In [171]:

    

import pandas as pd
import pandautils as pu
import numpy as np
import matplotlib.pyplot as plt
%matplotlib inline
import ROOT
ROOT.gSystem.Load('./linkdef_C.so') # in root: .L linkdef.C+


    

    
Out[171]:





1

In [3]:

    

input_vars = ['jet_pt', 'jet_eta', 'jet_phi', 'lep_pt', 'lep_eta', 'lep_phi', 'metx', 'mety', 'ME_Bkg_Likelihoods_Available',
             'ME_Sig_Likelihoods_Available', 'ME_Bkg_Mod_Likelihoods', 'ME_Sig_Mod_Likelihoods']


    

In [4]:

    

bkg_df = pu.root2panda('../samples/background.root', 
                       'Muon/6jetin252525252525_0elecex25topcommonetacommon_1muonex25topcommonetacommon/0btagin0.7892MV1/atree', 
                       branches = input_vars )


    

In [38]:

    

signal_df = pu.root2panda('../samples/signal.root', 
                          'Muon/6jetin252525252525_0elecex25topcommonetacommon_1muonex25topcommonetacommon/0btagin0.7892MV1/atree',
                         branches = input_vars)


    

In [6]:

    

background_file = ROOT.TFile('../samples/background.root')


    

In [28]:

    

signal_file = ROOT.TFile('../samples/signal.root')


    

In [7]:

    

bkgT = background_file.Get('Muon/6jetin252525252525_0elecex25topcommonetacommon_1muonex25topcommonetacommon/0btagin0.7892MV1/atree')


    

In [30]:

    

signalT = signal_file.Get('Muon/6jetin252525252525_0elecex25topcommonetacommon_1muonex25topcommonetacommon/0btagin0.7892MV1/atree')


    

In [8]:

    

bkg_perm = []
Nentries = bkgT.GetEntries()
for i in xrange(Nentries):
    bkgT.GetEntry(i)
    if i % 10000 == 0:
        print 'Event {} of {}'.format(i, Nentries)
    sp = []
    for j in xrange(bkgT.ME_Sig_Mod_Permutations[0][0].size()):
        sp.append(bkgT.ME_Sig_Mod_Permutations[0][0][j])
    bkg_perm.append(sp)


    

    




Event 0
Event 10000
Event 20000
Event 30000
Event 40000
Event 50000
Event 60000
Event 70000
Event 80000
Event 90000
Event 100000
Event 110000
Event 120000
Event 130000
Event 140000
Event 150000
Event 160000
Event 170000
Event 180000
Event 190000
Event 200000
Event 210000
Event 220000
Event 230000
Event 240000
Event 250000
Event 260000
Event 270000
Event 280000
Event 290000
Event 300000
Event 310000
Event 320000
Event 330000
Event 340000
Event 350000
Event 360000
Event 370000
Event 380000
Event 390000
Event 400000
Event 410000
Event 420000
Event 430000
Event 440000
Event 450000
Event 460000
Event 470000
Event 480000
Event 490000
Event 500000
Event 510000
Event 520000
Event 530000
Event 540000
Event 550000
Event 560000
Event 570000
Event 580000
Event 590000
Event 600000
Event 610000
Event 620000
Event 630000
Event 640000
Event 650000
Event 660000
Event 670000
Event 680000
Event 690000
Event 700000
Event 710000
Event 720000
Event 730000
Event 740000
Event 750000
Event 760000
Event 770000
Event 780000
Event 790000
Event 800000
Event 810000
Event 820000
Event 830000
Event 840000
Event 850000
Event 860000
Event 870000
Event 880000
Event 890000
Event 900000

In [31]:

    

signal_perm = []
Nentries = signalT.GetEntries()
for i in xrange(Nentries):
    signalT.GetEntry(i)
    if i % 10000 == 0:
        print 'Event {} of {}'.format(i, Nentries)
    sp = []
    for j in xrange(signalT.ME_Sig_Mod_Permutations[0][0].size()):
        sp.append(signalT.ME_Sig_Mod_Permutations[0][0][j])
    signal_perm.append(sp)


    

    




Event 0 of 113057
Event 10000 of 113057
Event 20000 of 113057
Event 30000 of 113057
Event 40000 of 113057
Event 50000 of 113057
Event 60000 of 113057
Event 70000 of 113057
Event 80000 of 113057
Event 90000 of 113057
Event 100000 of 113057
Event 110000 of 113057

In [40]:

    

bkg_df['jet_order'] = bkg_perm;
signal_df['jet_order'] = signal_perm;


    

In [64]:

    

bkg_df['metx/mety'] = bkg_df['metx']/bkg_df['mety'];
signal_df['metx/mety'] = signal_df['metx']/signal_df['mety'];


    

In [107]:

    

bkg_df['truth'] = 0;
signal_df['truth'] = 1;


    

In [108]:

    

# -- save to pickle, if you want
bkg_df.to_pickle('./bkg_df.pkl')
signal_df.to_pickle('./signal_df.pkl')
# there should be a better way of doing this using cPickle but I forget how to do it


    

In [109]:

    

# "Ask for ME_Bkg_Likelihoods_Available[0] == true && ME_Sig_Likelihoods_Available[0] == true"
bkg_df1 = bkg_df[(bkg_df['ME_Bkg_Likelihoods_Available']==1) & (bkg_df['ME_Sig_Likelihoods_Available']==1)];
signal_df1 = signal_df[(signal_df['ME_Bkg_Likelihoods_Available']==1) & (signal_df['ME_Sig_Likelihoods_Available']==1)];


    

In [208]:

    

tot_df = pd.concat([bkg_df1, signal_df1], ignore_index=True)


    

In [209]:

    

y = np.array(tot_df['truth'])


    

In [210]:

    

N_COL = 22 # number of features
X = np.zeros((bkg_df1.shape[0] + signal_df1.shape[0], N_COL))


    

In [ ]:

    

def pseudo(f, x):
    return np.sign(x) * f(np.abs(x))


    

In [211]:

    

X[:, 0] = tot_df['lep_pt'];
X[:, 0] = np.log(X[:, 0])
X[:, 1] = tot_df['lep_eta'];
X[:, 2] = tot_df['lep_phi'];
X[:, 3] = pseudo(np.log, tot_df['metx/mety']);
# X[:, 4] = pseudo(np.log, tot_df['metx']);
# X[:, 5] = pseudo(np.log, tot_df['mety']);


    

In [212]:

    

PT_COLUMN = 4
N_JETS = 6

ETA_COLUMN = PT_COLUMN + N_JETS
PHI_COLUMN = ETA_COLUMN + N_JETS
        
for event in xrange(tot_df.shape[0]):
    if event % 10000 == 0:
        print 'Event {} of {}'.format(event, tot_df.shape[0])
    reordering = np.array(tot_df['jet_order'][event][:-2]).astype('int') # -- -2 chops off the zero shit
    X[event, PT_COLUMN : (PT_COLUMN + N_JETS)] = np.log(tot_df['jet_pt'][event][reordering])
    X[event, ETA_COLUMN : (ETA_COLUMN + N_JETS)] = tot_df['jet_eta'][event][reordering]
    X[event, PHI_COLUMN : (PHI_COLUMN + N_JETS)] = tot_df['jet_phi'][event][reordering]


    

    




Event 0 of 1016336
Event 10000 of 1016336
Event 20000 of 1016336
Event 30000 of 1016336
Event 40000 of 1016336
Event 50000 of 1016336
Event 60000 of 1016336
Event 70000 of 1016336
Event 80000 of 1016336
Event 90000 of 1016336
Event 100000 of 1016336
Event 110000 of 1016336
Event 120000 of 1016336
Event 130000 of 1016336
Event 140000 of 1016336
Event 150000 of 1016336
Event 160000 of 1016336
Event 170000 of 1016336
Event 180000 of 1016336
Event 190000 of 1016336
Event 200000 of 1016336
Event 210000 of 1016336
Event 220000 of 1016336
Event 230000 of 1016336
Event 240000 of 1016336
Event 250000 of 1016336
Event 260000 of 1016336
Event 270000 of 1016336
Event 280000 of 1016336
Event 290000 of 1016336
Event 300000 of 1016336
Event 310000 of 1016336
Event 320000 of 1016336
Event 330000 of 1016336
Event 340000 of 1016336
Event 350000 of 1016336
Event 360000 of 1016336
Event 370000 of 1016336
Event 380000 of 1016336
Event 390000 of 1016336
Event 400000 of 1016336
Event 410000 of 1016336
Event 420000 of 1016336
Event 430000 of 1016336
Event 440000 of 1016336
Event 450000 of 1016336
Event 460000 of 1016336
Event 470000 of 1016336
Event 480000 of 1016336
Event 490000 of 1016336
Event 500000 of 1016336
Event 510000 of 1016336
Event 520000 of 1016336
Event 530000 of 1016336
Event 540000 of 1016336
Event 550000 of 1016336
Event 560000 of 1016336
Event 570000 of 1016336
Event 580000 of 1016336
Event 590000 of 1016336
Event 600000 of 1016336
Event 610000 of 1016336
Event 620000 of 1016336
Event 630000 of 1016336
Event 640000 of 1016336
Event 650000 of 1016336
Event 660000 of 1016336
Event 670000 of 1016336
Event 680000 of 1016336
Event 690000 of 1016336
Event 700000 of 1016336
Event 710000 of 1016336
Event 720000 of 1016336
Event 730000 of 1016336
Event 740000 of 1016336
Event 750000 of 1016336
Event 760000 of 1016336
Event 770000 of 1016336
Event 780000 of 1016336
Event 790000 of 1016336
Event 800000 of 1016336
Event 810000 of 1016336
Event 820000 of 1016336
Event 830000 of 1016336
Event 840000 of 1016336
Event 850000 of 1016336
Event 860000 of 1016336
Event 870000 of 1016336
Event 880000 of 1016336
Event 890000 of 1016336
Event 900000 of 1016336
Event 910000 of 1016336
Event 920000 of 1016336
Event 930000 of 1016336
Event 940000 of 1016336
Event 950000 of 1016336
Event 960000 of 1016336
Event 970000 of 1016336
Event 980000 of 1016336
Event 990000 of 1016336
Event 1000000 of 1016336
Event 1010000 of 1016336

In [213]:

    

# -- get number of available examples
ix = range(X.shape[0])
# -- shuffle the indices to shuffle X and y
np.random.shuffle(ix)
X, y = X[ix], y[ix]
# -- divide the sample in half for training and testing
n = int(0.6*X.shape[0] )


    

In [214]:

    

from sklearn.preprocessing import StandardScaler
scaler = StandardScaler()
Z = scaler.fit_transform(np.array(X))


    

In [215]:

    

for col in Z.T:
    bins = np.linspace(col.min(), col.max(), 50)
    plt.hist(col[y==1], bins=bins,color = 'red', histtype='step', normed=True, label = 'Signal')
    plt.hist(col[y==0], bins=bins,color = 'blue', histtype='step', normed=True, label = 'Background')
    plt.legend()
    plt.show()


    

In [232]:

    

# Keras NN:
from keras.models import Sequential
from keras.layers.core import Dense, MaxoutDense, Dropout, Activation
from keras.callbacks import EarlyStopping, ModelCheckpoint

net = Sequential()

net.add(MaxoutDense(64, 10, input_shape=(22, ), init='he_uniform'))
net.add(Dropout(0.6))

net.add(MaxoutDense(32, 10, init='he_uniform'))
net.add(Dropout(0.3))

net.add(Dense(32))#, init='he_uniform'))
net.add(Activation('relu'))
net.add(Dropout(0.2))

net.add(Dense(16))#, init='he_uniform'))
net.add(Activation('relu'))
net.add(Dropout(0.2))

net.add(Dense(8))#, init='he_uniform'))
net.add(Activation('relu'))
# net.add(Dropout(0.1))

net.add(Dense(1))#, init='he_uniform'))
net.add(Activation('sigmoid'))
#MD(64)-DO(0.6)-MD(32)-DO(0.3)-D(16)-relu-DO(0.2)-D(8)-relu-DO(0.1)-D(1)-sigmoid

net.compile('adam', 'binary_crossentropy', class_mode='binary')


    

In [234]:

    

weights = np.ones(Z.shape[0])
weights[y == 1] = 1 / (y.mean())


    

In [235]:

    

net.fit(Z[:n], y[:n], verbose=True, batch_size=128, show_accuracy=True, sample_weight=weights[:n], 
        callbacks = [
            EarlyStopping(verbose=True, patience=10, monitor='val_loss'),
            ModelCheckpoint('mlmem_net.h5', 
                            monitor='val_loss', verbose=True, save_best_only=True)
        ],
        nb_epoch=70, validation_data=(Z[n:], y[n:], weights[n:]))


    

    




Train on 609801 samples, validate on 406535 samples
Epoch 1/70
609801/609801 [==============================] - 43s - loss: 0.6746 - acc: 0.4928 - val_loss: 0.6590 - val_acc: 0.5111
Epoch 00000: val_loss improved from inf to 0.65902, saving model to mlmem_net.h5
Epoch 2/70
609801/609801 [==============================] - 44s - loss: 0.6601 - acc: 0.5568 - val_loss: 0.6560 - val_acc: 0.5064
Epoch 00001: val_loss improved from 0.65902 to 0.65604, saving model to mlmem_net.h5
Epoch 3/70
609801/609801 [==============================] - 46s - loss: 0.6567 - acc: 0.5714 - val_loss: 0.6557 - val_acc: 0.5015
Epoch 00002: val_loss improved from 0.65604 to 0.65569, saving model to mlmem_net.h5
Epoch 4/70
609801/609801 [==============================] - 46s - loss: 0.6540 - acc: 0.5855 - val_loss: 0.6505 - val_acc: 0.5692
Epoch 00003: val_loss improved from 0.65569 to 0.65048, saving model to mlmem_net.h5
Epoch 5/70
609801/609801 [==============================] - 50s - loss: 0.6531 - acc: 0.5967 - val_loss: 0.6509 - val_acc: 0.5461
Epoch 00004: val_loss did not improve
Epoch 6/70
609801/609801 [==============================] - 52s - loss: 0.6522 - acc: 0.6021 - val_loss: 0.6493 - val_acc: 0.6426
Epoch 00005: val_loss improved from 0.65048 to 0.64932, saving model to mlmem_net.h5
Epoch 7/70
609801/609801 [==============================] - 52s - loss: 0.6513 - acc: 0.6051 - val_loss: 0.6556 - val_acc: 0.5731
Epoch 00006: val_loss did not improve
Epoch 8/70
609801/609801 [==============================] - 53s - loss: 0.6511 - acc: 0.6134 - val_loss: 0.6502 - val_acc: 0.6184
Epoch 00007: val_loss did not improve
Epoch 9/70
609801/609801 [==============================] - 53s - loss: 0.6505 - acc: 0.5970 - val_loss: 0.6536 - val_acc: 0.6138
Epoch 00008: val_loss did not improve
Epoch 10/70
609801/609801 [==============================] - 53s - loss: 0.6494 - acc: 0.5899 - val_loss: 0.6500 - val_acc: 0.6031
Epoch 00009: val_loss did not improve
Epoch 11/70
609801/609801 [==============================] - 53s - loss: 0.6492 - acc: 0.5997 - val_loss: 0.6501 - val_acc: 0.5761
Epoch 00010: val_loss did not improve
Epoch 12/70
609801/609801 [==============================] - 53s - loss: 0.6492 - acc: 0.5846 - val_loss: 0.6489 - val_acc: 0.5153
Epoch 00011: val_loss improved from 0.64932 to 0.64890, saving model to mlmem_net.h5
Epoch 13/70
609801/609801 [==============================] - 53s - loss: 0.6491 - acc: 0.5815 - val_loss: 0.6513 - val_acc: 0.5497
Epoch 00012: val_loss did not improve
Epoch 14/70
609801/609801 [==============================] - 52s - loss: 0.6487 - acc: 0.5886 - val_loss: 0.6477 - val_acc: 0.5565
Epoch 00013: val_loss improved from 0.64890 to 0.64766, saving model to mlmem_net.h5
Epoch 15/70
609801/609801 [==============================] - 51s - loss: 0.6479 - acc: 0.5769 - val_loss: 0.6488 - val_acc: 0.5634
Epoch 00014: val_loss did not improve
Epoch 16/70
609801/609801 [==============================] - 52s - loss: 0.6481 - acc: 0.5747 - val_loss: 0.6484 - val_acc: 0.5818
Epoch 00015: val_loss did not improve
Epoch 17/70
609801/609801 [==============================] - 51s - loss: 0.6482 - acc: 0.5714 - val_loss: 0.6508 - val_acc: 0.5428
Epoch 00016: val_loss did not improve
Epoch 18/70
609801/609801 [==============================] - 49s - loss: 0.6474 - acc: 0.5661 - val_loss: 0.6470 - val_acc: 0.5687
Epoch 00017: val_loss improved from 0.64766 to 0.64704, saving model to mlmem_net.h5
Epoch 19/70
609801/609801 [==============================] - 49s - loss: 0.6474 - acc: 0.5820 - val_loss: 0.6491 - val_acc: 0.5488
Epoch 00018: val_loss did not improve
Epoch 20/70
609801/609801 [==============================] - 49s - loss: 0.6471 - acc: 0.5758 - val_loss: 0.6496 - val_acc: 0.5040
Epoch 00019: val_loss did not improve
Epoch 21/70
609801/609801 [==============================] - 51s - loss: 0.6470 - acc: 0.5676 - val_loss: 0.6495 - val_acc: 0.5241
Epoch 00020: val_loss did not improve
Epoch 22/70
609801/609801 [==============================] - 51s - loss: 0.6470 - acc: 0.5666 - val_loss: 0.6493 - val_acc: 0.5124
Epoch 00021: val_loss did not improve
Epoch 23/70
609801/609801 [==============================] - 50s - loss: 0.6464 - acc: 0.5682 - val_loss: 0.6463 - val_acc: 0.5398
Epoch 00022: val_loss improved from 0.64704 to 0.64633, saving model to mlmem_net.h5
Epoch 24/70
609801/609801 [==============================] - 50s - loss: 0.6463 - acc: 0.5711 - val_loss: 0.6479 - val_acc: 0.5445
Epoch 00023: val_loss did not improve
Epoch 25/70
609801/609801 [==============================] - 51s - loss: 0.6462 - acc: 0.5646 - val_loss: 0.6473 - val_acc: 0.5317
Epoch 00024: val_loss did not improve
Epoch 26/70
609801/609801 [==============================] - 51s - loss: 0.6459 - acc: 0.5761 - val_loss: 0.6476 - val_acc: 0.5632
Epoch 00025: val_loss did not improve
Epoch 27/70
609801/609801 [==============================] - 52s - loss: 0.6459 - acc: 0.5665 - val_loss: 0.6456 - val_acc: 0.5359
Epoch 00026: val_loss improved from 0.64633 to 0.64561, saving model to mlmem_net.h5
Epoch 28/70
609801/609801 [==============================] - 52s - loss: 0.6455 - acc: 0.5702 - val_loss: 0.6463 - val_acc: 0.5664
Epoch 00027: val_loss did not improve
Epoch 29/70
609801/609801 [==============================] - 50s - loss: 0.6456 - acc: 0.5689 - val_loss: 0.6461 - val_acc: 0.4939
Epoch 00028: val_loss did not improve
Epoch 30/70
609801/609801 [==============================] - 51s - loss: 0.6459 - acc: 0.5655 - val_loss: 0.6456 - val_acc: 0.5598
Epoch 00029: val_loss improved from 0.64561 to 0.64558, saving model to mlmem_net.h5
Epoch 31/70
609801/609801 [==============================] - 38s - loss: 0.6455 - acc: 0.5736 - val_loss: 0.6462 - val_acc: 0.5808
Epoch 00030: val_loss did not improve
Epoch 32/70
609801/609801 [==============================] - 38s - loss: 0.6454 - acc: 0.5639 - val_loss: 0.6460 - val_acc: 0.5648
Epoch 00031: val_loss did not improve
Epoch 33/70
609801/609801 [==============================] - 38s - loss: 0.6452 - acc: 0.5745 - val_loss: 0.6454 - val_acc: 0.5457
Epoch 00032: val_loss improved from 0.64558 to 0.64543, saving model to mlmem_net.h5
Epoch 34/70
609801/609801 [==============================] - 38s - loss: 0.6452 - acc: 0.5658 - val_loss: 0.6454 - val_acc: 0.5652
Epoch 00033: val_loss did not improve
Epoch 35/70
609801/609801 [==============================] - 38s - loss: 0.6450 - acc: 0.5633 - val_loss: 0.6461 - val_acc: 0.5803
Epoch 00034: val_loss did not improve
Epoch 36/70
609801/609801 [==============================] - 38s - loss: 0.6448 - acc: 0.5667 - val_loss: 0.6484 - val_acc: 0.5308
Epoch 00035: val_loss did not improve
Epoch 37/70
609801/609801 [==============================] - 38s - loss: 0.6450 - acc: 0.5668 - val_loss: 0.6455 - val_acc: 0.5379
Epoch 00036: val_loss did not improve
Epoch 38/70
609801/609801 [==============================] - 38s - loss: 0.6451 - acc: 0.5696 - val_loss: 0.6470 - val_acc: 0.5338
Epoch 00037: val_loss did not improve
Epoch 39/70
609801/609801 [==============================] - 37s - loss: 0.6447 - acc: 0.5625 - val_loss: 0.6453 - val_acc: 0.5426
Epoch 00038: val_loss improved from 0.64543 to 0.64530, saving model to mlmem_net.h5
Epoch 40/70
609801/609801 [==============================] - 36s - loss: 0.6444 - acc: 0.5635 - val_loss: 0.6464 - val_acc: 0.5613
Epoch 00039: val_loss did not improve
Epoch 41/70
609801/609801 [==============================] - 36s - loss: 0.6450 - acc: 0.5708 - val_loss: 0.6459 - val_acc: 0.5655
Epoch 00040: val_loss did not improve
Epoch 42/70
609801/609801 [==============================] - 36s - loss: 0.6443 - acc: 0.5727 - val_loss: 0.6451 - val_acc: 0.5434
Epoch 00041: val_loss improved from 0.64530 to 0.64512, saving model to mlmem_net.h5
Epoch 43/70
609801/609801 [==============================] - 36s - loss: 0.6440 - acc: 0.5683 - val_loss: 0.6449 - val_acc: 0.5480
Epoch 00042: val_loss improved from 0.64512 to 0.64489, saving model to mlmem_net.h5
Epoch 44/70
609801/609801 [==============================] - 36s - loss: 0.6444 - acc: 0.5628 - val_loss: 0.6447 - val_acc: 0.5715
Epoch 00043: val_loss improved from 0.64489 to 0.64474, saving model to mlmem_net.h5
Epoch 45/70
609801/609801 [==============================] - 36s - loss: 0.6444 - acc: 0.5642 - val_loss: 0.6475 - val_acc: 0.5212
Epoch 00044: val_loss did not improve
Epoch 46/70
609801/609801 [==============================] - 36s - loss: 0.6441 - acc: 0.5622 - val_loss: 0.6439 - val_acc: 0.5879
Epoch 00045: val_loss improved from 0.64474 to 0.64387, saving model to mlmem_net.h5
Epoch 47/70
609801/609801 [==============================] - 36s - loss: 0.6437 - acc: 0.5678 - val_loss: 0.6478 - val_acc: 0.5091
Epoch 00046: val_loss did not improve
Epoch 48/70
609801/609801 [==============================] - 36s - loss: 0.6444 - acc: 0.5676 - val_loss: 0.6460 - val_acc: 0.5603
Epoch 00047: val_loss did not improve
Epoch 49/70
609801/609801 [==============================] - 36s - loss: 0.6443 - acc: 0.5671 - val_loss: 0.6459 - val_acc: 0.5609
Epoch 00048: val_loss did not improve
Epoch 50/70
609801/609801 [==============================] - 36s - loss: 0.6438 - acc: 0.5664 - val_loss: 0.6442 - val_acc: 0.5894
Epoch 00049: val_loss did not improve
Epoch 51/70
609801/609801 [==============================] - 36s - loss: 0.6437 - acc: 0.5723 - val_loss: 0.6448 - val_acc: 0.5727
Epoch 00050: val_loss did not improve
Epoch 52/70
609801/609801 [==============================] - 36s - loss: 0.6435 - acc: 0.5815 - val_loss: 0.6471 - val_acc: 0.5587
Epoch 00051: val_loss did not improve
Epoch 53/70
609801/609801 [==============================] - 37s - loss: 0.6437 - acc: 0.5734 - val_loss: 0.6463 - val_acc: 0.5638
Epoch 00052: val_loss did not improve
Epoch 54/70
609801/609801 [==============================] - 36s - loss: 0.6433 - acc: 0.5683 - val_loss: 0.6472 - val_acc: 0.5392
Epoch 00053: val_loss did not improve
Epoch 55/70
609801/609801 [==============================] - 36s - loss: 0.6432 - acc: 0.5623 - val_loss: 0.6460 - val_acc: 0.5738
Epoch 00054: val_loss did not improve
Epoch 56/70
609801/609801 [==============================] - 36s - loss: 0.6432 - acc: 0.5680 - val_loss: 0.6460 - val_acc: 0.5160
Epoch 00055: val_loss did not improve
Epoch 57/70
609801/609801 [==============================] - 36s - loss: 0.6434 - acc: 0.5655 - val_loss: 0.6465 - val_acc: 0.5584
Epoch 00056: early stopping
Epoch 00056: val_loss did not improve






    




/usr/local/lib/python2.7/site-packages/ipykernel/__main__.py:1: DeprecationWarning: using a non-integer number instead of an integer will result in an error in the future
  if __name__ == '__main__':
/usr/local/lib/python2.7/site-packages/ipykernel/__main__.py:7: DeprecationWarning: using a non-integer number instead of an integer will result in an error in the future






    
Out[235]:





<keras.callbacks.History at 0x1763169d0>

In [238]:

    

net.load_weights('./mlmem_net.h5')


    

In [284]:

    

yhat = net.predict(Z[n:], verbose=True)


    

    




406535/406535 [==============================] - 7s     

In [285]:

    

bins = np.linspace(0, 1, 50)
plt.hist(yhat[y[n:] == 1], bins=bins,color = 'red', histtype='step', normed=True, label = 'Signal')
plt.hist(yhat[y[n:] == 0], bins=bins,color = 'blue', histtype='step', normed=True, label = 'Background')
plt.legend()
plt.show()


    

In [243]:

    

soft_target = tot_df['ME_Sig_Mod_Likelihoods'] / (tot_df['ME_Sig_Mod_Likelihoods'] + 0.23 * tot_df['ME_Bkg_Mod_Likelihoods'])


    

In [245]:

    

soft_target = soft_target[ix]


    

In [287]:

    

_ = plt.hist(pu.flatten(soft_target)[y == 1], color = 'red', histtype = 'step', bins = np.linspace(0,1,10), normed=True)
_ = plt.hist(pu.flatten(soft_target)[y == 0], color = 'blue', histtype = 'step', bins = np.linspace(0,1,10), normed=True)


    

In [258]:

    

%run viz/performance.py


    

In [281]:

    

n = int(n)
discs = {}
add_curve(r'MEM', 'black', calculate_roc(y, pu.flatten(soft_target), weights=weights), discs)
add_curve(r'Deep Network', 'red', calculate_roc(y[n:], yhat.ravel(), weights=weights[n:], bins=1000000), discs)


    

In [278]:

    

# y_pert = net.predict(Z[n:] + np.random.normal(0, 0.2, Z[n:].shape), verbose=True)


    

    




406535/406535 [==============================] - 7s     

In [279]:

    

# add_curve(r'Deep Network, perturbed', 'blue', calculate_roc(y[n:], y_pert.ravel(), weights=weights[n:], bins=1000000), discs)


    

In [282]:

    

fg = ROC_plotter(discs, title=r'MEM vs DNN', min_eff = 0.1, max_eff=0.9, logscale=False)


    

In [270]:

    

fg = ROC_plotter(discs, title=r'MEM vs DNN', min_eff = 0.3, max_eff=0.7, logscale=False)


    

In [ ]: