spectrometer-cnn

using CNN for spectrum reconstruction using all fiber spectromter

folder contents

• Model, saving caffe models

• Data, transmission matrices

• results, trained model, plots, etc

• archive, old results

• obsolete, old codes

analysis stream line

• DataLoader, contains different methods for loading data

• CaffeModel, contains wrapper for training using different caffe models

• Scheduler, orgnize batch model traning and model validation

• OptModel, using optimization, \sum (T*S - I)^2 + \lambda * \sum S^2

• RI sufix, refractive index data.

results

old

• in single line reconstruction, cnn is able to reconstruct sub correlation images with noiseless image
• in multiple line (5 lines) reconstruction, cnn perform similar to linear reconstruction

nerual network training

• for the last layer, using softmax cant converge (may be try L1 normalize)
• for the last layer, without relu reaches much better loss

new

• linear reconstruction does not work well, not nearly as good as optimization

facts

• noise is about 1e-3 of the whole mean value.

control experiments include

• noise options: yes or no (done)
• method options: linear, neural net, and optimization (done)
• spectrum options: single, multiple, and continuous (need continous)
• spectrum density: correlated and uncorrelated (done, using different transimission matrix)

next steps