Bong-HoonLee/ai_human_counseling

Non-face-to-face psychological counseling service using AI human

• Purpose of the project

  • Establishing a business plan using an Facial Expression Recognition classification model
  • Non-face-to-face psychological counseling treatment
  • Relieves social isolation by feeling like talking to real people

• A major customer base

  1. A person with social phobia
  2. A person with depression
  3. Elder who lives alone
  4. someone in need of a conversation

Dev env

IDE	GPU	Programing Language

Directory

• .streamlit : Folder used for streamlit
• archive: EDA, data pre-processing
• config_f: Auto-training config folder
• contents : source
• docs: documents, images, reports
• models: models
• tools : Other Architectures
• utils : Metric and other files
• requirements.txt: required libraries and packages
• trainer.py: main train&test logics

download pt, pth file

• google Drive : https://drive.google.com/drive/u/1/folders/1wu5isYVQtwvYDPkegjZm8EFXciWLCmhP

How to Run & Debug

1. pip install -r requirements.txt to install required packages
2. download pt, pth file and best.pt save "./archive/models/yolo/pt/bt" and other save "./tools/Wav2Lip/chechpoints"
3. streamlit run main.py
4. Drag image to sidebar uploading
5. And Try Psychological counseling

Dataset

1. Image data AI Hub::한국인 감정인식을 위한 복합 영상

• Face photo by each emotion (joy, panic, anger, anxiety, hurt, sadness, neutral)
• Total number of data: 500,000 source data
  • Train Data Count: 14000=2000*7
  • Test Data Count: 70000=1000*7

2. llm data

• Psychological counseling paper in severance hospital

EDA

1. Classifying pain as a psychological and expression of physical pain.
   • Psychological pain was similar to a sad expression, so it was judged that it was difficult to discern from sadness.
2. Learn 14000 by classifying it into train and validation in an 8:2 ratio, respectively.

Models

You can check the list at config.py

1. List of Neural Network models used to train models (total:18)

   • alexnet
   • convnext_tiny
   • densenet121
   • efficientnet_v2_s
   • googlenet
   • inception_v3
   • mnasnet0_5
   • mobilenet_v3_large
   • resnet18
   • resnet34
   • resnet50
   • resnet101
   • vgg11_bn
   • vgg13_bn
   • vgg16_bn
   • vit_b_16
   • swin_t
   • custom

2. Finally selected neural network models (total:10)

• Select by considering the appropriate model and performance for the chatbot
  • Yolo v8
  • AlexNet
  • DenseNet121
  • EfficientNet
  • VGG
  • ResNet
  • ViT
  • swin_t
  • MobileNet
  • Custom model

result

• we chose the YOLO with the fastest and the highest accuracy
  
  
• prams : 3.2M
• Accuracy: 69%

service

Psychological counseling chatbot