def get_conv_dense_model():
graph = Graph((32, 32, 3), False)
output_node_id = 0
output_node_id = graph.add_layer(StubReLU(), output_node_id)
output_node_id = graph.add_layer(StubConv(3, 3, 3), output_node_id)
output_node_id = graph.add_layer(StubBatchNormalization(3), output_node_id)
output_node_id = graph.add_layer(StubReLU(), output_node_id)
output_node_id = graph.add_layer(StubConv(3, 3, 3), output_node_id)
output_node_id = graph.add_layer(StubBatchNormalization(3), output_node_id)
output_node_id = graph.add_layer(StubFlatten(), output_node_id)
output_node_id = graph.add_layer(StubDropout(Constant.DENSE_DROPOUT_RATE),
output_node_id)
output_node_id = graph.add_layer(StubReLU(), output_node_id)
output_node_id = graph.add_layer(
StubDense(graph.node_list[output_node_id].shape[0], 5), output_node_id)
output_node_id = graph.add_layer(StubReLU(), output_node_id)
output_node_id = graph.add_layer(StubDense(5, 5), output_node_id)
graph.add_layer(StubSoftmax(), output_node_id)
graph.produce_model().set_weight_to_graph()
return graph
def deeper_conv_block(conv_layer, kernel_size, weighted=True):
filter_shape = (kernel_size,) * 2
n_filters = conv_layer.filters
weight = np.zeros((n_filters, n_filters) + filter_shape)
center = tuple(map(lambda x: int((x - 1) / 2), filter_shape))
for i in range(n_filters):
filter_weight = np.zeros((n_filters,) + filter_shape)
index = (i,) + center
filter_weight[index] = 1
weight[i, ...] = filter_weight
bias = np.zeros(n_filters)
new_conv_layer = StubConv(conv_layer.filters, n_filters, kernel_size=kernel_size)
bn = StubBatchNormalization(n_filters)
if weighted:
new_conv_layer.set_weights((add_noise(weight, np.array([0, 1])), add_noise(bias, np.array([0, 1]))))
new_weights = [add_noise(np.ones(n_filters, dtype=np.float32), np.array([0, 1])),
add_noise(np.zeros(n_filters, dtype=np.float32), np.array([0, 1])),
add_noise(np.zeros(n_filters, dtype=np.float32), np.array([0, 1])),
add_noise(np.ones(n_filters, dtype=np.float32), np.array([0, 1]))]
bn.set_weights(new_weights)
return [StubReLU(),
new_conv_layer,
bn]
def generate(self,
model_len=Constant.MODEL_LEN,
model_width=Constant.MODEL_WIDTH):
pooling_len = int(model_len / 4)
graph = Graph(self.input_shape, False)
temp_input_channel = self.input_shape[-1]
output_node_id = 0
for i in range(model_len):
output_node_id = graph.add_layer(StubReLU(), output_node_id)
output_node_id = graph.add_layer(
StubConv(temp_input_channel, model_width, kernel_size=3),
output_node_id)
output_node_id = graph.add_layer(
StubBatchNormalization(model_width), output_node_id)
temp_input_channel = model_width
if pooling_len == 0 or ((i + 1) % pooling_len == 0
and i != model_len - 1):
output_node_id = graph.add_layer(StubPooling(), output_node_id)
output_node_id = graph.add_layer(StubGlobalPooling(), output_node_id)
output_node_id = graph.add_layer(
StubDropout(Constant.CONV_DROPOUT_RATE), output_node_id)
output_node_id = graph.add_layer(
StubDense(graph.node_list[output_node_id].shape[0], model_width),
output_node_id)
output_node_id = graph.add_layer(StubReLU(), output_node_id)
graph.add_layer(StubDense(model_width, self.n_output_node),
output_node_id)
return graph
def wider_next_conv(layer, start_dim, total_dim, n_add, weighted=True):
if not weighted:
return StubConv(layer.input_channel + n_add, layer.filters, kernel_size=layer.kernel_size)
n_filters = layer.filters
teacher_w, teacher_b = layer.get_weights()
new_weight_shape = list(teacher_w.shape)
new_weight_shape[1] = n_add
new_weight = np.zeros(tuple(new_weight_shape))
student_w = np.concatenate((teacher_w[:, :start_dim, ...].copy(),
add_noise(new_weight, teacher_w),
teacher_w[:, start_dim:total_dim, ...].copy()), axis=1)
new_layer = StubConv(layer.input_channel + n_add, n_filters, layer.kernel_size)
new_layer.set_weights((student_w, teacher_b))
return new_layer
def wider_pre_conv(layer, n_add_filters, weighted=True):
if not weighted:
return StubConv(layer.input_channel, layer.filters + n_add_filters, kernel_size=layer.kernel_size)
n_pre_filters = layer.filters
rand = np.random.randint(n_pre_filters, size=n_add_filters)
teacher_w, teacher_b = layer.get_weights()
student_w = teacher_w.copy()
student_b = teacher_b.copy()
# target layer update (i)
for i in range(len(rand)):
teacher_index = rand[i]
new_weight = teacher_w[teacher_index, ...]
new_weight = new_weight[np.newaxis, ...]
student_w = np.concatenate((student_w, new_weight), axis=0)
student_b = np.append(student_b, teacher_b[teacher_index])
new_pre_layer = StubConv(layer.input_channel, n_pre_filters + n_add_filters, layer.kernel_size)
new_pre_layer.set_weights((add_noise(student_w, teacher_w), add_noise(student_b, teacher_b)))
return new_pre_layer
def test_layer_distance():
layer1 = StubConv(5, 5, 3, 2)
layer2 = StubConv(5, 1, 1, 1)
assert layer_distance(layer1, layer2) == 5.9 / 9