kunakl07/al-mlresearch

AL-MLresearch

These Repository contains the necessary code, research work and steps for preprocessing and developing Orcacall detection model with some research focused on Active Learning. The part two of this repo which contains the Active learning tool and its deployment is located in this repo. Most of the data present in the world are unlabeled. Even though labeling data is an expensive, difficult, and slow process, it is an essential part of the Machine Learning system. But what if a model could achieve a similar accuracy just by annotating a small amount of dataset. With the help of Active Learning, you can spend 10-20% of the time annotating data and still get the same performance. Therefore, we going to build an active learning tool that would label the vast amounts of unlabeled data coming in real-time streams from ocean observing systems.

Dataset

This github repository consists of the necessary steps taken for the backend i.e. preprocessing, building CNN models, and active learning. Along with this, this repository also contains different accuracy achieved by different models with the steps necessary steps taken to achieve that accuracy. I have used the podcast round 2 and podcast round 3 for training the model. The dataset was neatly labeled by the Orcasound organization that helped me a lot in preprocessing it.

Implementation

The implementation part consists of the stages:

• Data extraction and preprocessing
• Model building and training
• Active learning

General project flow

The directory structure of our project looks like this:

├── experiments
│   ├── experiments/Grayscale_magspectrogram_scipy.ipynb <- CNN model trained on grayscale power_spectral_density spectrograms
│   ├── experiments/Grayscale_specs_mag.ipynb  <- Differenet CNN model trained on grayscale power_spectral_density spectrograms
│   ├── experiments/Preprocess_grayscale.ipynb <- Preprocessing of the grayscale spectrograms
│   ├── experiments/README.md
│   ├── experiments/Sir_Val's_method_color.ipynb <- Preprocessing of the power_spectral_density spectrograms as suggested by Val
│   └── experiments/Spectrograms.ipynb <- Different spectrograms that are generated and their code
├── notebooks
│   ├── notebooks/active_learning
│   │   └── notebooks/active_learning/active_learning_pipeline.ipynb <- Active learning pipelne notebook
│   ├── notebooks/case_one
│   │   ├── notebooks/case_one/Preprocessing 
│   │   │   └── notebooks/case_one/Preprocessing/preprocess_case_one.ipynb <- Jupyter notebook in colab performing preprocessing
│   │   ├── notebooks/case_one/README.md
│   │   └── notebooks/case_one/Training
│   │       ├── notebooks/case_one/Training/README.md
│   │       └── notebooks/case_one/Training/training_on_case_one_on_model_VGG_and_CNN.ipynb <- Jupyter notebook in colab performing training
│   ├── notebooks/case_three
│   │   ├── notebooks/case_three/Preprocess
│   │   │   ├── notebooks/case_three/Preprocess/preprocessing_PCEN_and_Wavlet_denoising.ipynb <- Jupyter notebook in colab performing preprocessing process.
│   │   │   └── notebooks/case_three/Preprocess/README.md
│   │   ├── notebooks/case_three/README.md
│   │   └── notebooks/case_three/Training
│   │       ├── notebooks/case_three/Training/CNN_for_SRKW.ipynb <- Jupyter notebook in colab performing training stage on case three preprocessed spectrograms.
│   │       ├── notebooks/case_three/Training/README.md
│   │       ├── notebooks/case_three/Training/Resnet152_pre_pcen_wd_train.ipynb <- Jupyter notebook in colab performing training stage using Resnet152
│   │       └── notebooks/case_three/Training/Training.ipynb <- Jupyter notebook in colab performing training stage using VGGG-16
│   ├── notebooks/case_two
│   │   ├── notebooks/case_two/Preprocess
│   │   │   └── notebooks/case_two/Preprocess/preprocessing_using_PCEN.ipynb <- Jupyter notebook in colab performing preprocessing process using case two
│   │   └── notebooks/case_two/Training
│   │       └── notebooks/case_two/Training/training_on_PCEN_spectrograms.ipynb <- Jupyter notebook in colab performing training on stage two spectrograms.
│   ├── notebooks/preprocessing_and_training_using_Ketos
│   │   ├── notebooks/preprocessing_and_training_using_Ketos/Preprocessing
│   │   │   ├── notebooks/preprocessing_and_training_using_Ketos/Preprocessing/preprocessing_using_Ketos.ipynb 
│   │   │   └── notebooks/preprocessing_and_training_using_Ketos/Preprocessing/README.md
│   │   └── notebooks/preprocessing_and_training_using_Ketos/Training
│   │       ├── notebooks/preprocessing_and_training_using_Ketos/Training/training_on_RNN_using_Ketos.ipynb
│   │       └── notebooks/preprocessing_and_training_using_Ketos/Training/Training_SRKWs_Ketos.ipynb
│   │    
│   └── notebooks/README.md
├── requirements.txt
├── src
│   ├── src/preprocessing_script
│   │   ├── src/preprocessing_script/Dockerfile   <- The Dockerfile of the preprocessing script. 
│   │   ├── src/preprocessing_script/preprocess.py <- The preprocessing script. 
│   │   ├── src/preprocessing_script/README.md  
│   │   ├── src/preprocessing_script/requirements.txt <- Requirement file
│   │   └── src/preprocessing_script/selection_table.py <- Selection table for generating background noises
│   └── src/training_script
│       ├── src/training_script/Dockerfile <- The Dockerfile of the training script. 
│       ├── src/training_script/model_build_and_training.py <- The model building and training script.
│       ├── src/training_script/README.md
│       └── src/training_script/requirements.txt <- Requirement file
└── trained_models
   └── trained_models/README.md <- Links to the different models that could be used

Contributing to AL-MLresearch

To contribute to AL-MLresearch, follow these steps:

1. Fork this repository.
2. Create a branch: git checkout -b <branch_name>.
3. Make your changes and commit them: git commit -m '<commit_message>'
4. Push to the original branch: git push origin <AL-MLresearch>/<location>
5. Create the pull request.

Alternatively see the GitHub documentation on creating a pull request.

Contributors

Thanks to the following people who have contributed to this project without whom the project would never have been completed.

My Mentors:

• @jesselopez
• @valentina-s
• @scottveirs
• @abhisheksingh
• @valveirs

My partner who is working on the front-end part

• @jorgediego

Thanks to OrcaCNN and Ketos for their thorough documentation and code which helped not only me develop preprocessing and training stage but also helped me develop scripts.

License

This project uses the following license: MIT.