Ecoder

ECON-T autoencoder model

Setup

On VM `klijnsma-gpu3` or any other Linux VM

Get data from FNAL LPC:

/uscms/home/kkwok/eos/ecoder/V11

Setup environment using miniconda3

source install_miniconda3.sh #if your first time
source setup.sh #also if your first time
conda activate ecoder-env

If you are working on the LPC cluster working node, check out a CMSSW_11_1_X and use the tensorflow version that comes with cmsenv. If you have python default to run in python 2.7, run the scripts with python3.

To obtain qkeras (for training with quantized bit constrains), clone the repo locally: https://github.com/google/qkeras

Juypter notebook demos

Following files illustrates prototypes of different autoencoder architectures

auto.ipynb - 1D deep NN autoencoder demo
auto_CNN.ipynb - 2D CNN autoencoder demo
Auto_qCNN.ipynb - 2D quantized CNN autoencoder, trained with qKeras demo

Training scripts

Scripts to explore hyperparameters choices:

denseCNN.py - model class for constructing conv2D-dense architectures
qDenseCNN.py- model class for constructing conv2D-dense architectures with qKeras
train.py - train(or load weights) and evaluate models

Example usage:

## edit parameters setting inside train.py
## train with 1 epoch to make sure model parameters are OK, output to a trainning folder
python train.py -i /uscms/home/kkwok/eos/ecoder/V11/signal/nElinks_5/ -o ./test/ --epoch 1 --AEonly 1 --nELinks 5
## train the weights with max 150 epoch 
python train.py -i /uscms/home/kkwok/eos/ecoder/V11/signal/nElinks_5/ -o ./test/ --epoch 150 --AEonly 1 --nELinks 5

## After producing a `.hdf5` file from trainning, you can re-run the model skipping the trainning phase.
## Do so by simply setting the model parameter 'ws' to `modelname.hdf5`

Convert to a constant tensorflow graph

Using tensorflow 2.4.0 and keras 2.2.4-tf. Can be obtained from CMSSW_11_1_0.

### convert the decoder model
python3 converttoTF.py -o ./graphs/ -i decoder_Jul24_keras.json --outputGraph decoder --outputLayer decoder_output/Sigmoid 
### convert the encoder model
python3 converttoTF.py -o ./graphs/ -i encoder_Jul24_keras.json --outputGraph encoder --outputLayer encoded_vector/Relu

kakwok / Ecoder