Source code for autoaug.autoaugment_learners.EvoLearner

import torch
import torch.nn as nn
import pygad
import pygad.torchga as torchga
import torchvision
import torch

from autoaug.autoaugment_learners.AaLearner import AaLearner
import autoaug.controller_networks as cont_n


[docs]class EvoLearner(AaLearner): """Evolutionary Strategy learner This learner generates neural networks that predict optimal augmentation policies. Hence, there is no backpropagation or gradient descent. Instead, training is done by randomly changing weights of the 'parent' networks, where parents are determined by their ability to produce policies that increase the accuracy of the child network. Args: num_sub_policies (int, optional): number of subpolicies per policy. Defaults to 5. p_bins (int, optional): number of bins we divide the interval [0,1] for probabilities. e.g. (0.0, 0.1, ... 1.0) Defaults to 1. m_bins (int, optional): number of bins we divide the magnitude space. Defaults to 1. exclude_method (list, optional): list of names(:type:str) of image operations the user wants to exclude from the search space. Defaults to []. batch_size (int, optional): child_network training parameter. Defaults to 32. toy_size (int, optional): child_network training parameter. ratio of original dataset used in toy dataset. Defaults to 0.1. learning_rate (float, optional): child_network training parameter. Defaults to 1e-1. max_epochs (Union[int, float], optional): child_network training parameter. Defaults to float('inf'). early_stop_num (int, optional): child_network training parameter. Defaults to 20. num_solutions (int, optional): Number of offspring spawned at each generation of the algorithm. Default 5 num_parents_mating (int, optional): Number of networks chosen as parents for the next generation of networks Defaults to 3 controller (nn.Module, optional): Controller network for the evolutionary algorithm. Defaults to cont_n.EvoController Notes ----- The Evolutionary algorithm runs in generations, and so batches of child networks are trained at specific time intervals. Examples -------- from autoaug.autoaugment_learners.EvlLearner import EvoLearner evo_learner = EvoLearner() """ def __init__(self, # search space settings num_sub_policies=5, p_bins=1, m_bins=1, exclude_method=[], # child network settings learning_rate=1e-1, max_epochs=float('inf'), early_stop_num=20, batch_size=8, toy_size=1, # evolutionary learner specific settings num_solutions=5, num_parents_mating=3, controller=cont_n.EvoController ): super().__init__( num_sub_policies=num_sub_policies, p_bins=p_bins, m_bins=m_bins, discrete_p_m=False, batch_size=batch_size, toy_size=toy_size, learning_rate=learning_rate, max_epochs=max_epochs, early_stop_num=early_stop_num, exclude_method=exclude_method ) self.controller = controller( fun_num=self.fun_num, p_bins=self.p_bins, m_bins=self.m_bins, sub_num_pol=self.num_sub_policies ) # self.controller = controller self.num_solutions = num_solutions self.torch_ga = torchga.TorchGA(model=self.controller, num_solutions=num_solutions) self.num_parents_mating = num_parents_mating self.initial_population = self.torch_ga.population_weights # store our logs self.policy_dict = {} self.running_policy = [] self.first_run = True self.fun_num = len(self.augmentation_space) # evolutionary algorithm settings assert num_solutions > num_parents_mating, 'Number of solutions must be larger than the number of parents mating!' def _get_single_policy_cov(self, x, alpha = 0.5): """ Selects policy using population and covariance matrices. For this method we require p_bins = 1, num_sub_pol = 1, m_bins = 1. Parameters ------------ x -> PyTorch Tensor Input data for the AutoAugment network alpha -> float Proportion for covariance and population matrices Returns ----------- Subpolicy -> [(String, float, float), (String, float, float)] Subpolicy consisting of two tuples of policies, each with a string associated to a transformation, a float for a probability, and a float for a magnittude """ section = self.fun_num + self.p_bins + self.m_bins y = self.controller.forward(x) y_1 = torch.softmax(y[:,:self.fun_num], dim = 1) y[:,:self.fun_num] = y_1 y_2 = torch.softmax(y[:,section:section+self.fun_num], dim = 1) y[:,section:section+self.fun_num] = y_2 concat = torch.cat((y_1, y_2), dim = 1) cov_mat = torch.cov(concat.T) cov_mat = cov_mat[:self.fun_num, self.fun_num:] shape_store = cov_mat.shape counter, prob1, prob2, mag1, mag2 = (0, 0, 0, 0, 0) prob_mat = torch.zeros(shape_store) for idx in range(y.shape[0]): prob_mat[torch.argmax(y_1[idx])][torch.argmax(y_2[idx])] += 1 prob_mat = prob_mat / torch.sum(prob_mat) cov_mat = (alpha * cov_mat) + ((1 - alpha)*prob_mat) cov_mat = torch.reshape(cov_mat, (1, -1)).squeeze() max_idx = torch.argmax(cov_mat) val = (max_idx//shape_store[0]) max_idx = (val, max_idx - (val * shape_store[0])) if not self.augmentation_space[max_idx[0]][1]: mag1 = None if not self.augmentation_space[max_idx[1]][1]: mag2 = None for idx in range(y.shape[0]): if (torch.argmax(y_1[idx]) == max_idx[0]) and (torch.argmax(y_2[idx]) == max_idx[1]): prob1 += torch.sigmoid(y[idx, self.fun_num]).item() prob2 += torch.sigmoid(y[idx, section+self.fun_num]).item() if mag1 is not None: mag1 += min(9, 10 * torch.sigmoid(y[idx, self.fun_num+1]).item()) if mag2 is not None: mag2 += min(9, 10 * torch.sigmoid(y[idx, self.fun_num+1]).item()) counter += 1 prob1 = round(prob1/counter, 1) if counter != 0 else 0 prob2 = round(prob2/counter, 1) if counter != 0 else 0 if mag1 is not None: mag1 = int(mag1/counter) if mag2 is not None: mag2 = int(mag2/counter) return [((self.augmentation_space[max_idx[0]][0], prob1, mag1), (self.augmentation_space[max_idx[1]][0], prob2, mag2))]
[docs] def learn(self, train_dataset, test_dataset, child_network_architecture, iterations = 15, return_weights = False): """ Runs the GA instance and returns the model weights as a dictionary Parameters ------------ return_weights -> bool Determines if the weight of the GA network should be returned Returns ------------ If return_weights: Network weights -> dict Else: Solution -> Best GA instance solution Solution fitness -> float Solution_idx -> int """ self.num_generations = iterations self.history_best = [] self._set_up_instance(train_dataset, test_dataset, child_network_architecture) self.ga_instance.run() solution, solution_fitness, solution_idx = self.ga_instance.best_solution() if return_weights: return torchga.model_weights_as_dict(model=self.controller, weights_vector=solution) else: return solution, solution_fitness, solution_idx
def _in_pol_dict(self, new_policy): """ Checks if a potential subpolicy has already been testing by the agent Parameters ------------ new_policy -> subpolicy Returns ------------ if subpolicy has been tested: -> True else: -> False """ new_policy = new_policy[0] trans1, trans2 = new_policy[0][0], new_policy[1][0] new_set = {new_policy[0][1], new_policy[0][2], new_policy[1][1], new_policy[1][2]} if trans1 in self.policy_dict: if trans2 in self.policy_dict[trans1]: for test_pol in self.policy_dict[trans1][trans2]: if new_set == test_pol: return True self.policy_dict[trans1][trans2].append(new_set) else: self.policy_dict[trans1] = {trans2: [new_set]} return False def _set_up_instance(self, train_dataset, test_dataset, child_network_architecture): """ Initialises GA instance, as well as the fitness and 'on generation' functions """ def _fitness_func(solution, sol_idx): """ Defines the fitness function for the parent selection Parameters -------------- solution -> GA solution instance (parsed automatically) sol_idx -> GA solution index (parsed automatically) Returns -------------- fit_val -> float """ model_weights_dict = torchga.model_weights_as_dict(model=self.controller, weights_vector=solution) self.controller.load_state_dict(model_weights_dict) train_dataset.transform = torchvision.transforms.ToTensor() self.train_loader = torch.utils.data.DataLoader(train_dataset, batch_size=500) count = 0 new_pol = True for idx, (test_x, label_x) in enumerate(self.train_loader): count += 1 sub_pol = self._get_single_policy_cov(test_x) print("subpol: ", sub_pol) if idx == 0: break print("start test") self.print_every_epoch = True self.early_stop_num = 10 if new_pol: fit_val = self._test_autoaugment_policy(sub_pol,child_network_architecture,train_dataset,test_dataset) print("fit_val: ", fit_val) print("end test") self.running_policy.append((sub_pol, fit_val)) if len(self.running_policy) > self.num_sub_policies: self.running_policy = sorted(self.running_policy, key=lambda x: x[1], reverse=True) self.running_policy = self.running_policy[:self.num_sub_policies] if len(self.history_best) == 0: self.history_best.append(fit_val) self.new_pop = self.torch_ga.population_weights elif fit_val > self.history_best[-1]: self.history_best.append(fit_val) self.new_pop = self.torch_ga.population_weights else: self.history_best.append(self.history_best[-1]) self.first_run = False return fit_val def _on_generation(ga_instance): """ Prints information of generation's fitness Parameters ------------- ga_instance -> GA instance Returns ------------- None """ print("Generation = {generation}".format(generation=ga_instance.generations_completed)) print("Fitness = {fitness}".format(fitness=ga_instance.best_solution()[1])) return if self.first_run: self.ga_instance = pygad.GA(num_generations=self.num_generations, num_parents_mating=self.num_parents_mating, initial_population=self.initial_population, mutation_percent_genes = 0.1, fitness_func=_fitness_func, on_generation = _on_generation) else: self.ga_instance = pygad.GA(num_generations=self.num_generations, num_parents_mating=self.num_parents_mating, initial_population=self.new_pop, mutation_percent_genes = 0.1, fitness_func=_fitness_func, on_generation = _on_generation)