ziquan b37e3b50f3 add east		2 years ago
ascend310_infer	add east	2 years ago

evaluate	add east	2 years ago

lanms	add east	2 years ago

scripts	add east	2 years ago

src	add east	2 years ago

README.md	add	2 years ago

detect.py	add east	2 years ago

eval.py	add east	2 years ago

export.py	add east	2 years ago

postprocess.py	add east	2 years ago

train.py	add east	2 years ago

README.md

EAST for Ascend

EAST Description
Model Architecture
Dataset
Features
- Mixed Precision
Environment Requirements
Script Description
Model Description
- Performance
  - Training Performance
  - Inference Performance
Description of Random Situation
ModelZoo Homepage

EAST Description

EAST is an efficient and accurate neural network architecture for scene text detection pipeline. The method is divided into two stages: the fully convolutional network stage and the network management system fusion stage. FCN directly generates the text area, excluding redundant and time-consuming intermediate steps. This idea was proposed in the paper "EAST: An Efficient and Accurate Scene Text Detector.", published in 2017.

Paper Xinyu Zhou, Cong Yao, He Wen, Yuzhi Wang, Shuchang Zhou, Weiran He, and Jiajun Liang Megvii Technology Inc., Beijing, China, Published in CVPR 2017.

Model architecture

The network structure can be decomposed into three parts: feature extraction, feature merging and output layer.Use VGG, Resnet50 and other networks in the feature extraction layer to obtain feature map,In the feature merging part, the author actually borrowed the idea of U-net to obtain different levels of information,finally, score map and geometry map are obtained in the output layer part.

Dataset

Dataset used ICDAR 2015

Dataset: ICDAR 2015: Focused Scene Text
- Train: 88.5MB, 1000 images
- Test:43.3MB, 500 images

Features

Environment Requirements

Hardware（Ascend）
- Prepare hardware environment with Ascend processor. If you want to try Ascend , please send the application form to ascend@huawei.com. Once approved, you can get the resources.
Framework
- MindSpore
For more information, please check the resources below：
- MindSpore Tutorials
- MindSpore Python API

Script description

Script and sample code

.
└─east
  ├─README.md
  ├─scripts
    ├─run_standalone_train.sh     # launch standalone training with ascend platform(1p)
    ├─run_distribute.sh           # launch distributed training with ascend platform(8p)
    └─run_eval.sh                 # launch evaluating with ascend platform
  ├─src
    ├─dataset.py                      # data proprocessing
    ├─lr_schedule.py                  # learning rate scheduler
    ├─east.py                         # network definition
    └─utils.py                        # some functions which is commonly used
    └─distributed_sampler.py          # distributed train
    └─initializer.py                  # init
    └─logger.py                       # logger output
  ├─eval.py                           # eval net
  └─train.py                          # train net

Training process

Usage

Ascend:

# distribute training example(8p)
sh run_distribute_train.sh [DATASET_PATH] [PRETRAINED_BACKBONE] [RANK_TABLE_FILE]
# standalone training
sh run_standalone_train_ascend.sh [DATASET_PATH] [PRETRAINED_BACKBONE] [DEVICE_ID]
# evaluation:
sh run_eval_ascend.sh [DATASET_PATH] [CKPT_PATH] [DEVICE_ID]

Notes:
RANK_TABLE_FILE can refer to Link , and the device_ip can be got as Link. For large models like InceptionV4, it's better to export an external environment variable export HCCL_CONNECT_TIMEOUT=600 to extend hccl connection checking time from the default 120 seconds to 600 seconds. Otherwise, the connection could be timeout since compiling time increases with the growth of model size.

This is processor cores binding operation regarding the device_num and total processor numbers. If you are not expect to do it, remove the operations taskset in scripts/run_distribute_train.sh

The pretrained_path should be a checkpoint of vgg16 trained on Imagenet2012. The name of weight in dict should be totally the same, also the batch_norm should be enabled in the trainig of vgg16, otherwise fails in further steps.

Launch

# training example
  shell:
    Ascend:
      # distribute training example(8p)
      sh run_distribute_train.sh [DATASET_PATH] [PRETRAINED_BACKBONE] [RANK_TABLE_FILE]
      # standalone training
      sh run_standalone_train_ascend.sh [DATASET_PATH] [PRETRAINED_BACKBONE] [DEVICE_ID]

Result

Training result will be stored in the example path. Checkpoints will be stored at ckpt_path by default, and training log will be redirected to ./log

(8p)
...
epoch: 397 step: 1, loss is 0.2616188
epoch: 397 step: 2, loss is 0.38392675
epoch: 397 step: 3, loss is 0.21342245
epoch: 397 step: 4, loss is 0.29853413
epoch: 397 step: 5, loss is 0.2697169
epoch time: 4432.678 ms, per step time: 886.536 ms
epoch: 398 step: 1, loss is 0.32656515
epoch: 398 step: 2, loss is 0.28596723
epoch: 398 step: 3, loss is 0.24983373
epoch: 398 step: 4, loss is 0.29556546
epoch: 398 step: 5, loss is 0.28608245
epoch time: 5230.462 ms, per step time: 1046.092 ms
epoch: 399 step: 1, loss is 0.24444203
epoch: 399 step: 2, loss is 0.24407807
epoch: 399 step: 3, loss is 0.29774582
epoch: 399 step: 4, loss is 0.2569809
epoch: 399 step: 5, loss is 0.25168353
epoch time: 2595.220 ms, per step time: 519.044 ms
epoch: 400 step: 1, loss is 0.21435773
epoch: 400 step: 2, loss is 0.2563093
epoch: 400 step: 3, loss is 0.23374572
epoch: 400 step: 4, loss is 0.457117
epoch: 400 step: 5, loss is 0.28918257
epoch time: 4661.479 ms, per step time: 932.296 ms
epoch: 401 step: 1, loss is 0.26602226
epoch: 401 step: 2, loss is 0.267757
epoch: 401 step: 3, loss is 0.27752787
epoch: 401 step: 4, loss is 0.28883433
epoch: 401 step: 5, loss is 0.20567583
epoch time: 4297.705 ms, per step time: 859.541 ms
...
(1p)
...
epoch time: 20190.564 ms, per step time: 492.453 ms
epoch: 23 step: 1, loss is 1.4938335
epoch: 23 step: 2, loss is 1.7320133
epoch: 23 step: 3, loss is 1.3432003
epoch: 23 step: 4, loss is 1.375334
epoch: 23 step: 5, loss is 1.2183237
epoch: 23 step: 6, loss is 1.152751
epoch: 23 step: 7, loss is 1.1234403
epoch: 23 step: 8, loss is 1.1597326
epoch: 23 step: 9, loss is 1.390804
epoch: 23 step: 10, loss is 1.2011471
epoch: 23 step: 11, loss is 1.7939932
epoch: 23 step: 12, loss is 1.7997816
epoch: 23 step: 13, loss is 1.4836912
epoch: 23 step: 14, loss is 1.3689598
epoch: 23 step: 15, loss is 1.3506227
epoch: 23 step: 16, loss is 2.132399
epoch: 23 step: 17, loss is 1.4153867
epoch: 23 step: 18, loss is 1.351174
epoch: 23 step: 19, loss is 1.9559281
epoch: 23 step: 20, loss is 1.317142
epoch: 23 step: 21, loss is 1.4965435
epoch: 23 step: 22, loss is 1.2664857
epoch: 23 step: 23, loss is 1.7235017
epoch: 23 step: 24, loss is 1.4537313
epoch: 23 step: 25, loss is 1.7973338
epoch: 23 step: 26, loss is 1.583169
epoch: 23 step: 27, loss is 1.5295832
epoch: 23 step: 28, loss is 2.0665898
epoch: 23 step: 29, loss is 1.3507215
epoch: 23 step: 30, loss is 1.2847648
epoch: 23 step: 31, loss is 1.5181551
epoch: 23 step: 32, loss is 1.4159863
epoch: 23 step: 33, loss is 1.4176369
epoch: 23 step: 34, loss is 1.4142565
epoch: 23 step: 35, loss is 1.3644646
epoch: 23 step: 36, loss is 1.1788905
epoch: 23 step: 37, loss is 1.4377214
epoch: 23 step: 38, loss is 1.108615
epoch: 23 step: 39, loss is 1.2742603
epoch: 23 step: 40, loss is 1.3961313
epoch: 23 step: 41, loss is 1.3044286
...

Eval process

Usage

You can start training using python or shell scripts. The usage of shell scripts as follows:

Ascend:

  sh run_eval_ascend.sh [DATASET_PATH] [CKPT_PATH] [DEVICE_ID]

Launch

A fast Locality-Aware NMS in C++ provided by the paper's author.(g++/gcc version 6.0 + will be ok), you can click here get it.

You can download evaluation tool before evaluate . rename the tool as evaluate and make directory like following:

 ├─lnms                               # lnms tool
    ├─evaluate
    └─gt.zip                          # test ground Truth  
    └─rrc_evaluation_funcs_1_1.py     # evaluate Tool from icdar2015
    └─script.py                       # evaluate Tool from icdar2015  
  ├─eval.py                           # eval net

The evaluation scripts are from ICDAR Offline evaluation and have been modified to run successfully with Python 3.7.1.
Change the evaluate/gt.zip if you test on other datasets.
Modify the parameters in eval.py and run:

# eval example
  shell:
      Ascend:
            sh run_eval_ascend.sh [DATASET_PATH] [CKPT_PATH] [DEVICE_ID]

checkpoint can be produced in training process.

Result

Evaluation result will be stored in the example path, you can find result like the followings in log.

Calculated {"precision": 0.8329088130412634, "recall": 0.7871930669234473, "hmean": 0.8094059405940593, "AP": 0}

Model description

Performance

Training Performance

Parameters	Ascend
Model Version	EAST
Resource	Ascend 910, cpu:2.60GHz 192cores, memory:755G
uploaded Date	04/27/2021
MindSpore Version	1.1.1
Dataset	1000 images
Batch_size	24
Training Parameters	epoch=600, batch_size=24, lr=0.001
Optimizer	Adam
Loss Function	Dice for classification, Iou for bbox regression
Loss	~0.27
Total time (8p)	1h20m
Scripts	east script

Inference Performance

Parameters	Ascend
Model Version	EAST
Resource	Ascend 910, cpu:2.60GHz 192cores, memory:755G
Uploaded Date	12/27/2021
MindSpore Version	1.1.1
Dataset	500 images
Batch_size	1
Accuracy	"precision": 0.8329088130412634, "recall": 0.7871930669234473, "hmean": 0.8094059405940593
Total time	2 min
Model for inference	276M (.ckpt file)

Training performance results

Ascend	train performance
1p	51.25 img/s

Ascend	train performance
8p	300 img/s

Description of Random Situation

We set seed to 1 in train.py.

ModelZoo Homepage

Please check the official homepage.

No Description

mindspore

C++ Python Shell other

How to access data resources in code

README.md

EAST for Ascend

Usage

Launch

Result

Usage

Launch

Result

Training Performance

Inference Performance

Training performance results

Contributors (1) All

Contributors (1)
All