def buildmodel(opts, trn):
noiselayer = None
micalculator = None
# Build model
inputs = Input(shape=(opts['INPUT_DIM'], ))
model_layers = []
cbs = []
HIDDEN_DIMS = map(int, opts['encoder'].split('-'))
HIDDEN_ACTS = opts['encoder_acts'].split('-')
if opts['decoder']:
DECODER_DIMS = map(int, opts['decoder'].split('-'))
else:
DECODER_DIMS = []
for hndx, hdim in enumerate(HIDDEN_DIMS):
if opts['mode'] == 'dropout':
model_layers.append(Dropout(.2 if hndx == 0 else .5))
layer_args = {}
layer_args['activation'] = HIDDEN_ACTS[hndx]
if layer_args['activation'] == 'relu':
layer_args['init'] = 'he_uniform'
else:
layer_args['init'] = 'glorot_uniform'
#if args.maxnorm is not None:
# import keras.constraints
# layer_args['W_constraint'] = keras.constraints.maxnorm(args.maxnorm)
model_layers.append(Dense(hdim, **layer_args))
if opts['mode'] in ['nlIB', 'nlIBnokde', 'vIB']:
test_phase_noise = not opts['no_test_phase_noise']
if opts['mode'] == 'nlIB' or opts['mode'] == 'nlIBnokde':
micalculator = layers.MICalculator(
opts['beta'],
model_layers,
data=trn.X,
miN=opts['miN'],
init_kde_logvar=opts['init_kde_logvar'])
if opts['mode'] != 'nlIBnokde':
cbs.append(training.KDETrain(mi_calculator=micalculator))
noiselayer = layers.NoiseLayer(
init_logvar=opts['init_noise_logvar'],
logvar_trainable=opts['noise_logvar_grad_trainable'],
test_phase_noise=test_phase_noise)
else:
micalculator = vib.MICalculatorVIB(opts['beta'])
noiselayer = vib.NoiseLayerVIB(mean_dims=HIDDEN_DIMS[-1] / 2,
test_phase_noise=test_phase_noise)
micalculator.set_noiselayer(noiselayer)
cur_layer = inputs
for l in model_layers:
cur_layer = l(cur_layer)
noise_input_layer = layers.IdentityMap(
activity_regularizer=micalculator)(cur_layer)
del cur_layer
cur_layer = noiselayer(noise_input_layer)
#if not opts['noise_logvar_grad_trainable']:
# cbs.append(training.NoiseTrain(traindata=trn, noiselayer=noiselayer))
else:
cur_layer = inputs
for l in model_layers:
cur_layer = l(cur_layer)
for hndx, hdim in enumerate(DECODER_DIMS):
layer_args = {}
layer_args['activation'] = 'relu' # opts['DECODING_ACTS'][hndx]
if layer_args['activation'] == 'relu':
layer_args['init'] = 'he_uniform'
else:
layer_args['init'] = 'glorot_uniform'
#if args.maxnorm is not None:
# import keras.constraints
# layer_args['W_constraint'] = keras.constraints.maxnorm(args.maxnorm)
cur_layer = Dense(hdim, **layer_args)(cur_layer)
predictions = Dense(trn.nb_classes,
init='glorot_uniform',
activation='softmax')(cur_layer)
model = Model(input=inputs, output=predictions)
return model, cbs, noiselayer, micalculator
]
# *** The following creates the layers and callbacks necessary to run nonlinearIB ***
micalculator = layers.MICalculator(BETA_VAL,
model_layers=hidden_layers_to_add,
data=trn.X,
miN=1000)
noiselayer = layers.NoiseLayer(logvar_trainable=True, test_phase_noise=False)
micalculator.set_noiselayer(noiselayer)
# Start hooking up the layers together
cur_hidden_layer = input_layer
for l in hidden_layers_to_add:
cur_hidden_layer = l(cur_hidden_layer)
noise_input_layer = layers.IdentityMap(
activity_regularizer=micalculator)(cur_hidden_layer)
nonlinearIB_output_layer = noiselayer(noise_input_layer)
nonlinearIB_callback = training.KDETrain(mi_calculator=micalculator)
# *** Done setting up nonlinearIB stuff ***
decoder = Dense(800, activation='relu')(nonlinearIB_output_layer)
outputs = Dense(trn.nb_classes, activation='softmax')(decoder)
model = Model(inputs=input_layer, outputs=outputs)
model.compile(loss='categorical_crossentropy',
optimizer='adam',
metrics=['accuracy'])
init_lr = 0.001