Physics, and more.
Holod – Hologram autofocus trainer
Holod Digital Holographic Reconstruction & Autofocus Toolkit Overview Holod is a Python package for automated focus estimation and amplitude/phase reconstruction of digital holograms. It provides: Dataset preparation: easily build and validate metadata CSV for hologram datasets. Autofocus training: classification or regression models (EfficientNet, ResNet50, ViT, FocusNet) to predict depth bins or continuous depth. Evaluation & plotting: residual vs true depth, confusion matrices, density plots, violin plots, and hexbin visualizations. Reconstruction: Fresnel‐based reconstruction of amplitude and phase. Command‐line interface (CLI): easy-to-use commands to train, plot results, reconstruct holograms, and generate metadata. Demonstration of running holo train × Demonstration of running holo train Features Flexible analysis: choose between classification (--bins > 1) or regression (--bins 1). Backbones supported: efficientnet, resnet50, vit, focusnet. Configurable training: batch size, epochs, learning rate, validation split, device (CPU/GPU). Visualizations: Plotly and Rich functions for analysis. Code Downloads ...
Archive: Holod V1
Holod Digital Holographic Reconstruction & Autofocus Toolkit Overview Holod is a Python package for automated focus estimation and amplitude/phase reconstruction of digital holograms. It provides: Dataset preparation: easily build and validate metadata CSV for hologram datasets. Autofocus training: classification or regression models (EfficientNet, ResNet, ViT) to predict depth bins or continuous depth. Evaluation & plotting: residual vs true depth, confusion matrices, density plots, violin plots, and hexbin visualizations. Reconstruction: Fresnel‐based reconstruction of amplitude and phase. Command‐line interface (CLI): easy-to-use commands to train, plot results, reconstruct holograms, and generate metadata. Demonstration of running holo train × Demonstration of running holo train Features Flexible analysis: choose between classification (--bins > 1) or regression (--bins 1). Backbones supported: efficientnet, resnet50, vit. Configurable training: batch size, epochs, learning rate, validation split, device (CPU/GPU). Visualizations: Plotly and Rich functions for analysis. To be Implemented Features Setup an addition to the cli, the option to display real time metrics of a plot being updated in real time, via plotly/nicegui. Option to expand the base of information by artificially creating focus/defocus. Optimization using JIT methods for PyTorch To see a more in depth explanation of the features: Methods page. Installation Downloads ...
Comparing Pixel-wise Neural Networks and PCA+SVM for Alginate Detection in Raman Maps
Repository Project-Pyspec Class Spec to Spec Pixel-wise and secondarily scene-level alginate detection from Raman spectral maps. This detection is done primarily as a comparison between machine learning (ML) via a neural network (NN) and principal component analysis (PCA) combined with support vector machine (PCA+SVM). These both yield classification probabilities of positive or negative for presence of alginate. To evaluate scene-level presence, both the ML and PCA+SVM pipelines split the data into cross-validation (CV) folds.To measure the accuracy of the classification, the results are compared with an intersection over union (IoU) metric, and confusion matrices. PCA+SVM methods match labels almost perfectly under current labeling / preprocessing, however, it seems that a machine learning method could allow for a supplement or even perfect classification. ...
Holod Further Work
This page is a work in progress! Will be updated soon!
Holod Neural Network Learning
This page is a work in progress! Will be updated soon!