- For the main repository you can go to this repository
Trasholution-ML
├── ML-FastAPI-app
│ ├── app
│ │ ├── model
│ │ │ ├── [PLACE THE MODEL IN THIS FOLDER]
│ │ │ └── README.md
│ │ ├── __init__.py
│ │ └── main.py
│ ├── .dockerignore
│ ├── Dockerfile
│ └── requirements.txt
├── Training Notebooks
│ ├── training_model_waste_classification_v2_1.ipynb
│ └── training_model_waste_classification_v2_2.ipynb
├── .gitignore
└── README.md
- Datasets used to train the model can be accessed on Google Drive and Kaggle Dataset.
- In this dataset there are 8 class of trash/waste type; Electronic_Waste, Food_Scraps, Glass, Metalic_Materials, Organic_Vegetation_Waste, Paper, Plastic, and Textile.
- This number of image in this dataset are split into 3 part per image class:
- Training data : 80% of 520 images (416 images per image class).
- Validation data : 10% of 520 images (52 images per image class).
- Testing data : 10% of 520 images (52 images per image class).
- This dataset consist of image taken from scraping the internet.
- At first, we develop two kind of model with version 1 created from scratch and version 2 are from transfer learning.
- From our early test we found that the version 1 only able to reach the testing accuracy of 73%, meanwhile the first complete iteration of version were able to reach testing accuracy of approx. 89%.
- Therefore we decided to drop the version 1 model to keep improving the version 2 model.
- When we first tried to develop the version 2 model, we tried to compare the result of 3 base model that is the VGG16, Inception V3, and Xception to our dataset. And we found that the with Xception as base model have the highest score out of the three.
- Jupyter notebooks used in training the model are in the
/Training Notebooks/folder:training_model_waste_classification_v2_1.ipynbused to train themodel-waste_classification-v2_1.h5model- This model architecture are a little bit different from the previous model in terms of the number of hidden layer in the fully reconected layer (Dense layer).
training_model_waste_classification_v2_2.ipynbused to train themodel-waste_classification-v2_2.h5model- This model architecture differ from the previous version in terms of fine tuning the transfer learning model with the last 6 layers (the block 14 group of Xception model layers) are fine-tuned to the this dataset.
- The exported model can be accessed here (Google Drive):
model-waste_classification-v2.zip(TensorFlow SavedModel format) has an accuracy score of:- training : 99.82%
- validation : 89.90%
- testing : 89.18%
model-waste_classification-v2_1.h5(Keras HDF5 format) has an accuracy score of:- training : 99.07%
- validation : 89.90%
- testing : 90.38%
model-waste_classification-v2_2.h5(Keras HDF5 format) has an accuracy score of:- training : 99.07%
- validation : 92.07%
- testing : 92.06%
- The model that is currently in use for deployment is the
model-waste_classification-v2_2.h5.
/ML-FastAPI-app/folder contains the web app FastAPI implementation for the model deployment in GCP Cloud Run./ML-FastAPI-app/model/folder contain the model for creating Docker Image. Due to Git file size limit at 100MB, you can download the exported model in theExported Modelssection above.- The Docker Image for model deployment in Cloud Run can be found on Docker Hub.
- Docker Image that is currently in deployment in GCP Cloud Run is the
latesttag (rafaelmahesa/ml-waste-image:latest).
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent