1 card on ModelArts
# (1) choose a or b
# a. in default_config.yaml set "enable_modelarts=True"
# in default_config.yaml set "coco_root='/cache/data'"
# in default_config.yaml set "epoch_size=500"
# (optional)in default_config.yaml set "checkpoint_path='s3://dir_to_your_pretrained/'"
# in default_config.yaml set other parameters
# b. Set on the web page "enable_modelarts=True"
# Set on the web page "coco_root='/cache/data'"
# Set on the web page "epoch_size=500"
# (optional)Set on the web page "checkpoint_path='s3://dir_to_your_pretrained/'"
# Set on the web page other parameters
# (2) Prepare model code
# (3) If you choose to fine-tune your model, upload your pretrained model to the S3 bucket
# (4) choose a or b (recommended choice a)
# a. First, compress the dataset into a ".zip" file.
# Second, upload your compressed dataset to the S3 bucket (you can also upload the uncompressed dataset, but that may be slow.)
# b. Upload the original dataset to the S3 bucket.
# (Data set conversion occurs during the training process, which takes more time. It will be re-transformed every time you train.)
# (5) Set your code path on the web page to "/path/retinanet"
# (6) Set the startup file on the web page as "train.py"
# (7) Set "training dataset", "training output file path", "job log path", etc. on the webpage
# (8) Create a training job
#
# Eval on ModelArts
# (1) choose a or b
# a. in default_config.yaml set "enable_modelarts=True"
# in default_config.yaml set "checkpoint_path='s3://dir_to_your_trained_model/'"
# in default_config.yaml set "mindrecord_dir='./MindRecord_COCO'"
# in default_config.yaml set "coco_root='/cache/data'"
# in default_config.yaml set other parameters
# b. Set on the web page "enable_modelarts=True"
# Set on the web page "checkpoint_path='s3://dir_to_your_trained_model/'"
# Set on the web page "mindrecord_dir='./MindRecord_COCO'"
# Set on the web page "coco_root='/cache/data'"
# Set on the web page other parameters
# (2) Prepare model code
# (3) Upload your pretrained model to the S3 bucket
# (4) choose a or b (recommended choice a)
# a. First, compress the dataset into a ".zip" file.
# Second, upload your compressed dataset to the S3 bucket (you can also upload the uncompressed dataset, but that may be slow.)
# b. Upload the original dataset to the S3 bucket.
# (Data set conversion occurs during the training process, which takes more time. It will be re-transformed every time you train.)
# (5) Set your code path on the web page to "/path/retinanet"
# (6) Set the startup file on the web page as "eval.py"
# (7) Set "training dataset", "training output file path", "job log path", etc. on the webpage
# (8) Create a training job
```
- Export in ModelArts (If you want to run on ModelArts, you can refer to the following [modelarts documents](https://support.huaweicloud.com/modelarts/))
```python
# (1) choose a or b
# a. in default_config.yaml set "enable_modelarts=True"
# in default_config.yaml set "file_name='retinanet'"
# in default_config.yaml set "file_format='MINDIR'"
# in base_config.yaml set "checkpoint_path='/The path of checkpoint in S3/'"
# in base_config.yaml set other parameters
# b. Set on the web page "enable_modelarts=True"
# Set on the web page "file_name='retinanet'"
# Set on the web page "file_format='MINDIR'"
# Set on the web page "checkpoint_path='/The path of checkpoint in S3/'"
# Set on the web page other parameters
# (2) Upload your pretrained model to the S3 bucket
# (3) Set your code path on the web page to "/path/retinanet"
# (4) Set the startup file on the web page as "export.py"
# (5) Set "training dataset", "training output file path", "job log path", etc. on the webpage
# (6) Create a training job
```
## [Script Description](#content)
### [Script and Sample Code](#content)
```text
.
└─Retinanet_resnet152
├─README.md
├─README_CN.md
├─ascend310_infer
├─scripts
├─run_single_train.sh # training on Ascend single card
├─run_single_train_gpu.sh # training on GPU single card
├─run_distribute_train.sh # training on Ascend multiple card
├─run_distribute_train_gpu.sh # training on GPU multiple card
├─run_eval.sh # inference on Ascend
├─run_eval_gpu.sh # inference on GPU
└─run_infer_310.sh
├─src
├─backbone.py
├─bottleneck.py
├─dataset.py
├─retinahead.py
├─init_params.py
├─lr_schedule.py
├─coco_eval
├─box_utils.py
└─model_utils
├─config.py
├─device_adapter.py
├─local_adapter.py
└─moxing_adapter.py
├─default_config.yaml
├─export.py
├─postprocess.py
├─requirements.txt
├─train.py
└─eval.py
```
### [Script Parameters](#content)
```text
Main parameteres used in train.py and config.py:
"img_shape": [640, 640], # image size
"num_retinanet_boxes": 76725, # The total number of a priori boxes set
"match_thershold": 0.5,
"softnms_sigma": 0.5,
"nms_thershold": 0.6,
"min_score": 0.1,
"max_boxes": 100, # Maximum number of detection frames
"global_step": 0,
"lr_init": 1e-6,
"lr_end_rate": 5e-3, # The ratio of the final learning rate to the maximum learning rate
"warmup_epochs1": 2, # Number of cycles of the 1st stage warmup
"warmup_epochs2": 5, # Number of cycles of the 2nd stage warmup
"warmup_epochs3": 23, # Number of cycles of the 3d stage warmup
"warmup_epochs4": 60, # Number of cycles of the 4th stage warmup
"warmup_epochs5": 160, # Number of cycles of the 5th stage warmup
"momentum": 0.9, # momentum
"weight_decay": 1.5e-4,
"num_default": [9, 9, 9, 9, 9], # The number of a priori boxes in a single grid
"extras_out_channels": [256, 256, 256, 256, 256], # Feature layer output channels
"feature_size": [75, 38, 19, 10, 5],
"aspect_ratios": [(0.5,1.0,2.0), (0.5,1.0,2.0), (0.5,1.0,2.0), (0.5,1.0,2.0), (0.5,1.0,2.0)], # Priori box size change ratio
"steps": ( 8, 16, 32, 64, 128), # Priori box setting step size
"anchor_size":(32, 64, 128, 256, 512), # A priori box size
"prior_scaling": (0.1, 0.2), # Used to adjust the ratio of regression and regression in loss
"gamma": 2.0, # focal loss parameter
"alpha": 0.75, # focal loss parameter
"mindrecord_dir": "/opr/root/data/MindRecord_COCO",
"coco_root": "/opr/root/data/",
"train_data_type": "train2017",
"val_data_type": "val2017",
"instances_set": "annotations_trainval2017/annotations/instances_{}.json",
"coco_classes": ('background', 'person', 'bicycle', 'car', 'motorcycle', 'airplane', 'bus',
'train', 'truck', 'boat', 'traffic light', 'fire hydrant',
'stop sign', 'parking meter', 'bench', 'bird', 'cat', 'dog',
'horse', 'sheep', 'cow', 'elephant', 'bear', 'zebra',
'giraffe', 'backpack', 'umbrella', 'handbag', 'tie',
'suitcase', 'frisbee', 'skis', 'snowboard', 'sports ball',
'kite', 'baseball bat', 'baseball glove', 'skateboard',
'surfboard', 'tennis racket', 'bottle', 'wine glass', 'cup',
'fork', 'knife', 'spoon', 'bowl', 'banana', 'apple',
'sandwich', 'orange', 'broccoli', 'carrot', 'hot dog', 'pizza',
'donut', 'cake', 'chair', 'couch', 'potted plant', 'bed',
'dining table', 'toilet', 'tv', 'laptop', 'mouse', 'remote',
'keyboard', 'cell phone', 'microwave', 'oven', 'toaster', 'sink',
'refrigerator', 'book', 'clock', 'vase', 'scissors',
'teddy bear', 'hair drier', 'toothbrush'),
"num_classes": 81,
"voc_root": "",
"voc_dir": "",
"image_dir": "",
"anno_path": "",
"save_checkpoint": True,
"save_checkpoint_epochs": 1,
"keep_checkpoint_max":1,
"save_checkpoint_path": "./model", # Path to save checkpoints
"finish_epoch":0, # number of epoch that have been run
"checkpoint_path":"/opr/root/reretina/retinanet2/LOG0/model/retinanet-400_458.ckpt" # checkpoint path for evaluation
```
### [Training process](#content)
#### Usage
You can use python or shell scripts for training. The usage of the shell script is as follows:
- Ascend:
```bash
# data and the path to store the mindrecord file are set in default_config.yaml
# before training convert dataset to MindRecord format:
python train.py --only_create_dataset=True --run_platform="Ascend"
# Eight-card parallel training example：
# create RANK_TABLE_FILE
bash run_distribute_train.sh DEVICE_NUM EPOCH_SIZE LR DATASET RANK_TABLE_FILE PRE_TRAINED(optional) PRE_TRAINED_EPOCH_SIZE(optional)
# Single card training example：
bash run_distribute_train.sh DEVICE_ID EPOCH_SIZE LR DATASET PRE_TRAINED(optional) PRE_TRAINED_EPOCH_SIZE(optional)
```
> Note: RANK_TABLE_FILE related reference materials see in this [link](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/distributed_training_ascend.html),
> for details on how to get device_ip check this [link](https://gitee.com/mindspore/models/tree/master/utils/hccl_tools).
- GPU:
```bash
# data and the path to store the mindrecord file are set in default_config.yaml
# convert dataset to MindRecord format:
python train.py --only_create_dataset=True --run_platform="GPU"
# Eight-card parallel training example：
bash run_distribute_train_gpu.sh [DEVICE_NUM] [EPOCH_SIZE] [LR] [DATASET] [PRE_TRAINED](optional) [PRE_TRAINED_EPOCH_SIZE](optional)
# Single card training example：
bash run_single_train_gpu.sh [DEVICE_ID] [EPOCH_SIZE] [LR] [DATASET] [PRE_TRAINED](optional) [PRE_TRAINED_EPOCH_SIZE](optional)
```
#### Run
- Ascend:
```bash
# Ascend 8-card parallel training example (run in the retinanet directory):
bash scripts/run_distribute_train.sh 8 500 0.1 coco scripts/rank_table_8pcs.json /dataset/retinanet-322_458.ckpt 322
# Ascend single card training example (run in the retinanet directory):
bash scripts/run_single_train.sh 0 500 0.1 coco /dataset/retinanet-322_458.ckpt 322
```
- GPU:
```bash
# Ascend 8-card parallel training example (run in the retinanet directory):
bash run_distribute_train_gpu.sh 8 500 0.1 coco /dataset/retinanet-322_458.ckpt 322
# Ascend single card training example (run in the retinanet directory):
bash run_single_train_gpu.sh 0 500 0.1 coco /dataset/retinanet-322_458.ckpt 322
```
#### Result
Paths are set in `default_config.yaml`. Checkpoints will be save in `./model`. The training log will be recorded to LOG/train.log，an example of the training log is as follows:
```text
epoch: 117 step: 916, loss is 0.8391866
lr:[0.025187]
epoch time: 444315.944 ms, per step time: 485.061 ms
epoch: 118 step: 916, loss is 1.070719
lr:[0.025749]
epoch time: 444288.450 ms, per step time: 485.031 ms
epoch: 119 step: 916, loss is 1.1607553
lr:[0.026312]
epoch time: 444339.538 ms, per step time: 485.087 ms
epoch: 120 step: 916, loss is 1.174742
lr:[0.026874]
epoch time: 444237.851 ms, per step time: 484.976 ms
```
### [Evaluation process](#content)
#### Usage
You can use python or shell scripts for training. The usage of the shell script is as follows:
- Ascend
```bash
bash scripts/run_eval.sh [DATASET] [DEVICE_ID]
```
- GPU
```bash
bash run_eval_gpu.sh [DATASET] [DEVICE_ID] [CHECKPOINT_PATH]
```
#### Run
- Ascend
```bash
bash scripts/run_eval.sh coco 0
```
- GPU
```bash
bash run_eval_gpu.sh coco 0 LOG/model/retinanet-500_610.ckpt
```
> checkpoint can be generated during training.
#### Result
The calculation results will be stored in the example path, which you can view in `eval.log`.
```text
Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.357
Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.503
Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.400
Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.141
Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.381
Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.498
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 1 ] = 0.307
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 10 ] = 0.447
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.458
Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.172
Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.487
Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.643
========================================
mAP: 0.3571988469737286
```
### [Model Export](#content)
#### Usage
Before exporting the model, modify the checkpoint_path configuration item in the `default_config.yaml` file. The value is the path of the checkpoint.
```shell
python export.py --file_name [RUN_PLATFORM] --file_format[EXPORT_FORMAT] --checkpoint_path [CHECKPOINT PATH]
```
`EXPORT_FORMAT` choose from ["AIR", "MINDIR"]
#### Run
```bash
python export.py
```
### [Inference Process](#content)
#### Usage
Before inference, model needs to be exported to the Ascend 910 environment. Pictures with iscrowd set to true should be excluded. The image id after removal is saved in the ascend310_infer directory.
You also need to modify the coco_root, val_data_type, and instances_set configuration items in the config.py file. The values are respectively taken as the directory of the coco data set, the directory name of the data set used for inference, and the annotation file used to calculate the accuracy after the inference is completed. The instances_set is spliced with val_data_type to ensure that the file is correct and exists.
```shell
# Ascend310 inference
bash run_infer_310.sh [MINDIR_PATH] [DATA_PATH] [ANN_FILE] [DEVICE_ID]
```
#### Run
```shell
bash run_infer_310.sh [MINDIR_PATH] [DATASET_PATH] [DEVICE_ID]
```
#### Result
The result of the inference is saved in acc.log in the current directory, and the result similar to the following:
```text
Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.356
Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.499
Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.396
Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.145
Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.380
Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.506
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 1 ] = 0.308
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 10 ] = 0.446
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.457
Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.179
Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.483
Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.647
mAP: 0.35625723922139957
```
## [Model Description](#content)
### [Performance](#content)
#### Training Performance
| Parameters | Ascend (8 pcs) | GPU Tesla V100 (1 pcs) | GPU Tesla V100 (8 pcs) |
| -------------------------- |----------------------------------------------------------------------------------------|----------------------------------------------------------------------------------------|----------------------------------------------------------------------------------------|
| Model | Retinanet-resnet-152 | Retinanet-resnet-152 | Retinanet-resnet-152 |
| Environment | Huawei cloud Modelarts | Ubuntu 18.04.6, Tesla V100, CPU 2.7 GHz, 56 cores, RAM 504 GB | Ubuntu 18.04.6, Tesla V100 (8 pcs), CPU 2.7 GHz, 56 cores, RAM 504 GB |
| Uploaded Date | 10/03/2021 | 27/12/2021 | 27/12/2021 |
| MindSpore version | 1.0.1 | 1.6.0 | 1.6.0 |
| Dataset | 118287 images | 118287 images | 118287 images |
| Training Parameters | batch_size=16 | batch_size=12, epochs=200, lr=0.1, steps_per_epoch=9772 | batch_size=12*8, epochs=180, lr=0.1, steps_per_epoch=1221 |
| Other training parameters | default_config.yaml | default_config.yaml | default_config.yaml |
| Optimizer | Momentum | Momentum | Momentum |
| Loss function | Focal loss | Focal loss | Focal loss |
| Final loss | 0.69 | 0.84 | 0.84 |
| Speed | | 860 ms/step | 1205 ms/step |
| Total training time (8p) | 41h 32m 20s | 440h | 72h |
| Script | [Link](https://gitee.com/mindspore/models/tree/master/research/cv/retinanet_resnet152) | [Link](https://gitee.com/mindspore/models/tree/master/research/cv/retinanet_resnet152) | [Link](https://gitee.com/mindspore/models/tree/master/research/cv/retinanet_resnet152) |
#### Evaluation Performance
| Parameters | Ascend | GPU |
|-------------------| --------------------------- |---------------------------------------------------------------|
| Model | Retinanet-resnet-152 | Retinanet-resnet-152 |
| Environment | Huawei cloud Modelarts | Ubuntu 18.04.6, Tesla V100, CPU 2.7 GHz, 56 cores, RAM 504 GB |
| Uploaded Date | 10/03/2021 | 27/12/2021 |
| MindSpore version | 1.0.1 | 1.6.0 |
| Dataset | 5k images | 5k images |
| Batch_size | 1 | 1 |
| Accuracy | mAP[0.3571] | |
| Total time | 12 min and 03 seconds | 12 min |
# [Description of Random State](#content)
Random seed is set in function `create_dataset` of `dataset.py` script. It's also set in the `train.py` script.
# [ModelZoo Homepage](#content)
Please check the official [homepage](https://gitee.com/mindspore/models).
research/cv/retinanet_resnet152/README_CN.md
+ 94
− 75
View file @ 3941f823
...@@ -2,6 +2,8 @@ ...@@ -2,6 +2,8 @@
<!-- TOC --> <!-- TOC -->
[View English](./README.md)
- [Retinanet 描述](#retinanet描述) - [Retinanet 描述](#retinanet描述)
- [模型架构](#模型架构) - [模型架构](#模型架构)
- [数据集](#数据集) - [数据集](#数据集)
...@@ -48,19 +50,14 @@ Retinanet的整体网络架构如下所示： ...@@ -48,19 +50,14 @@ Retinanet的整体网络架构如下所示：
## [数据集](#content) ## [数据集](#content)
数据集可参考文献. 数据集可参考文献. [COCO2017](https://cocodataset.org/#download)
MSCOCO2017
- 数据集大小: 19.3G, 123287张80类彩色图像 - 数据集大小: 19.3G, 123287张80类彩色图像
- 训练:19.3G, 118287张图片 - 训练:19.3G, 118287张图片
- 测试:1814.3M, 5000张图片 - 测试:1814.3M, 5000张图片
- 数据格式:RGB图像
- 数据格式:RGB图像. >注意：数据将在src/dataset.py 中被处理
- 注意：数据将在src/dataset.py 中被处理
## [环境要求](#content) ## [环境要求](#content)
...@@ -175,7 +172,7 @@ MSCOCO2017 ...@@ -175,7 +172,7 @@ MSCOCO2017
### [脚本和示例代码](#content) ### [脚本和示例代码](#content)
```shell ```text
. .
└─Retinanet_resnet152 └─Retinanet_resnet152
├─README.md ├─README.md
...@@ -201,12 +198,11 @@ MSCOCO2017 ...@@ -201,12 +198,11 @@ MSCOCO2017
├─default_config.yaml # 参数配置 ├─default_config.yaml # 参数配置
├─train.py # 网络训练脚本 ├─train.py # 网络训练脚本
└─eval.py # 网络推理脚本 └─eval.py # 网络推理脚本
``` ```
### [脚本参数](#content) ### [脚本参数](#content)
```python ```text
在train.py和default_config.yaml脚本中使用到的主要参数是: 在train.py和default_config.yaml脚本中使用到的主要参数是:
"img_shape": [640, 640], # 图像尺寸 "img_shape": [640, 640], # 图像尺寸
"num_retinanet_boxes": 76725, # 设置的先验框总数 "num_retinanet_boxes": 76725, # 设置的先验框总数
...@@ -275,52 +271,65 @@ MSCOCO2017 ...@@ -275,52 +271,65 @@ MSCOCO2017
- Ascend: - Ascend:
```训练 ```bash
# 八卡并行训练示例： # data和存储mindrecord文件的路径在default_config.yaml里设置
创建 RANK_TABLE_FILE # 训练以前， 请运行：
python train.py --only_create_dataset=True --run_platform="Ascend"
# 八卡并行训练示例：
# 创建 RANK_TABLE_FILE
bash run_distribute_train.sh DEVICE_NUM EPOCH_SIZE LR DATASET RANK_TABLE_FILE PRE_TRAINED(optional) PRE_TRAINED_EPOCH_SIZE(optional) bash run_distribute_train.sh DEVICE_NUM EPOCH_SIZE LR DATASET RANK_TABLE_FILE PRE_TRAINED(optional) PRE_TRAINED_EPOCH_SIZE(optional)
# 单卡训练示例： # 单卡训练示例：
bash run_distribute_train.sh DEVICE_ID EPOCH_SIZE LR DATASET PRE_TRAINED(optional) PRE_TRAINED_EPOCH_SIZE(optional) bash run_distribute_train.sh DEVICE_ID EPOCH_SIZE LR DATASET PRE_TRAINED(optional) PRE_TRAINED_EPOCH_SIZE(optional)
``` ```
> 注意: > 注意: RANK_TABLE_FILE相关参考资料见[链接](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/distributed_training_ascend.html),
> 获取device_ip方法详见[链接](https://gitee.com/mindspore/models/tree/master/utils/hccl_tools).
RANK_TABLE_FILE相关参考资料见[链接](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/distributed_training_ascend.html), 获取device_ip方法详见[链接](https://gitee.com/mindspore/models/tree/master/utils/hccl_tools). - GPU:
#### 运行 ```bash
# data和存储mindrecord文件的路径在default_config.yaml里设置
``` 运行 # 训练以前， 请运行：
# 训练示例 python train.py --only_create_dataset=True --run_platform="GPU"
python: # 八卡并行训练示例：
data和存储mindrecord文件的路径在config里设置 bash run_distribute_train_gpu.sh [DEVICE_NUM] [EPOCH_SIZE] [LR] [DATASET] [PRE_TRAINED](optional) [PRE_TRAINED_EPOCH_SIZE](optional)
# 单卡训练示例： # 单卡训练示例：
bash run_single_train_gpu.sh [DEVICE_ID] [EPOCH_SIZE] [LR] [DATASET] [PRE_TRAINED](optional) [PRE_TRAINED_EPOCH_SIZE](optional)
```
#### 运行
python train.py - Ascend
shell:
Ascend:
```bash
# 八卡并行训练示例(在retinanet目录下运行)： # 八卡并行训练示例(在retinanet目录下运行)：
bash scripts/run_distribute_train.sh 8 500 0.1 coco scripts/rank_table_8pcs.json /dataset/retinanet-322_458.ckpt 322
bash scripts/run_distribute_train.sh 8 500 0.1 coco RANK_TABLE_FILE(创建的RANK_TABLE_FILE的地址) PRE_TRAINED(预训练checkpoint地址) PRE_TRAINED_EPOCH_SIZE（预训练EPOCH大小）
例如：bash scripts/run_distribute_train.sh 8 500 0.1 coco scripts/rank_table_8pcs.json /dataset/retinanet-322_458.ckpt 322
# 单卡训练示例(在retinanet目录下运行)： # 单卡训练示例(在retinanet目录下运行)：
bash scripts/run_single_train.sh 0 500 0.1 coco /dataset/retinanet-322_458.ckpt 322 bash scripts/run_single_train.sh 0 500 0.1 coco /dataset/retinanet-322_458.ckpt 322
```
- GPU
```bash
# 八卡并行训练示例(在retinanet目录下运行)：
bash run_distribute_train_gpu.sh 8 500 0.1 coco /dataset/retinanet-322_458.ckpt 322
# 单卡训练示例(在retinanet目录下运行)：
bash run_single_train_gpu.sh 0 500 0.1 coco /dataset/retinanet-322_458.ckpt 322
``` ```
#### 结果 #### 结果
训练结果将存储在示例路径中。checkpoint将存储在 `./model` 路径下，训练日志将被记录到 `./log.txt` 中，训练日志部分示例如下： 路径在`default_config.yaml`里设置。checkpoint将存储在 `./model` 路径下，训练日志将被记录到 `./train.log` 中，训练日志部分示例如下：
``` 训练日志 ```text
epoch: 117 step: 916, loss is 0.8391866 epoch: 117 step: 916, loss is 0.8391866
lr:[0.025187] lr:[0.025187]
epoch time: 444315.944 ms, per step time: 485.061 ms epoch time: 444315.944 ms, per step time: 485.061 ms
...@@ -341,20 +350,30 @@ epoch time: 444237.851 ms, per step time: 484.976 ms ...@@ -341,20 +350,30 @@ epoch time: 444237.851 ms, per step time: 484.976 ms
您可以使用python或shell脚本进行训练。shell脚本的用法如下: 您可以使用python或shell脚本进行训练。shell脚本的用法如下:
```eval - Ascend
```bash
bash scripts/run_eval.sh [DATASET] [DEVICE_ID] bash scripts/run_eval.sh [DATASET] [DEVICE_ID]
``` ```
- GPU
```bash
bash run_eval_gpu.sh [DATASET] [DEVICE_ID] [CHECKPOINT_PATH]
```
#### 运行 #### 运行
```eval运行 - Ascend
# 验证示例
```bash
bash scripts/run_eval.sh coco 0
```
- GPU
python: ```bash
Ascend: python eval.py bash run_eval_gpu.sh coco 0 LOG/model/retinanet-500_610.ckpt
checkpoint 的路径在config里设置
shell:
Ascend: bash scripts/run_eval.sh coco 0
``` ```
> checkpoint 可以在训练过程中产生. > checkpoint 可以在训练过程中产生.
...@@ -363,7 +382,7 @@ bash scripts/run_eval.sh [DATASET] [DEVICE_ID] ...@@ -363,7 +382,7 @@ bash scripts/run_eval.sh [DATASET] [DEVICE_ID]
计算结果将存储在示例路径中，您可以在 `eval.log` 查看. 计算结果将存储在示例路径中，您可以在 `eval.log` 查看.
``` mAP ```text
Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.357 Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.357
Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.503 Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.503
Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.400 Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.400
...@@ -386,7 +405,7 @@ mAP: 0.3571988469737286 ...@@ -386,7 +405,7 @@ mAP: 0.3571988469737286
#### 用途 #### 用途
导出模型前要修改default_config.yaml文件中的checkpoint_path配置项，值为checkpoint的路径。 导出模型前要修改`default_config.yaml`文件中的checkpoint_path配置项，值为checkpoint的路径。
```shell ```shell
python export.py --file_name [RUN_PLATFORM] --file_format[EXPORT_FORMAT] --checkpoint_path [CHECKPOINT PATH] python export.py --file_name [RUN_PLATFORM] --file_format[EXPORT_FORMAT] --checkpoint_path [CHECKPOINT PATH]
...@@ -396,7 +415,7 @@ python export.py --file_name [RUN_PLATFORM] --file_format[EXPORT_FORMAT] --check ...@@ -396,7 +415,7 @@ python export.py --file_name [RUN_PLATFORM] --file_format[EXPORT_FORMAT] --check
#### 运行方式 #### 运行方式
```运行 ```python
python export.py python export.py
``` ```
...@@ -414,7 +433,7 @@ bash run_infer_310.sh [MINDIR_PATH] [DATA_PATH] [ANN_FILE] [DEVICE_ID] ...@@ -414,7 +433,7 @@ bash run_infer_310.sh [MINDIR_PATH] [DATA_PATH] [ANN_FILE] [DEVICE_ID]
#### 运行命令 #### 运行命令
```运行 ```shell
bash run_infer_310.sh [MINDIR_PATH] [DATASET_PATH] [DEVICE_ID] bash run_infer_310.sh [MINDIR_PATH] [DATASET_PATH] [DEVICE_ID]
``` ```
...@@ -422,7 +441,7 @@ bash run_infer_310.sh [MINDIR_PATH] [DATASET_PATH] [DEVICE_ID] ...@@ -422,7 +441,7 @@ bash run_infer_310.sh [MINDIR_PATH] [DATASET_PATH] [DEVICE_ID]
推理的结果保存在当前目录下，在acc.log日志文件中可以找到类似以下的结果。 推理的结果保存在当前目录下，在acc.log日志文件中可以找到类似以下的结果。
```mAP ```text
Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.356 Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.356
Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.499 Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.499
Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.396 Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.396
...@@ -444,34 +463,34 @@ mAP: 0.35625723922139957 ...@@ -444,34 +463,34 @@ mAP: 0.35625723922139957
#### 训练性能 #### 训练性能
| 参数 | Ascend | | 参数 | Ascend (8pcs) | GPU Tesla V100 (1 pcs) | GPU Tesla V100 (8 pcs) |
| -------------------------- | ------------------------------------- | | -------------------------- | --------------------- | ---- | ---- |
| 模型名称 | Retinanet | | 模型名称 | Retinanet-resnet-152 | Retinanet-resnet-152 | Retinanet-resnet-152 |
| 运行环境 | 华为云 Modelarts | | 运行环境 | 华为云 Modelarts | Ubuntu 18.04.6, Tesla V100, CPU 2.7 GHz, 56 cores, RAM 504 GB | Ubuntu 18.04.6, Tesla V100 (8 pcs), CPU 2.7 GHz, 56 cores, RAM 504 GB |
| 上传时间 | 10/03/2021 | | 上传时间 | 10/03/2021 | 27/12/2021 | 27/12/2021 |
| MindSpore 版本 | 1.0.1 | | MindSpore 版本 | 1.0.1 | 1.6.0 | 1.6.0 |
| 数据集 | 123287 张图片 | | 数据集 | 123287 张图片 | 123287 张图片 | 123287 张图片 |
| Batch_size | 16 | | 训练参数 | 16 | batch_size=12, epochs=180, lr=0.1, steps_per_epoch=1221 | batch_size=12*8, epochs=180, lr=0.1, steps_per_epoch=1221 |
| 训练参数 | default_config.yaml | | 其他训练参数 | default_config.yaml | default_config.yaml | default_config.yaml |
| 优化器 | Momentum | | 优化器 | Momentum | Momentum | Momentum |
| 损失函数 | Focal loss | | 损失函数 | Focal loss | Focal loss | Focal loss |
| 最终损失 | 0.69 | | 最终损失 | 0.69 | 0.84 | 0.84 |
| 精确度 (8p) | mAP[0.3571] | | 速度 (8p) | | 860 毫秒/步 | 1205 毫秒/步 |
| 训练总时间 (8p) | 41h32m20s | | 训练总时间 (8p) | 41h32m20s | 440小时 | 72小时 |
| 脚本 | [Retianet script](https://gitee.com/mindspore/models/tree/master/research/cv/retinanet_resnet152) | | 脚本 | [Retianet script](https://gitee.com/mindspore/models/tree/master/research/cv/retinanet_resnet152) |
#### 推理性能 #### 推理性能
| 参数 | Ascend | | 参数 | Ascend | GPU |
| ------------------- | :-------------------------- | | ------------------- | :-------------------------- | --- |
| 模型名称 | Retinanet | | 模型名称 | Retinanet-resnet-152 | Retinanet-resnet-152 |
| 运行环境 | 华为云 Modelarts | | 运行环境 | 华为云 Modelarts | Ubuntu 18.04.6, Tesla V100, CPU 2.7 GHz, 56 cores, RAM 504 GB |
| 上传时间 | 10/03/2021 | | 上传时间 | 10/03/2021 | 27/12/2021 |
| MindSpore 版本 | 1.0.1 | | MindSpore 版本 | 1.0.1 | 1.6.0 |
| 数据集 | 5k 张图片 | | 数据集 | 5k 张图片 | 5k 张图片 |
| Batch_size | 1 | | Batch_size | 1 | 1 |
| 精确度 | mAP[0.3571] | | 精确度 | mAP[0.3571] | |
| 总时间 | 12m3s | | 总时间 | 12m3s | 12m |
# [随机情况的描述](#内容) # [随机情况的描述](#内容)
......
research/cv/retinanet_resnet152/default_config.yaml
+ 6
− 6
View file @ 3941f823
...@@ -49,8 +49,8 @@ gamma: 2.0 ...@@ -49,8 +49,8 @@ gamma: 2.0
alpha: 0.75 alpha: 0.75
# `mindrecord_dir` and `coco_root` are better to use absolute path. # `mindrecord_dir` and `coco_root` are better to use absolute path.
mindrecord_dir: "/cache/train/MindRecord_COCO" mindrecord_dir: "/data/Datasets/COCO2017/MindRecord_COCO"
coco_root: "/cache/data" coco_root: "/data/Datasets/COCO2017"
train_data_type: "train2017" train_data_type: "train2017"
val_data_type: "val2017" val_data_type: "val2017"
instances_set: "annotations/instances_{}.json" instances_set: "annotations/instances_{}.json"
...@@ -78,7 +78,7 @@ voc_dir: "" ...@@ -78,7 +78,7 @@ voc_dir: ""
image_dir: "" image_dir: ""
anno_path: "" anno_path: ""
save_checkpoint: True save_checkpoint: True
keep_checkpoint_max: 30 keep_checkpoint_max: 50
save_checkpoint_path: "./model" save_checkpoint_path: "./model"
finish_epoch: 0 finish_epoch: 0
checkpoint_path: "/cache/train/model/retinanet-10_916.ckpt" checkpoint_path: "/cache/train/model/retinanet-10_916.ckpt"
...@@ -91,7 +91,7 @@ device_num: 1 ...@@ -91,7 +91,7 @@ device_num: 1
lr: 0.1 lr: 0.1
mode: "sink" mode: "sink"
dataset: "coco" dataset: "coco"
epoch_size: 500 epoch_size: 180
batch_size: 16 batch_size: 16
pre_trained: '' pre_trained: ''
pre_trained_epoch_size: 0 pre_trained_epoch_size: 0
...@@ -126,7 +126,7 @@ pre_trained_epoch_size: 'Pretrained epoch size.' ...@@ -126,7 +126,7 @@ pre_trained_epoch_size: 'Pretrained epoch size.'
save_checkpoint_epochs: 'Save checkpoint epochs, default is 1.' save_checkpoint_epochs: 'Save checkpoint epochs, default is 1.'
loss_scale: 'Loss scale, default is 1024.' loss_scale: 'Loss scale, default is 1024.'
filter_weight: 'Filter weight parameters, default is False.' filter_weight: 'Filter weight parameters, default is False.'
run_platform: 'Run platform, only support Ascend in training and export, only support Ascend and GPU in evaling.' run_platform: 'Run platform, only support Ascend and GPU.'
file_format: 'file format' file_format: 'file format'
file_name: 'output file name.' file_name: 'output file name.'
result_path: 'result file path.' result_path: 'result file path.'
...@@ -139,5 +139,5 @@ train_url: 'Training output url for obs' ...@@ -139,5 +139,5 @@ train_url: 'Training output url for obs'
data_path: 'Dataset path for local' data_path: 'Dataset path for local'
--- ---
run_platform: ['Ascend'] run_platform: ['Ascend', 'GPU']
file_format: ["AIR", "MINDIR"] file_format: ["AIR", "MINDIR"]
research/cv/retinanet_resnet152/eval.py
+ 3
− 0
View file @ 3941f823
...@@ -38,6 +38,7 @@ def retinanet_eval(dataset_path, ckpt_path): ...@@ -38,6 +38,7 @@ def retinanet_eval(dataset_path, ckpt_path):
net = retinahead(backbone, config) net = retinahead(backbone, config)
net = retinanetInferWithDecoder(net, Tensor(default_boxes), config) net = retinanetInferWithDecoder(net, Tensor(default_boxes), config)
print("Load Checkpoint!") print("Load Checkpoint!")
print("ckpt_path: ", ckpt_path)
param_dict = load_checkpoint(ckpt_path) param_dict = load_checkpoint(ckpt_path)
net.init_parameters_data() net.init_parameters_data()
load_param_into_net(net, param_dict) load_param_into_net(net, param_dict)
...@@ -77,6 +78,7 @@ def modelarts_process(): ...@@ -77,6 +78,7 @@ def modelarts_process():
config.coco_root = os.path.join(config.coco_root, config.modelarts_dataset_unzip_name) config.coco_root = os.path.join(config.coco_root, config.modelarts_dataset_unzip_name)
print(os.listdir(os.path.join(config.data_path, config.modelarts_dataset_unzip_name))) print(os.listdir(os.path.join(config.data_path, config.modelarts_dataset_unzip_name)))
@moxing_wrapper(pre_process=modelarts_process) @moxing_wrapper(pre_process=modelarts_process)
def eval_retinanet_resnet152(): def eval_retinanet_resnet152():
""" eval_retinanet_resnet152 """ """ eval_retinanet_resnet152 """
...@@ -115,5 +117,6 @@ def eval_retinanet_resnet152(): ...@@ -115,5 +117,6 @@ def eval_retinanet_resnet152():
print("Start Eval!") print("Start Eval!")
retinanet_eval(mindrecord_file, config.checkpoint_path) retinanet_eval(mindrecord_file, config.checkpoint_path)
if __name__ == '__main__': if __name__ == '__main__':
eval_retinanet_resnet152() eval_retinanet_resnet152()
research/cv/retinanet_resnet152/export.py
+ 2
− 0
View file @ 3941f823
...@@ -28,6 +28,7 @@ from src.backbone import resnet152 ...@@ -28,6 +28,7 @@ from src.backbone import resnet152
def modelarts_process(): def modelarts_process():
pass pass
@moxing_wrapper(pre_process=modelarts_process) @moxing_wrapper(pre_process=modelarts_process)
def export_retinanet_resnet152(): def export_retinanet_resnet152():
""" export_retinanet_resnet152 """ """ export_retinanet_resnet152 """
...@@ -44,5 +45,6 @@ def export_retinanet_resnet152(): ...@@ -44,5 +45,6 @@ def export_retinanet_resnet152():
input_data = Tensor(np.zeros(shape), mstype.float32) input_data = Tensor(np.zeros(shape), mstype.float32)
export(net, input_data, file_name=config.file_name, file_format=config.file_format) export(net, input_data, file_name=config.file_name, file_format=config.file_format)
if __name__ == '__main__': if __name__ == '__main__':
export_retinanet_resnet152() export_retinanet_resnet152()
research/cv/retinanet_resnet152/postprocess.py
+ 4
− 0
View file @ 3941f823
...@@ -30,10 +30,12 @@ def get_pred(result_path, img_id): ...@@ -30,10 +30,12 @@ def get_pred(result_path, img_id):
scores = np.fromfile(scores_file, dtype=np.float32).reshape(76725, config.num_classes) scores = np.fromfile(scores_file, dtype=np.float32).reshape(76725, config.num_classes)
return boxes, scores return boxes, scores
def get_img_size(file_name): def get_img_size(file_name):
img = Image.open(file_name) img = Image.open(file_name)
return img.size return img.size
def get_img_id(img_id_file): def get_img_id(img_id_file):
f = open(img_id_file) f = open(img_id_file)
lines = f.readlines() lines = f.readlines()
...@@ -44,6 +46,7 @@ def get_img_id(img_id_file): ...@@ -44,6 +46,7 @@ def get_img_id(img_id_file):
return ids return ids
def cal_acc(result_path, img_path, img_id_file): def cal_acc(result_path, img_path, img_id_file):
"""Calculate acc""" """Calculate acc"""
ids = get_img_id(img_id_file) ids = get_img_id(img_id_file)
...@@ -66,5 +69,6 @@ def cal_acc(result_path, img_path, img_id_file): ...@@ -66,5 +69,6 @@ def cal_acc(result_path, img_path, img_id_file):
mAP = metrics(pred_data) mAP = metrics(pred_data)
print(f"mAP: {mAP}") print(f"mAP: {mAP}")
if __name__ == '__main__': if __name__ == '__main__':
cal_acc(config.result_path, config.img_path, config.img_id_file) cal_acc(config.result_path, config.img_path, config.img_id_file)
research/cv/retinanet_resnet152/requirements.txt 0 → 100644
+ 1
− 0
View file @ 3941f823
pycocotools
research/cv/retinanet_resnet152/scripts/run_distribute_train.sh
+ 3
− 0
View file @ 3941f823
...@@ -56,9 +56,11 @@ do ...@@ -56,9 +56,11 @@ do
export RANK_ID=$i export RANK_ID=$i
echo "start training for rank $i, device $DEVICE_ID" echo "start training for rank $i, device $DEVICE_ID"
env > env.log env > env.log
if [ $# == 5 ] if [ $# == 5 ]
then then
python train.py \ python train.py \
--run_platform="Ascend" \
--distribute=True \ --distribute=True \
--lr=$LR \ --lr=$LR \
--dataset=$DATASET \ --dataset=$DATASET \
...@@ -70,6 +72,7 @@ do ...@@ -70,6 +72,7 @@ do
if [ $# == 7 ] if [ $# == 7 ]
then then
python train.py \ python train.py \
--run_platform="Ascend" \
--distribute=True \ --distribute=True \
--lr=$LR \ --lr=$LR \
--dataset=$DATASET \ --dataset=$DATASET \
......
research/cv/retinanet_resnet152/scripts/run_distribute_train_gpu.sh 0 → 100644
+ 77
− 0
View file @ 3941f823
#!/bin/bash
# Copyright 2022 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.
# ============================================================================
echo "=============================================================================================================="
echo "Please run the script as: "
echo "sh run_distribute_train_gpu.sh DEVICE_NUM EPOCH_SIZE LR DATASET RANK_TABLE_FILE PRE_TRAINED PRE_TRAINED_EPOCH_SIZE"
echo "for example: sh run_distribute_train_gpu.sh 8 500 0.1 coco /data/hccl.json /opt/retinanet-500_458.ckpt(optional) 200(optional)"
echo "It is better to use absolute path."
echo "================================================================================================================="
if [ $# != 4 ] && [ $# != 6 ]
then
echo "Usage: sh run_distribute_train_gpu.sh [DEVICE_NUM] [EPOCH_SIZE] [LR] [DATASET] \
[PRE_TRAINED](optional) [PRE_TRAINED_EPOCH_SIZE](optional)"
exit 1
fi
echo "After running the script, the network runs in the background. The log will be generated in LOG/train_log.txt"
export DEVICE_NUM=$1
EPOCH_SIZE=$2
LR=$3
DATASET=$4
PRE_TRAINED=$5
PRE_TRAINED_EPOCH_SIZE=$6
rm -rf LOG
mkdir ./LOG
cp ../*.py ./LOG
cp ../*.yaml ./LOG
cp -r ../src ./LOG
cd ./LOG || exit
echo "start training on GPU $DEVICE_NUM devices"
env > env.log
if [ $# == 4 ]
then
mpirun -n $DEVICE_NUM --output-filename log_output --merge-stderr-to-stdout \
python train.py \
--run_platform="GPU" \
--batch_size=12 \
--distribute=True \
--lr=$LR \
--dataset=$DATASET \
--device_num=$DEVICE_NUM \
--epoch_size=$EPOCH_SIZE > train_log.txt 2>&1 &
fi
if [ $# == 6 ]
then
mpirun -n $DEVICE_NUM --output-filename log_output --merge-stderr-to-stdout \
python train.py \
--run_platform="GPU" \
--batch_size=12 \
--distribute=True \
--lr=$LR \
--dataset=$DATASET \
--device_num=$DEVICE_NUM \
--pre_trained=$PRE_TRAINED \
--pre_trained_epoch_size=$PRE_TRAINED_EPOCH_SIZE \
--epoch_size=$EPOCH_SIZE > train_log.txt 2>&1 &
fi
cd ..
\ No newline at end of file
research/cv/retinanet_resnet152/scripts/run_eval.sh
+ 2
− 0
View file @ 3941f823
...@@ -44,7 +44,9 @@ cp -r ./src ./eval$2 ...@@ -44,7 +44,9 @@ cp -r ./src ./eval$2
cd ./eval$2 || exit cd ./eval$2 || exit
env > env.log env > env.log
echo "start inferring for device $DEVICE_ID" echo "start inferring for device $DEVICE_ID"
python eval.py \ python eval.py \
--run_platform="Ascend" \
--dataset=$DATASET \ --dataset=$DATASET \
--device_id=$2 > log.txt 2>&1 & --device_id=$2 > log.txt 2>&1 &
cd .. cd ..
\ No newline at end of file
research/cv/retinanet_resnet152/scripts/run_eval_gpu.sh 0 → 100644
+ 80
− 0
View file @ 3941f823
#!/bin/bash
# Copyright 2022 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.
# ============================================================================
if [ $# != 2 ] && [ $# != 3 ]
then
echo "Usage: sh run_eval_gpu.sh [DATASET] [DEVICE_ID] [CHECKPOINT_PATH](optional)"
exit 1
fi
DATASET=$1
echo $DATASET
export DEVICE_NUM=1
export DEVICE_ID=$2
export RANK_SIZE=$DEVICE_NUM
export RANK_ID=0
get_real_path() {
if [ "${1:0:1}" == "/" ]; then
echo "$1"
else
echo "$(realpath -m $PWD/$1)"
fi
}
if [ $# == 3 ]
then
CHECKPOINT_PATH=$(get_real_path $3)
if [ ! -f $CHECKPOINT_PATH ]
then
echo "error: CHECKPOINT_PATH=$CHECKPOINT_PATH is not a file"
exit 1
fi
fi
if [ -d "eval$2" ];
then
rm -rf ./eval$2
fi
mkdir ./eval$2
cp ../*.py ./eval$2
cp ../*.yaml ./eval$2
cp -r ../src ./eval$2
cd ./eval$2 || exit
env > env.log
echo "start inferring for device $DEVICE_ID"
if [ $# == 2 ]
then
python eval.py \
--run_platform="GPU" \
--dataset=$DATASET \
--device_id=$2 > eval_log.txt 2>&1 &
fi
if [ $# == 3 ]
then
python eval.py \
--run_platform="GPU" \
--dataset=$DATASET \
--checkpoint_path=$CHECKPOINT_PATH \
--device_id=$2 > eval_log.txt 2>&1 &
fi
cd ..
\ No newline at end of file
research/cv/retinanet_resnet152/scripts/run_single_train.sh
+ 29
− 7
View file @ 3941f823
...@@ -38,6 +38,8 @@ export DEVICE_ID=$1 ...@@ -38,6 +38,8 @@ export DEVICE_ID=$1
EPOCH_SIZE=$2 EPOCH_SIZE=$2
LR=$3 LR=$3
DATASET=$4 DATASET=$4
PRE_TRAINED=$5
PRE_TRAINED_EPOCH_SIZE=$6
rm -rf LOG$1 rm -rf LOG$1
mkdir ./LOG$1 mkdir ./LOG$1
...@@ -45,13 +47,33 @@ cp ./*.py ./LOG$1 ...@@ -45,13 +47,33 @@ cp ./*.py ./LOG$1
cp ./*.yaml ./LOG$1 cp ./*.yaml ./LOG$1
cp -r ./src ./LOG$1 cp -r ./src ./LOG$1
cd ./LOG$1 || exit cd ./LOG$1 || exit
echo "start training for device $1" echo "start training for device $1"
env > env.log env > env.log
if [ $# == 4 ]
then
python train.py \ python train.py \
--run_platform="Ascend" \
--distribute=False \ --distribute=False \
--lr=$LR \ --lr=$LR \
--dataset=$DATASET \ --dataset=$DATASET \
--device_num=1 \ --device_num=1 \
--device_id=$DEVICE_ID \ --device_id=$DEVICE_ID \
--epoch_size=$EPOCH_SIZE > log.txt 2>&1 & --epoch_size=$EPOCH_SIZE > log.txt 2>&1 &
fi
if [ $# == 6 ]
then
python train_retinanet.py \
--run_platform="Ascend" \
--distribute=False \
--lr=$LR \
--dataset=$DATASET \
--device_num=1 \
--device_id=$DEVICE_ID \
--pre_trained=$PRE_TRAINED \
--pre_trained_epoch_size=$PRE_TRAINED_EPOCH_SIZE \
--epoch_size=$EPOCH_SIZE > train_log.txt 2>&1 &
fi
cd ../ cd ../
research/cv/retinanet_resnet152/scripts/run_single_train_gpu.sh 0 → 100644
+ 76
− 0
View file @ 3941f823
#!/bin/bash
# Copyright 2022 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
# 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.
# ============================================================================
echo "=============================================================================================================="
echo "Please run the script as: "
echo "sh run_single_train_gpu.sh DEVICE_ID EPOCH_SIZE LR DATASET PRE_TRAINED PRE_TRAINED_EPOCH_SIZE"
echo "for example: sh run_single_train_gpu.sh 0 500 0.1 coco /opt/retinanet-500_458.ckpt(optional) 200(optional)"
echo "It is better to use absolute path."
echo "================================================================================================================="
if [ $# != 4 ] && [ $# != 6 ]
then
echo "Usage: sh run_single_train_gpu.sh [DEVICE_ID] [EPOCH_SIZE] [LR] [DATASET] \
[PRE_TRAINED](optional) [PRE_TRAINED_EPOCH_SIZE](optional)"
exit 1
fi
echo "After running the script, the network runs in the background. The log will be generated in LOGx/train_log.txt"
export DEVICE_ID=$1
EPOCH_SIZE=$2
LR=$3
DATASET=$4
PRE_TRAINED=$5
PRE_TRAINED_EPOCH_SIZE=$6
rm -rf LOG$1
mkdir ./LOG$1
cp ../*.py ./LOG$1
cp ../*.yaml ./LOG$1
cp -r ../src ./LOG$1
cd ./LOG$1 || exit
echo "start training for device $1"
env > env.log
if [ $# == 4 ]
then
python train.py \
--run_platform="GPU" \
--batch_size=12 \
--distribute=False \
--lr=$LR \
--dataset=$DATASET \
--device_num=1 \
--device_id=$DEVICE_ID \
--epoch_size=$EPOCH_SIZE > train_log.txt 2>&1 &
fi
if [ $# == 6 ]
then
python train_retinanet.py \
--run_platform="GPU" \
--batch_size=12 \
--distribute=False \
--lr=$LR \
--dataset=$DATASET \
--device_num=1 \
--device_id=$DEVICE_ID \
--pre_trained=$PRE_TRAINED \
--pre_trained_epoch_size=$PRE_TRAINED_EPOCH_SIZE \
--epoch_size=$EPOCH_SIZE > train_log.txt 2>&1 &
fi
cd ../
research/cv/retinanet_resnet152/src/backbone.py
+ 2
− 1
View file @ 3941f823
...@@ -93,7 +93,7 @@ class ResidualBlock(nn.Cell): ...@@ -93,7 +93,7 @@ class ResidualBlock(nn.Cell):
self.down_sample_layer = nn.Conv2dBnAct(in_channel, out_channel, self.down_sample_layer = nn.Conv2dBnAct(in_channel, out_channel,
kernel_size=1, stride=stride, kernel_size=1, stride=stride,
pad_mode='same', padding=0, has_bn=True, activation='relu') pad_mode='same', padding=0, has_bn=True, activation='relu')
self.add = P.TensorAdd() self.add = P.Add()
self.relu = P.ReLU() self.relu = P.ReLU()
def construct(self, x): def construct(self, x):
...@@ -111,6 +111,7 @@ class ResidualBlock(nn.Cell): ...@@ -111,6 +111,7 @@ class ResidualBlock(nn.Cell):
return out return out
class resnet(nn.Cell): class resnet(nn.Cell):
""" """
ResNet architecture. ResNet architecture.
......
research/cv/retinanet_resnet152/src/bottleneck.py
+ 1
− 0
View file @ 3941f823
...@@ -18,6 +18,7 @@ ...@@ -18,6 +18,7 @@
import mindspore.nn as nn import mindspore.nn as nn
from mindspore.ops import operations as P from mindspore.ops import operations as P
class FPN(nn.Cell): class FPN(nn.Cell):
"""FPN""" """FPN"""
def __init__(self, config, backbone, is_training=True): def __init__(self, config, backbone, is_training=True):
......
research/cv/retinanet_resnet152/src/box_utils.py
+ 1
− 1
View file @ 3941f823
...@@ -74,7 +74,7 @@ def retinanet_bboxes_encode(boxes): ...@@ -74,7 +74,7 @@ def retinanet_bboxes_encode(boxes):
Labels anchors with ground truth inputs. Labels anchors with ground truth inputs.
Args: Args:
boxex: ground truth with shape [N, 5], for each row, it stores [y, x, h, w, cls]. boxes: ground truth with shape [N, 5], for each row, it stores [y, x, h, w, cls].
Returns: Returns:
gt_loc: location ground truth with shape [num_anchors, 4]. gt_loc: location ground truth with shape [num_anchors, 4].
......
research/cv/retinanet_resnet152/src/dataset.py
+ 16
− 8
View file @ 3941f823
...@@ -215,10 +215,13 @@ def create_voc_label(is_training): ...@@ -215,10 +215,13 @@ def create_voc_label(is_training):
if not is_training: if not is_training:
o_width = abs(x_max - x_min) o_width = abs(x_max - x_min)
o_height = abs(y_max - y_min) o_height = abs(y_max - y_min)
ann = {'area': o_width * o_height, 'iscrowd': 0, 'image_id': \ ann = {'area': o_width * o_height,
img_id, 'bbox': [x_min, y_min, o_width, o_height], \ 'iscrowd': 0,
'category_id': cls_map[cls_name], 'id': bnd_id, \ 'image_id': img_id,
'ignore': 0, \ 'bbox': [x_min, y_min, o_width, o_height],
'category_id': cls_map[cls_name],
'id': bnd_id,
'ignore': 0,
'segmentation': []} 'segmentation': []}
json_dict['annotations'].append(ann) json_dict['annotations'].append(ann)
bnd_id = bnd_id + 1 bnd_id = bnd_id + 1
...@@ -390,8 +393,8 @@ def data_to_mindrecord_byte_image(dataset="coco", is_training=True, prefix="reti ...@@ -390,8 +393,8 @@ def data_to_mindrecord_byte_image(dataset="coco", is_training=True, prefix="reti
def create_retinanet_dataset(mindrecord_file, batch_size, repeat_num, device_num=1, rank=0, def create_retinanet_dataset(mindrecord_file, batch_size, repeat_num, device_num=1, rank=0,
is_training=True, num_parallel_workers=64): is_training=True, num_parallel_workers=8):
"""Creatr retinanet dataset with MindDataset.""" """Create retinanet dataset with MindDataset."""
ds = de.MindDataset(mindrecord_file, columns_list=["img_id", "image", "annotation"], num_shards=device_num, ds = de.MindDataset(mindrecord_file, columns_list=["img_id", "image", "annotation"], num_shards=device_num,
shard_id=rank, num_parallel_workers=num_parallel_workers, shuffle=is_training) shard_id=rank, num_parallel_workers=num_parallel_workers, shuffle=is_training)
decode = C.Decode() decode = C.Decode()
...@@ -427,6 +430,8 @@ def create_mindrecord(dataset="coco", prefix="retinanet.mindrecord", is_training ...@@ -427,6 +430,8 @@ def create_mindrecord(dataset="coco", prefix="retinanet.mindrecord", is_training
mindrecord_dir = config.mindrecord_dir mindrecord_dir = config.mindrecord_dir
mindrecord_file = os.path.join(mindrecord_dir, prefix + "0") mindrecord_file = os.path.join(mindrecord_dir, prefix + "0")
if dataset == "voc":
config.coco_root = config.voc_root
if not os.path.exists(mindrecord_file): if not os.path.exists(mindrecord_file):
if not os.path.isdir(mindrecord_dir): if not os.path.isdir(mindrecord_dir):
os.makedirs(mindrecord_dir) os.makedirs(mindrecord_dir)
...@@ -438,12 +443,12 @@ def create_mindrecord(dataset="coco", prefix="retinanet.mindrecord", is_training ...@@ -438,12 +443,12 @@ def create_mindrecord(dataset="coco", prefix="retinanet.mindrecord", is_training
else: else:
print("coco_root not exits.") print("coco_root not exits.")
elif dataset == "voc": elif dataset == "voc":
if os.path.isdir(config.voc_dir): if os.path.isdir(config.voc_dir) and os.path.isdir(config.voc_root):
print("Create Mindrecord.") print("Create Mindrecord.")
voc_data_to_mindrecord(mindrecord_dir, is_training, prefix) voc_data_to_mindrecord(mindrecord_dir, is_training, prefix)
print("Create Mindrecord Done, at {}".format(mindrecord_dir)) print("Create Mindrecord Done, at {}".format(mindrecord_dir))
else: else:
print("voc_dir not exits.") print("voc_root or voc_dir not exits.")
else: else:
if os.path.isdir(config.image_dir) and os.path.exists(config.anno_path): if os.path.isdir(config.image_dir) and os.path.exists(config.anno_path):
print("Create Mindrecord.") print("Create Mindrecord.")
...@@ -451,4 +456,7 @@ def create_mindrecord(dataset="coco", prefix="retinanet.mindrecord", is_training ...@@ -451,4 +456,7 @@ def create_mindrecord(dataset="coco", prefix="retinanet.mindrecord", is_training
print("Create Mindrecord Done, at {}".format(mindrecord_dir)) print("Create Mindrecord Done, at {}".format(mindrecord_dir))
else: else:
print("image_dir or anno_path not exits.") print("image_dir or anno_path not exits.")
else:
print("Mindrecord file exists.")
return mindrecord_file return mindrecord_file
research/cv/retinanet_resnet152/src/model_utils/config.py
+ 1
− 1
View file @ 3941f823
...@@ -119,9 +119,9 @@ def get_config(): ...@@ -119,9 +119,9 @@ def get_config():
help="Config file path") help="Config file path")
path_args, _ = parser.parse_known_args() path_args, _ = parser.parse_known_args()
default, helper, choices = parse_yaml(path_args.config_path) default, helper, choices = parse_yaml(path_args.config_path)
pprint(default)
args = parse_cli_to_yaml(parser=parser, cfg=default, helper=helper, choices=choices, cfg_path=path_args.config_path) args = parse_cli_to_yaml(parser=parser, cfg=default, helper=helper, choices=choices, cfg_path=path_args.config_path)
final_config = merge(args, default) final_config = merge(args, default)
pprint(final_config)
return Config(final_config) return Config(final_config)
config = get_config() config = get_config()
research/cv/retinanet_resnet152/train.py
+ 14
− 3
View file @ 3941f823
...@@ -19,7 +19,7 @@ import os ...@@ -19,7 +19,7 @@ import os
import mindspore import mindspore
import mindspore.nn as nn import mindspore.nn as nn
from mindspore import context, Tensor from mindspore import context, Tensor
from mindspore.communication.management import init, get_rank from mindspore.communication.management import init, get_rank, get_group_size
from mindspore.train.callback import CheckpointConfig, ModelCheckpoint, LossMonitor, TimeMonitor, Callback from mindspore.train.callback import CheckpointConfig, ModelCheckpoint, LossMonitor, TimeMonitor, Callback
from mindspore.train import Model from mindspore.train import Model
from mindspore.context import ParallelMode from mindspore.context import ParallelMode
...@@ -65,11 +65,12 @@ def modelarts_process(): ...@@ -65,11 +65,12 @@ def modelarts_process():
config.coco_root = os.path.join(config.coco_root, config.modelarts_dataset_unzip_name) config.coco_root = os.path.join(config.coco_root, config.modelarts_dataset_unzip_name)
print(os.listdir(os.path.join(config.data_path, config.modelarts_dataset_unzip_name))) print(os.listdir(os.path.join(config.data_path, config.modelarts_dataset_unzip_name)))
@moxing_wrapper(pre_process=modelarts_process) @moxing_wrapper(pre_process=modelarts_process)
def train_retinanet_resnet152(): def train_retinanet_resnet152():
""" train_retinanet_resnet152 """ """ train_retinanet_resnet152 """
context.set_context(mode=context.GRAPH_MODE, device_target=config.run_platform)
if config.run_platform == "Ascend": if config.run_platform == "Ascend":
context.set_context(mode=context.GRAPH_MODE, device_target="Ascend")
if config.distribute: if config.distribute:
if os.getenv("DEVICE_ID", "not_set").isdigit(): if os.getenv("DEVICE_ID", "not_set").isdigit():
context.set_context(device_id=int(os.getenv("DEVICE_ID"))) context.set_context(device_id=int(os.getenv("DEVICE_ID")))
...@@ -83,6 +84,15 @@ def train_retinanet_resnet152(): ...@@ -83,6 +84,15 @@ def train_retinanet_resnet152():
device_num = 1 device_num = 1
context.set_context(device_id=get_device_id()) context.set_context(device_id=get_device_id())
elif config.run_platform == "GPU":
rank = config.device_id
device_num = config.device_num
if config.distribute:
init()
rank = get_rank()
device_num = get_group_size()
context.set_auto_parallel_context(parallel_mode=ParallelMode.DATA_PARALLEL, gradients_mean=True,
device_num=device_num)
else: else:
raise ValueError("Unsupported platform.") raise ValueError("Unsupported platform.")
...@@ -102,7 +112,7 @@ def train_retinanet_resnet152(): ...@@ -102,7 +112,7 @@ def train_retinanet_resnet152():
backbone = resnet152(config.num_classes) backbone = resnet152(config.num_classes)
retinanet = retinahead(backbone, config) retinanet = retinahead(backbone, config)
net = retinanetWithLossCell(retinanet, config) net = retinanetWithLossCell(retinanet, config)
net.to_float(mindspore.float16) net.to_float(mindspore.float32)
init_net_param(net) init_net_param(net)
if config.pre_trained: if config.pre_trained:
...@@ -137,5 +147,6 @@ def train_retinanet_resnet152(): ...@@ -137,5 +147,6 @@ def train_retinanet_resnet152():
cb += [ckpt_cb] cb += [ckpt_cb]
model.train(config.epoch_size, dataset, callbacks=cb, dataset_sink_mode=True) model.train(config.epoch_size, dataset, callbacks=cb, dataset_sink_mode=True)
if __name__ == '__main__': if __name__ == '__main__':
train_retinanet_resnet152() train_retinanet_resnet152()
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Please register or to comment