This is the source code for LG-HACKATHON hosted by DACON. The main objective of the competition was to predict the SMILES sequence given a chemical structure image. Our prediction achieved 0.99567 in private leaderboard(Tanimoto-Similarity based)

You can download the model weights from this dropbox link

Dependencies (with python =3.6)

conda env create -f utils/LG_CONDA_ENV.yml

pip install utils/pip_requirements.txt

python training_data_generation/dataframe_generation_by_group.py 
python training_data_generation/train_image_generation.py 

python training_data_generation/sequence_dataframe_generation.py

python training_data_generation/data_sampling.py

You will need to modify the src/config.py to accustom your directory setting.

# The data_dir should point the data directory folder. All training and testing files should be placed below
data_dir = Path('/home/jaeho_ubuntu/SMILES/data/')
train_dir = data_dir / 'train' # all train_<number>.png files should be under the train folder
test_dir = data_dir / 'test'  # all test_<number>.png files should be under the test folder
train_csv_dir = data_dir /'train.csv'  #train.csv file provided by LG 
# Sample submission directory
sample_submission_dir = data_dir /'sample_submission.csv'

# train_modified contains a modifed version of train.csv. 
# Information of the train/validation split is stored here. I saved in pickle just for efficiency
train_pickle_dir = data_dir /'train_modified.pkl'

### Data directory containing .hdf5, .json file files.
input_data_dir = data_dir / 'input_data'
base_file_name = 'seed_123_max75smiles'

### seed for train/val split
random_seed = 123

### Reversed_token_file used to map numbers to string. 
reversed_token_map_dir = input_data_dir/ f'REVERSED_TOKENMAP_{base_file_name}.json'

Running below scripts will return a .hdf5, .json files needed for training and testing.

# Only if you need to make training and validation set 
python --split True --train_file True
# If you need to make a test set provided by LG
python --test_file True

python main.py --work_type train

if you want to train again form the checkpoint(saved model weight)

python main.py --work_type train \
               --model_load_path <path where the model is saved> \
               --mode_load_num <model number>

When you run prediction with our model, you should choose whether to predict with a single model or an ensamble model. Turn either --work_type single_test or --work_type ensemble_test flags True.

single model test requires three flags: --model load path, --model load number, and --test file path

python main.py --work_type single_test \
               --model_load_path <path where the model is saved> \
               --model_load_num <model number> \
               --test_file_path <path where the test images are saved>

ensemble model contains weight file of 5 single models. The test requires two flags: --model load path and --test file path Running this script also requires ./model/prediction_models.yaml which contains hyperparameters of the model and the type of model used. The models in the --model load path should match with this prediction_models.yaml file configurations.

python main.py --work_type ensemble_test \
               --model_load_path <path where the model is saved> \
               --test_file_path <path where the test images are saved>