Solution for Bristol-Myers Squibb – Molecular Translation

Source code of 70th place solution for Bristol-Myers Squibb – Molecular Translation.

You can read about the solution in detail here.

Solution

The main goal is to translate chemical structure images into InChI transcription - machine-readable format. This is image captioning task and solved as CNN+RNN architecture.

Key points:

• EfficientNet as encoder, and LSTM+Attention as decoder.
• Adaptive batchsampler (the higher sample loss, the higher probability to add the sample in a batch).
• Predict Smile notation and than convert it to InChI (Smiles is much more simpler to predict - shorter notation and less token-classes).
• Freeze the encoder and train the lstm-decoder separately.

What other experiments could be done:

• More experiments with synthetic images generation.
• Split InChI up to 8 indepenpent string layers (separated by the "/" notation: "/b", "/t", "/m", "/s", etc.) and train the separate models for each layer. Each layer in the InChI describes different information about the molecule, and several models trained on separate inchi-layers can get good results.
• Add rotation transoform to different angles.
• Add beam search.
• Try to replace LSTM to Transformer.

Quick setup and start

• Nvidia drivers >= 465, CUDA >= 11.3
• Docker, nvidia-docker

The provided Dockerfile is supplied to build an image with CUDA support and cuDNN.

Preparations

• Clone the repo.

  git clone git@github.com:skalinin/BMS-Molecular-Translation.git
  cd BMS-Molecular-Translation

• Download and extract dataset to the data/bms-molecular-translation folder.

• sudo make all to build a docker image and create a container.

• Preprocess dataset csv, generate additional synth images and tokenize data.

  python src/scripts/data_preprocess/prepare_csv/BMS_preprocess.py
  python src/scripts/data_preprocess/prepare_csv/external_synth_data.py
  python src/scripts/data_preprocess/tokenize_csv.py

Run

• Train model

  python src/scripts/train.py

• Make submission csv

  python src/scripts/submission.py \
    --encoder_pretrain /path/to/encoder-weigths \
    --decoder_pretrain /path/to/decoder-weigths