diff --git a/exps/algos-v2/random_wo_share.py b/exps/algos-v2/random_wo_share.py index 774dfd4..5fd1d73 100644 --- a/exps/algos-v2/random_wo_share.py +++ b/exps/algos-v2/random_wo_share.py @@ -4,6 +4,8 @@ # Random Search for Hyper-Parameter Optimization, JMLR 2012 ################## ############################################################################## # python ./exps/algos-v2/random_wo_share.py --dataset cifar10 --search_space tss +# python ./exps/algos-v2/random_wo_share.py --dataset cifar100 --search_space tss +# python ./exps/algos-v2/random_wo_share.py --dataset ImageNet16-120 --search_space tss ############################################################################## import os, sys, time, glob, random, argparse import numpy as np, collections @@ -20,7 +22,7 @@ from utils import get_model_infos, obtain_accuracy from log_utils import AverageMeter, time_string, convert_secs2time from models import get_search_spaces from nas_201_api import NASBench201API, NASBench301API -from .regularized_ea import random_topology_func, random_size_func +from regularized_ea import random_topology_func, random_size_func def main(xargs, api): @@ -28,16 +30,18 @@ def main(xargs, api): prepare_seed(xargs.rand_seed) logger = prepare_logger(args) + logger.log('{:} use api : {:}'.format(time_string(), api)) + api.reset_time() + search_space = get_search_spaces(xargs.search_space, 'nas-bench-301') if xargs.search_space == 'tss': random_arch = random_topology_func(search_space) else: random_arch = random_size_func(search_space) - x_start_time = time.time() - logger.log('{:} use nas_bench : {:}'.format(time_string(), nas_bench)) best_arch, best_acc, total_time_cost, history = None, -1, [], [] - while total_time_cost[-1] < xargs.time_budget: + current_best_index = [] + while len(total_time_cost) == 0 or total_time_cost[-1] < xargs.time_budget: arch = random_arch() accuracy, _, _, total_cost = api.simulate_train_eval(arch, xargs.dataset, '12') total_time_cost.append(total_cost) @@ -45,13 +49,14 @@ def main(xargs, api): if best_arch is None or best_acc < accuracy: best_acc, best_arch = accuracy, arch logger.log('[{:03d}] : {:} : accuracy = {:.2f}%'.format(len(history), arch, accuracy)) - logger.log('{:} best arch is {:}, accuracy = {:.2f}%, visit {:} archs with {:.1f} s (real-cost = {:.3f} s).'.format(time_string(), best_arch, best_acc, len(history), total_time_cost, time.time()-x_start_time)) + current_best_index.append(api.query_index_by_arch(best_arch)) + logger.log('{:} best arch is {:}, accuracy = {:.2f}%, visit {:} archs with {:.1f} s.'.format(time_string(), best_arch, best_acc, len(history), total_time_cost[-1])) info = api.query_info_str_by_arch(best_arch, '200' if xargs.search_space == 'tss' else '90') logger.log('{:}'.format(info)) logger.log('-'*100) logger.close() - return logger.log_dir, total_time_cost, history + return logger.log_dir, current_best_index, total_time_cost if __name__ == '__main__': @@ -62,7 +67,7 @@ if __name__ == '__main__': parser.add_argument('--time_budget', type=int, default=20000, help='The total time cost budge for searching (in seconds).') parser.add_argument('--loops_if_rand', type=int, default=500, help='The total runs for evaluation.') # log - parser.add_argument('--save_dir', type=str, help='Folder to save checkpoints and log.') + parser.add_argument('--save_dir', type=str, default='./output/search', help='Folder to save checkpoints and log.') parser.add_argument('--rand_seed', type=int, default=-1, help='manual seed') args = parser.parse_args() @@ -77,7 +82,7 @@ if __name__ == '__main__': print('save-dir : {:}'.format(args.save_dir)) if args.rand_seed < 0: - save_dir, all_info = None, {} + save_dir, all_info = None, collections.OrderedDict() for i in range(args.loops_if_rand): print ('{:} : {:03d}/{:03d}'.format(time_string(), i, args.loops_if_rand)) args.rand_seed = random.randint(1, 100000) diff --git a/exps/algos-v2/regularized_ea.py b/exps/algos-v2/regularized_ea.py index 4e0a3bd..845bd28 100644 --- a/exps/algos-v2/regularized_ea.py +++ b/exps/algos-v2/regularized_ea.py @@ -155,7 +155,7 @@ def regularized_evolution(cycles, population_size, sample_size, time_budget, ran population = collections.deque() api.reset_time() history, total_time_cost = [], [] # Not used by the algorithm, only used to report results. - + current_best_index = [] # Initialize the population with random models. while len(population) < population_size: model = Model() @@ -163,8 +163,9 @@ def regularized_evolution(cycles, population_size, sample_size, time_budget, ran model.accuracy, _, _, total_cost = api.simulate_train_eval(model.arch, dataset, '12') # Append the info population.append(model) - history.append(model) + history.append((model.accuracy, model.arch)) total_time_cost.append(total_cost) + current_best_index.append(api.query_index_by_arch(max(history, key=lambda x: x[0])[1])) # Carry out evolution in cycles. Each cycle produces a model and removes another. while total_time_cost[-1] < time_budget: @@ -183,15 +184,16 @@ def regularized_evolution(cycles, population_size, sample_size, time_budget, ran # Create the child model and store it. child = Model() child.arch = mutate_arch(parent.arch) - child.accuracy, _, _, total_cost = api.simulate_train_eval(model.arch, dataset, '12') + child.accuracy, _, _, total_cost = api.simulate_train_eval(child.arch, dataset, '12') # Append the info population.append(child) - history.append(child) + history.append((child.accuracy, child.arch)) + current_best_index.append(api.query_index_by_arch(max(history, key=lambda x: x[0])[1])) total_time_cost.append(total_cost) # Remove the oldest model. population.popleft() - return history, total_time_cost + return history, current_best_index, total_time_cost def main(xargs, api): @@ -210,7 +212,7 @@ def main(xargs, api): x_start_time = time.time() logger.log('{:} use api : {:}'.format(time_string(), api)) logger.log('-'*30 + ' start searching with the time budget of {:} s'.format(xargs.time_budget)) - history, total_times = regularized_evolution(xargs.ea_cycles, xargs.ea_population, xargs.ea_sample_size, xargs.time_budget, random_arch, mutate_arch, api, xargs.dataset) + history, current_best_index, total_times = regularized_evolution(xargs.ea_cycles, xargs.ea_population, xargs.ea_sample_size, xargs.time_budget, random_arch, mutate_arch, api, xargs.dataset) logger.log('{:} regularized_evolution finish with history of {:} arch with {:.1f} s (real-cost={:.2f} s).'.format(time_string(), len(history), total_times[-1], time.time()-x_start_time)) best_arch = max(history, key=lambda i: i.accuracy) best_arch = best_arch.arch @@ -220,7 +222,7 @@ def main(xargs, api): logger.log('{:}'.format(info)) logger.log('-'*100) logger.close() - return logger.log_dir, [api.query_index_by_arch(x.arch) for x in history], total_times + return logger.log_dir, current_best_index, total_times if __name__ == '__main__': @@ -249,7 +251,7 @@ if __name__ == '__main__': print('save-dir : {:}'.format(args.save_dir)) if args.rand_seed < 0: - save_dir, all_info = None, {} + save_dir, all_info = None, collections.OrderedDict() for i in range(args.loops_if_rand): print ('{:} : {:03d}/{:03d}'.format(time_string(), i, args.loops_if_rand)) args.rand_seed = random.randint(1, 100000) diff --git a/exps/algos-v2/reinforce.py b/exps/algos-v2/reinforce.py index 400f1ef..11babe4 100644 --- a/exps/algos-v2/reinforce.py +++ b/exps/algos-v2/reinforce.py @@ -145,6 +145,7 @@ def main(xargs, api): x_start_time = time.time() logger.log('Will start searching with time budget of {:} s.'.format(xargs.time_budget)) total_steps, total_costs, trace = 0, [], [] + current_best_index = [] while len(total_costs) == 0 or total_costs[-1] < xargs.time_budget: start_time = time.time() log_prob, action = select_action( policy ) @@ -162,9 +163,8 @@ def main(xargs, api): # accumulate time total_steps += 1 logger.log('step [{:3d}] : average-reward={:.3f} : policy_loss={:.4f} : {:}'.format(total_steps, baseline.value(), policy_loss.item(), policy.genotype())) - #logger.log('----> {:}'.format(policy.arch_parameters)) - #logger.log('') - + # to analyze + current_best_index.append(api.query_index_by_arch(max(trace, key=lambda x: x[0])[1])) # best_arch = policy.genotype() # first version best_arch = max(trace, key=lambda x: x[0])[1] logger.log('REINFORCE finish with {:} steps and {:.1f} s (real cost={:.3f}).'.format(total_steps, total_costs[-1], time.time()-x_start_time)) @@ -173,7 +173,7 @@ def main(xargs, api): logger.log('-'*100) logger.close() - return logger.log_dir, [api.query_index_by_arch(x[1]) for x in trace], total_costs + return logger.log_dir, current_best_index, total_costs if __name__ == '__main__': @@ -203,7 +203,7 @@ if __name__ == '__main__': print('save-dir : {:}'.format(args.save_dir)) if args.rand_seed < 0: - save_dir, all_info = None, {} + save_dir, all_info = None, collections.OrderedDict() for i in range(args.loops_if_rand): print ('{:} : {:03d}/{:03d}'.format(time_string(), i, args.loops_if_rand)) args.rand_seed = random.randint(1, 100000) diff --git a/exps/algos-v2/run-all.sh b/exps/algos-v2/run-all.sh index 3f2f01d..41a907b 100644 --- a/exps/algos-v2/run-all.sh +++ b/exps/algos-v2/run-all.sh @@ -13,5 +13,6 @@ do do python ./exps/algos-v2/reinforce.py --dataset ${dataset} --search_space ${search_space} --learning_rate 0.001 python ./exps/algos-v2/regularized_ea.py --dataset ${dataset} --search_space ${search_space} --ea_cycles 200 --ea_population 10 --ea_sample_size 3 + python ./exps/algos-v2/random_wo_share.py --dataset ${dataset} --search_space ${search_space} done done diff --git a/exps/experimental/vis-bench-algos.py b/exps/experimental/vis-bench-algos.py index f0a4b1b..2cc1f51 100644 --- a/exps/experimental/vis-bench-algos.py +++ b/exps/experimental/vis-bench-algos.py @@ -3,7 +3,7 @@ ############################################################### # Copyright (c) Xuanyi Dong [GitHub D-X-Y], 2020.06 # ############################################################### -# Usage: python exps/experimental/vis-bench-algos.py +# Usage: python exps/experimental/vis-bench-algos.py # ############################################################### import os, sys, time, torch, argparse import numpy as np @@ -30,6 +30,7 @@ def fetch_data(root_dir='./output/search', search_space='tss', dataset=None): alg2name, alg2path = OrderedDict(), OrderedDict() alg2name['REA'] = 'R-EA-SS3' alg2name['REINFORCE'] = 'REINFORCE-0.001' + # alg2name['RANDOM'] = 'RANDOM' for alg, name in alg2name.items(): alg2path[alg] = os.path.join(ss_dir, dataset, name, 'results.pth') assert os.path.isfile(alg2path[alg]) @@ -62,14 +63,15 @@ def visualize_curve(api, vis_save_dir, search_space, max_time): vis_save_dir = vis_save_dir.resolve() vis_save_dir.mkdir(parents=True, exist_ok=True) - dpi, width, height = 250, 4700, 1500 + dpi, width, height = 250, 5100, 1500 figsize = width / float(dpi), height / float(dpi) LabelSize, LegendFontsize = 14, 14 def sub_plot_fn(ax, dataset): alg2data = fetch_data(search_space=search_space, dataset=dataset) alg2accuracies = OrderedDict() - time_tickets = [float(i) / 100 * max_time for i in range(100)] + total_tickets = 150 + time_tickets = [float(i) / total_tickets * max_time for i in range(total_tickets)] colors = ['b', 'g', 'c', 'm', 'y'] for idx, (alg, data) in enumerate(alg2data.items()): print('plot alg : {:}'.format(alg)) @@ -78,7 +80,10 @@ def visualize_curve(api, vis_save_dir, search_space, max_time): accuracy = query_performance(api, data, dataset, ticket) accuracies.append(accuracy) alg2accuracies[alg] = accuracies - ax.plot(time_tickets, accuracies, c=colors[idx], label='{:}'.format(alg)) + ax.plot([x/100 for x in time_tickets], accuracies, c=colors[idx], label='{:}'.format(alg)) + ax.set_xlabel('Estimated wall-clock time (1e2 seconds)', fontsize=LabelSize) + ax.set_ylabel('Test accuracy on {:}'.format(dataset), fontsize=LabelSize) + ax.set_title('Searching results on {:}'.format(dataset), fontsize=LabelSize+4) ax.legend(loc=4, fontsize=LegendFontsize) fig, axs = plt.subplots(1, 3, figsize=figsize) @@ -104,4 +109,3 @@ if __name__ == '__main__': visualize_curve(api201, save_dir, 'tss', args.max_time) api301 = NASBench301API(verbose=False) visualize_curve(api301, save_dir, 'sss', args.max_time) -