Numerical instability occurs when `--float` flag is used for conversion

[mbarbetti]

To reproduce the problem, download the files contained in this public folder: https://pandora.infn.it/public/335590

Folder contents:

• symptoms.npz (data)
• tX.pkl (input sklearn scaler)
• saved_model (keras sequential model)
• CompiledModel_sim10-2016MU_latest.so (compiled pipeline)

The following lines allow to reproduce the problem, in particular, all the instances contained in symptoms.npz produce quantile and absolute errors greater than $10^{-3}$.

import pickle
import ctypes
import numpy as np
import pandas as pd
from tensorflow import keras
from scipy.stats import percentileofscore as pos

# Symptomatic data

npz_file = np.load("/path/to/symptoms.npz")
x = npz_file["x"].astype(np.float32)
print(x.shape)  # (80, 4)

# Python output

with open("/path/to/tX.pkl", "rb") as file:
  x_scaler = pickle.load(file)
x_prep = x_scaler.transform(x)

model = keras.models.load_model("/path/to/saved_model")
ismuon_py = model.predict(x_prep)

# C output

dll = ctypes.CDLL("/path/to/CompiledModel_sim10-2016MU_latest.so")
float_p = ctypes.POINTER(ctypes.c_float)

ismuon_c = []
for input_row in x:
  in_f = input_row.astype(ctypes.c_float)
  out_f = np.empty(1, dtype=ctypes.c_float)
  getattr(dll, f"isMuon_pion_pipe")(
    out_f.ctypes.data_as(float_p), 
    in_f.ctypes.data_as(float_p),
  )
  ismuon_c.append(out_f)

# Numerical instability

ismuon_py = ismuon_py.flatten()
ismuon_c = np.array(ismuon_c).flatten()

ismuon_err_df = pd.DataFrame()
ismuon_q_err_py = pos(ismuon_py, ismuon_py)
ismuon_q_err_c = pos(ismuon_py, ismuon_c)
ismuon_err_df["q_err_isMuon"] = (ismuon_q_err_py - ismuon_q_err_c) / 1e2
ismuon_err_df["abs_err_isMuon"] = abs(ismuon_py - ismuon_c)

err_counts = np.count_nonzero(
  (abs(ismuon_err_df["q_err_isMuon"]) > 0.001)
  & (ismuon_err_df["abs_err_isMuon"] > 0.001)
)
print(err_counts)  # 80

[mbarbetti]

Deployment results for Lamarr available here: https://codimd.web.cern.ch/fuTqvOcOSJSanLqraS7nUw