Final project for Winter 2020 iteration of EE 367: Computational Imaging at Stanford.

Author: Cindy Nguyen, cindyn at stanford.edu

This repo contains code to perform end-to-end optimization of a plastic phase mask placed close to the sensor (<= 25 mm focal distance).

We implement an optics module, sensor module, and Wiener deconvolution for image reconstruction. The loss is backpropagated into the heightmap to optimize a coded phase mask.

• train.py is the main training script (implements sequence of experiments as a queue, each experiment means training on one particular set of hyperparameters)
• params.json is used to load hyperparameters relevant for training and initialization (see README.md for details)
• ranges.json is where you can specify the hyperparameters to switch out in between experiments (each experiment will stop depending on early stopping criteria)
• dataio.py contains the dataloader that will load the SBDataset for training
• denoising_unet.py contains the model and model helper functions
• notebooks/analyze_models.ipynb loads models and plots losses

• git clone this repository
• Edit params.json and ranges.json as needed for experiment (see PARAMS.md for details)
  • If running script to test is things are set up properly, you can run with the settings as is. This will load in a single image and beginning optimizing a height map and damping factor starting from an in-focus Fresnel lens height map.
  • If you want to run with the dataset, set the parameter download_data to be true. After first run, set this to false.
• In console, run

$ CUDA_VISIBLE_DEVICES=# python3 train.py

where # specifies GPU device number. If running on CPU, you can simply run python3 train.py.

• Data generated from the experiment (saved models and Tensorboard files) will be specified in runs/exp_name/exp_name_# where exp_name is as specified in hyperaparameters and # is automatically determined.
• The training script is set up so that a new experiment is created for each hyperparameter in ranges.json, run each sequentially to completion, and save model checkpoints during training in the runs folder.
• Data files are not included to save space.

Results from optimizing the height map only.

Results from optimizing both the height map and Wiener deconvolution damping factor.

• pytorch
• numpy
• tensorboard
• U-Net repo
• Propagation and utils repo