Adding synthetic_complexity_score

molgenevalmetric.py

@@ -38,7 +38,13 @@ from fcd_torch import FCD
 
 # from SCScore import SCScorer
 
-import math 
+'''
+This is a standalone, importable SCScorer model. It does not have tensorflow as a
+dependency and is a more attractive option for deployment. The calculations are
+fast enough that there is no real reason to use GPUs (via tf) instead of CPUs (via np)
+'''
+
+import math, sys, random, os
 import numpy as np
 import time
 import rdkit.Chem as Chem
@@ -47,6 +53,9 @@ import json
 import gzip
 import six
 
+import os
+project_root = os.path.dirname(os.path.dirname(__file__))
+
 score_scale = 5.0
 min_separation = 0.25
 
@@ -62,7 +71,7 @@ class SCScorer():
         self.score_scale = score_scale
         self._restored = False
 
-    def restore(self, weight_path=os.path.join('model.ckpt-10654.as_numpy.json.gz'), FP_rad=FP_rad, FP_len=FP_len):
+    def restore(self, weight_path=os.path.join(project_root, 'models', 'full_reaxys_model_1024bool', 'model.ckpt-10654.as_numpy.json.gz'), FP_rad=FP_rad, FP_len=FP_len):
         self.FP_len = FP_len; self.FP_rad = FP_rad
         self._load_vars(weight_path)
         # print('Restored variables from {}'.format(weight_path))