def build_wrapped_model(self, model, sparse=False):
if sparse:
wrapper = create_model_wrapper(model, sparse=True, num_sparse_params=6)
else:
wrapper = create_model_wrapper(model)
params = Input(shape=(wrapper.num_params,))
trainable_params = Input(shape=(wrapper.num_trainable_params,))
x = Input(shape=K.int_shape(model.inputs[0]))
y = wrapper([params, trainable_params, x])
return Model(inputs=[params, trainable_params, x], outputs=[y])
def testGetAllWeights(self):
model = self.build_model()
model.set_weights((np.ones(2), np.zeros(2), np.eye(2), np.zeros(2), np.ones(2)))
wrapper = create_model_wrapper(model)
expected_weights = np.array([1., 1., 0., 0., 1., 0., 0., 1., 0., 0., 1., 1.])
np.testing.assert_allclose(expected_weights, get_model_weights(model))
def testGetParamGroupsWithTrainableParameters(self):
model = self.build_model()
for l in model.layers:
l.trainable = l.name.startswith("dense")
wrapper = create_model_wrapper(model)
expected_groups = [(0, 6)]
self.assertEqual(expected_groups, list(wrapper.param_groups()))
def build_wrapped_model(self, model, batch_size=1):
wrapper = create_model_wrapper(model, batch_size=batch_size)
params = Input(shape=(wrapper.num_params,))
trainable_params = Input(shape=(wrapper.num_trainable_params,))
x = Input(shape=K.int_shape(model.inputs[0]))
y = wrapper([params, trainable_params, x])
return Model(inputs=[params, trainable_params, x], outputs=[y])
def build_batchnorm_model_wrapper(self, batchnorm_training=False, batch_size=1):
model = Sequential()
model.add(BatchNormalization(input_shape=(None, 2)))
model.compile(loss='mse', optimizer='SGD')
wrapper = create_model_wrapper(model, batch_size=batch_size)
params = Input(shape=(wrapper.num_params,))
x = Input(shape=K.int_shape(model.inputs[0]))
y = wrapper([params, params, x], training=batchnorm_training)
wrapped_model = Model(inputs=[params, x], outputs=[y])
wrapped_model.compile(loss='mse', optimizer='SGD')
return model, wrapped_model
def testGetParamGroups(self):
model = self.build_model()
wrapper = create_model_wrapper(model)
expected_groups = [(0, 4), (4, 10), (10, 12)]
self.assertEqual(expected_groups, list(wrapper.param_groups()))