adding generate_actmax.py script

scripts/generate_actmax.py

+#!/usr/bin/env python
+
+import sys, argparse, re, random
+
+parser = argparse.ArgumentParser(description="generate events that maximise the activation for a given neuron")
+parser.add_argument("input_project")
+parser.add_argument("output_file")
+parser.add_argument("-n", "--nevents", default=100000, type=int)
+parser.add_argument("-j", "--mask-jets", action="store_true",
+                    help="mask variables called jet*Pt/Eta/Phi and generate a random uniform distribution of the number of jets (only nescessary for non-recurrent NN)")
+args = parser.parse_args()
+
+import logging
+logging.basicConfig()
+logging.getLogger().setLevel(logging.INFO)
+
+import h5py
+import numpy as np
+
+from KerasROOTClassification.utils import (
+    weighted_quantile,
+    get_max_activation_events,
+    create_random_event,
+    get_ranges
+)
+from KerasROOTClassification import load_from_dir
+import meme
+
+meme.setOptions(deactivated=True)
+
+input_project = args.input_project
+output_file = args.output_file
+
+c = load_from_dir(input_project)
+
+ranges, mask_probs = get_ranges(c.transform(c.x_train), [0.01, 0.99], c.w_train_tot, mask_value=c.mask_value, max_evts=10000)
+
+def mask_uniform(x, mask_value, recurrent_field_idx):
+    """
+    Mask recurrent fields with a random (uniform) number of objects. Works in place.
+    """
+    for rec_idx in recurrent_field_idx:
+        for evt in x:
+            masked = False
+            nobj = int(random.random()*(rec_idx.shape[1]+1))
+            for obj_number, line_idx in enumerate(rec_idx.reshape(*rec_idx.shape[1:])):
+                if obj_number == nobj:
+                    masked=True
+                if masked:
+                    evt[line_idx] = mask_value
+
+def get_input_flat(x):
+    return x[0].reshape(-1, len(c.fields))
+
+
+if args.mask_jets:
+    jet_fields = {}
+    for field_name in c.fields:
+        if any(field_name.startswith("jet") and field_name.endswith(suffix) for suffix in ["Pt", "Eta", "Phi"]):
+            jet_number = re.findall("[0-9]+", field_name)[0]
+            if not jet_number in jet_fields:
+                jet_fields[jet_number] = []
+            jet_fields[jet_number].append(c.fields.index(field_name))
+    jet_fields = [np.array([[v for k, v in sorted(jet_fields.items(), key=lambda l:l[0])]])]
+
+
+def input_transform(x):
+    x = np.array(x)
+    if hasattr(c, "mask_uniform"):
+        c.mask_uniform(x)
+        return c.get_input_list(x)
+    elif args.mask_jets:
+        mask_uniform(x, c.mask_value, jet_fields)
+        return x
+
+
+opt_kwargs = dict()
+if hasattr(c, "mask_uniform"):
+    opt_kwargs["input_transform"] = input_transform
+    opt_kwargs["input_inverse_transform"] = c.get_input_flat
+if args.mask_jets:
+    opt_kwargs["input_transform"] = input_transform
+    opt_kwargs["input_inverse_transform"] = get_input_flat
+
+
+evts = get_max_activation_events(
+    c.model, ranges,
+    ntries=args.nevents,
+    layer=len(c.model.layers)-1,
+    neuron=0,
+    maxit=10,
+    seed=45,
+    threshold=0,
+    **opt_kwargs
+)
+
+with h5py.File(output_file, "w") as f:
+    f.create_dataset("actmax", data=evts[1])