def wider_pre_dense(layer, n_add, weighted=True): '''wider previous dense layer. ''' 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): '''deeper dense layer. ''' 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): '''wider next dense layer. ''' 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