norm = BatchNormalization(momentum=0.6, name='lstmdensenorm')(dense)
for i in xrange(5):
dense = Dense(50, activation='relu', name='dense%i' % i)(norm)
norm = BatchNormalization(momentum=0.6, name='densenorm%i' % i)(dense)
output_p = Dense(config.n_truth, activation='softmax')(norm)
# model = Model(inputs=[input_charged, input_inclusive, input_sv], outputs=[output_p, output_b])
model = Model(inputs=input_inclusive, outputs=output_p)
model.compile(optimizer=Adam(lr=0.0005),
loss='categorical_crossentropy',
metrics=['accuracy'])
print model.summary()
train_generator = obj.generatePF(data,
partition='train',
batch=100,
mask=False)
validation_generator = obj.generatePF(data,
partition='validate',
batch=100,
mask=False)
test_generator = obj.generatePF(data,
partition='validate',
batch=1000,
mask=False)
test_i, test_o, test_w = next(test_generator)
pred = model.predict(test_i)
print test_o[:5]
print pred[:5]
print test_o[-5:]
print pred[-5:]
'tau32' : utils.NH1(np.arange(0,1.1,0.05)),
'dnn' : utils.NH1(np.arange(0,1.1,0.05))
}
hists_qcd = {
'tau32' : utils.NH1(np.arange(0,1.1,0.05)),
'dnn' : utils.NH1(np.arange(0,1.1,0.05))
}
hists_top['tau32'] = top_4.draw_singletons([('tau32', hists_top['tau32'].bins)], partition='test')['tau32']
hists_qcd['tau32'] = qcd_0.draw_singletons([('tau32', hists_qcd['tau32'].bins)], partition='test')['tau32']
top_4.refresh(partitions=['test'])
qcd_0.refresh(partitions=['test'])
# test_generator = obj.generatePFSV([top_4], partition='test', batch=100)
test_generator = obj.generatePF(data, partition='test', batch=10000, repartition=False)
while True:
try:
i, o, w = next(test_generator)
pred = model.predict(i)[:,4]
o = np.array(o)
mask_signal = (o[:,4] == 1)
mask_background = np.logical_not(mask_signal)
hists_top['dnn'].fill_array(pred[mask_signal], w[mask_signal])
hists_qcd['dnn'].fill_array(pred[mask_background], w[mask_background])
except StopIteration:
break
OUTPUT = '/home/snarayan/public_html/figs/testplots/test/'
system('mkdir -p '+OUTPUT)
dims = data[0].objects['train']['pf'].data.data.shape
else:
dims = (None, obj.limit, 9) # override
'''
first build the classifier!
'''
# set up data
opts = {
'learn_mass': LEARNMASS,
'learn_pt': LEARNPT,
'learn_rho': LEARNRHO,
'normalize': False
}
classifier_train_gen = obj.generatePF(data,
partition='train',
batch=501,
**opts)
classifier_validation_gen = obj.generatePF(data,
partition='validate',
batch=1001,
**opts)
classifier_test_gen = obj.generatePF(data, partition='test', batch=2, **opts)
test_i, test_o, test_w = next(classifier_test_gen)
#print test_i
inputs = Input(shape=(dims[1], dims[2]), name='input')
mass_inputs = Input(shape=(1, ), name='mass_input')
rho_inputs = Input(shape=(1, ), name='rho_input')
pt_inputs = Input(shape=(1, ), name='pt_input')
norm = BatchNormalization(momentum=0.6, name='input_bnorm')(inputs)
conv = Conv1D(32,
if obj.limit is None:
data[0].objects['train']['pf'].load(memory=False)
dims = data[0].objects['train']['pf'].data.data.shape
else:
dims = (None, obj.limit, 9) # override
'''
first build the classifier!
'''
# set up data
opts = {'learn_mass':LEARNMASS,
'learn_pt':LEARNPT,
'learn_rho':LEARNRHO,
'normalize':False}
classifier_train_gen = obj.generatePF(data, partition='train', batch=502, **opts)
classifier_validation_gen = obj.generatePF(data, partition='validate', batch=1002, **opts)
classifier_test_gen = obj.generatePF(data, partition='test', batch=2, **opts)
test_i, test_o, test_w = next(classifier_test_gen)
#print test_i
inputs = Input(shape=(dims[1], dims[2]), name='input')
mass_inputs = Input(shape=(1,), name='mass_input')
rho_inputs = Input(shape=(1,), name='rho_input')
pt_inputs = Input(shape=(1,), name='pt_input')
norm = BatchNormalization(momentum=0.6, name='input_bnorm') (inputs)
conv = Conv1D(32, 2, activation='relu', name='conv0', kernel_initializer='lecun_uniform', padding='same')(norm)
norm = BatchNormalization(momentum=0.6, name='conv0_bnorm') (conv)
conv = Conv1D(16, 4, activation='relu', name='conv1', kernel_initializer='lecun_uniform', padding='same')(norm)
norm = BatchNormalization(momentum=0.6, name='conv1_bnorm') (conv)
lstm = LSTM(100, go_backwards=True, implementation=2, name='lstm') (norm)
# preload some data just to get the dimensions
data[0].objects['train']['pf'].load(memory=False)
dims = data[0].objects['train']['pf'].data.data.shape
# obj.limit = 20
# dims = (None, 20, 9) # override
'''
some global definitions
'''
'''
first build the classifier!
'''
# set up data
classifier_train_gen = obj.generatePF(data,
partition='train',
batch=1000,
normalize=False)
classifier_validation_gen = obj.generatePF(data,
partition='validate',
batch=100)
classifier_test_gen = obj.generatePF(data, partition='validate', batch=1000)
test_i, test_o, test_w = next(classifier_test_gen)
inputs = Input(shape=(dims[1], dims[2]), name='input')
norm = BatchNormalization(momentum=0.6, name='input_bnorm')(inputs)
lstm = LSTM(100, go_backwards=True, implementation=2, name='lstm')(norm)
norm = BatchNormalization(momentum=0.6, name='lstm_norm')(lstm)
dense = Dense(100,
activation='relu',
name='lstmdense',
kernel_initializer='lecun_uniform')(norm)