DECAF

This is the code repo for the paper 'Recovery of Continuous 3D Refractive Index Maps from Discrete Intensity-Only Measurements using Neural Fields' (previously known as 'Zero-Shot Learning of Continuous 3D Refractive Index Maps from Discrete Intensity-Only Measurements').

Project page.

Download datasets

Available datasets:

• Algae (Figure 2)
• Diatom (aidt, Figure 3)
• Diatom_midt (mdit, Extended Figure 1)
• Cells_b (Figure 4)
• Cells_c (Figure 4)
• Celegans_head (Figure 5)
• Celegans_body/middle (Figure 5)
• Simulated granulocyte cluster (Figure 6)
• Yanny's data (Supplementary)

Data avaliable at https://drive.google.com/drive/folders/1maQxPFHFcouoEFOUo7e5FJSN2hBT78eB?usp=sharing. Please move data to datasets/DATASET_NAME/input/DATASET_NAME.mat

Setup the environment

Setup the environment

conda env create --file decaf.yml

To activate this environment, use

conda activate decaf-env

To deactivate an active environment, use

conda deactivate

Run the code

Run inference:

python predict.py --flagfile=datasets/DATASET_NAME/pred_config.txt

NOTE: We already provide the pre-trained models for each sample in the folder datasets/DATASET_NAME/trained_model/

Run training:

python main.py --flagfile=datasets/DATASET_NAME/train_config.txt

NOTE: We trained the model on a machine equipped with one AMD Threadripper 3960X 24-core CPU and four Nvidia RTX 3090 GPUs. We parallelized the training of DeCAF over two GPUs to accelerate the convergence. Under this setup, it approximately takes one day to train the model.

Example:

python predict.py --flagfile=datasets/Cells_c/pred_config.txt

Expected outputs

After inference, the results will be saved in the folder datasets/DATASET_NAME/inference/, including

one data file (.mat)
one video (.mp4)
one stack of images (.tif)

File structure

DECAF
  |-datasets
    |-Algae
	  |-inference : inference results (avaliable after running predict.py).
		|- includes .mat, .mp4, and .tif images
	  |-models: training models (avaliable after running main.py).
	  |-input: forward model and measurements (requires additional download).
	  |-trained_model: trained neural representaion weights for the data set.
    |...
  |-model: DECAF model
  |-trained_regularizer: Trained DnCNN denoiser.
  |-main.py : training main function.
  |-predict.py: inference main function.