Variation of scores of different runs of inference.py

[sri480673]

Hi @gcorso,
I have tested for CPU enabled inference.py of recent version(1.1.3) of Diffdock, where the output scores are different by rerunning the inference.py And number of failed cases also different.

Below shows the two run outputs:
ls results_batch_size_10/user_predictions_small/5zcu/
rank10_confidence-3.96.sdf rank2_confidence-0.51.sdf rank5_confidence-1.50.sdf rank8_confidence-3.18.sdf
rank1_confidence-0.37.sdf rank3_confidence-0.53.sdf rank6_confidence-1.71.sdf rank9_confidence-3.21.sdf

ls results_batch_size_10_rerun/user_predictions_small/5zcu:
rank10_confidence-4.67.sdf rank2_confidence-0.34.sdf rank5_confidence-0.64.sdf rank8_confidence-1.94.sdf
rank1_confidence-0.12.sdf rank3_confidence-0.44.sdf rank6_confidence-1.92.sdf rank9_confidence-3.66.sdf
rank1.sdf rank4_confidence-0.49.sdf rank7_confidence-1.92.sdf

After I set the no_random=True in randomize_position() and sampling() called from inference.py Still getting variation in scores by rerunning the inference.py and different no. of failed cases also.

Below is output of one of the score of protein 5zcu,
rank10_confidence-4.34.sdf rank2_confidence-1.38.sdf rank5_confidence-2.71.sdf rank8_confidence-3.67.sdf
rank1_confidence-1.35.sdf rank3_confidence-2.27.sdf rank6_confidence-2.92.sdf rank9_confidence-3.80.sdf
rank1.sdf rank4_confidence-2.48.sdf rank7_confidence-3.65.sdf

ls results_1_norandom/user_predictions_small/5zcu:
rank10_confidence-4.07.sdf rank2_confidence-2.11.sdf rank5_confidence-2.78.sdf rank8_confidence-3.71.sdf
rank1_confidence-1.86.sdf rank3_confidence-2.30.sdf rank6_confidence-3.30.sdf rank9_confidence-3.81.sdf
rank1.sdf rank4_confidence-2.30.sdf rank7_confidence-3.60.sdf

Along with no_random=True and fix the seed for numpy,torch and random package, still getting different failed cases as well as scores are also different.
Please give a suggestion how to overcome the above issue