Analytic initialisation

n3fit/runcards/examples/simunet_examples/quick_check.yaml

@@ -0,0 +1,189 @@
+#Configuration file for NNPDF++
+#
+############################################################
+description: "Test fit for the simunet release project. Oll scaled by 10.0, other operators by 1.0 and constant initialisation about 0.0 used."
+
+############################################################
+# frac: training fraction
+# ewk: apply ewk k-factors
+# sys: systematics treatment (see systypes)
+
+dataset_inputs:
+# # EWPO
+- {dataset: LEP_ZDATA, simu_fac: "EFT_LO", use_fixed_predictions: True}
+- {dataset: LEP_BRW, simu_fac: "EFT_LO", use_fixed_predictions: True}
+- {dataset: LEP_BHABHA, simu_fac: "EFT_LO", use_fixed_predictions: True}
+- {dataset: ALPHAEW_WITHTHUNC, simu_fac: "EFT_LO", use_fixed_predictions: True}
+
+# # Diboson
+- {dataset: LEP_EEWW_182GEV, simu_fac: "EFT_LO", use_fixed_predictions: True}
+- {dataset: LEP_EEWW_189GEV, simu_fac: "EFT_LO", use_fixed_predictions: True}
+- {dataset: LEP_EEWW_198GEV, simu_fac: "EFT_LO", use_fixed_predictions: True}
+- {dataset: LEP_EEWW_206GEV, simu_fac: "EFT_LO", use_fixed_predictions: True}
+- {dataset: ATLAS_WW_13TEV_2016_MEMU, simu_fac: "EFT_NLO", use_fixed_predictions: True}
+- {dataset: ATLAS_WZ_13TEV_2016_MTWZ, simu_fac: "EFT_NLO", use_fixed_predictions: True}
+- {dataset: CMS_WZ_13TEV_2016_PTZ, simu_fac: "EFT_NLO", use_fixed_predictions: True}
+
+
+fixed_pdf_fit: True
+load_weights_from_fit: 221103-jmm-no_top_1000_iterated # If this is uncommented, training starts here.
+analytic_initialisation_pdf: 221103-jmm-no_top_1000_iterated
+use_th_covmat: False
+
+simu_parameters:
+# Dipoles
+# - {name: 'OtZ', scale: 0.01, initialisation: {type: uniform, minval: -1, maxval: 1}}
+# - {name: 'OtW', scale: 0.01, initialisation: {type: uniform, minval: -1, maxval: 1}}
+# - {name: 'OtG', scale: 0.01, initialisation: {type: uniform, minval: -1, maxval: 1}}
+# Quark Currents
+# - {name: 'Opt', scale: 0.01, initialisation: {type: uniform, minval: -1, maxval: 1}}
+# - {name: 'O3pQ3', scale: 0.01, initialisation: {type: uniform, minval: -1, maxval: 1}}
+- {name: 'O3pq', scale: 1.0, initialisation: {type: analytic}}
+- {name: 'OpQM', scale: 1.0, initialisation: {type: analytic}}
+- {name: 'OpqMi', scale: 1.0, initialisation: {type: analytic}}
+- {name: 'Opui', scale: 1.0, initialisation: {type: analytic}}
+- {name: 'Opdi', scale: 1.0, initialisation: {type: analytic}}
+# Lepton currents
+- {name: 'O3pl', scale: 1.0, initialisation: {type: analytic}}
+- {name: 'Opl', scale: 1.0, initialisation: {type: analytic}}
+- {name: 'Ope', scale: 1.0, initialisation: {type: analytic}}
+# 4 Fermions 4Q
+# - {name: 'O1qd', scale: 0.01, initialisation: {type: uniform, minval: -1, maxval: 1}}
+# - {name: 'O1qu', scale: 0.01, initialisation: {type: uniform, minval: -1, maxval: 1}}
+# - {name: 'O1dt', scale: 0.01, initialisation: {type: uniform, minval: -1, maxval: 1}} 
+# - {name: 'O1qt', scale: 0.01, initialisation: {type: uniform, minval: -1, maxval: 1}} 
+# - {name: 'O1ut', scale: 0.01, initialisation: {type: uniform, minval: -1, maxval: 1}} 
+# - {name: 'O11qq', scale: 0.01, initialisation: {type: uniform, minval: -1, maxval: 1}} 
+# - {name: 'O13qq', scale: 0.01, initialisation: {type: uniform, minval: -1, maxval: 1}} 
+# - {name: 'O8qd', scale: 0.01, initialisation: {type: uniform, minval: -1, maxval: 1}}
+# - {name: 'O8qu', scale: 0.01, initialisation: {type: uniform, minval: -1, maxval: 1}}
+# - {name: 'O8dt', scale: 0.01, initialisation: {type: uniform, minval: -1, maxval: 1}} 
+# - {name: 'O8qt', scale: 0.01, initialisation: {type: uniform, minval: -1, maxval: 1}} 
+# - {name: 'O8ut', scale: 0.01, initialisation: {type: uniform, minval: -1, maxval: 1}} 
+# - {name: 'O81qq', scale: 0.01, initialisation: {type: uniform, minval: -1, maxval: 1}} 
+# - {name: 'O83qq', scale: 0.01, initialisation: {type: uniform, minval: -1, maxval: 1}}
+# 4 Fermions 4HeavyQ
+# - {name: 'OQQ1', scale: 0.01, initialisation: {type: uniform, minval: -1, maxval: 1}}
+# - {name: 'OQQ8', scale: 0.01, initialisation: {type: uniform, minval: -1, maxval: 1}}
+# - {name: 'OQt1', scale: 0.01, initialisation: {type: uniform, minval: -1, maxval: 1}}
+# - {name: 'OQt8', scale: 0.01, initialisation: {type: uniform, minval: -1, maxval: 1}}
+# - {name: 'Ott1', scale: 0.01, initialisation: {type: uniform, minval: -1, maxval: 1}}
+# 4 Fermions 2L2Q
+# - {name: 'Oeu', scale: 0.01, initialisation: {type: uniform, minval: -1, maxval: 1}}
+# - {name: 'Olu', scale: 0.01, initialisation: {type: uniform, minval: -1, maxval: 1}}
+# - {name: 'Oed', scale: 0.01, initialisation: {type: uniform, minval: -1, maxval: 1}}
+# - {name: 'Olq3', scale: 0.01, initialisation: {type: uniform, minval: -1, maxval: 1}}
+# - {name: 'Olq1', scale: 0.01, initialisation: {type: uniform, minval: -1, maxval: 1}}
+# - {name: 'Oqe', scale: 0.01, initialisation: {type: uniform, minval: -1, maxval: 1}}
+# - {name: 'Old', scale: 0.01, initialisation: {type: uniform, minval: -1, maxval: 1}}
+# 4 Fermions 4L
+- {name: 'Oll', scale: 10.0, initialisation: {type: analytic}}
+# Yukawa
+# - {name: 'Omup', scale: 0.01, initialisation: {type: uniform, minval: -1, maxval: 1}}
+# - {name: 'Otap', scale: 0.01, initialisation: {type: uniform, minval: -1, maxval: 1}}
+# - {name: 'Otp', scale: 0.01, initialisation: {type: uniform, minval: -1, maxval: 1}}
+# - {name: 'Obp', scale: 0.01, initialisation: {type: uniform, minval: -1, maxval: 1}}
+# - {name: 'Ocp', scale: 0.01, initialisation: {type: uniform, minval: -1, maxval: 1}}
+# Bosonic
+# - {name: 'OG', scale: 0.01, initialisation: {type: uniform, minval: -1, maxval: 1}}
+- {name: 'OWWW', scale: 1.0, initialisation: {type: analytic}}
+# - {name: 'OpG', scale: 0.01, initialisation: {type: uniform, minval: -1, maxval: 1}}
+# - {name: 'OpW', scale: 0.01, initialisation: {type: uniform, minval: -1, maxval: 1}}
+# - {name: 'OpB', scale: 0.01, initialisation: {type: uniform, minval: -1, maxval: 1}}
+- {name: 'OpWB', scale: 1.0, initialisation: {type: analytic}}
+# - {name: 'Opd', scale: 0.01, initialisation: {type: uniform, minval: -1, maxval: 1}}
+- {name: 'OpD', scale: 1.0, initialisation: {type: analytic}}
+
+
+############################################################
+datacuts:
+  t0pdfset: 221103-jmm-no_top_1000_iterated # PDF set to generate t0 covmat
+  q2min: 3.49                        # Q2 minimum
+  w2min: 12.5                        # W2 minimum
+
+############################################################
+theory:
+  theoryid: 200     # database id
+
+############################################################
+trvlseed: 475038818
+nnseed: 2394641471
+mcseed: 1831662593
+save: "weights.h5"
+genrep: true      # true = generate MC replicas, false = use real data
+
+############################################################
+
+
+parameters: # This defines the parameter dictionary that is passed to the Model Trainer
+  nodes_per_layer: [25, 20, 8]
+  activation_per_layer: [tanh, tanh, linear]
+  initializer: glorot_normal
+  optimizer:
+    clipnorm: 6.073e-6
+    learning_rate: 2.621e-3
+    optimizer_name: Nadam
+  epochs: 30000
+  positivity:
+    initial: 184.8
+    multiplier:
+  integrability:
+    initial: 184.8
+    multiplier:
+  stopping_patience: 1.0
+  layer_type: dense
+  dropout: 0.0
+  threshold_chi2: 3.5
+
+fitting:
+  # EVOL(QED) = sng=0,g=1,v=2,v3=3,v8=4,t3=5,t8=6,(pht=7)
+  # EVOLS(QED)= sng=0,g=1,v=2,v8=4,t3=4,t8=5,ds=6,(pht=7)
+  # FLVR(QED) = g=0, u=1, ubar=2, d=3, dbar=4, s=5, sbar=6, (pht=7)
+  fitbasis: EVOL  # EVOL (7), EVOLQED (8), etc.
+  basis:
+  - {fl: sng, pos: false, trainable: false, mutsize: [15], mutprob: [0.05], smallx: [
+      1.093, 1.121], largex: [1.486, 3.287]}
+  - {fl: g, pos: false, trainable: false, mutsize: [15], mutprob: [0.05], smallx: [
+      0.8329, 1.071], largex: [3.084, 6.767]}
+  - {fl: v, pos: false, trainable: false, mutsize: [15], mutprob: [0.05], smallx: [
+      0.5202, 0.7431], largex: [1.556, 3.639]}
+  - {fl: v3, pos: false, trainable: false, mutsize: [15], mutprob: [0.05], smallx: [
+      0.1205, 0.4839], largex: [1.736, 3.622]}
+  - {fl: v8, pos: false, trainable: false, mutsize: [15], mutprob: [0.05], smallx: [
+      0.5864, 0.7987], largex: [1.559, 3.569]}
+  - {fl: t3, pos: false, trainable: false, mutsize: [15], mutprob: [0.05], smallx: [
+      -0.5019, 1.126], largex: [1.754, 3.479]}
+  - {fl: t8, pos: false, trainable: false, mutsize: [15], mutprob: [0.05], smallx: [
+      0.6305, 0.8806], largex: [1.544, 3.481]}
+  - {fl: t15, pos: false, trainable: false, mutsize: [15], mutprob: [0.05], smallx: [
+      1.087, 1.139], largex: [1.48, 3.365]}
+
+############################################################
+positivity:
+  posdatasets:
+  - {dataset: POSF2U, maxlambda: 1e6}        # Positivity Lagrange Multiplier
+  - {dataset: POSF2DW, maxlambda: 1e6}
+  - {dataset: POSF2S, maxlambda: 1e6}
+  - {dataset: POSFLL, maxlambda: 1e6}
+  - {dataset: POSDYU, maxlambda: 1e10}
+  - {dataset: POSDYD, maxlambda: 1e10}
+  - {dataset: POSDYS, maxlambda: 1e10}
+  - {dataset: POSF2C, maxlambda: 1e6}
+  - {dataset: POSXUQ, maxlambda: 1e6}        # Positivity of MSbar PDFs
+  - {dataset: POSXUB, maxlambda: 1e6}
+  - {dataset: POSXDQ, maxlambda: 1e6}
+  - {dataset: POSXDB, maxlambda: 1e6}
+  - {dataset: POSXSQ, maxlambda: 1e6}
+  - {dataset: POSXSB, maxlambda: 1e6}
+  - {dataset: POSXGL, maxlambda: 1e6}
+
+############################################################
+integrability:
+  integdatasets:
+  - {dataset: INTEGXT8, maxlambda: 1e2}
+  - {dataset: INTEGXT3, maxlambda: 1e2}
+
+############################################################
+debug: false
+maxcores: 4
+

n3fit/src/n3fit/layers/CombineCfac.py

@@ -6,6 +6,7 @@
 
 from validphys import initialisation_specs
 
+
 class CombineCfacLayer(Layer):
     """
     Creates the combination layer of SIMUnet.
@@ -16,6 +17,7 @@ def __init__(
         scales,
         linear_names,
         initialisations,
+        analytic_initialisation,
         initialisation_seed,
         replica_number,
     ):
@@ -37,12 +39,15 @@ def __init__(
         # At this point, create a tf object with the correct random initialisation.
         initial_values = []
         assert len(initialisations) == len(linear_names)
+        index = 0
         for ini, name in zip(initialisations, linear_names):
-            hash_value = int(hashlib.sha1(name.encode("utf-8")).hexdigest(), 16) % (10 ** 18)
+            hash_value = int(hashlib.sha1(name.encode("utf-8")).hexdigest(), 16) % (
+                10**18
+            )
             seed = np.int32((initialisation_seed + replica_number) ^ hash_value)
 
             if isinstance(ini, initialisation_specs.ConstantInitialisation):
-                val = tf.constant(ini.value, dtype='float32', shape=(1,))
+                val = tf.constant(ini.value, dtype="float32", shape=(1,))
             elif isinstance(ini, initialisation_specs.UniformInitialisation):
                 val = tf.random_uniform_initializer(
                     minval=ini.minval,
@@ -52,31 +57,40 @@ def __init__(
             elif isinstance(ini, initialisation_specs.GaussianInitialisation):
                 tf.random.set_seed(seed)
                 val = tf.random.normal([1], ini.mean, ini.std_dev, tf.float32)
+            elif isinstance(ini, initialisation_specs.AnalyticInitialisation):
+                val = np.array([float(analytic_initialisation[index])])
             else:
                 raise RuntimeError(
                     "Invalid initialisation: choose form constant, uniform or Gaussian."
                 )
+            index += 1
             initial_values.append(val)
 
+        num_initial = len(initial_values)
+
         self.scales = np.array(scales, dtype=np.float32)
-        if len(initial_values) > 0:
+        if num_initial > 0:
             initial_values = tf.concat(initial_values, 0)
-            initial_values = tf.math.divide(initial_values, self.scales)
+            initial_values = tf.math.multiply(initial_values, self.scales)
 
+        if num_initial > 0:
+            initial_values = tf.concat(initial_values, 0)
+            initial_values = tf.reshape(initial_values, (num_initial,))
             self.w = tf.Variable(
                 initial_value=initial_values,
                 trainable=True,
             )
         else:
             self.w = tf.Variable(
-                initial_value=tf.zeros(shape=(len(initial_values),), dtype="float32"),
+                initial_value=tf.zeros(shape=(num_initial,), dtype="float32"),
                 trainable=True,
             )
+
         self.scales = np.array(scales, dtype=np.float32)
         self.linear_names = linear_names
 
     def _compute_linear(self, linear_values):
-        scaled_values = linear_values/self.scales[:, np.newaxis]
+        scaled_values = linear_values / self.scales[:, np.newaxis]
         return tf.reduce_sum(self.w[:, tf.newaxis] * scaled_values, axis=0)
 
     def call(self, inputs, linear_values):
@@ -112,6 +126,8 @@ def call(self, inputs, linear_values):
                 dtype=np.float32,
             )
             linear = self._compute_linear(linear_values)
+            inputs = tf.cast(inputs, tf.float64)
+
             return (1 + linear) * inputs
 
         return inputs

n3fit/src/n3fit/model_trainer.py

@@ -124,6 +124,7 @@ def __init__(
         simu_parameters_names=None,
         simu_parameters_scales=None,
         bsm_fac_initialisations=None,
+        analytic_initialisation=None,
         bsm_initialisation_seed=0,
         debug=False,
         kfold_parameters=None,
@@ -187,6 +188,7 @@ def __init__(
         self.simu_parameters_names=simu_parameters_names
         self.simu_parameters_scales = simu_parameters_scales
         self.bsm_fac_initialisations = bsm_fac_initialisations
+        self.analytic_initialisation = analytic_initialisation
         self.bsm_initialisation_seed = bsm_initialisation_seed
         self.fixed_pdf = fixed_pdf
         self.replicas = replicas
@@ -511,6 +513,7 @@ def _generate_observables(
             scales=np.array(self.simu_parameters_scales, dtype=np.float32),
             linear_names=self.simu_parameters_names,
             initialisations=self.bsm_fac_initialisations,
+            analytic_initialisation=self.analytic_initialisation,
             initialisation_seed=self.bsm_initialisation_seed,
             replica_number=self.replicas[0],
         )