Have a look of Organ sample/model generated smiles file! about moses HOT 2 CLOSED

Seems weird, indeed. We'll investigate it, but such "chainy" molecules can be a solution to high uniqueness—just add a few carbons and you get a new molecule. Probably you can achieve better results by trying other reward functions.

from moses.

True, using QED can avoid this type of novel molecules. But do you have any data to support the advantage of using FrechetMetrics over direct calling QED? Calling FrechetMetrics is very slow and the GPU usage will be easily lost after certain pg optimizations and turned back to CPU calculation.
The worst case is to call FCD metrics as the reward function. FCD metrics calls Keras/tensorflow several times for one pg iteration. Hope someone can rewrite it with pytorch and then we can see the advantage of FCDMetrics (as a reward function).

from moses.