The most convenient way to get the code up and running is by setting up a conda
environment and install the necesary packages in a new environment. The yml
file can be found at conda_environment.yml.
conda env create -f conda_environment.yml
conda activate spectra2structure
The above would create and activate a conda environment with the name spectra2structure.
After installing python 3.8.3, you can install all the dependensies. The dependendies can
be found at pip_requirements.txt.
python -m pip install -r pip_requirements.txt
.
├── data
│ └── RPNMRData
│ └── graph_conv_many_nuc_pipeline.datasets
├── requirements
├── source
│ ├── test
│ │ ├── environment
│ │ │ └── RPNMR
│ │ ├── test
│ │ └── utils
│ └── train
│ ├── environment
│ │ ├── RPNMR
│ ├── saved_models
│ └── utils
└── trainedModels
├── RPNMRModels
└── crossvalidationModels
If you have a slurm workload manager then the framework can be run by simply
modifying and running the batch script provided in the train directory.
If not, then you can begin training by the following set of commands:
ulimit -n 40960
ray start --head --num-cpus=40 --num-gpus=2 --object-store-memory 50000000000
python parallel_agent.py
Change --num-cpus to the number of cores that you have available in your machine
and also change the value in parallel_agent.py for the variable num_processes.
num_processes = 1.5 * num-cpus is usually handlable by any machine.
The models would be saved inside the directory saved_models and the running
loss values would be appended to the files prior_log.txt and value_log.txt
for the prior model and the value model respectively.
For each episode, the environment chooses a target spectra randomly from the training dataset and creates datapoint through MCTS Runs. This training dataset can be found in the directory data. Already trainined model can be found in the directory trainedModels.
If you have a slurm workload manager then the framework can be run by simply
modifying and running batch.sh file in the test directory. If not,
you can begin testing on target spectra by the following set of commands:
ulimit -n 40960
ray start --head --num-cpus=40 --num-gpus=4 --object-store-memory 50000000000
python parallel_agent.py
ray stop
Change --num-cpus and --num-gpus to the number of cores and gpus available.
Also change the value of variable NUM_GPU and num_processes in the file
parallel_agent.py.
The target spectra to best tested are specified by the array smiles_idx in the
file parallel_agent.py. These indices are passed as a parameter to the function
env.reset(idx) present in the file
environment/environment.py which then
chooses the target spectra from the test dataset based on the passed parameter.
To run test on custom molecules, one needs to only modify the function reset in
the environment class to load the target information appropriately.
After a series of episodes, the resulting guesses are pickled and stored for each
target spectra in the directory source/test/test/.
The information of each episode completed can also be seen in the log_file
generated in the source/test/ directory.
One can also run source/test/rescheck.py from the test directory to see a
summary of the testing where in, the accuracy and the closest molecule to each
target molecule(by rdkit similarity) are reported.
A well trained forward model is reqd to run the agent for training and testing.
You can train the model by running the batch script in the directory
source/forwardModel/batch.sh, this training is tracked by wandb if you have
configured with your account, the model would be saved in the directory
source/forwardModel/checkpoints/best_model.00000000.state. Copy this model to
the location: trainedModels/RPNMRModels/best_model.00000000.state
Once saved, you can go ahead with training and testing the inverse framework.
Information about hyperparameters used for the experiments can be found in the file hyper_params.txt.