def test_shape_sanity_check(in_dim_base, dim1, dim2, num_nodes, bond_dim):
model = Sequential([
Input(in_dim_base**num_nodes),
DenseMPO(dim1**num_nodes, num_nodes=num_nodes, bond_dim=bond_dim),
DenseMPO(dim2**num_nodes, num_nodes=num_nodes, bond_dim=bond_dim),
])
# Hard code batch size.
result = model.predict(np.ones((32, in_dim_base**num_nodes)))
assert result.shape == (32, dim2**num_nodes)
def test_config(make_model):
# Disable the redefined-outer-name violation in this function
# pylint: disable=redefined-outer-name
model = make_model
expected_num_parameters = model.layers[0].count_params()
# Serialize model and use config to create new layer
model_config = model.get_config()
layer_config = model_config['layers'][1]['config']
if 'mpo' in model.layers[0].name:
new_model = DenseMPO.from_config(layer_config)
elif 'decomp' in model.layers[0].name:
new_model = DenseDecomp.from_config(layer_config)
elif 'condenser' in model.layers[0].name:
new_model = DenseCondenser.from_config(layer_config)
elif 'expander' in model.layers[0].name:
new_model = DenseExpander.from_config(layer_config)
elif 'entangler' in model.layers[0].name:
new_model = DenseEntangler.from_config(layer_config)
# Build the layer so we can count params below
new_model.build(layer_config['batch_input_shape'])
# Check that original layer had same num params as layer built from config
np.testing.assert_equal(expected_num_parameters, new_model.count_params())
def test_mpo_num_parameters(dummy_data):
# Disable the redefined-outer-name violation in this function
# pylint: disable=redefined-outer-name
data, _ = dummy_data
output_dim = data.shape[1]
num_nodes = int(math.log(data.shape[1], 8))
bond_dim = 8
model = Sequential()
model.add(
DenseMPO(output_dim,
num_nodes=num_nodes,
bond_dim=bond_dim,
use_bias=True,
activation='relu',
input_shape=(data.shape[1], )))
in_leg_dim = math.ceil(data.shape[1]**(1. / num_nodes))
out_leg_dim = math.ceil(output_dim**(1. / num_nodes))
# num_params = num_edge_node_params + num_middle_node_params + bias_params
expected_num_parameters = (2 * in_leg_dim * bond_dim * out_leg_dim) + (
(num_nodes - 2) * in_leg_dim * bond_dim * bond_dim *
out_leg_dim) + output_dim
np.testing.assert_equal(expected_num_parameters, model.count_params())
def make_high_dim_model(high_dim_data, request):
# Disable the redefined-outer-name violation in this function
# pylint: disable=redefined-outer-name
data = high_dim_data
if request.param == 'DenseMPO':
model = Sequential()
model.add(
DenseMPO(data.shape[-1],
num_nodes=int(math.log(int(data.shape[-1]), 8)),
bond_dim=8,
use_bias=True,
activation='relu',
input_shape=(data.shape[-1], )))
elif request.param == 'DenseDecomp':
model = Sequential()
model.add(
DenseDecomp(512,
decomp_size=128,
use_bias=True,
activation='relu',
input_shape=(data.shape[-1], )))
elif request.param == 'DenseCondenser':
model = Sequential()
model.add(
DenseCondenser(exp_base=2,
num_nodes=3,
use_bias=True,
activation='relu',
input_shape=(data.shape[-1], )))
elif request.param == 'DenseExpander':
model = Sequential()
model.add(
DenseExpander(exp_base=2,
num_nodes=3,
use_bias=True,
activation='relu',
input_shape=(data.shape[-1], )))
elif request.param == 'DenseEntangler':
num_legs = 3
leg_dim = round(data.shape[-1]**(1. / num_legs))
assert leg_dim**num_legs == data.shape[-1]
model = Sequential()
model.add(
DenseEntangler(leg_dim**num_legs,
num_legs=num_legs,
num_levels=3,
use_bias=True,
activation='relu',
input_shape=(data.shape[-1], )))
return data, model