def account_one_arch(arch_index, arch_str, checkpoints, datasets, dataloader_dict): information = ArchResults(arch_index, arch_str) for checkpoint_path in checkpoints: checkpoint = torch.load(checkpoint_path, map_location='cpu') used_seed = checkpoint_path.name.split('-')[-1].split('.')[0] for dataset in datasets: assert dataset in checkpoint, 'Can not find {:} in arch-{:} from {:}'.format(dataset, arch_index, checkpoint_path) results = checkpoint[dataset] assert results['finish-train'], 'This {:} arch seed={:} does not finish train on {:} ::: {:}'.format(arch_index, used_seed, dataset, checkpoint_path) arch_config = {'channel': results['channel'], 'num_cells': results['num_cells'], 'arch_str': arch_str, 'class_num': results['config']['class_num']} xresult = ResultsCount(dataset, results['net_state_dict'], results['train_acc1es'], results['train_losses'], \ results['param'], results['flop'], arch_config, used_seed, results['total_epoch'], None) if dataset == 'cifar10-valid': xresult.update_eval('x-valid' , results['valid_acc1es'], results['valid_losses']) elif dataset == 'cifar10': xresult.update_eval('ori-test', results['valid_acc1es'], results['valid_losses']) elif dataset == 'cifar100' or dataset == 'ImageNet16-120': xresult.update_eval('ori-test', results['valid_acc1es'], results['valid_losses']) net_config = dict2config({'name': 'infer.tiny', 'C': arch_config['channel'], 'N': arch_config['num_cells'], 'genotype': CellStructure.str2structure(arch_config['arch_str']), 'num_classes':arch_config['class_num']}, None) network = get_cell_based_tiny_net(net_config) network.load_state_dict(xresult.get_net_param()) network = network.cuda() loss, top1, top5, latencies = pure_evaluate(dataloader_dict['{:}@{:}'.format(dataset, 'valid')], network) xresult.update_eval('x-valid', {results['total_epoch']-1: top1}, {results['total_epoch']-1: loss}) loss, top1, top5, latencies = pure_evaluate(dataloader_dict['{:}@{:}'.format(dataset, 'test')], network) xresult.update_eval('x-test' , {results['total_epoch']-1: top1}, {results['total_epoch']-1: loss}) xresult.update_latency(latencies) else: raise ValueError('invalid dataset name : {:}'.format(dataset)) information.update(dataset, int(used_seed), xresult) return information
def create_result_count(used_seed, dataset, arch_config, results, dataloader_dict): xresult = ResultsCount(dataset, results['net_state_dict'], results['train_acc1es'], results['train_losses'], \ results['param'], results['flop'], arch_config, used_seed, results['total_epoch'], None) net_config = dict2config( { 'name': 'infer.tiny', 'C': arch_config['channel'], 'N': arch_config['num_cells'], 'genotype': CellStructure.str2structure(arch_config['arch_str']), 'num_classes': arch_config['class_num'] }, None) network = get_cell_based_tiny_net(net_config) network.load_state_dict(xresult.get_net_param()) if 'train_times' in results: # new version xresult.update_train_info(results['train_acc1es'], results['train_acc5es'], results['train_losses'], results['train_times']) xresult.update_eval(results['valid_acc1es'], results['valid_losses'], results['valid_times']) else: if dataset == 'cifar10-valid': xresult.update_OLD_eval('x-valid', results['valid_acc1es'], results['valid_losses']) loss, top1, top5, latencies = pure_evaluate( dataloader_dict['{:}@{:}'.format('cifar10', 'test')], network.cuda()) xresult.update_OLD_eval('ori-test', {results['total_epoch'] - 1: top1}, {results['total_epoch'] - 1: loss}) xresult.update_latency(latencies) elif dataset == 'cifar10': xresult.update_OLD_eval('ori-test', results['valid_acc1es'], results['valid_losses']) loss, top1, top5, latencies = pure_evaluate( dataloader_dict['{:}@{:}'.format(dataset, 'test')], network.cuda()) xresult.update_latency(latencies) elif dataset == 'cifar100' or dataset == 'ImageNet16-120': xresult.update_OLD_eval('ori-test', results['valid_acc1es'], results['valid_losses']) loss, top1, top5, latencies = pure_evaluate( dataloader_dict['{:}@{:}'.format(dataset, 'valid')], network.cuda()) xresult.update_OLD_eval('x-valid', {results['total_epoch'] - 1: top1}, {results['total_epoch'] - 1: loss}) loss, top1, top5, latencies = pure_evaluate( dataloader_dict['{:}@{:}'.format(dataset, 'test')], network.cuda()) xresult.update_OLD_eval('x-test', {results['total_epoch'] - 1: top1}, {results['total_epoch'] - 1: loss}) xresult.update_latency(latencies) else: raise ValueError('invalid dataset name : {:}'.format(dataset)) return xresult
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 {:} <finish> other procedures need {:}'.format(_is, len(seeds), seed, need_time)) logger.close()
def check_unique_arch(meta_file): api = API(str(meta_file)) arch_strs = deepcopy(api.meta_archs) xarchs = [CellStructure.str2structure(x) for x in arch_strs] def get_unique_matrix(archs, consider_zero): UniquStrs = [arch.to_unique_str(consider_zero) for arch in archs] print("{:} create unique-string ({:}/{:}) done".format( time_string(), len(set(UniquStrs)), len(UniquStrs))) Unique2Index = dict() for index, xstr in enumerate(UniquStrs): if xstr not in Unique2Index: Unique2Index[xstr] = list() Unique2Index[xstr].append(index) sm_matrix = torch.eye(len(archs)).bool() for _, xlist in Unique2Index.items(): for i in xlist: for j in xlist: sm_matrix[i, j] = True unique_ids, unique_num = [-1 for _ in archs], 0 for i in range(len(unique_ids)): if unique_ids[i] > -1: continue neighbours = sm_matrix[i].nonzero().view(-1).tolist() for nghb in neighbours: assert unique_ids[nghb] == -1, "impossible" unique_ids[nghb] = unique_num unique_num += 1 return sm_matrix, unique_ids, unique_num print("There are {:} valid-archs".format( sum(arch.check_valid() for arch in xarchs))) sm_matrix, uniqueIDs, unique_num = get_unique_matrix(xarchs, None) print( "{:} There are {:} unique architectures (considering nothing).".format( time_string(), unique_num)) sm_matrix, uniqueIDs, unique_num = get_unique_matrix(xarchs, False) print("{:} There are {:} unique architectures (not considering zero).". format(time_string(), unique_num)) sm_matrix, uniqueIDs, unique_num = get_unique_matrix(xarchs, True) print("{:} There are {:} unique architectures (considering zero).".format( time_string(), unique_num))
def test_issue_81_82(api): results = api.query_by_index(0, 'cifar10-valid', hp='12') results = api.query_by_index(0, 'cifar10-valid', hp='200') print(list(results.keys())) print(results[888].get_eval('valid')) print(results[888].get_eval('x-valid')) result_dict = api.get_more_info(index=0, dataset='cifar10-valid', iepoch=11, hp='200', is_random=False) info = api.query_by_arch( '|nor_conv_3x3~0|+|skip_connect~0|nor_conv_3x3~1|+|skip_connect~0|none~1|nor_conv_3x3~2|', '200') print(info) structure = CellStructure.str2structure( '|nor_conv_3x3~0|+|skip_connect~0|nor_conv_3x3~1|+|skip_connect~0|none~1|nor_conv_3x3~2|' ) info = api.query_by_arch(structure, '200') print(info)
def main(save_dir, workers, datasets, xpaths, splits, use_less, srange, arch_index, seeds, cover_mode, meta_info, arch_config): assert torch.cuda.is_available(), 'CUDA is not available.' torch.backends.cudnn.enabled = True #torch.backends.cudnn.benchmark = True torch.backends.cudnn.deterministic = True torch.set_num_threads( workers ) assert len(srange) == 2 and 0 <= srange[0] <= srange[1], 'invalid srange : {:}'.format(srange) if use_less: sub_dir = Path(save_dir) / '{:06d}-{:06d}-C{:}-N{:}-LESS'.format(srange[0], srange[1], arch_config['channel'], arch_config['num_cells']) else: sub_dir = Path(save_dir) / '{:06d}-{:06d}-C{:}-N{:}'.format(srange[0], srange[1], arch_config['channel'], arch_config['num_cells']) logger = Logger(str(sub_dir), 0, False) all_archs = meta_info['archs'] assert srange[1] < meta_info['total'], 'invalid range : {:}-{:} vs. {:}'.format(srange[0], srange[1], meta_info['total']) assert arch_index == -1 or srange[0] <= arch_index <= srange[1], 'invalid range : {:} vs. {:} vs. {:}'.format(srange[0], arch_index, srange[1]) if arch_index == -1: to_evaluate_indexes = list(range(srange[0], srange[1]+1)) else: to_evaluate_indexes = [arch_index] logger.log('xargs : seeds = {:}'.format(seeds)) logger.log('xargs : arch_index = {:}'.format(arch_index)) logger.log('xargs : cover_mode = {:}'.format(cover_mode)) logger.log('-'*100) logger.log('Start evaluating range =: {:06d} vs. {:06d} vs. {:06d} / {:06d} with cover-mode={:}'.format(srange[0], arch_index, srange[1], meta_info['total'], 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('--->>> architecture config : {:}'.format(arch_config)) start_time, epoch_time = time.time(), AverageMeter() for i, index in enumerate(to_evaluate_indexes): arch = all_archs[index] logger.log('\n{:} evaluate {:06d}/{:06d} ({:06d}/{:06d})-th architecture [seeds={:}] {:}'.format('-'*15, i, len(to_evaluate_indexes), index, meta_info['total'], seeds, '-'*15)) #logger.log('{:} {:} {:}'.format('-'*15, arch.tostr(), '-'*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 = sub_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, use_less, seed, \ arch_config, workers, logger) torch.save(results, to_save_name) logger.log('{:} --evaluate-- {:06d}/{:06d} ({:06d}/{:06d})-th seed={:} done, save into {:}'.format('-'*15, i, len(to_evaluate_indexes), index, meta_info['total'], seed, 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, meta_info['total'], 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 {:} <finish> other procedures need {:}".format( _is, len(seeds), seed, need_time ) ) logger.close()
def create_result_count( used_seed: int, dataset: Text, arch_config: Dict[Text, Any], results: Dict[Text, Any], dataloader_dict: Dict[Text, Any], ) -> ResultsCount: xresult = ResultsCount( dataset, results["net_state_dict"], results["train_acc1es"], results["train_losses"], results["param"], results["flop"], arch_config, used_seed, results["total_epoch"], None, ) net_config = dict2config( { "name": "infer.tiny", "C": arch_config["channel"], "N": arch_config["num_cells"], "genotype": CellStructure.str2structure(arch_config["arch_str"]), "num_classes": arch_config["class_num"], }, None, ) if "train_times" in results: # new version xresult.update_train_info( results["train_acc1es"], results["train_acc5es"], results["train_losses"], results["train_times"], ) xresult.update_eval(results["valid_acc1es"], results["valid_losses"], results["valid_times"]) else: network = get_cell_based_tiny_net(net_config) network.load_state_dict(xresult.get_net_param()) if dataset == "cifar10-valid": xresult.update_OLD_eval("x-valid", results["valid_acc1es"], results["valid_losses"]) loss, top1, top5, latencies = pure_evaluate( dataloader_dict["{:}@{:}".format("cifar10", "test")], network.cuda()) xresult.update_OLD_eval( "ori-test", {results["total_epoch"] - 1: top1}, {results["total_epoch"] - 1: loss}, ) xresult.update_latency(latencies) elif dataset == "cifar10": xresult.update_OLD_eval("ori-test", results["valid_acc1es"], results["valid_losses"]) loss, top1, top5, latencies = pure_evaluate( dataloader_dict["{:}@{:}".format(dataset, "test")], network.cuda()) xresult.update_latency(latencies) elif dataset == "cifar100" or dataset == "ImageNet16-120": xresult.update_OLD_eval("ori-test", results["valid_acc1es"], results["valid_losses"]) loss, top1, top5, latencies = pure_evaluate( dataloader_dict["{:}@{:}".format(dataset, "valid")], network.cuda()) xresult.update_OLD_eval( "x-valid", {results["total_epoch"] - 1: top1}, {results["total_epoch"] - 1: loss}, ) loss, top1, top5, latencies = pure_evaluate( dataloader_dict["{:}@{:}".format(dataset, "test")], network.cuda()) xresult.update_OLD_eval( "x-test", {results["total_epoch"] - 1: top1}, {results["total_epoch"] - 1: loss}, ) xresult.update_latency(latencies) else: raise ValueError("invalid dataset name : {:}".format(dataset)) return xresult