RAD

1. Input Data and preprocessing: https://github.com/uboone/OpenSamples Use the hdf5 files, 'Inclusive, WithWire'

Use ubopendata conda environment for this part

The wire_table contains (Wire, Timetick, ADC) values. I used plane2, the collection plane The input data is preprocessed to reduce the model size. My naming scheme is AXB, where A and B are the division factors for Timetick and Wire respectively. The Timetick dimension is downsampled by a factor of 10 before division

ex: 10X4 Original input size values for single event is (Wire, Timetick) = (3456, 6400) --> (3456,640) after downsampling --> (864,64) after division

hdf5 files are preprocessed into npy files through 01_data2npy_full.py. Samples are inside inputData

2. Training Teacher Use 02_train_teacher.py Use ubqkeras conda environment for step 2 to 4

3. Training Student Use 03_train_student.py Requires Teacher from step 2.

Trained models are saved in savedModel

4. Evaluating Loss (or 'Anomaly Score') Use 04_getLoss.py Evaluates loss (or Anomlay Score in our language) and saves into npy file

5. Dependencies Use conda_envs If one has to install the environments from scratch, the .txt folders have dependencies.

file.py, microboone_utils.py are obtained from https://github.com/uboone/OpenSamples QDenseBatchnorm.py is in this branch of qkeras google/qkeras#74

6. Models: Taken from https://github.com/Princeton-AD/cicada/blob/main/models.py

Defined in models.py

teacher_reshape and teacher_dense, teacher_reshape2 values need to be changed according with the prepocessed input image size