rocks-classifier 0.1.2

Rock classifier deployed on railway and monitored using Weights and Biases!

Whats-this-rock

This project deploys a telegram bot that classifies rock images into 1 of 7 types.

This package uses tensorflow to accelerate deep learning experimentation.

MLOps workflow like

• Experiment Tracking
• Model Management
• Hyperparameter Tuning

was all done using Weights & Biases

Additionally, nbdev was used to

• develop the package
• produce documentation based on a series of notebooks.
• CI
• publishing to PyPi

Inspiration

The common complaint that you need massive amounts of data to do deep learning  can be a very long way from the truth!

You very often don’t need much data at all, a lot of people are looking for ways to share data and aggregate data, but that’s unnecessary.They assume they need more data than they do, cause they’re not familiar with the basics of transfer learning which is this critical technique for needing orders of magnitudes less data.

Jeremy Howards

Documentation

Documentation for the project has been created using nbdev, and is available at udaylunawat.github.io/Whats-this-rock.

nbdev is a notebook-driven development platform. Simply write notebooks with lightweight markup and get high-quality documentation, tests, continuous integration, and packaging for free!

Once I discovered nbdev, I couldn’t help myself but redo the whole project from scratch.

It’s just makes me 10x more productive and makes the whole process streamlined and more enjoyable.

Installation

You can directly install using pip

pip install rocks_classifier

Install - Directly from Github (latest beta version)

pip install git+https://github.com/udaylunawat/Whats-this-rock.git

Download and process data

%%bash
rocks_process_data  --config-dir configs \
                    remove_bad= True \
                    remove_misclassified= True \
                    remove_duplicates= True \
                    remove_corrupted= True \
                    remove_unsupported= True \
                    sampling=None \
                    train_split=0.8 \

Train Model

Train model using default parameters in configs/config.yaml.

rocks_train_model   --config-dir configs

You can try different models and parameters by editing configs/config.yaml, or you can override it by passing arguments like this:-

By using Hydra it’s now much more easier to override parameters like this:-

rocks_train_model   --config-dir configs \
                    wandb.project=Whats-this-rock \
                    wandb.mode=offline \
                    wandb.use=False \
                    dataset_id=[1,2] \
                    epochs=30 \
                    lr=0.005 \
                    augmentation=None \
                    monitor=val_loss \
                    loss=categorical_crossentropy \
                    backbone=resnet \
                    lr_schedule=cosine_decay_restarts \
                    lr_decay_steps=300 \
                    trainable=False \

Wandb Sweeps (Hyperparameter Tuning)

Edit configs/sweep.yaml

wandb sweep \
--project Whats-this-rock \
--entity udaylunawat \
configs/sweep.yaml

This will return a command with $sweepid, run it to start running sweeps!

wandb agent udaylunawat/Whats-this-rock/$sweepid

Telegram Bot

You can try the bot here on Telegram.

Type /help to get instructions in chat.

Or deploy it yourself

rocks_deploy_bot

Demo

Colab	GitHub	Download
Run in Colab	View Source on GitHub	Download Notebook

Features

& Things I’ve Experimented with

Feature		Feature
Wandb	- Experiment Tracking - System Tracking - Model Tracking - Hyperparameter Tuning	Datasets	- Dataset 1 - Dataset 2
Augmentation	- Keras-CV - MixUp - CutMix - Normal	Models	- ConvNextTiny - Efficientnet - Resnet101 - MobileNetv1 - MobileNetv2 - Xception
Optimisers	- Adam - Adamax - SGD - RMSProp	LR Scheduler	- CosineDecay - ExponentialDecay - CosineDecayRestarts
Remove Images	- Duplicate Images - Corrupt Images - Bad Images - Misclassified	Configuration Management	- hydra - ml-collections - argparse -google-fire
Generators	- tf.data.DataSet - ImageDataGenerator	Deployment	- Heroku - Railway
Evaluation	- Classification Report - Confusion Matrix	GitHub Actions (CICD)	- GitHub Super Linter - Deploy to Telegram - Deploy to Railway - nbdev tests CI - GitHub Pages(Documentation)
Linting	- Flake8 - Pydocstyle	Telegram Bot	- Greet - Info - Predict Image
Formatting	- Black - yapf	Documentation	- Code Description - Code comments - Source link - Doclinks
Badges	- Build - Issues - Lint Codebase	Docker
Publishing	- PyPi

Planned Features

Feature		Feature
Deploy	- HuggingFaces	Backend	- FastAPI
Coding Style	- Object Oriented	Frontend	- Streamlit
WandB	- Group Runs - WandB Tables	Badges	- Railway

Technologies Used

	.

Directory Tree

├── imgs                              <- Images for skill banner, project banner and other images
│
├── configs                           <- Configuration files
│   ├── configs.yaml                  <- config for single run
│   └── sweeps.yaml                   <- confguration file for sweeps hyperparameter tuning
│
├── data
│   ├── corrupted_images              <- corrupted images will be moved to this directory
│   ├── misclassified_images          <- misclassified images will be moved to this directory
│   ├── bad_images                    <- Bad images will be moved to this directory
│   ├── duplicate_images              <- Duplicate images will be moved to this directory
│   ├── sample_images                 <- Sample images for inference
│   ├── 0_raw                         <- The original, immutable data dump.
│   ├── 1_external                    <- Data from third party sources.
│   ├── 2_interim                     <- Intermediate data that has been transformed.
│   └── 3_processed                   <- The final, canonical data sets for modeling.
│
├── notebooks                         <- Jupyter notebooks. Naming convention is a number (for ordering),
│                                        the creator's initials, and a short `-` delimited description, e.g.
│                                        1.0-jqp-initial-data-exploration`.
│
│
├── rocks_classifier                  <- Source code for use in this project.
│   │
│   ├── data                          <- Scripts to download or generate data
│   │   ├── download.py
│   │   ├── preprocess.py
│   │   └── utils.py
│   │
│   ├── callbacks                     <- functions that are executed during training at given stages of the training procedure
│   │   └── callbacks.py
│   │
│   ├── models                        <- Scripts to train models and then use trained models to make
│   │   │                                predictions
│   │   ├── evaluate.py
│   │   ├── models.py
│   │   ├── predict.py
│   │   ├── train.py
│   │   └── utils.py
│   │
│   │
│   └── visualization                 <- Scripts for visualizations
│
├── .dockerignore                     <- Docker ignore
├── .gitignore                        <- GitHub's excellent Python .gitignore customized for this project
├── LICENSE                           <- Your project's license.
├── README.md                         <- The top-level README for developers using this project.
├── CHANGELOG.md                      <- Release changes.
├── CODE_OF_CONDUCT.md                <- Code of conduct.
├── CONTRIBUTING.md                   <- Contributing Guidelines.
├── settings.ini                      <- configuration.
├── README.md                         <- The top-level README for developers using this project.
├── requirements.txt                  <- The requirements file for reproducing the analysis environment, e.g.
│                                        generated with `pip freeze > requirements.txt`
└── setup.py                          <- makes project pip installable (pip install -e .) so src can be imported

Bug / Feature Request

If you find a bug (the site couldn’t handle the query and / or gave undesired results), kindly open an issue here by including your search query and the expected result.

If you’d like to request a new function, feel free to do so by opening an issue here. Please include sample queries and their corresponding results.

Contributing

• Contributions make the open source community such an amazing place to learn, inspire, and create.
• Any contributions you make are greatly appreciated.
• Check out our contribution guidelines for more information.

License

LinkFree is licensed under the MIT License - see the LICENSE file for details.

Credits

• Dataset 1 - by Mahmoud Alforawi
• Dataset 2 - by salmaneunus
• nbdev inspiration - tmabraham

Support

This project needs a ⭐️ from you. Don’t forget to leave a star ⭐️

Walt might be the one who knocks
but Hank is the one who rocks.