• Phase-1-Documentation
• Phase-2-EDA-Documentation
• Phase-3-Modelling
• Phase-4-Advanced-Modelling
• Phase-5-Deployment

• The app is deployed on share.streamlit.io.
• For making predictions using the Vanilla_Net, the tensorflow model is served at heroku container services using tensorflow serving.
• The app makes calls to the served model to get predictions.

Dry bean is the most popular pulse produced in the world. The main problem dry bean producers and marketers face is in ascertaining good seed quality. Lower quality of seeds leads to lower quality of produce. Seed quality is the key to bean cultivation in terms of yield and disease. Manual classification and sorting of bean seeds is a difficult process. Our objective is to use Machine learning techniques to do the automatic classification of seeds.

Ascertaining seed quality is important for producers and marketers. Doing this manually would require a lot of effort and is a difficult process. This is why we try to use machine learning techniques to do the automatic classification of seeds.

• Saves hours of manual sorting and classification of seeds.
• We can do it in real-time.

View the noteboks phase-wise following the links:

• Data Acquisition
• EDA - Exploratory Data Analysis
• Feature Selection & Modelling
• Advanced Modelling

• The best model we have is a Tuned - Light Gradient Boosting Classifer without any preprocessing on the data. It has an accuracy of 93 % and a F1-score of 0.929
• The best parameters are found with RandomizedSearchCV:

  {
    'boosting_type': 'gbdt',
    'class_weight': None,
    'colsample_bytree': 1.0,
    'importance_type': 'split',
    'learning_rate': 0.05,
    'max_depth': -1,
    'min_child_samples': 31,
    'min_child_weight': 0.001,
    'min_split_gain': 0.2,
    'n_estimators': 200,
    'n_jobs': -1,
    'num_leaves': 90,
    'objective': None,
    'random_state': 8144,
    'reg_alpha': 0.7,
    'reg_lambda': 0.4,
    'silent': 'warn',
    'subsample': 1.0,
    'subsample_for_bin': 200000,
    'subsample_freq': 0,
    'feature_fraction': 0.4,
    'bagging_freq': 0,
    'bagging_fraction': 1.0
  }
  

Confusion Matrix for Tuned-LightGBM:

• I used a 2 layer NN with relu activation.

  • The first hidden layer has 512 nodes
  • The second one has 256 nodes
  • Both the layers use relu activation
  • Optimizer: Adam with a lr=3e-4
  • loss: SparseCategoricalCrossEntropy(logits=True)
  • Epochs: 20

• Without resampling: accuracy -> 93.09%, f1-score -> 0.9304

With resampling (Oversampling with SMOTE) We even beat the best accuracy and f1-score in the paper with an accuracy of 93.39% & f1-score of 0.9340

Confusion Matrix for Vanilla-Net

You can directly serve the model and make api calls to get the predictions. Pre-requisite is just have docker installed

• Build the container docker build -t app .

• Run the container: docker run -p 8501:8501 -e PORT=8501 app

• First, serve the model using the above instructions

• Next, in another terminal first install the requirements using: pip install -r requirements.txt

• The model is being served at http://localhost:8501/saved_model. To make predictions all you have to do is make calls to the endpoint: http://localhost:8501/saved_model:predict. A basic example of it is the test_files/test_server.py

• you can just run it: cd test_files && python test_server.py or use it as a template for your own application.

• Install all the requirements: pip install -r requirements.txt

• Now, just do: streamlit run app.py and voila it's up on local host!

• Multiclass classification of dry beans using computer vision and machine learning techniques
• Dry Beans Classification Using Machine Learning