Timing for MSA not reported/documented

[andrejberg]

Hi, I would kindly ask you to report the timing for features/MSA. Your current table about prediction speed with 4.9 s for 100 AA is highly misleading. This is what I get for 100 AA on an GCloud instance with 32 CPU cores and a single NVIDIA T4:

{
    "features": 1481.1342334747314,
    "process_features_model_1_pred_0": 3.2654552459716797,
    "predict_and_compile_model_1_pred_0": 162.3107635974884,
    "predict_benchmark_model_1_pred_0": 25.0131938457489,
    "relax_model_1_pred_0": 12.343136310577393,
    "process_features_model_2_pred_0": 1.918269395828247,
    "predict_and_compile_model_2_pred_0": 160.3056800365448,
    "predict_benchmark_model_2_pred_0": 22.53654146194458,
    "relax_model_2_pred_0": 5.992400169372559,
    "process_features_model_3_pred_0": 1.6289887428283691,
    "predict_and_compile_model_3_pred_0": 130.60518503189087,
    "predict_benchmark_model_3_pred_0": 24.988179683685303,
    "relax_model_3_pred_0": 6.115017414093018,
    "process_features_model_4_pred_0": 1.7446517944335938,
    "predict_and_compile_model_4_pred_0": 130.17664051055908,
    "predict_benchmark_model_4_pred_0": 24.91871929168701,
    "relax_model_4_pred_0": 6.450090408325195,
    "process_features_model_5_pred_0": 1.7851057052612305,
    "predict_and_compile_model_5_pred_0": 129.50175189971924,
    "predict_benchmark_model_5_pred_0": 22.370880603790283,
    "relax_model_5_pred_0": 6.818777799606323
}

I spend days testing and "debugging" because I was certain that you would mention it if the feature step would take 10 times longer than the prediction itself.

Edit: Yes, it gets faster with the reduced data set, 10% faster....