CNN4Jets - Convolutional neural network (CNN) for jet classification

Pretrained Model:

A pretrained model can be downloaded from: https://drive.google.com/drive/folders/1sKZEqtc8Gj44FUUwBgmfDq0N7OGoiQsi?usp=sharing

Quickstart:

To create and train a model, execute the main.py file in the shell:

$ python3 main.py

Now the following things will happen:

1. The model loads the data from the specified txt files, e.g.:

'input_paths' : {'ttbar' : './conv_out_ttbar.txt', 'qcdJets' : './conv_out_qcdJets.txt'}.

2. The model will create a CNN network.
3. The model creates a dataset from the loaded txt files.
4. A test set of images will be saved as numpy arrays in: ./model/image_data for each of the classes.
5. Furthermore these images will be plottet and saved to: ./model/images.
6. Then a training loop will start.
7. The model will be saved in a model file in the folder: ./model

Now a model exists, that can be used for the classification itself. Creating and training a model will not be done on the computer in the museum. Instead, a model file and all additional data will be provided beforehand.

To classify an image, two parser arguments exist:

--model=<path> and --classify=<path>.

Using these arguments, one can then choose an image file from ./model/image_data and classify it. For example, we can classify an image of the ttbar dataset:

$ python3 main.py --model=./model/model --classify=./model/image_data/ttbar/X_test_i.npy.

The index i must be replaced by an existing image in the ./model/image_data/ttbar folder. The model then predicts the probabilities of the image to be in each of the classes and saves these probabilities as a numpy array in a tmp_prob.npy file in the ./model folder

Todo:

-[x] install Madgraph, ROOT & Delphes

-[x] Make calorimater images from delphes output

-[x] Write params card for constit2images

-[x] Preprocessing of the jet images

-[x] Train/test split in Dataset

-[x] Write model architecture

-[x] Write training loop

-[x] Write validation

-[x] Additional learning rate scheduler