############################################################################## # NATS-Bench: Benchmarking NAS algorithms for Architecture Topology and Size # ############################################################################## # Copyright (c) Xuanyi Dong [GitHub D-X-Y], 2020.07 # ############################################################################## # This file is used to train (all) architecture candidate in the topology # # search space in NATS-Bench (tss) with different hyper-parameters. # # When use mode=new, it will automatically detect whether the checkpoint of # # a trial exists, if so, it will skip this trial. When use mode=cover, it # # will ignore the (possible) existing checkpoint, run each trial, and save. # ############################################################################## # Please use the script of scripts/NATS-Bench/train-topology.sh to run. # # bash scripts/NATS-Bench/train-topology.sh 00000-15624 12 777 # # bash scripts/NATS-Bench/train-topology.sh 00000-15624 200 '777 888 999' # # # ################ # # [Deprecated Function: Generate the meta information] # # python ./exps/NATS-Bench/main-tss.py --mode meta # ############################################################################## import os, sys, time, torch, random, argparse from typing import List, Text, Dict, Any from PIL import ImageFile ImageFile.LOAD_TRUNCATED_IMAGES = True from copy import deepcopy from pathlib import Path lib_dir = (Path(__file__).parent / '..' / '..' / 'lib').resolve() if str(lib_dir) not in sys.path: sys.path.insert(0, str(lib_dir)) from config_utils import dict2config, load_config from procedures import bench_evaluate_for_seed from procedures import get_machine_info from datasets import get_datasets from log_utils import Logger, AverageMeter, time_string, convert_secs2time from models import CellStructure, CellArchitectures, get_search_spaces from utils import split_str2indexes def evaluate_all_datasets(arch: Text, datasets: List[Text], xpaths: List[Text], splits: List[Text], config_path: Text, seed: int, raw_arch_config, workers, logger): machine_info, raw_arch_config = get_machine_info(), deepcopy(raw_arch_config) all_infos = {'info': machine_info} all_dataset_keys = [] # look all the datasets for dataset, xpath, split in zip(datasets, xpaths, splits): # train valid data train_data, valid_data, xshape, class_num = get_datasets(dataset, xpath, -1) # load the configuration if dataset == 'cifar10' or dataset == 'cifar100': split_info = load_config('configs/nas-benchmark/cifar-split.txt', None, None) elif dataset.startswith('ImageNet16'): split_info = load_config('configs/nas-benchmark/{:}-split.txt'.format(dataset), None, None) else: raise ValueError('invalid dataset : {:}'.format(dataset)) config = load_config(config_path, dict(class_num=class_num, xshape=xshape), logger) # check whether use splited validation set if bool(split): assert dataset == 'cifar10' ValLoaders = {'ori-test': torch.utils.data.DataLoader(valid_data, batch_size=config.batch_size, shuffle=False, num_workers=workers, pin_memory=True)} assert len(train_data) == len(split_info.train) + len(split_info.valid), 'invalid length : {:} vs {:} + {:}'.format(len(train_data), len(split_info.train), len(split_info.valid)) train_data_v2 = deepcopy(train_data) train_data_v2.transform = valid_data.transform valid_data = train_data_v2 # data loader train_loader = torch.utils.data.DataLoader(train_data, batch_size=config.batch_size, sampler=torch.utils.data.sampler.SubsetRandomSampler(split_info.train), num_workers=workers, pin_memory=True) valid_loader = torch.utils.data.DataLoader(valid_data, batch_size=config.batch_size, sampler=torch.utils.data.sampler.SubsetRandomSampler(split_info.valid), num_workers=workers, pin_memory=True) ValLoaders['x-valid'] = valid_loader else: # data loader train_loader = torch.utils.data.DataLoader(train_data, batch_size=config.batch_size, shuffle=True , num_workers=workers, pin_memory=True) valid_loader = torch.utils.data.DataLoader(valid_data, batch_size=config.batch_size, shuffle=False, num_workers=workers, pin_memory=True) if dataset == 'cifar10': ValLoaders = {'ori-test': valid_loader} elif dataset == 'cifar100': cifar100_splits = load_config('configs/nas-benchmark/cifar100-test-split.txt', None, None) ValLoaders = {'ori-test': valid_loader, 'x-valid' : torch.utils.data.DataLoader(valid_data, batch_size=config.batch_size, sampler=torch.utils.data.sampler.SubsetRandomSampler(cifar100_splits.xvalid), num_workers=workers, pin_memory=True), 'x-test' : torch.utils.data.DataLoader(valid_data, batch_size=config.batch_size, sampler=torch.utils.data.sampler.SubsetRandomSampler(cifar100_splits.xtest ), num_workers=workers, pin_memory=True) } elif dataset == 'ImageNet16-120': imagenet16_splits = load_config('configs/nas-benchmark/imagenet-16-120-test-split.txt', None, None) ValLoaders = {'ori-test': valid_loader, 'x-valid' : torch.utils.data.DataLoader(valid_data, batch_size=config.batch_size, sampler=torch.utils.data.sampler.SubsetRandomSampler(imagenet16_splits.xvalid), num_workers=workers, pin_memory=True), 'x-test' : torch.utils.data.DataLoader(valid_data, batch_size=config.batch_size, sampler=torch.utils.data.sampler.SubsetRandomSampler(imagenet16_splits.xtest ), num_workers=workers, pin_memory=True) } else: raise ValueError('invalid dataset : {:}'.format(dataset)) dataset_key = '{:}'.format(dataset) if bool(split): dataset_key = dataset_key + '-valid' logger.log('Evaluate ||||||| {:10s} ||||||| Train-Num={:}, Valid-Num={:}, Train-Loader-Num={:}, Valid-Loader-Num={:}, batch size={:}'.format(dataset_key, len(train_data), len(valid_data), len(train_loader), len(valid_loader), config.batch_size)) logger.log('Evaluate ||||||| {:10s} ||||||| Config={:}'.format(dataset_key, config)) for key, value in ValLoaders.items(): logger.log('Evaluate ---->>>> {:10s} with {:} batchs'.format(key, len(value))) arch_config = dict2config(dict(name='infer.tiny', C=raw_arch_config['channel'], N=raw_arch_config['num_cells'], genotype=arch, num_classes=config.class_num), None) results = bench_evaluate_for_seed(arch_config, config, train_loader, ValLoaders, seed, logger) all_infos[dataset_key] = results all_dataset_keys.append( dataset_key ) all_infos['all_dataset_keys'] = all_dataset_keys return all_infos def main(save_dir: Path, workers: int, datasets: List[Text], xpaths: List[Text], splits: List[int], seeds: List[int], nets: List[str], opt_config: Dict[Text, Any], to_evaluate_indexes: tuple, cover_mode: bool, arch_config: Dict[Text, Any]): log_dir = save_dir / 'logs' log_dir.mkdir(parents=True, exist_ok=True) logger = Logger(str(log_dir), os.getpid(), False) logger.log('xargs : seeds = {:}'.format(seeds)) logger.log('xargs : cover_mode = {:}'.format(cover_mode)) logger.log('-' * 100) logger.log( 'Start evaluating range =: {:06d} - {:06d}'.format(min(to_evaluate_indexes), max(to_evaluate_indexes)) +'({:} in total) / {:06d} with cover-mode={:}'.format(len(to_evaluate_indexes), len(nets), cover_mode)) for i, (dataset, xpath, split) in enumerate(zip(datasets, xpaths, splits)): logger.log( '--->>> Evaluate {:}/{:} : dataset={:9s}, path={:}, split={:}'.format(i, len(datasets), dataset, xpath, split)) logger.log('--->>> optimization config : {:}'.format(opt_config)) start_time, epoch_time = time.time(), AverageMeter() for i, index in enumerate(to_evaluate_indexes): arch = nets[index] logger.log('\n{:} evaluate {:06d}/{:06d} ({:06d}/{:06d})-th arch [seeds={:}] {:}'.format(time_string(), i, len(to_evaluate_indexes), index, len(nets), seeds, '-' * 15)) logger.log('{:} {:} {:}'.format('-' * 15, arch, '-' * 15)) # test this arch on different datasets with different seeds has_continue = False for seed in seeds: to_save_name = save_dir / 'arch-{:06d}-seed-{:04d}.pth'.format(index, seed) if to_save_name.exists(): if cover_mode: logger.log('Find existing file : {:}, remove it before evaluation'.format(to_save_name)) os.remove(str(to_save_name)) else: logger.log('Find existing file : {:}, skip this evaluation'.format(to_save_name)) has_continue = True continue results = evaluate_all_datasets(CellStructure.str2structure(arch), datasets, xpaths, splits, opt_config, seed, arch_config, workers, logger) torch.save(results, to_save_name) logger.log('\n{:} evaluate {:06d}/{:06d} ({:06d}/{:06d})-th arch [seeds={:}] ===>>> {:}'.format(time_string(), i, len(to_evaluate_indexes), index, len(nets), seeds, to_save_name)) # measure elapsed time if not has_continue: epoch_time.update(time.time() - start_time) start_time = time.time() need_time = 'Time Left: {:}'.format(convert_secs2time(epoch_time.avg * (len(to_evaluate_indexes)-i-1), True) ) logger.log('This arch costs : {:}'.format(convert_secs2time(epoch_time.val, True) )) logger.log('{:}'.format('*' * 100)) logger.log('{:} {:74s} {:}'.format('*' * 10, '{:06d}/{:06d} ({:06d}/{:06d})-th done, left {:}'.format(i, len( to_evaluate_indexes), index, len(nets), need_time), '*' * 10)) logger.log('{:}'.format('*' * 100)) logger.close() def train_single_model(save_dir, workers, datasets, xpaths, splits, use_less, seeds, model_str, arch_config): assert torch.cuda.is_available(), 'CUDA is not available.' torch.backends.cudnn.enabled = True torch.backends.cudnn.deterministic = True #torch.backends.cudnn.benchmark = True torch.set_num_threads( workers ) save_dir = Path(save_dir) / 'specifics' / '{:}-{:}-{:}-{:}'.format('LESS' if use_less else 'FULL', model_str, arch_config['channel'], arch_config['num_cells']) logger = Logger(str(save_dir), 0, False) if model_str in CellArchitectures: arch = CellArchitectures[model_str] logger.log('The model string is found in pre-defined architecture dict : {:}'.format(model_str)) else: try: arch = CellStructure.str2structure(model_str) except: raise ValueError('Invalid model string : {:}. It can not be found or parsed.'.format(model_str)) assert arch.check_valid_op(get_search_spaces('cell', 'full')), '{:} has the invalid op.'.format(arch) logger.log('Start train-evaluate {:}'.format(arch.tostr())) logger.log('arch_config : {:}'.format(arch_config)) start_time, seed_time = time.time(), AverageMeter() for _is, seed in enumerate(seeds): logger.log('\nThe {:02d}/{:02d}-th seed is {:} ----------------------<.>----------------------'.format(_is, len(seeds), seed)) to_save_name = save_dir / 'seed-{:04d}.pth'.format(seed) if to_save_name.exists(): logger.log('Find the existing file {:}, directly load!'.format(to_save_name)) checkpoint = torch.load(to_save_name) else: logger.log('Does not find the existing file {:}, train and evaluate!'.format(to_save_name)) checkpoint = evaluate_all_datasets(arch, datasets, xpaths, splits, use_less, seed, arch_config, workers, logger) torch.save(checkpoint, to_save_name) # log information logger.log('{:}'.format(checkpoint['info'])) all_dataset_keys = checkpoint['all_dataset_keys'] for dataset_key in all_dataset_keys: logger.log('\n{:} dataset : {:} {:}'.format('-'*15, dataset_key, '-'*15)) dataset_info = checkpoint[dataset_key] #logger.log('Network ==>\n{:}'.format( dataset_info['net_string'] )) logger.log('Flops = {:} MB, Params = {:} MB'.format(dataset_info['flop'], dataset_info['param'])) logger.log('config : {:}'.format(dataset_info['config'])) logger.log('Training State (finish) = {:}'.format(dataset_info['finish-train'])) last_epoch = dataset_info['total_epoch'] - 1 train_acc1es, train_acc5es = dataset_info['train_acc1es'], dataset_info['train_acc5es'] valid_acc1es, valid_acc5es = dataset_info['valid_acc1es'], dataset_info['valid_acc5es'] logger.log('Last Info : Train = Acc@1 {:.2f}% Acc@5 {:.2f}% Error@1 {:.2f}%, Test = Acc@1 {:.2f}% Acc@5 {:.2f}% Error@1 {:.2f}%'.format(train_acc1es[last_epoch], train_acc5es[last_epoch], 100-train_acc1es[last_epoch], valid_acc1es[last_epoch], valid_acc5es[last_epoch], 100-valid_acc1es[last_epoch])) # measure elapsed time seed_time.update(time.time() - start_time) start_time = time.time() need_time = 'Time Left: {:}'.format( convert_secs2time(seed_time.avg * (len(seeds)-_is-1), True) ) logger.log('\n<<<***>>> The {:02d}/{:02d}-th seed is {:} other procedures need {:}'.format(_is, len(seeds), seed, need_time)) logger.close() def generate_meta_info(save_dir, max_node, divide=40): aa_nas_bench_ss = get_search_spaces('cell', 'nas-bench-201') archs = CellStructure.gen_all(aa_nas_bench_ss, max_node, False) print ('There are {:} archs vs {:}.'.format(len(archs), len(aa_nas_bench_ss) ** ((max_node-1)*max_node/2))) random.seed( 88 ) # please do not change this line for reproducibility random.shuffle( archs ) # to test fixed-random shuffle #print ('arch [0] : {:}\n---->>>> {:}'.format( archs[0], archs[0].tostr() )) #print ('arch [9] : {:}\n---->>>> {:}'.format( archs[9], archs[9].tostr() )) assert archs[0 ].tostr() == '|avg_pool_3x3~0|+|nor_conv_1x1~0|skip_connect~1|+|nor_conv_1x1~0|skip_connect~1|skip_connect~2|', 'please check the 0-th architecture : {:}'.format(archs[0]) assert archs[9 ].tostr() == '|avg_pool_3x3~0|+|none~0|none~1|+|skip_connect~0|none~1|nor_conv_3x3~2|', 'please check the 9-th architecture : {:}'.format(archs[9]) assert archs[123].tostr() == '|avg_pool_3x3~0|+|avg_pool_3x3~0|nor_conv_1x1~1|+|none~0|avg_pool_3x3~1|nor_conv_3x3~2|', 'please check the 123-th architecture : {:}'.format(archs[123]) total_arch = len(archs) num = 50000 indexes_5W = list(range(num)) random.seed( 1021 ) random.shuffle( indexes_5W ) train_split = sorted( list(set(indexes_5W[:num//2])) ) valid_split = sorted( list(set(indexes_5W[num//2:])) ) assert len(train_split) + len(valid_split) == num assert train_split[0] == 0 and train_split[10] == 26 and train_split[111] == 203 and valid_split[0] == 1 and valid_split[10] == 18 and valid_split[111] == 242, '{:} {:} {:} - {:} {:} {:}'.format(train_split[0], train_split[10], train_split[111], valid_split[0], valid_split[10], valid_split[111]) splits = {num: {'train': train_split, 'valid': valid_split} } info = {'archs' : [x.tostr() for x in archs], 'total' : total_arch, 'max_node' : max_node, 'splits': splits} save_dir = Path(save_dir) save_dir.mkdir(parents=True, exist_ok=True) save_name = save_dir / 'meta-node-{:}.pth'.format(max_node) assert not save_name.exists(), '{:} already exist'.format(save_name) torch.save(info, save_name) print ('save the meta file into {:}'.format(save_name)) def traverse_net(max_node): aa_nas_bench_ss = get_search_spaces('cell', 'nats-bench') archs = CellStructure.gen_all(aa_nas_bench_ss, max_node, False) print ('There are {:} archs vs {:}.'.format(len(archs), len(aa_nas_bench_ss) ** ((max_node-1)*max_node/2))) random.seed( 88 ) # please do not change this line for reproducibility random.shuffle( archs ) assert archs[0 ].tostr() == '|avg_pool_3x3~0|+|nor_conv_1x1~0|skip_connect~1|+|nor_conv_1x1~0|skip_connect~1|skip_connect~2|', 'please check the 0-th architecture : {:}'.format(archs[0]) assert archs[9 ].tostr() == '|avg_pool_3x3~0|+|none~0|none~1|+|skip_connect~0|none~1|nor_conv_3x3~2|', 'please check the 9-th architecture : {:}'.format(archs[9]) assert archs[123].tostr() == '|avg_pool_3x3~0|+|avg_pool_3x3~0|nor_conv_1x1~1|+|none~0|avg_pool_3x3~1|nor_conv_3x3~2|', 'please check the 123-th architecture : {:}'.format(archs[123]) return [x.tostr() for x in archs] def filter_indexes(xlist, mode, save_dir, seeds): all_indexes = [] for index in xlist: if mode == 'cover': all_indexes.append(index) else: for seed in seeds: temp_path = save_dir / 'arch-{:06d}-seed-{:04d}.pth'.format(index, seed) if not temp_path.exists(): all_indexes.append(index) break print('{:} [FILTER-INDEXES] : there are {:}/{:} architectures in total'.format(time_string(), len(all_indexes), len(xlist))) return all_indexes if __name__ == '__main__': # mode_choices = ['meta', 'new', 'cover'] + ['specific-{:}'.format(_) for _ in CellArchitectures.keys()] parser = argparse.ArgumentParser(description='NATS-Bench (topology search space)', formatter_class=argparse.ArgumentDefaultsHelpFormatter) parser.add_argument('--mode' , type=str, required=True, help='The script mode.') parser.add_argument('--save_dir', type=str, default='output/NATS-Bench-topology', help='Folder to save checkpoints and log.') parser.add_argument('--max_node', type=int, default=4, help='The maximum node in a cell (please do not change it).') # use for train the model parser.add_argument('--workers', type=int, default=8, help='number of data loading workers (default: 2)') parser.add_argument('--srange' , type=str, required=True, help='The range of models to be evaluated') parser.add_argument('--datasets', type=str, nargs='+', help='The applied datasets.') parser.add_argument('--xpaths', type=str, nargs='+', help='The root path for this dataset.') parser.add_argument('--splits', type=int, nargs='+', help='The root path for this dataset.') parser.add_argument('--hyper', type=str, default='12', choices=['01', '12', '200'], help='The tag for hyper-parameters.') parser.add_argument('--seeds' , type=int, nargs='+', help='The range of models to be evaluated') parser.add_argument('--channel', type=int, default=16, help='The number of channels.') parser.add_argument('--num_cells', type=int, default=5, help='The number of cells in one stage.') parser.add_argument('--check_N', type=int, default=15625, help='For safety.') args = parser.parse_args() assert args.mode in ['meta', 'new', 'cover'] or args.mode.startswith('specific-'), 'invalid mode : {:}'.format(args.mode) if args.mode == 'meta': generate_meta_info(args.save_dir, args.max_node) elif args.mode.startswith('specific'): assert len(args.mode.split('-')) == 2, 'invalid mode : {:}'.format(args.mode) model_str = args.mode.split('-')[1] train_single_model(args.save_dir, args.workers, args.datasets, args.xpaths, args.splits, args.use_less>0, \ tuple(args.seeds), model_str, {'channel': args.channel, 'num_cells': args.num_cells}) else: nets = traverse_net(args.max_node) if len(nets) != args.check_N: raise ValueError('Pre-num-check failed : {:} vs {:}'.format(len(nets), args.check_N)) opt_config = './configs/nas-benchmark/hyper-opts/{:}E.config'.format(args.hyper) if not os.path.isfile(opt_config): raise ValueError('{:} is not a file.'.format(opt_config)) save_dir = Path(args.save_dir) / 'raw-data-{:}'.format(args.hyper) save_dir.mkdir(parents=True, exist_ok=True) to_evaluate_indexes = split_str2indexes(args.srange, args.check_N, 5) if not len(args.seeds): raise ValueError('invalid length of seeds args: {:}'.format(args.seeds)) if not (len(args.datasets) == len(args.xpaths) == len(args.splits)): raise ValueError('invalid infos : {:} vs {:} vs {:}'.format(len(args.datasets), len(args.xpaths), len(args.splits))) if args.workers <= 0: raise ValueError('invalid number of workers : {:}'.format(args.workers)) target_indexes = filter_indexes(to_evaluate_indexes, args.mode, save_dir, args.seeds) assert torch.cuda.is_available(), 'CUDA is not available.' torch.backends.cudnn.enabled = True torch.backends.cudnn.deterministic = True torch.set_num_threads(args.workers) main(save_dir, args.workers, args.datasets, args.xpaths, args.splits, tuple(args.seeds), nets, opt_config, target_indexes, args.mode == 'cover', \ {'name': 'infer.tiny', 'channel': args.channel, 'num_cells': args.num_cells})