def wider_pre_dense(layer, n_add, weighted=True):
if not weighted:
return StubDense(layer.input_units, layer.units + n_add)
n_units2 = layer.units
teacher_w, teacher_b = layer.get_weights()
rand = np.random.randint(n_units2, size=n_add)
student_w = teacher_w.copy()
student_b = teacher_b.copy()
# target layer update (i)
for i in range(n_add):
teacher_index = rand[i]
new_weight = teacher_w[teacher_index, :]
new_weight = new_weight[np.newaxis, :]
student_w = np.concatenate(
(student_w, add_noise(new_weight, student_w)), axis=0)
student_b = np.append(student_b,
add_noise(teacher_b[teacher_index], student_b))
new_pre_layer = StubDense(layer.input_units, n_units2 + n_add)
new_pre_layer.set_weights((student_w, student_b))
return new_pre_layer
def dense_to_deeper_block(dense_layer, weighted=True):
units = dense_layer.units
weight = np.eye(units)
bias = np.zeros(units)
new_dense_layer = StubDense(units, units)
if weighted:
new_dense_layer.set_weights((add_noise(weight, np.array([0, 1])), add_noise(bias, np.array([0, 1]))))
return [StubReLU(), new_dense_layer]
def wider_next_dense(layer, start_dim, total_dim, n_add, weighted=True):
if not weighted:
return StubDense(layer.input_units + n_add, layer.units)
teacher_w, teacher_b = layer.get_weights()
student_w = teacher_w.copy()
n_units_each_channel = int(teacher_w.shape[1] / total_dim)
new_weight = np.zeros((teacher_w.shape[0], n_add * n_units_each_channel))
student_w = np.concatenate((student_w[:, :start_dim * n_units_each_channel],
add_noise(new_weight, student_w),
student_w[:, start_dim * n_units_each_channel:total_dim * n_units_each_channel]),
axis=1)
new_layer = StubDense(layer.input_units + n_add, layer.units)
new_layer.set_weights((student_w, teacher_b))
return new_layer
def wider_next_dense(layer, start_dim, total_dim, n_add, weighted=True):
if not weighted:
return StubDense(layer.input_units + n_add, layer.units)
teacher_w, teacher_b = layer.get_weights()
student_w = teacher_w.copy()
n_units_each_channel = int(teacher_w.shape[1] / total_dim)
new_weight = np.zeros((teacher_w.shape[0], n_add * n_units_each_channel))
student_w = np.concatenate((student_w[:, :start_dim * n_units_each_channel],
add_noise(new_weight, student_w),
student_w[:, start_dim * n_units_each_channel:total_dim * n_units_each_channel]),
axis=1)
new_layer = StubDense(layer.input_units + n_add, layer.units)
new_layer.set_weights((student_w, teacher_b))
return new_layer
def wider_pre_dense(layer, n_add, weighted=True):
if not weighted:
return StubDense(layer.input_units, layer.units + n_add)
n_units2 = layer.units
teacher_w, teacher_b = layer.get_weights()
rand = np.random.randint(n_units2, size=n_add)
student_w = teacher_w.copy()
student_b = teacher_b.copy()
# target layer update (i)
for i in range(n_add):
teacher_index = rand[i]
new_weight = teacher_w[teacher_index, :]
new_weight = new_weight[np.newaxis, :]
student_w = np.concatenate((student_w, add_noise(new_weight, student_w)), axis=0)
student_b = np.append(student_b, add_noise(teacher_b[teacher_index], student_b))
new_pre_layer = StubDense(layer.input_units, n_units2 + n_add)
new_pre_layer.set_weights((student_w, student_b))
return new_pre_layer
