BNN optimization.

This tackles three issues:

• The aleatoric uncertainty was a single number in the past for the full spectrum. In this version, a number per output is used, allowing the uncertainty to adapt to each case.
• There was a major bug fixed in the hyperprior for the aleatoric uncertainty, which used incorrectly the weight's prior instead of the uncertainty variable. This has now been fixed in the hyperprior calculation.
• The number of neurons have been reduced to simplify the network.

pes_to_spec/test/offline_analysis.py

@@ -144,7 +144,7 @@ def main():
     parser.add_argument('-o', '--offset', type=int, metavar='INT', default=0, help='Train ID offset')
     parser.add_argument('-c', '--xgm_cut', type=float, metavar='INTENSITY', default=500, help='XGM intensity threshold in uJ.')
     parser.add_argument('-e', '--bnn', action="store_true", default=False, help='Use BNN?')
-    parser.add_argument('-w', '--weight', action="store_true", default=True, help='Whether to reweight data as a function of the pulse energy to make it invariant to that.')
+    parser.add_argument('-w', '--weight', action="store_true", default=False, help='Whether to reweight data as a function of the pulse energy to make it invariant to that.')
 
     args = parser.parse_args()