def minimodel(Inputs,feature_dropout=-1.):
x = Inputs[0] #this is the self.x list from the TrainData data structure
energy_raw = SelectFeatures(0,3)(x)
x = BatchNormalization(momentum=0.6)(x)
feat=[x]
for i in range(6):
#add global exchange and another dense here
x = GlobalExchange()(x)
x = Dense(64, activation='elu')(x)
x = Dense(64, activation='elu')(x)
x = BatchNormalization(momentum=0.6)(x)
x = Dense(64, activation='elu')(x)
x = GravNet_simple(n_neighbours=10,
n_dimensions=4,
n_filters=128,
n_propagate=64)(x)
x = BatchNormalization(momentum=0.6)(x)
feat.append(Dense(32, activation='elu')(x))
x = Concatenate()(feat)
x = Dense(64, activation='elu')(x)
return Model(inputs=Inputs, outputs=output_block(x,checkids(Inputs),energy_raw))
def gravnet(Inputs, nclasses, nregressions, otheroption=0):
'''
Just from the top of my head...
'''
nfilters = 48
nprop = 22
ndims = 4
x = Inputs[0] #this is the self.x list from the TrainData data structure
x = BatchNormalization(momentum=0.9)(x)
feat = []
for i in range(4):
x = GlobalExchange()(x)
x = Conv1D(64, 1, activation='elu')(x)
x = Conv1D(64, 1, activation='elu')(x)
x = Conv1D(64, 1, activation='tanh')(x)
x = GravNet(n_neighbours=40,
n_dimensions=ndims,
n_filters=nfilters,
n_propagate=nprop,
name='GravNet0_' + str(i))(x)
x = BatchNormalization(momentum=0.9)(x)
feat.append(x)
x = Concatenate()(feat)
x = Conv1D(128, 1, activation='elu', name='last_conv_a')(x)
x = Conv1D(3, 1, activation='elu', name='last_conv_b')(x)
x = Conv1D(2, 1, activation='softmax',
name="last_conv_output")(x) #that should be a 2
predictions = [x]
return Model(inputs=Inputs, outputs=predictions)
def gravnet_model(Inputs, nclasses, nregressions, feature_dropout=0.1):
x = Inputs[0] #this is the self.x list from the TrainData data structure
print('x', x.shape)
coords = []
etas = SelectFeatures(1, 2)(x) #just to propagate to the prediction
mask = CreateZeroMask(0)(x)
x = BatchNormalization(momentum=0.9)(x)
x = Multiply()([x, mask])
x, coord = GravNet(n_neighbours=40,
n_dimensions=4,
n_filters=80,
n_propagate=16,
name='gravnet_pre',
also_coordinates=True)(x)
coords.append(coord)
x = BatchNormalization(momentum=0.9)(x)
x = Multiply()([x, mask])
feats = []
for i in range(n_gravnet_layers):
x = GlobalExchange()(x)
x = Multiply()([x, mask])
x = Dense(64, activation='tanh')(x)
x = Dense(64, activation='tanh')(x)
x = BatchNormalization(momentum=0.9)(x)
x = Multiply()([x, mask])
x = Dense(64, activation='sigmoid')(x)
x = Multiply()([x, mask])
x, coord = GravNet(n_neighbours=40,
n_dimensions=4,
n_filters=80,
n_propagate=16,
name='gravnet_' + str(i),
also_coordinates=True,
feature_dropout=feature_dropout)(x)
coords.append(coord)
x = BatchNormalization(momentum=0.9)(x)
x = Multiply()([x, mask])
feats.append(x)
x = Concatenate()(feats)
x = Dense(64, activation='elu', name='pre_last_correction')(x)
x = BatchNormalization(momentum=0.9)(x)
x = Multiply()([x, mask])
x = Dense(nregressions, activation=None, kernel_initializer='zeros')(x)
#x = Clip(-0.5, 1.5) (x)
x = Multiply()([x, mask])
#x = SortPredictionByEta(input_energy_index=0, input_eta_index=1)([x,Inputs[0]])
x = Concatenate()([x] + coords + [etas])
predictions = [x]
return Model(inputs=Inputs, outputs=predictions)
def gravnet_model(Inputs, nclasses, nregressions, otheroption):
x = Inputs[0] #this is the self.x list from the TrainData data structure
print('x', x.shape)
coords = []
mask = CreateZeroMask(0)(x)
x = BatchNormalization(momentum=0.9)(x)
x = Multiply()([x, mask])
x, coord = GravNet(n_neighbours=40,
n_dimensions=4,
n_filters=80,
n_propagate=16,
name='gravnet_pre',
also_coordinates=True)(x)
coords.append(coord)
x = BatchNormalization(momentum=0.9)(x)
x = Multiply()([x, mask])
feats = []
for i in range(n_gravnet_layers):
x = GlobalExchange()(x)
x = Multiply()([x, mask])
x = Dense(64, activation='tanh')(x)
x = Dense(64, activation='tanh')(x)
x = BatchNormalization(momentum=0.9)(x)
x = Multiply()([x, mask])
x = Dense(64, activation='sigmoid')(x)
x = Multiply()([x, mask])
x, coord = GravNet(n_neighbours=40,
n_dimensions=4,
n_filters=80,
n_propagate=16,
name='gravnet_' + str(i),
also_coordinates=True)(x)
coords.append(coord)
x = BatchNormalization(momentum=0.9)(x)
x = Multiply()([x, mask])
feats.append(x)
x = Concatenate()(feats)
x = Dense(64, activation='elu', name='pre_last_correction')(x)
x = BatchNormalization(momentum=0.9)(x)
x = Multiply()([x, mask])
x = Dense(nregressions,
activation=None,
kernel_initializer=keras.initializers.RandomNormal(
mean=0.0, stddev=0.001))(x)
#x = Clip(-0.5, 1.5) (x)
x = Multiply()([x, mask])
x = Concatenate()([x] + coords)
predictions = [x]
return Model(inputs=Inputs, outputs=predictions)
def gravnet_model(Inputs, nclasses, nregressions, otheroption):
x = Inputs[0] #this is the self.x list from the TrainData data structure
print('x', x.shape)
mask = CreateZeroMask(0)(x)
x = BatchNormalization(momentum=0.6)(x)
coords = []
feats = []
for i in range(n_gravnet_layers):
x = GlobalExchange()(x)
x = Multiply()([x, mask])
x = Dense(64, activation='elu')(x)
x = Dense(64, activation='elu')(x)
x = BatchNormalization(momentum=0.3)(x)
x = GarNet(n_aggregators=8, n_filters=64, n_propagate=16)(x)
x = Dense(64, activation='tanh')(x)
x = BatchNormalization(momentum=0.3)(x)
x, coord = GravNet(n_neighbours=40,
n_dimensions=4,
n_filters=64,
n_propagate=16,
also_coordinates=True)(x)
coords.append(coord)
x = BatchNormalization(momentum=0.3)(x)
x = Multiply()([x, mask])
feats.append(x)
x = Concatenate()(feats)
x = Dense(32, activation='elu', name='pre_last_correction')(x)
x = Dense(nregressions,
activation=None,
kernel_initializer='zeros',
use_bias=False)(x) #max 1 shower here
x = Clip(-0.2, 1.2)(x)
x = Concatenate()([x] + coords)
predictions = [x]
return Model(inputs=Inputs, outputs=predictions)
def garnet_model(Inputs, nclasses, nregressions, otheroption):
nfilters = 48
x = Inputs[0] #this is the self.x list from the TrainData data structure
print('x', x.shape)
mask = CreateZeroMask(0)(x)
x = BatchNormalization(momentum=0.6)(x)
x = Multiply()([x, mask])
feats = []
for i in range(n_garnet_layers):
x = GlobalExchange()(x)
x = Multiply()([x, mask])
x = Dense(nfilters, activation='tanh')(x)
x = BatchNormalization(momentum=0.6)(x)
x = Multiply()([x, mask])
x = GarNet(
n_aggregators=8,
n_filters=nfilters,
n_propagate=8,
name='garnet_' + str(i),
)(x)
x = BatchNormalization(momentum=0.6)(x)
x = Multiply()([x, mask])
if n_garnet_layers <= 10 or i % 2 == 0:
feats.append(x)
x = Concatenate()(feats)
x = Dense(32, activation='tanh', name='pre_last_correction')(x)
x = Dense(nregressions, activation=None, name='pre_last_correction_sm')(x)
#x = SelectFeatures(1,nregressions+1)(x) #zero is the no-simcluster index
#this cannot be used for the final trainings, but to get an idea of what is happening this might be useful
#x = Clip(-0.5, 1.5) (x)
#x = Concatenate()([x]+coords)
predictions = [x]
return Model(inputs=Inputs, outputs=predictions)
def gravnet_model(Inputs, nclasses, nregressions, feature_dropout=-1.):
coords = []
feats = []
x = Inputs[0] #this is the self.x list from the TrainData data structure
x = CenterPhi(2)(x)
mask = CreateZeroMask(0)(x)
x_in = norm_and_mask(x, mask)
etas_phis = SelectFeatures(1, 3)(
x) #eta, phi, just to propagate to the prediction
r_coordinate = SelectFeatures(4, 5)(x)
energy = SelectFeatures(0, 1)(x)
x = Concatenate()([etas_phis, r_coordinate,
x]) #just for the kernel initializer
x = norm_and_mask(x, mask)
x, coord = GravNet(n_neighbours=40,
n_dimensions=3,
n_filters=80,
n_propagate=16,
name='gravnet_pre',
fix_coordinate_space=True,
also_coordinates=True,
masked_coordinate_offset=-10)([x, mask])
x = norm_and_mask(x, mask)
coords.append(coord)
feats.append(x)
for i in range(n_gravnet_layers):
x = GlobalExchange()(x)
x = Dense(64, activation='elu', name='dense_a_' + str(i))(x)
x = norm_and_mask(x, mask)
x = Dense(64, activation='elu', name='dense_b_' + str(i))(x)
#x = Concatenate()([TransformCoordinates()(x),x])
x = Dense(64, activation='elu', name='dense_c_' + str(i))(x)
x = norm_and_mask(x, mask)
x, coord = GravNet(
n_neighbours=40,
n_dimensions=4,
n_filters=80,
n_propagate=16,
name='gravnet_' + str(i),
also_coordinates=True,
feature_dropout=feature_dropout,
masked_coordinate_offset=-10)([
x, mask
]) #shift+activation makes it impossible to mix real with zero-pad
x = norm_and_mask(x, mask)
coords.append(coord)
feats.append(x)
x = Concatenate()(feats)
x = Dense(64, activation='elu', name='dense_a_last')(x)
x = Dense(64, activation='elu', name='dense_b_last')(x)
x = norm_and_mask(x, mask)
x = Dense(64, activation='elu', name='dense_c_last')(x)
x = norm_and_mask(x, mask)
n_showers = AveragePoolVertices(keepdims=True)(x)
n_showers = Dense(64, activation='elu',
name='dense_n_showers_a')(n_showers)
n_showers = Dense(1, activation=None, name='dense_n_showers')(n_showers)
x = Dense(nregressions, activation=None, name='dense_pre_fracs')(x)
x = Concatenate()([x, x_in])
x = Dense(64, activation='elu', name='dense_last_correction')(x)
x = Dense(nregressions,
activation=None,
name='dense_fracs',
kernel_initializer=keras.initializers.RandomNormal(
mean=0.0, stddev=0.01))(x)
x = Concatenate(name="concatlast",
axis=-1)([x] + coords + [n_showers] + [etas_phis])
x = Multiply()([x, mask])
predictions = [x]
return Model(inputs=Inputs, outputs=predictions)