Implementation of the Differentiable Neural Computer used by Simon Johansson and Oleksii Prykhodko in this paper on ChemRxiv.
================================================================================
Based on the RNN model used in this paper, with its implementation here.
This was later expanded upon to use random SMILES, but some optimizations used is incompatible with our DNC modules. Please refer to this paper for instructions on how to create randomized SMILES for the training files.
A Conda environment.yml is supplied with all the required libraries.
$> conda env create -f environment.yml
$> source activate reinvent-dnc
(reinvent-dnc) $> ./create_model.py -h
usage: create_model.py [-h] --input-smiles-path INPUT_SMILES_PATH
[--output-model-path OUTPUT_MODEL_PATH]
[--num-layers NUM_LAYERS]
[--num-controller-layers NUM_CONTROLLER_LAYERS]
[--layer-size LAYER_SIZE]
[--embedding-layer-size EMBEDDING_LAYER_SIZE]
[--dropout DROPOUT]
[--max-string-size MAX_STRING_SIZE]
[--memory-cells MEMORY_CELLS]
[--cell-size CELL_SIZE]
[--read-heads READ_HEADS]
[--model-type MODEL_TYPE]
[--controller-type CONTROLLER_TYPE]
Create a model with the vocabulary extracted from a SMILES file.
optional arguments:
-h, --help show this help message and exit
--input-smiles-path INPUT_SMILES_PATH, -i INPUT_SMILES_PATH
SMILES to calculate the vocabulary from. The SMILES
are taken as-is, no processing is done.
--output-model-path OUTPUT_MODEL_PATH, -o OUTPUT_MODEL_PATH
Prefix to the output model.
--num-layers NUM_LAYERS, -g NUM_LAYERS
Number of layers of the model [DEFAULT: 1]
--num-controller-layers NUM_CONTROLLER_LAYERS, -cg NUM_CONTROLLER_LAYERS
Number of layers in the controller (if using a DNC) [DEFAULT: 3]
--layer-size LAYER_SIZE, -s LAYER_SIZE
Size of each of the hidden layers [DEFAULT: 512]
--embedding-layer-size EMBEDDING_LAYER_SIZE, -e EMBEDDING_LAYER_SIZE
Size of the embedding layer [DEFAULT: 128]
--dropout DROPOUT, -d DROPOUT
Dropout constant [DEFAULT: 0.0]
--max-string-size MAX_STRING_SIZE
Maximum size of the strings [DEFAULT: 256]
--memory-cells MEMORY_CELLS
Amount of memory cells in DNC [DEFAULT: 32]
--cell-size CELL_SIZE
The size of the cell in DNC [DEFAULT: 20]
--read-heads READ_HEADS
Amount of read heads in DNC [DEFAULT: 8]
--model-type MODEL_TYPE
The model to use for training [DEFAULT: dnc]
--controller-type CONTROLLER_TYPE
The cell types in the controller [DEFAULT: lstm]
Note that, if you opt to run an RNN model insted of dnc, the proper input is to use 1 controller layer, and let num-layers decide the number of hidden layers in the model.
General Usage
-------------
1) Create Model (`create_model.py`): Creates a blank model file. This will by default create a directory called `storage/modelname/`, where the modelname lists tour parameteres in the format of `day_month_modeltype_controllertype_numlayers_numcontrollerlayers_hiddensize_memorycells_memorylength_readheads`
2) Train Model (`train_model.py`): Trains a created model with the specified parameters. Input is the model directory.
3) Sample Model (`sample_from_model.py`): Samples an already trained model for a given number of SMILES. Input is the full path to the model checkpoint you want to use.
4) Calculate NLLs (`calculate_nlls.py`): Calculates the NLLs of input `.smi` file using specified model checkpoint.