def build_feature_model(input_shape,
name='',
dense_layers=[1000, 1000],
activation='relu',
residual=False,
dropout_rate=0,
permanent_dropout=True):
print('input_shape', input_shape)
x_input = Input(shape=input_shape)
h = x_input
for i, layer in enumerate(dense_layers):
x = h
h = Dense(layer, activation=activation)(h)
if dropout_rate > 0:
if permanent_dropout:
h = PermanentDropout(dropout_rate)(h)
else:
h = Dropout(dropout_rate)(h)
if residual:
try:
h = Add([h, x])
except ValueError:
pass
model = Model(x_input, h, name=name)
return model
def build_model(loader, args, permanent_dropout=True, silent=False):
input_shapes = {}
dropout_rate = args.dropout
for fea_type, shape in loader.feature_shapes.items():
base_type = fea_type.split('.')[0]
if base_type in ['cell', 'drug']:
if not silent:
print('Feature encoding submodel for %s:', fea_type)
#box.summary(print_fn=logger.debug)
input_shapes[fea_type] = shape
inputs = []
encoded_inputs = []
for fea_name, fea_type in loader.input_features.items():
shape = loader.feature_shapes[fea_type]
fea_input = Input(shape, name='input.' + fea_name)
inputs.append(fea_input)
if fea_type in input_shapes:
if args.dense_cell_feature_layers is not None and base_type == 'cell':
dense_feature_layers = args.dense_cell_feature_layers
elif args.dense_drug_feature_layers is not None and base_type == 'drug':
dense_feature_layers = args.dense_drug_feature_layers
else:
dense_feature_layers = args.dense_feature_layers
input_model = build_feature_model(
input_shape=shape,
name=fea_type,
dense_layers=dense_feature_layers,
dropout_rate=dropout_rate,
permanent_dropout=permanent_dropout)
encoded = input_model(fea_input)
else:
encoded = fea_input
encoded_inputs.append(encoded)
merged = Concatenate(axis=1)(encoded_inputs)
h = merged
for i, layer in enumerate(args.dense):
x = h
h = Dense(layer, activation=args.activation)(h)
if dropout_rate > 0:
if permanent_dropout:
h = PermanentDropout(dropout_rate)(h)
else:
h = Dropout(dropout_rate)(h)
if args.residual:
try:
h = Add([h, x])
except ValueError:
pass
output = Dense(1)(h)
return Model(inputs, output)
def add_test(): input1 = Input(shape=(16, ), dtype="float32") x1 = Dense(8, activation='relu')(input1) input2 = Input(shape=(32,), dtype="float32") x2 = Dense(8, activation='relu')(input2) subtracted = Add()([x1, x2]) out = Dense(4)(subtracted) model = Model([input1, input2], out) opt = flexflow.keras.optimizers.SGD(learning_rate=0.01) model.compile(optimizer=opt, loss='sparse_categorical_crossentropy', metrics=['accuracy', 'sparse_categorical_crossentropy']) print(model.summary()) model.ffmodel.init_layers()