Support regularization

.github/workflows/main.yml

@@ -236,18 +236,35 @@ jobs:
           rabbit_fit.py $RABBIT_OUTDIR/test_tensor.hdf5 -o $RABBIT_OUTDIR/
           -t 0 --postfix blinded --setConstraintMinimum bkg_2_norm 0.5 --doImpacts --globalImpacts
 
+      - name: pseudodata fit
+        run: >- 
+          rabbit_fit.py $RABBIT_OUTDIR/test_tensor.hdf5 -o $RABBIT_OUTDIR/ --postfix pseudodata
+          -t 0 --pseudoData original --doImpacts --globalImpacts --unblind ^sig$
+
+
+  alternative-fits:
+    runs-on: [self-hosted, linux, x64]
+    needs: [setenv, make-tensor]
+
+    steps:
+      - env:
+          RABBIT_OUTDIR: ${{ needs.setenv.outputs.RABBIT_OUTDIR }}
+          PYTHONPATH: ${{ needs.setenv.outputs.PYTHONPATH }}
+          PATH: ${{ needs.setenv.outputs.PATH }}
+        run: |
+          echo "RABBIT_OUTDIR=${RABBIT_OUTDIR}" >> $GITHUB_ENV
+          echo "PYTHONPATH=${PYTHONPATH}" >> $GITHUB_ENV
+          echo "PATH=${PATH}" >> $GITHUB_ENV
+
+      - uses: actions/checkout@v4
+
       - name: sparse tensor fit 
         run: >- 
           rabbit_fit.py $RABBIT_OUTDIR/test_tensor_sparse.hdf5 -o $RABBIT_OUTDIR/ --postfix sparse
           -t 0 --noBinByBinStat --doImpacts --globalImpacts --computeVariations
           --saveHists --saveHistsPerProcess --computeHistErrors --computeHistErrorsPerProcess
           -m Project ch1 a -m Project ch1 b
 
-      - name: pseudodata fit
-        run: >- 
-          rabbit_fit.py $RABBIT_OUTDIR/test_tensor.hdf5 -o $RABBIT_OUTDIR/ --postfix pseudodata
-          -t 0 --pseudoData original --doImpacts --globalImpacts --unblind ^sig$
-
       - name: covariance fit
         run: >- 
           rabbit_fit.py $RABBIT_OUTDIR/test_tensor.hdf5 -o $RABBIT_OUTDIR/ --postfix covariance
@@ -262,11 +279,44 @@ jobs:
           --saveHists --saveHistsPerProcess --computeHistErrors --computeHistErrorsPerProcess
           --computeHistImpacts --computeHistCov -m Project ch1 a -m Project ch1 b 
 
-      - name: chi2 fit with Barlow--Beeston full
+      - name: Barlow--Beeston full (gamma)
+        run: >- 
+          rabbit_fit.py $RABBIT_OUTDIR/test_tensor_symmetric.hdf5 -o $RABBIT_OUTDIR/ --postfix nonlinear
+          -t 0 --doImpacts --globalImpacts --globalImpactsDisableJVP --binByBinStatType gamma --binByBinStatMode full
+          --saveHists --saveHistsPerProcess --computeHistErrors --computeHistErrorsPerProcess
+          --computeHistImpacts --computeHistCov -m Project ch1 a -m Project ch1 b 
+
+      - name: chi2 fit with Barlow--Beeston full (normal-multiplicative)
         run: >- 
           rabbit_fit.py $RABBIT_OUTDIR/test_tensor_symmetric.hdf5 -o $RABBIT_OUTDIR/ --postfix bb_full
           -t 0 --globalImpacts --binByBinStatType normal-multiplicative --binByBinStatMode full
 
+  regularization:
+    runs-on: [self-hosted, linux, x64]
+    needs: [setenv, make-tensor]
+    steps:
+      - env:
+          RABBIT_OUTDIR: ${{ needs.setenv.outputs.RABBIT_OUTDIR }}
+          PYTHONPATH: ${{ needs.setenv.outputs.PYTHONPATH }}
+          PATH: ${{ needs.setenv.outputs.PATH }}
+          WEB_DIR: ${{ needs.setenv.outputs.WEB_DIR }}
+          PLOT_DIR: ${{ needs.setenv.outputs.PLOT_DIR }}
+        run: |
+          echo "RABBIT_OUTDIR=${RABBIT_OUTDIR}" >> $GITHUB_ENV
+          echo "PYTHONPATH=${PYTHONPATH}" >> $GITHUB_ENV
+          echo "PATH=${PATH}" >> $GITHUB_ENV
+          echo "WEB_DIR=${WEB_DIR}" >> $GITHUB_ENV
+          echo "PLOT_DIR=${PLOT_DIR}" >> $GITHUB_ENV
+
+      - name: lcurve scan
+        run: >-
+          rabbit_fit.py $RABBIT_OUTDIR/test_tensor.hdf5 -o $RABBIT_OUTDIR/ --postfix lcurve_scan
+          -t 0 --unblind -r SVD Select 'ch0_masked' -r SVD Select 'ch0_masked'--earlyStopping 15
+
+      - name: plot lcurve
+        run: >-
+          python tests/plot_epoch_loss_time.py results/fitresults_lcurve_scan.hdf5 -o $WEB_DIR/$PLOT_DIR --types lcurve --title Experiment --subtitle 'Work in progress'
+
   bsm:
     runs-on: [self-hosted, linux, x64]
     needs: [setenv, make-tensor]
@@ -431,7 +481,7 @@ jobs:
 
   copy-clean:
     runs-on: [self-hosted, linux, x64]
-    needs: [setenv, symmerizations, bsm, plotting, likelihoodscans]
+    needs: [setenv, symmerizations, alternative-fits, bsm, regularization, plotting, likelihoodscans]
     if: always() 
     steps:
       - env:

bin/rabbit_fit.py

@@ -13,6 +13,8 @@
 from rabbit.mappings import helpers as mh
 from rabbit.mappings import mapping as mp
 from rabbit.poi_models import helpers as ph
+from rabbit.regularization import helpers as rh
+from rabbit.regularization.lcurve import l_curve_optimize_tau, l_curve_scan_tau
 from rabbit.tfhelpers import edmval_cov
 
 from wums import output_tools, logging  # isort: skip
@@ -21,7 +23,7 @@
 
 
 def make_parser():
-    parser = parsing.common_parser()
+    parser = parsing.common_parser("Perform binned profilme maximum likelihood fits")
     parser.add_argument("--outname", default="fitresults.hdf5", help="output file name")
     parser.add_argument(
         "--fullNll",
@@ -149,6 +151,12 @@ def make_parser():
         type=str,
         help="Specify result from external postfit file",
     )
+    parser.add_argument(
+        "--noFit",
+        default=False,
+        action="store_true",
+        help="Do not not perform the minimization.",
+    )
     parser.add_argument(
         "--noPostfitProfileBB",
         default=False,
@@ -182,7 +190,24 @@ def make_parser():
         action="store_true",
         help="compute impacts of frozen (non-profiled) systematics",
     )
-
+    parser.add_argument(
+        "--lCurveScan",
+        default=False,
+        action="store_true",
+        help="For use with regularization, scan the L curve versus values for tau",
+    )
+    parser.add_argument(
+        "--lCurveOptimize",
+        default=False,
+        action="store_true",
+        help="For use with regularization, find the value of tau that maximizes the curvature",
+    )
+    parser.add_argument(
+        "--regularizationStrength",
+        default=0.0,
+        type=float,
+        help="For use with regularization, set the regularization strength (tau)",
+    )
     return parser.parse_args()
 
 
@@ -281,7 +306,21 @@ def fit(args, fitter, ws, dofit=True):
     edmval = None
 
     if args.externalPostfit is not None:
-        fitter.load_fitresult(args.externalPostfit, args.externalPostfitResult)
+        fitter.load_fitresult(
+            args.externalPostfit,
+            args.externalPostfitResult,
+            profile=not args.noPostfitProfileBB,
+        )
+
+    if args.lCurveScan:
+        tau_values, l_curve_values = l_curve_scan_tau(fitter)
+        ws.add_1D_integer_hist(tau_values, "step", "tau")
+        ws.add_1D_integer_hist(l_curve_values, "step", "lcurve")
+
+    if args.lCurveOptimize:
+        best_tau, max_curvature = l_curve_optimize_tau(fitter)
+        ws.add_1D_integer_hist([best_tau], "best", "tau")
+        ws.add_1D_integer_hist([max_curvature], "best", "lcurve")
 
     if dofit:
         cb = fitter.minimize()
@@ -293,7 +332,8 @@ def fit(args, fitter, ws, dofit=True):
             fitter._profile_beta()
 
         if cb is not None:
-            ws.add_loss_time_hist(cb.loss_history, cb.time_history)
+            ws.add_1D_integer_hist(cb.loss_history, "epoch", "loss")
+            ws.add_1D_integer_hist(cb.time_history, "epoch", "time")
 
     if not args.noHessian:
         # compute the covariance matrix and estimated distance to minimum
@@ -477,6 +517,14 @@ def main():
             mp.CompositeMapping(mappings),
         ]
 
+    ifitter.tau.assign(args.regularizationStrength)
+    regularizers = []
+    for margs in args.regularization:
+        mapping = mh.load_mapping(margs[1], indata, *margs[2:])
+        regularizer = rh.load_regularizer(margs[0], mapping, dtype=indata.dtype)
+        regularizers.append(regularizer)
+    ifitter.regularizers = regularizers
+
     np.random.seed(args.seed)
     tf.random.set_seed(args.seed)
 
@@ -560,7 +608,7 @@ def main():
 
                 if not args.prefitOnly:
                     ifitter.set_blinding_offsets(blind=blinded_fits[i])
-                    fit(args, ifitter, ws, dofit=ifit >= 0)
+                    fit(args, ifitter, ws, dofit=ifit >= 0 and not args.noFit)
                     fit_time.append(time.time())
 
                     if args.saveHists: