This library is designed as a supporting library to my other research efforts in Deep Learning using PyTorch. It has many features and implementations but some of the most relevant ones:
- 🧪 Experiment Management - Human friendly classes to wrap the complexity of running experiments. Experiments can be loaded, providing convient access to model, data and logged results.
- 🚀 Flexible Compilation - Experiment options are specified via hierarchical YAML configs that are implicitly verified. No more
argparse. - ☎️ Two-way Callbacks - To modify experiments in a flexible and modular way, Pylot uses callbacks to extend the behavior of experiments.
- 💾 Simplified I/O - Pylot removes the pains of loading/saving data through its many OOP interfaces in
util. - 📊 Results Analysis - Result aggregation primitives such as Pareto frontiers, bootstrapping, and recursive selection are included.
- 📈 Metrics and Losses - Implementation of common Computer Vision metrics and losses in pure PyTorch
- 🐘 Model/Layer/Optim ZOO - Pylot includes implementations of popular deep learning models, building blocks and optimization algorithms, such as: Sharpness Aware Minimization (SAM), Exponential Moving Average Weights (EMA), U-Net models, Voxelmorph networks, Hypernetworks, SqueezeExcitation layers and many many more.
- ✅ Extensive Type Validation - Type hints are included and enforced thanks to pydantic.
- 🐼 Pandas Extensions - Pandas DataFrame API is extended to better handle result dataframes, with methods such as
.select,.drop_constant,.unique_per_color theUnixAccessorfunctionality. - 🏎️ Designed for Speed - From careful I/O considerations to CUDA optimizations, pylot is painstakingly designed with efficiency and speed in mind.