compare.py

@@ -22,6 +22,7 @@ def overlay_ROC(filename, *projects, **kwargs):
     threshold_log = kwargs.pop("threshold_log", True)
     lumifactor = kwargs.pop("lumifactor", None)
     tight_layout = kwargs.pop("tight_layout", False)
+    show_auc = kwargs.pop("show_auc", True)
     if kwargs:
         raise KeyError("Unknown kwargs: {}".format(kwargs))
 
@@ -43,7 +44,7 @@ def overlay_ROC(filename, *projects, **kwargs):
     colors = prop_cycle.by_key()['color']
 
     for p, color in zip(projects, colors):
-        fpr, tpr, threshold = roc_curve(p.y_test, p.scores_test, sample_weight = p.w_test)
+        fpr, tpr, threshold = roc_curve(p.l_test, p.scores_test, sample_weight = p.w_test)
         fpr = 1.0 - fpr
         try:
             roc_auc = auc(tpr, fpr)
@@ -52,12 +53,16 @@ def overlay_ROC(filename, *projects, **kwargs):
             roc_auc = auc(tpr, fpr, reorder=True)
 
         ax.grid(color='gray', linestyle='--', linewidth=1)
-        ax.plot(tpr,  fpr, label=str(p.name+" (AUC = {:.3f})".format(roc_auc)), color=color)
+        if show_auc:
+            label = str(p.name+" (AUC = {:.3f})".format(roc_auc))
+        else:
+            label = p.name
+        ax.plot(tpr,  fpr, label=label, color=color)
         if plot_thresholds:
             ax2.plot(tpr, threshold, "--", color=color)
         if lumifactor is not None:
-            sumw_b = p.w_test[p.y_test==0].sum()*lumifactor
-            sumw_s = p.w_test[p.y_test==1].sum()*lumifactor
+            sumw_b = p.w_test[p.l_test==0].sum()*lumifactor
+            sumw_s = p.w_test[p.l_test==1].sum()*lumifactor
             ax_abs_b.plot(tpr, (1.-fpr)*sumw_b, alpha=0)
             ax_abs_b.invert_yaxis()
             ax_abs_s.plot(tpr*sumw_s, fpr, alpha=0)

keras_visualize_activations/read_activations.py

 import keras.backend as K
 
-from keras.layers import Masking, InputLayer
 
 def get_activations(model, model_inputs, print_shape_only=False, layer_name=None):
     print('----- activations -----')
@@ -13,12 +12,8 @@ def get_activations(model, model_inputs, print_shape_only=False, layer_name=None
         inp = [inp]
         model_multi_inputs_cond = False
 
-    # all layer outputs
-    # skip input and masking layers
     outputs = [layer.output for layer in model.layers if
-               (layer.name == layer_name or layer_name is None)
-               and not isinstance(layer, InputLayer)
-               and not isinstance(layer, Masking)]
+               layer.name == layer_name or layer_name is None]  # all layer outputs
 
     funcs = [K.function(inp + [K.learning_phase()], [out]) for out in outputs]  # evaluation functions
 

plotting.py

@@ -20,9 +20,9 @@ logger.addHandler(logging.NullHandler())
 Some further plotting functions
 """
 
-def save_show(plt, fig, filename):
+def save_show(plt, fig, filename, **kwargs):
     "Save a figure and show it in case we are in ipython or jupyter notebook."
-    fig.savefig(filename)
+    fig.savefig(filename, **kwargs)
     try:
         get_ipython
         plt.show()
@@ -234,6 +234,114 @@ def plot_NN_vs_var_2D_all(plotname, model, means,
     save_show(plt, fig, plotname, bbox_inches='tight')
 
 
+def plot_profile_2D_all(plotname, model, events,
+                        valsx, valsy,
+                        nbinsx, xmin, xmax,
+                        nbinsy, ymin, ymax,
+                        transform_function=None,
+                        varx_label=None,
+                        vary_label=None,
+                        zrange=None, logz=False,
+                        plot_last_layer=False,
+                        log_default_ymin=1e-5,
+                        global_norm=True,
+                        cmap="inferno", **kwargs):
+
+    "Similar to plot_profile_2D, but creates a grid of plots for all neurons."
+
+    # transform
+    if transform_function is not None:
+        events = transform_function(events)
+
+    logger.info("Reading activations for all neurons")
+    acts = get_activations(model, events, print_shape_only=True)
+    logger.info("Done")
+
+    if plot_last_layer:
+        n_neurons = [len(i.reshape(i.shape[0], -1)[0]) for i in acts]
+    else:
+        n_neurons = [len(i.reshape(i.shape[0], -1)[0]) for i in acts[:-1]]
+    layers = len(n_neurons)
+
+    nrows_ncols = (layers, max(n_neurons))
+    fig = plt.figure(1, figsize=nrows_ncols)
+    grid = ImageGrid(fig, 111, nrows_ncols=nrows_ncols[::-1], axes_pad=0,
+                     label_mode="1",
+                     aspect=False,
+                     cbar_location="top",
+                     cbar_mode="single",
+                     cbar_pad=.2,
+                     cbar_size="5%",)
+    grid_array = np.array(grid)
+    grid_array = grid_array.reshape(*nrows_ncols[::-1])
+
+    global_min = None
+    global_max = None
+
+    logger.info("Creating profile histograms")
+    ims = []
+    reg_plots = []
+    for layer in range(layers):
+        neurons_acts = acts[layer]
+        neurons_acts = neurons_acts.reshape(neurons_acts.shape[0], -1)
+        for neuron in range(len(neurons_acts[0])):
+            acts_neuron = neurons_acts[:,neuron]
+            ax = grid_array[neuron][layer]
+            extra_opts = {}
+            if not (plot_last_layer and layer == layers-1):
+                # for hidden layers, plot the same z-scale
+                if logz:
+                    norm = matplotlib.colors.LogNorm
+                else:
+                    norm = matplotlib.colors.Normalize
+                if zrange is not None:
+                    extra_opts["norm"] = norm(vmin=zrange[0], vmax=zrange[1])
+                else:
+                    extra_opts["norm"] = norm(vmin=global_min, vmax=global_max)
+            hist, xedges, yedges = get_profile_2D(
+                valsx, valsy, acts_neuron,
+                nbinsx, xmin, xmax,
+                nbinsy, ymin, ymax,
+                **kwargs
+            )
+            if global_min is None or hist.min() < global_min:
+                global_min = hist.min()
+            if global_max is None or hist.max() > global_max:
+                global_max = hist.max()
+            X, Y = np.meshgrid(xedges, yedges)
+            reg_plots.append((layer, neuron, ax, (X, Y, hist), dict(cmap="inferno", linewidth=0, rasterized=True, **extra_opts)))
+    logger.info("Done")
+
+    logger.info("global_min: {}".format(global_min))
+    logger.info("global_max: {}".format(global_max))
+
+    if global_min <= 0 and logz:
+        global_min = log_default_ymin
+        logger.info("Changing global_min to {}".format(log_default_ymin))
+
+    for layer, neuron, ax, args, kwargs in reg_plots:
+        if zrange is None:
+            kwargs["norm"].vmin = global_min
+            kwargs["norm"].vmax = global_max
+        if not global_norm:
+            kwargs["norm"] = None
+        im = ax.pcolormesh(*args, **kwargs)
+        ax.set_facecolor("black")
+        if varx_label is not None:
+            ax.set_xlabel(varx_label)
+        if vary_label is not None:
+            ax.set_ylabel(vary_label)
+        ax.text(0., 0.5, "{}, {}".format(layer, neuron), transform=ax.transAxes, color="white")
+
+    cb = fig.colorbar(im, cax=grid[0].cax, orientation="horizontal")
+    cb.ax.xaxis.set_ticks_position('top')
+    cb.ax.xaxis.set_label_position('top')
+
+    logger.info("Rendering")
+    save_show(plt, fig, plotname, bbox_inches='tight')
+    logger.info("Done")
+
+
 def plot_hist_2D(plotname, xedges, yedges, hist, varx_label=None, vary_label=None, log=False, zlabel="# of events"):
     X, Y = np.meshgrid(xedges, yedges)
 

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)
+c._load_data()
+
+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])

scripts/write_parametrized.py

+#!/usr/bin/env python
+"""
+Write new TTrees with signal parameters as branches. For the
+backgrounds the parameters are generated following the total
+distribution for all signals. The discrete values for the whole ntuple
+of signal parameters are counted, such that correlations between
+signal parameters are taken into account.
+"""
+
+import argparse, re, os
+
+import ROOT
+
+from root_numpy import list_trees
+from root_pandas import read_root
+import numpy as np
+
+if __name__ == "__main__":
+
+    input_filename = "/project/etp4/nhartmann/trees/allTrees_m1.8_NoSys.root"
+    output_filename = "/project/etp4/nhartmann/trees/allTrees_m1.8_NoSys_parametrized.root"
+
+    param_names = ["mg", "mc", "mn"]
+
+    param_match = "GG_oneStep_(.*?)_(.*?)_(.*?)_NoSys"
+
+    output_signal_treename = "GG_oneStep_NoSys"
+
+    bkg_trees = [
+        "diboson_Sherpa221_NoSys",
+        "singletop_NoSys",
+        "ttbar_NoSys",
+        "ttv_NoSys",
+        "wjets_Sherpa221_NoSys",
+        "zjets_Sherpa221_NoSys",
+    ]
+
+    # read in the number of events for each combination of parameters
+    f = ROOT.TFile.Open(input_filename)
+    count_dict = {}
+    for key in f.GetListOfKeys():
+        tree_name = key.GetName()
+        match = re.match(param_match, tree_name)
+        if match is not None:
+            tree = f.Get(tree_name)
+            params = tuple([float(i) for i in match.groups()])
+            if not params in count_dict:
+                count_dict[params] = 0
+            # TODO: might be better to use sum of weights
+            count_dict[params] += tree.GetEntries()
+    f.Close()
+
+    # calculate cumulative sum of counts to sample signal parameters for background from
+    numbers = np.array(count_dict.keys(), dtype=np.float)
+    counts = np.array(count_dict.values(), dtype=np.float)
+    probs = counts/counts.sum()
+    prob_bins = np.cumsum(probs)
+
+    # read and write the rest in chunks
+    if os.path.exists(output_filename):
+        os.remove(output_filename)
+    for tree_name in list_trees(input_filename):
+        match_signal = re.match(param_match, tree_name)
+        if match_signal is not None or tree_name in bkg_trees:
+            print("Writing {}".format(tree_name))
+            nwritten = 0
+            for df in read_root(input_filename, tree_name, chunksize=100000):
+                print("Writing event {}".format(nwritten))
+                if match_signal is None:
+                    rnd = np.random.random(len(df))
+                    rnd_idx = np.digitize(rnd, prob_bins)
+                    param_values = numbers[rnd_idx]
+                    for param_idx, param_name in enumerate(param_names):
+                        df[param_name] = param_values[:,param_idx]
+                    df["training_weight"] = df["eventWeight"]*df["genWeight"]
+                else:
+                    for param_name, param_value in zip(param_names, match_signal.groups()):
+                        df[param_name] = float(param_value)
+                    df["training_weight"] = df["eventWeight"]
+                if match_signal is None:
+                    out_tree_name = tree_name
+                else:
+                    out_tree_name = output_signal_treename
+                df.to_root(output_filename, mode="a", key=out_tree_name)
+                nwritten += len(df)

test/test_toolkit.py

+import pytest
+import numpy as np
+import root_numpy
+import pandas as pd
+from sklearn.datasets import make_classification
+from keras.layers import GRU
+
+from KerasROOTClassification import ClassificationProject, ClassificationProjectRNN
+
+
+def create_dataset(path):
+
+    # create example dataset with (low-weighted) noise added
+    X, y = make_classification(n_samples=10000, random_state=1)
+    X2 = np.random.normal(size=20*10000).reshape(-1, 20)
+    y2 = np.concatenate([np.zeros(5000), np.ones(5000)])
+    X = np.concatenate([X, X2])
+    y = np.concatenate([y, y2])
+    w = np.concatenate([np.ones(10000), 0.01*np.ones(10000)])
+
+    # shift and scale randomly (to check if transformation is working)
+    shift = np.random.rand(20)*100
+    scale = np.random.rand(20)*1000
+    X *= scale
+    X += shift
+
+    # write to root files
+    branches = ["var_{}".format(i) for i in range(len(X[0]))]
+    df = pd.DataFrame(X, columns=branches)
+    df["class"] = y
+    df["weight"] = w
+    tree_path_bkg = str(path / "bkg.root")
+    tree_path_sig = str(path / "sig.root")
+    root_numpy.array2root(df[df["class"]==0].to_records(), tree_path_bkg)
+    root_numpy.array2root(df[df["class"]==1].to_records(), tree_path_sig)
+    return branches, tree_path_sig, tree_path_bkg
+
+
+def test_ClassificationProject(tmp_path):
+    branches, tree_path_sig, tree_path_bkg = create_dataset(tmp_path)
+    c = ClassificationProject(
+        str(tmp_path / "project"),
+        bkg_trees = [(tree_path_bkg, "tree")],
+        signal_trees = [(tree_path_sig, "tree")],
+        branches = branches,
+        weight_expr = "weight",
+        identifiers = ["index"],
+        optimizer="Adam",
+        earlystopping_opts=dict(patience=5),
+        dropout=0.5,
+        layers=3,
+        nodes=128,
+    )
+
+    c.train(epochs=200)
+    c.plot_all_inputs()
+    c.plot_loss()
+    assert min(c.history.history["val_loss"]) < 0.18
+
+
+def test_ClassificationProjectRNN(tmp_path):
+    branches, tree_path_sig, tree_path_bkg = create_dataset(tmp_path)
+    c = ClassificationProjectRNN(
+        str(tmp_path / "project"),
+        bkg_trees = [(tree_path_bkg, "tree")],
+        signal_trees = [(tree_path_sig, "tree")],
+        branches = branches,
+        recurrent_field_names=[
+            [
+                ["var_1", "var_2", "var_3"],
+                ["var_4", "var_5", "var_6"]
+            ],
+            [
+                ["var_10", "var_11", "var_12"],
+                ["var_13", "var_14", "var_15"]
+            ],
+        ],
+        weight_expr = "weight",
+        identifiers = ["index"],
+        optimizer="Adam",
+        earlystopping_opts=dict(patience=5),
+        dropout=0.5,
+        layers=3,
+        nodes=128,
+    )
+    assert sum([isinstance(layer, GRU) for layer in c.model.layers]) == 2
+    c.train(epochs=200)
+    c.plot_all_inputs()
+    c.plot_loss()
+    assert min(c.history.history["val_loss"]) < 0.18

utils.py

@@ -197,14 +197,26 @@ def weighted_quantile(values, quantiles, sample_weight=None, values_sorted=False
 
 class WeightedRobustScaler(RobustScaler):
 
-    def fit(self, X, y=None, weights=None):
-        if not np.isnan(X).any():
+    def fit(self, X, y=None, weights=None, mask_value=None):
+        if not np.isnan(X).any() and mask_value is not None and weights is None:
             # these checks don't work for nan values
-            super(WeightedRobustScaler, self).fit(X, y)
-        if weights is None:
-            return self
+            return super(WeightedRobustScaler, self).fit(X, y)
         else:
-            wqs = np.array([weighted_quantile(X[:,i][~np.isnan(X[:,i])], [0.25, 0.5, 0.75], sample_weight=weights) for i in range(X.shape[1])])
+            if weights is None:
+                weights = np.ones(len(self.X))
+            wqs = []
+            for i in range(X.shape[1]):
+                mask = ~np.isnan(X[:,i])
+                if mask_value is not None:
+                    mask &= (X[:,i] != mask_value)
+                wqs.append(
+                    weighted_quantile(
+                        X[:,i][mask],
+                        [0.25, 0.5, 0.75],
+                        sample_weight=weights[mask]
+                    )
+                )
+            wqs = np.array(wqs)
             self.center_ = wqs[:,1]
             self.scale_ = wqs[:,2]-wqs[:,0]
             self.scale_ = _handle_zeros_in_scale(self.scale_, copy=False)