Skip to content

Latest commit

 

History

History
97 lines (68 loc) · 2.91 KB

README.md

File metadata and controls

97 lines (68 loc) · 2.91 KB
Raw

Sequence VAE

Background

Deep generative models are rapidly becoming popular tools for generating new molecules and optimizing the chemical properties. In this work, we will introduce a VAE model based on molecular sequence - Sequence VAE.

Instructions

This repository will give you the instruction of training a sequence VAE model and sampling new molecules from the pretrained model.

How to get ?

Model link:

You can download the pre-trained model.

Data link

We use the data from ZINC Clean Leads collection [1].

You can download the data from [datalink] (https://baidu-nlp.bj.bcebos.com/PaddleHelix/datasets/molecular_generation/zinc_moses.tgz).

  1. create a folder './data'
  2. unzip the file into './data'

data |-- zinc_mose | |-- train.csv | |-- test.csv

Training Models

Model config

The model config is the parameters used for building the model graph. They are saved in the file: model_config.json

encoding

"max_length":the maximun length of the inout sequence "q_cell": the RNN cell type of encoding network "q_bidir": if it's bidirectional RNN or not "q_d_h": the hidden size of encoding RNN "q_n_layers": the layer numbers of encoding RNN "q_dropout": the drop out rate of encoding RNN

# decoding
"d_cell": the RNN cell type of decoding network
"d_n_layers": the layer numbers of decoding RNN
"d_dropout": the drop out rate of decoding RNN
"d_z": the hidden size of latent space
"d_d_h":the hidden size of decoding RNN
"freeze_embeddings": if freeze the embedding layer

Training setting

In order to train the model, we need to set the training parameters. The dafault paramaters are saved in file: args.py

# Train
'--n_epoch': number of trianing epoch, default=1000
'--n_batch': number of bach size, default=1000
'--lr_start': 'Initial lr value, default=3 * 1e-4

# kl annealing
'--kl_start': Epoch to start change kl weight from, default=0
'--kl_w_start': Initial kl weight value, default=0
'--kl_w_end': Maximum kl weight value, default=0.05

To run the trianing scripts:


CUDA_VISIBLE_DEVICES=0 python trainer.py \

--device='gpu' \

--dataset_dir='./data/zinc_moses/train.csv' \

--model_config='model_config.json' \

--model_save='./results/train_models/' \

--config_save='./results/config/' \

Sampling results from prior

Valid: 0.9765 Novelty: 0.731 Unique@1k: 0.993 Filters: 0.853 IntDiv: 0.846

Reference

[1] @article{polykovskiy2020molecular, title={Molecular sets (MOSES): a benchmarking platform for molecular generation models}, author={Polykovskiy, Daniil and Zhebrak, Alexander and Sanchez-Lengeling, Benjamin and Golovanov, Sergey and Tatanov, Oktai and Belyaev, Stanislav and Kurbanov, Rauf and Artamonov, Aleksey and Aladinskiy, Vladimir and Veselov, Mark and others}, journal={Frontiers in pharmacology}, volume={11}, year={2020}, publisher={Frontiers Media SA} }