def test_aggregated_weights_NormClip():
num_clients = 10
num_layers = 5
tams = [[128, 64], [64, 64], [64, 64], [64, 32], [32, 10]]
weights = []
for i in range(num_clients):
weights.append([
np.random.rand(tams[j][0], tams[j][1]) for j in range(num_layers)
])
clients_params = weights
avgfa = NormClipAggregator(clip=100)
aggregated_weights = avgfa.aggregate_weights(clients_params)
clients_params = np.array(weights)
own_agg = np.array([
np.mean(clients_params[:, layer], axis=0)
for layer in range(num_layers)
])
for i in range(num_layers):
assert np.array_equal(own_agg[i], aggregated_weights[i])
assert len(aggregated_weights) == num_layers
def test_aggregated_weights_multidimensional_2D_array_NormClip():
num_clients = 10
num_rows_params = 3
num_cols_params = 9
clients_params = []
for i in range(num_clients):
clients_params.append(np.random.rand(num_rows_params, num_cols_params))
clients_params = np.array(clients_params)
avgfa = NormClipAggregator(clip=10)
aggregated_weights = avgfa.aggregate_weights(clients_params)
own_agg = np.zeros((num_rows_params, num_cols_params))
for i_client in range(num_clients):
own_agg += clients_params[i_client]
own_agg = own_agg / num_clients
assert np.array_equal(own_agg, aggregated_weights)
assert aggregated_weights.shape == own_agg.shape
def test_aggregated_weights_list_of_arrays_NormClip():
num_clients = 10
clients_params = []
for i_client in range(num_clients):
clients_params.append([
np.random.rand(30, 20),
np.random.rand(20, 30),
np.random.rand(50, 40)
])
avgfa = NormClipAggregator(clip=100)
aggregated_weights = avgfa.aggregate_weights(clients_params)
own_agg = [np.zeros((30, 20)), np.zeros((20, 30)), np.zeros((50, 40))]
for i_client in range(num_clients):
for i_params in range(len(clients_params[0])):
own_agg[i_params] += clients_params[i_client][i_params]
for i_params in range(len(clients_params[0])):
own_agg[i_params] = own_agg[i_params] / num_clients
for i_params in range(len(clients_params[0])):
assert np.array_equal(own_agg[i_params], aggregated_weights[i_params])
assert aggregated_weights[i_params].shape == own_agg[i_params].shape