Model interpretation in Python script takes several minutes

Hello!

I am trying to implement model interpretation within my ML pipeline. The interpretation instructions provided in the README work perfectly fine when I tried them on command line. Then, I tried to include this into my code by writing a small function. The method works but it takes several minutes to provide an output unlike the command line which returns the result immediately. I tried this with just one compound: Clc1ccc(c(Cl)c1Cl)c2c(Cl)cc(Cl)c(Cl)c2Cl

I would really appreciate your help in fixing this. Code and output below.

Best, Vishal

Code:

import tempfile
import time
import pandas as pd
import sys
sys.path.insert(0, './predictors/chemprop')
from chemprop.args import InterpretArgs
from chemprop.interpret import interpret

def get_interpretation(kekule_smiles_df):

    start = time.time()
    intrprt_df = pd.DataFrame()

    with tempfile.NamedTemporaryFile(delete=True) as temp:

        kekule_smiles_df.to_csv(temp.name + '.csv', index=None)

        # interpretation arguments
        intrprt_args = [
        '--data_path', temp.name + '.csv',
        '--checkpoint_path', './models/pampa/gcnn_model.pt',
        '--property_id', '1',
        ]

        # slightly modified the interpret method to return a dataframe for further use
        intrprt_df = interpret(args=InterpretArgs().parse_args(intrprt_args))
        temp.close()
    end = time.time()
    print(f'{end - start} seconds to interpret {kekule_smiles_df.shape[0]} molecules')
    return intrprt_df

Output:

Loading pretrained parameter "encoder.encoder.cached_zero_vector".
Loading pretrained parameter "encoder.encoder.W_i.weight".
Loading pretrained parameter "encoder.encoder.W_h.weight".
Loading pretrained parameter "encoder.encoder.W_o.weight".
Loading pretrained parameter "encoder.encoder.W_o.bias".
Loading pretrained parameter "ffn.1.weight".
Loading pretrained parameter "ffn.1.bias".
Loading pretrained parameter "ffn.4.weight".
Loading pretrained parameter "ffn.4.bias".
smiles,score,rationale,rationale_score
ClC1=C(Cl)C(Cl)=C(C2=C(Cl)C(Cl)=C(Cl)C=C2Cl)C=C1,0.981,Clc1c[cH:1]c(Cl)[cH:1]c1[CH3:1],1.000
1469.1523442268372 seconds to interpret 1 molecules

[hesther]

I will need to run some tests. Unfortunately, the MCTS algorithm in general is not really fast. However, you can parallelize it manually by e.g. splitting up your test data into multiple files (command line), or running multiple processes within a python script.

I will try to reproduce the weird timing behavior and track it down (if I can reproduce it)

[iwwwish]

Thank you @hesther. I am not sure I’ll be able to do this myself completely. But if I were to get around it, I will definitely open a PR.