g2net_ml_dl

Workflows

Installing the library

cd python
pip install .

Then you can import it into whatever package you need. This makes it convenient to use outside of just this repo.

• Downside: If you need to make changes to the library, you must reinstall the package for the new changes to work

Full Workflow

git clone -b YOUR_BRANCH_NAME https://github.com/jchen42703/g2net_ml_dl
cd g2net_ml_dl/python
pip install .

# Then run your scripts from here or just import stuff from the library ^^.

If you're debugging a lot, I recommend:

import os
os.chdir("python")
import g2net
os.chdir("..")
os.getcwd()

A simple kernel restart will update your changes, so you won't need to reinstall the library with pip install .

Using the Neural network on HPC

Move all of the files to related to the network onto HPC and then find the environment.yaml file. There is an error in the environment.yaml file and you will need to vim into it and comment out

  - pytorch=1.9.0=py3.7_cuda11.1_cudnn8.0.5_0

And another to:

    - timm==0.4.12

To run it and set up the environement run the following commands:

module load miniconda3/4.9.2
conda env create -n kumaconda -f=environment.yaml
bash
conda activate kumaconda
conda install pytorch=1.9.0=py3.7_cuda11.1_cudnn8.0.5_0

Prep network with

python prep_data.py --cache

Make sure that kuma_utils is up to date. In particular, the files:

kuma_utils/torch/modules/pooling.py
kuma_utils/torch/trainer.py
kuma_utils/torch/clip_grad.py

You may need to change the batch size and hardware in configs.py:

HW_CFG = {
    'RTX3090': (16, 128, 1, 24), # CPU cores, RAM amount, GPU count, GPU RAM total
    'A100': (20, 100, 2, 40),
}
class Baseline:
    name = 'baseline'
    seed = 2021
    train_path = INPUT_DIR/'train.csv'
    test_path = INPUT_DIR/'test.csv'
    train_cache = INPUT_DIR/'train_cache.pickle' # here
    test_cache = INPUT_DIR
...
parallel = 'df'
...
class Nspec23arch3(Nspec23):
    name = 'nspec_23_arch_3'
    model_params = Nspec23.model_params.copy()
    model_params['spec_params'] = dict(
        base_filters=128,
        kernel_sizes=(64, 16, 4),
    )
    batch_size = 32
    model_params['model_name'] = 'tf_efficientnet_b6_ns'
    transforms = Nspec22aug1.transforms.copy()

To train the network:

python train.py --config Nspec23arch3 --gpu 0 1 --progress_bar

For output, we will use g2net-submission.ipynb, put it will need to be edited as our version will be only using one part of this scheme. More specifically, we will have to edit

prediction_list = [
    RESULT_DIR/'pseudo_12',
    RESULT_DIR/'pseudo_seq_04',
    RESULT_DIR/'pseudo_13',
    RESULT_DIR/'pseudo_14',
    RESULT_DIR/'pseudo_seq_07',
    RESULT_DIR/'pseudo_16',
    RESULT_DIR/'pseudo_17',
    RESULT_DIR/'pseudo_18',
    RESULT_DIR/'pseudo_19',
    RESULT_DIR/'pseudo_21',
    RESULT_DIR/'pseudo_23',
    #
    RESULT_DIR/'pseudo_24',
    RESULT_DIR/'pseudo_10',
    RESULT_DIR/'pseudo_26',
    RESULT_DIR/'pseudo_25',
    RESULT_DIR/'pseudo_07',
    RESULT_DIR/'pseudo_22',
    RESULT_DIR/'pseudo_15',
    RESULT_DIR/'pseudo_seq_03',
    RESULT_DIR/'pseudo_06',
]

since we will not have these files, but only one of them.

Authors & Acknowledgments

• Joseph Chen
• Richard Chen
• Takoda Denhof
• Owen Hill
• Julian Robinson

Thank you to Case Western Reserve University and Professor Gajurel for letting us use the HPC cluster to run our experiments.