Code Implementations and packages
Code Implementations and packages
Embedded Files
Probabilistic Diffusion
Probabilistic Diffusion
With the recent popularity of Stable-Diffusion image generation, I decided to implement the general purpose diffusion based training algorithm to work with any desired type of PyTorch model and any d-dimensional tensor dataset, not just images.
With the recent popularity of Stable-Diffusion image generation, I decided to implement the general purpose diffusion based training algorithm to work with any desired type of PyTorch model and any d-dimensional tensor dataset, not just images.
Training data to be learned by Diffusion training
Learned reverse diffusion process generating 1,000 new samples over 200 timesteps
MIxreg reinforcement learning regularization
MIxreg reinforcement learning regularization
Many modern Reinforcement Learning methods suffer from overfitting and lack of generalization. Mixture Regularization (MixReg) is a method proposed by Kaixin Wang, Bingyi Kang, Jie Shao and Jiashi Feng which addresses this. The method achieves improved generalizability on out of sample data. The idea is simple in nature and suggests mixing data together, augmenting each observation and producing more diverse data.
Many modern Reinforcement Learning methods suffer from overfitting and lack of generalization. Mixture Regularization (MixReg) is a method proposed by Kaixin Wang, Bingyi Kang, Jie Shao and Jiashi Feng which addresses this. The method achieves improved generalizability on out of sample data. The idea is simple in nature and suggests mixing data together, augmenting each observation and producing more diverse data.
I've implemented this regularization technique using the Python Stable-Baselines library.
I've implemented this regularization technique using the Python Stable-Baselines library.
Equation Learner neural network architecture
Equation Learner neural network architecture
Many functions follow cyclical patterns, specifically those found in a signal processing domain. The Equation Learner Neural Network based model introduced by Georg Martius & Christoph H. Lampert is built with these cyclical functions in mind, incorporating sinusoidal activation functions which allow for an accurate representation of such functions.
Many functions follow cyclical patterns, specifically those found in a signal processing domain. The Equation Learner Neural Network based model introduced by Georg Martius & Christoph H. Lampert is built with these cyclical functions in mind, incorporating sinusoidal activation functions which allow for an accurate representation of such functions.
I've created a Python TensorFlow implementation of this neural network architecture with a software engineering focus, allowing users to easily customize specific aspects of the architecture as well as specify training regiments with simple method calls.
I've created a Python TensorFlow implementation of this neural network architecture with a software engineering focus, allowing users to easily customize specific aspects of the architecture as well as specify training regiments with simple method calls.
Architecture of 2 layer EQL network
Example of EQL learning a sum of sines type function
Report abuse