AI model automatic generation software

Graduation Research

Table of Contents

• About
• Getting Started
  • Prerequisites
  • Usage
    • automatic_train kick off
    • Variables reference
    • sample_training
    • results
    • Demonstration
• License

---

About

It is the best algorithm to automate the process of building the optimal model for AI image recognition. The data is also returned to the web as an example of an application using the resources collected there, allowing for machine learning from the web.The system can provide "generated and compressed compact AI models," "weight information," and "classified images spit out by the generated models. It is also possible to retrieve desired information such as mAP (recognition rate) and Flops (computational cost) that are generated in the process of generation. Architecture This is an algorithm that automatically learns and determines which combination of backbone and neck & head is optimal for the collected annotated image data, and automatically outputs the optimal AI model. Compression of generated AI models The second is compression technology for AI models. Recent developments in hardware devices have enabled the storage of huge amounts of information and large-scale computation, which have brought about an evolution in AI technology. However, many existing AI models contain a lot of redundant computations that do not fully consider the use of resources in the pursuit of recognition accuracy. The AI models that have been generated at this stage of development, while not having significant accuracy problems (85-95% of conventional methods), are large and far from "compact. In this case, the key is to "compress the software to a compact size while maintaining accuracy," which is a difficult problem to solve, and indeed, we are facing challenges in this area. However, once this issue is overcome, the technology can be used in a wider range of applications because it can be realized on smaller devices at lower cost. Key features of automatic_train: img : save image file of 20 dish samples (private) templates : html file for web application utils : mishmash of resources used in learning yolomodels : net... yolo...many files describe the heart of yolo's network architecture, which leverages resources in utils for learning. static : It contains pre-annotated and post-annotated images. In addition, there are txt files of the training results and the deep learning transition results from the automatic training. model_data/dish_classes.txt : 20 dishs detection label is here Chopsticks-cover Chopsticks-one Chopsticks-two Coffee Coffee-cup Coffee-dish Cup Fish-dish Paper Rice-bowl Soup-bowl Spoon Square-bowl Tea-cup Tea-dish Towel Towel-dish Waster-paper Water-cup Wine-cup Additional info Image recognition includes techniques for visual inspection for defects, defects, dirt, etc., and object detection. In the past, general image processing was used for image recognition in appearance inspection and object detection. However, conventional image processing methods have difficulties in detecting defects in object contours and handling image recognition due to changes in light. This software was created to overcome these challenges. When an optimal model can be output as a result of combinatorial learning on an input image, it shows performance far beyond human judgment.

Getting Started

Prerequisites

The easiest way to install automatic_train is by running:

git clone https://github.com/ultra-supara/automatic_train.git

Usage

automatic_train kick off

After installing automatic_train , you need to do is to run the following command:

pip install -r requirements.txt

Variables reference

Please note that entered values are case-sensitive. Default values are provided as an example to help you figure out what should be entered.

On manual setup, you need to replace only values written in uppercase.

Name	Tips	Description
PROJECT_NAME	automatic_train	My project name
GITHUB_USERNAME	ultra-supara	My GitHub username
FULL_NAME	sada atsushi	My realname
OPEN_SOURCE_LICENSE	MIT license	Full OSS license
app.py	python3 app.py	url Home
first_option_auto_train	simply option	url autotrain
second_option_auto_train	more detailed opt	url manualtrain
upload_place	ihpc	upload.php ihpc server

launch Flask app

python3 app.py

then , you may see this url Home

The deep learning part can all be operated from the browser side! We are currently preparing a Docker environment in which it can run.

sample_training

This is an example of training with autotrain , One-Key Training (a mechanism that allows machine learning from a browser without the need to adjust parameters).

The training conditions are in train_onekey.py

# 学習パラメタの設定
if __name__ == "__main__":
    imagesize = 32  #N*32(N=(1,20))
    epochs = 1
    batch_size = 2
    evalu_param = 'map'
    eager           = False
    train_gpu       = [0,]
    classes_path    = 'model_data/voc_classes.txt'
    anchors_path    = 'model_data/yolo_anchors.txt'
    anchors_mask    = [[6, 7, 8], [3, 4, 5], [0, 1, 2]]
    model_path      = ''

    input_shape     = [imagesize, imagesize]
    mosaic              = True
    mosaic_prob         = 0.5
    mixup               = True
    mixup_prob          = 0.5
    special_aug_ratio   = 0.7

    #label_smoothing     = 0

    Init_Epoch          = 0
    UnFreeze_Epoch      = epochs
    Unfreeze_batch_size = batch_size

    Freeze_Train        = False

    Init_lr             = 1e-3
    Min_lr              = Init_lr * 0.01

    optimizer_type      = "adam"
    momentum            = 0.937
    weight_decay        = 0

    lr_decay_type       = 'cos'

    focal_loss          = False
    focal_alpha         = 0.25
    focal_gamma         = 2

    save_period         = 10

    eval_flag           = True
    eval_period         = 20

    num_workers         = 1

    train_annotation_path   = '2007_train.txt'
    val_annotation_path     = '2007_val.txt'

    os.environ["CUDA_VISIBLE_DEVICES"]  = ','.join(str(x) for x in train_gpu)
    ngpus_per_node                      = len(train_gpu)

    gpus = tf.config.experimental.list_physical_devices(device_type='GPU')

results

This is an example of the results (static/train.txt) when the combination of yolov7_v5CSPdarknet is turned around training with an epochs count of 50.

logs/yolov7_v5CSPdarknet/best_epoch_weights.h5 model, anchors, and classes loaded.
Configurations: ** each part of the network **
----------------------------------------------------------------------
|                     keys |                                   values|
----------------------------------------------------------------------
|                     neck |                                         |
|                 backbone |                                         |
|             classes_path |               model_data/voc_classes.txt|
|             anchors_path |              model_data/yolo_anchors.txt|
|             anchors_mask |        [[6, 7, 8], [3, 4, 5], [0, 1, 2]]|
|              input_shape |                                       []|
|               confidence |                                      0.5|
|                  nms_iou |                                      0.3|
|                max_boxes |                                      100|
|          letterbox_image |                                     True|
----------------------------------------------------------------------
Load model done.
Get predict result.
Get predict result done.
Get ground truth result.
Get ground truth result done.
Get map.
...
yolov7 + v5CSPdarknet done!
There are 6 backbones.
Current training model: yolov7 + xCSPdarknet!
Evaluation parameter: mAP
Total GFLOPs: 0.284G

Configurations: ** each parameters **
----------------------------------------------------------------------
|                     keys |                                   values|
----------------------------------------------------------------------
|             classes_path |               model_data/voc_classes.txt|
|             anchors_path |              model_data/yolo_anchors.txt|
|             anchors_mask |        [[6, 7, 8], [3, 4, 5], [0, 1, 2]]|
|               model_path |                                         |
|              input_shape |                                 [32, 32]|
|               Init_Epoch |                                        0|
|           UnFreeze_Epoch |                                       50|
|      Unfreeze_batch_size |                                        2|
|             Freeze_Train |                                    False|
|                  Init_lr |                                    0.001|
|                   Min_lr |                                    1e-05|
|           optimizer_type |                                     adam|
|                 momentum |                                    0.937|
|            lr_decay_type |                                      cos|
|              save_period |                                       10|
|                 save_dir |                  logs/yolov7_xCSPdarknet|
|              num_workers |                                        1|
|                num_train |                                      409|
|                  num_val |                                       46|
----------------------------------------------------------------------
Train on 409 samples, val on 46 samples, with batch size 2.
----------------------------------------------------------------------

yolov7 + v5CSPdarknet done!
There are 6 backbones.
Current training model: yolov7 + xCSPdarknet!

This part shows the yolov7 + v5CSPdarknet combination has been trained, and the next part shows the yolov7 + xCSPdarknet being trained in succession. These can be done fully automatically. We have succeeded in automating the recombination process that was previously done manually by humans.

detection best model : static/evalu_parames.txt

yolov7_v5CSPdarknet,mAP:100.0%,flops:0.284323918G,model_size:204.553464MB,FPS:20.251316360863697,Parameters:50.902497M,
yolov7_xCSPdarknet,mAP:95.2%,flops:0.284323918G,model_size:204.579944MB,FPS:17.79088361770951,Parameters:50.902497M,
yolov7_v3darknet,mAP:92.4%,flops:0.385888334G,model_size:258.473392MB,FPS:17.01340010345215,Parameters:64.411713M,
yolov7_v7backbone,mAP:93.2%,flops:0.262721614G,model_size:150.126088MB,FPS:25.20116813428497,Parameters:37.346753M,
yolov7_v4CSPdarknet53,mAP:93.2%,flops:0.29979451G,model_size:202.75124MB,FPS:18.02773895714467,Parameters:50.443585M,
yolov5_v5CSPdarknet,mAP:87.2%,flops:0.232378318G,model_size:154.303304MB,FPS:24.313955865774663,Parameters:38.357737M,
yolov5_xCSPdarknet,mAP:85.4%,flops:0.232378318G,model_size:154.306808MB,FPS:29.451885644243497,Parameters:38.357737M,
yolov5_v3darknet,mAP:86.8%,flops:0.333942734G,model_size:208.223072MB,FPS:20.756551659258147,Parameters:51.866953M,
yolov5_v7backbone,mAP:86.3%,flops:0.211431374G,model_size:100.07604MB,FPS:45.17792878841556,Parameters:24.853193M,
yolov5_v4CSPdarknet53,mAP:88.0%,flops:0.24784891G,model_size:152.497128MB,FPS:28.897781129573318,Parameters:37.898825M,
yolov3_v5CSPdarknet,mAP:80.0%,flops:0.28694587G,model_size:193.377224MB,FPS:21.945551351170344,Parameters:48.169441M,
yolov3_xCSPdarknet,mAP:74.6%,flops:0.28694587G,model_size:193.385488MB,FPS:21.715122896668202,Parameters:48.169441M,
yolov3_v3darknet,mAP:77.4%,flops:0.388510286G,model_size:247.291248MB,FPS:17.55582086893424,Parameters:61.678657M,
yolov3_v7backbone,mAP:72.3%,flops:0.265343566G,model_size:138.944872MB,FPS:29.82171052299465,Parameters:34.613697M,
yolov3_v4CSPdarknet53,mAP:75.9%,flops:0.302416462G,model_size:191.576568MB,FPS:21.516757871989906,Parameters:47.710529M,
![image](https://github.com/ultra-supara/automatic_train/assets/67861004/c51593b0-47ae-4daa-b73d-173fe02e6a15)

Demonstration

In this video , first_option_auto_train automation_demonstraion

License

This project is licensed under the MIT license. Feel free to edit and distribute this template as you like.

See LICENSE for more information.