def test_default_transform():
graphs = default_transform(
DefaultClassifierGenerator(10, (32, 32, 3)).generate())
model = graphs[0].produce_model()
model(torch.Tensor(get_conv_data()))
assert len(graphs) == 1
assert len(graphs[0].layer_list) == 44
def test_copy_model():
model = get_add_skip_model()
new_model = Graph(model).produce_model()
input_data = get_conv_data()
output1 = model.predict_on_batch(input_data).flatten()
output2 = new_model.predict_on_batch(input_data).flatten()
assert np.sum(output1 - output2) == 0
def test_dense_wider():
model = get_add_skip_model()
graph = Graph(model)
new_model = graph.to_wider_model(model.layers[-2], 3)
input_data = get_conv_data()
output1 = model.predict_on_batch(input_data).flatten()
output2 = new_model.predict_on_batch(input_data).flatten()
assert abs(np.sum(output1 - output2)) < 1e-4
def test_conv_deeper():
model = get_conv_model()
graph = Graph(model)
new_model = graph.to_conv_deeper_model(model.layers[4], 3)
input_data = get_conv_data()
output1 = model.predict_on_batch(input_data).flatten()
output2 = new_model.predict_on_batch(input_data).flatten()
assert abs(np.sum(output1 - output2)) < 0.2
def test_skip_concat_over_pooling():
model = get_pooling_model()
graph = Graph(model)
new_model = graph.to_concat_skip_model(model.layers[4], model.layers[11])
input_data = get_conv_data()
output1 = model.predict_on_batch(input_data).flatten()
output2 = new_model.predict_on_batch(input_data).flatten()
assert abs(np.sum(output1 - output2)) < 1e-4
def test_skip_add_over_pooling():
model = get_pooling_model()
graph = Graph(model)
new_model = graph.to_add_skip_model(model.layers[4], model.layers[11])
input_data = get_conv_data()
output1 = model.predict_on_batch(input_data).flatten()
output2 = new_model.predict_on_batch(input_data).flatten()
assert np.array_equal(output1, output2)
def test_deeper_conv_block():
model = get_conv_model()
layers = deeper_conv_block(model.layers[1], 3)
output_tensor = layers[0](model.outputs[0])
output_tensor = layers[1](output_tensor)
output_tensor = layers[2](output_tensor)
new_model = Model(inputs=model.inputs, outputs=output_tensor)
input_data = get_conv_data()
output1 = model.predict_on_batch(input_data).flatten()
output2 = new_model.predict_on_batch(input_data).flatten()
assert np.sum(np.abs(output1 - output2)) < 1e-1
def test_dense_wider():
model = get_add_skip_model()
graph = NetworkMorphismGraph(model)
graph.to_wider_model(graph.layer_to_id[model.layers[-2]], 3)
new_model = graph.produce_model()
input_data = get_conv_data()
output1 = model.predict_on_batch(input_data).flatten()
output2 = new_model.predict_on_batch(input_data).flatten()
assert np.sum(np.abs(output1 - output2)) < 1e-4
def test_skip_add_over_pooling():
model = get_pooling_model()
graph = NetworkMorphismGraph(model)
graph.to_add_skip_model(graph.layer_to_id[model.layers[4]], graph.layer_to_id[model.layers[11]])
new_model = graph.produce_model()
input_data = get_conv_data()
output1 = model.predict_on_batch(input_data).flatten()
output2 = new_model.predict_on_batch(input_data).flatten()
assert np.array_equal(output1, output2)
def test_dense_wider():
model = get_add_skip_model()
graph = Graph(model, True)
graph.to_wider_model(19, 3)
new_model = graph.produce_model()
input_data = get_conv_data()
output1 = model.predict_on_batch(input_data).flatten()
output2 = new_model.predict_on_batch(input_data).flatten()
assert np.sum(np.abs(output1 - output2)) < 1e-4
def test_skip_add_over_pooling():
model = get_pooling_model()
graph = Graph(model, True)
graph.to_add_skip_model(2, 10)
new_model = graph.produce_model()
input_data = get_conv_data()
output1 = model.predict_on_batch(input_data).flatten()
output2 = new_model.predict_on_batch(input_data).flatten()
assert np.array_equal(output1, output2)
def test_conv_deeper():
model = get_conv_model()
graph = Graph(model, True)
graph.to_conv_deeper_model(6, 3)
new_model = graph.produce_model()
input_data = get_conv_data()
output1 = model.predict_on_batch(input_data).flatten()
output2 = new_model.predict_on_batch(input_data).flatten()
assert np.sum(np.abs(output1 - output2)) < 4e-1
def test_skip_concat_over_pooling():
model = get_pooling_model()
graph = Graph(model, True)
graph.to_concat_skip_model(6, 11)
graph.to_concat_skip_model(6, 11)
graph = Graph(graph.produce_model(), True)
new_model = graph.produce_model()
input_data = get_conv_data()
output1 = model.predict_on_batch(input_data).flatten()
output2 = new_model.predict_on_batch(input_data).flatten()
assert np.sum(np.abs(output1 - output2)) < 4e-1
def test_skip_concat_over_pooling():
model = get_pooling_model()
graph = NetworkMorphismGraph(model)
graph.to_concat_skip_model(graph.layer_to_id[model.layers[4]], graph.layer_to_id[model.layers[8]])
graph.to_concat_skip_model(graph.layer_to_id[model.layers[4]], graph.layer_to_id[model.layers[8]])
graph = NetworkMorphismGraph(graph.produce_model())
new_model = graph.produce_model()
input_data = get_conv_data()
output1 = model.predict_on_batch(input_data).flatten()
output2 = new_model.predict_on_batch(input_data).flatten()
assert np.sum(np.abs(output1 - output2)) < 4e-1
def test_dense_deeper():
graph = get_conv_dense_model()
model = graph.produce_model()
graph = deepcopy(graph)
graph.to_dense_deeper_model(10)
new_model = graph.produce_model()
input_data = torch.Tensor(get_conv_data())
model.eval()
new_model.eval()
output1 = model(input_data)
output2 = new_model(input_data)
assert (output1 - output2).abs().sum() < 1e-3
def test_dense_deeper():
graph = get_conv_dense_model()
model = graph.produce_model()
graph = deepcopy(graph)
graph.to_dense_deeper_model(9)
new_model = graph.produce_model()
input_data = torch.Tensor(get_conv_data())
model.eval()
new_model.eval()
output1 = model(input_data)
output2 = new_model(input_data)
assert (output1 - output2).abs().sum() < 1e-3
def test_skip_add_over_pooling():
graph = get_pooling_model()
model = graph.produce_model()
graph = deepcopy(graph)
graph.to_add_skip_model(1, 10)
new_model = graph.produce_model()
input_data = torch.Tensor(get_conv_data())
model.eval()
new_model.eval()
output1 = model(input_data)
output2 = new_model(input_data)
assert (output1 - output2).abs().sum() < 1e-4
def test_skip_add_over_pooling():
graph = get_pooling_model()
model = graph.produce_model()
graph = deepcopy(graph)
graph.to_add_skip_model(1, 8)
new_model = graph.produce_model()
input_data = torch.Tensor(get_conv_data())
model.eval()
new_model.eval()
output1 = model(input_data)
output2 = new_model(input_data)
assert (output1 - output2).abs().sum() < 1e-4
def test_wider_conv():
model = get_conv_model()
layer1 = wider_pre_conv(model.layers[1], 3)
layer2 = wider_bn(model.layers[2], 3, 3, 3)
layer3 = wider_next_conv(model.layers[4], 3, 3, 3)
input_tensor = Input(shape=(5, 5, 3))
output_tensor = layer1(input_tensor)
output_tensor = layer2(output_tensor)
output_tensor = Activation('relu')(output_tensor)
output_tensor = layer3(output_tensor)
output_tensor = BatchNormalization()(output_tensor)
output_tensor = Activation('relu')(output_tensor)
model2 = Model(inputs=input_tensor, outputs=output_tensor)
random_input = get_conv_data()
output1 = model.predict_on_batch(random_input)
output2 = model2.predict_on_batch(random_input)
assert np.sum(np.abs(output1.flatten() - output2.flatten())) < 1e-1
def test_default_transform():
graphs = default_transform(CnnGenerator(10, (32, 32, 3)).generate())
model = graphs[0].produce_model()
model(torch.Tensor(get_conv_data()))
assert len(graphs) == 1
assert len(graphs[0].layer_list) == 43
