def evaluate(xargs): # start cudnn cudnn.enabled = True # make each conv is the same cudnn.benchmark = False # make sure the same seed has the same result cudnn.deterministic = True torch.set_num_threads(xargs.workers) prepare_seed(xargs.rand_seed) xargs.save_dir += 'eva/' logger = prepare_logger(xargs) train_data, valid_data, xshape, class_num = get_dataset( xargs.dataset, xargs.data_path, -1) logger.log('Train Config:') eva_config = load_config( xargs.eva_config, { 'class_num': class_num, 'xshape': xshape, 'tau_max': xargs.tau_max, 'tau_min': xargs.tau_min }, logger) search_loader, train_loader, test_loader = get_nas_search_loaders( train_data, valid_data, xargs.dataset, 'config/', eva_config.batch_size, xargs.workers) logger.log('dataset: {:} Train-Loader-length={:}, batch size={:}'.format( xargs.dataset, len(train_loader), eva_config.batch_size)) eva_dense_net(search_loader, test_loader, eva_config, logger, 'normal')
def train(xargs): # start cudnn cudnn.enabled = True # make each conv is the same cudnn.benchmark = False # make sure the same seed has the same result cudnn.deterministic = True torch.set_num_threads(xargs.workers) prepare_seed(xargs.rand_seed) logger = prepare_logger(xargs) # get original data(cifar10/cifar100/uci) train_data, valid_data, xshape, class_num = get_dataset(xargs.dataset, xargs.data_path, -1) logger.log('{:}Train Config{:}'.format("-" * 50, "-" * 50)) opt_config = load_config(xargs.opt_config, {'class_num': class_num, 'xshape': xshape, 'batch_size': xargs.batch_size, 'epochs': xargs.epochs, 'LR': xargs.opt_learning_rate}, logger) _, train_loader, _ = get_nas_search_loaders(train_data, valid_data, xargs.dataset, 'config/', opt_config.batch_size, xargs.workers) logger.log('dataset: {:} Train-Loader-length={:}, batch size={:}'.format(xargs.dataset, len(train_loader), opt_config.batch_size))
def get_nas_search_loaders(train_data, valid_data, dataset, config_root, batch_size, workers): # get search_loader, train_loader, valid_loader if isinstance(batch_size, (list, tuple)): batch, test_batch = batch_size else: batch, test_batch = batch_size, batch_size if dataset == 'cifar10': cifar_split = load_config('{:}/cifar10-split.txt'.format(config_root), None, None) train_split, valid_split = cifar_split.train, cifar_split.valid # search over the proposed training and validation set # To split data xvalid_data = deepcopy(train_data) if hasattr(xvalid_data, 'transforms'): # to avoid a print issue xvalid_data.transforms = valid_data.transform xvalid_data.transform = deepcopy(valid_data.transform) search_data = SearchDataset(dataset, train_data, train_split, valid_split) # data loader search_loader = DataLoader(search_data, batch_size=batch, shuffle=True, num_workers=workers, pin_memory=True) train_loader = DataLoader( train_data, batch_size=batch, sampler=sampler.SubsetRandomSampler(train_split), num_workers=workers, pin_memory=True) valid_loader = DataLoader( xvalid_data, batch_size=test_batch, sampler=sampler.SubsetRandomSampler(valid_split), num_workers=workers, pin_memory=True) elif dataset == 'cifar100': cifar100_test_split = load_config( '{:}/cifar100-split.txt'.format(config_root), None, None) search_train_data = train_data search_valid_data = deepcopy(valid_data) search_valid_data.transform = train_data.transform search_data = SearchDataset(dataset, [search_train_data, search_valid_data], list(range(len(search_train_data))), cifar100_test_split.xvalid) search_loader = DataLoader(search_data, batch_size=batch, shuffle=True, num_workers=workers, pin_memory=True) train_loader = DataLoader(train_data, batch_size=batch, shuffle=True, num_workers=workers, pin_memory=True) valid_loader = DataLoader(valid_data, batch_size=test_batch, sampler=sampler.SubsetRandomSampler( cifar100_test_split.xvalid), num_workers=workers, pin_memory=True) elif dataset == 'HAPT': HAPT_split = load_config('{:}HAPT-split.txt'.format(config_root), None, None) train_split, valid_split = HAPT_split.train, HAPT_split.valid search_data = GenDataset(dataset, train_data, train_split, valid_split) search_loader = DataLoader(search_data, batch_size=batch, shuffle=True, num_workers=workers, pin_memory=True) train_loader = DataLoader(NormalDataset(dataset, train_data), batch_size=batch, shuffle=True, num_workers=workers, pin_memory=True) valid_loader = DataLoader(NormalDataset(dataset, valid_data), batch_size=batch, shuffle=False, num_workers=workers, pin_memory=True) else: raise ValueError('invalid dataset : {:}'.format(dataset)) return search_loader, train_loader, valid_loader
def train(xargs): lib_dir = (Path(__file__).parent / '..' / '..' / 'lib').resolve() if str(lib_dir) not in sys.path: sys.path.insert(0, str(lib_dir)) assert (torch.cuda.is_available(), 'CUDA is not available.') # start cudnn cudnn.enabled = True # make each conv is the same cudnn.benchmark = False # make sure the same seed has the same result cudnn.deterministic = True torch.set_num_threads(xargs.workers) prepare_seed(xargs.rand_seed) logger = prepare_logger(xargs) # get original data(cifar10/cifar100/uci) train_data, valid_data, xshape, class_num = get_dataset( xargs.dataset, xargs.data_path, -1) logger.log('{:}Train Config{:}'.format("-" * 50, "-" * 50)) opt_config = load_config( xargs.opt_config, { 'class_num': class_num, 'xshape': xshape, 'batch_size': xargs.batch_size, 'epochs': xargs.epochs, 'LR': xargs.opt_learning_rate }, logger) search_loader, _, valid_loader = get_nas_search_loaders( train_data, valid_data, xargs.dataset, 'config/', opt_config.batch_size, xargs.workers) logger.log('dataset: {:} Search-Loader-length={:}, batch size={:}'.format( xargs.dataset, len(search_loader), opt_config.batch_size)) logger.log('{:}Arch Config{:}'.format("-" * 50, "-" * 50)) arch_config = load_config( xargs.arch_config, { 'class_num': class_num, 'space': HAPT_SPACE, 'affine': False, 'track_running_stats': bool(xargs.track_running_stats) }, logger) if xargs.dataset == 'HAPT': search_model = DNNModel(config=arch_config, logger=logger) elif xargs.dataset in ('cifar10', 'cifar100'): search_model = DNNModel(config=arch_config, logger=logger) else: raise NameError( "dataset must be in \"HAPT\", \"cifar100\", \"cifar100\"") if xargs.evaluate == 'test': search_model.load_state_dict( torch.load(logger.path('best'))['network']) network = search_model.cuda() test_loader = valid_loader test(test_loader, network, arch_config.C_out) return # logger.log('search-model :\n{:}'.format(search_model)) logger.log('{:}model-config{:}\n{:}'.format("-" * 50, "-" * 50, arch_config)) if opt_config.criterion == 'cross_entropy': criterion = nn.CrossEntropyLoss() else: raise NameError('unknown loss function {:}'.format( opt_config.criterion)) # criterion = nn.MSELoss() w_optimizer = torch.optim.SGD(params=search_model.get_weights(), lr=opt_config.LR, weight_decay=opt_config.w_decay) w_scheduler = torch.optim.lr_scheduler.CosineAnnealingLR( optimizer=w_optimizer, T_max=opt_config.epochs, eta_min=opt_config.eta_min) a_optimizer = torch.optim.Adam(params=search_model.get_alphas(), lr=xargs.arch_learning_rate, betas=(0.5, 0.999), weight_decay=opt_config.a_decay) logger.log('{:}w-optimizer{:}\n{:}'.format("-" * 50, "-" * 50, w_optimizer)) logger.log('{:}a-optimizer{:}\n{:}'.format("-" * 50, "-" * 50, a_optimizer)) logger.log('{:}w-scheduler{:}\n{:}'.format("-" * 50, "-" * 50, w_scheduler)) logger.log('{:}criterion{:}\n{:}'.format("-" * 50, "-" * 50, criterion)) flop, param = get_model_infos(search_model, xshape) logger.log('FLOP = {:.6f} M, Params = {:.6f} MB'.format(flop, param)) logger.log('search-space [{:} ops] : {:}'.format(len(HAPT_SPACE), HAPT_SPACE)) last_info, model_base_path, model_best_path = logger.path( 'info'), logger.path('model'), logger.path('best') # network, criterion = torch.nn.DataParallel(search_model).cuda(), criterion.cuda() network = search_model.cuda() criterion = criterion.cuda() if last_info.exists(): # automatically resume from previous checkpoint logger.log("=> loading checkpoint of the last-info '{:}' start".format( last_info)) last_info = torch.load(last_info) start_epoch = last_info['epoch'] checkpoint = torch.load(last_info['last_checkpoint']) genotypes = checkpoint['genotypes'] valid_accuracies = checkpoint['valid_accuracies'] network.load_state_dict(checkpoint['network']) w_scheduler.load_state_dict(checkpoint['w_scheduler']) w_optimizer.load_state_dict(checkpoint['w_optimizer']) a_optimizer.load_state_dict(checkpoint['a_optimizer']) logger.log( "=> loading checkpoint of the last-info '{:}' start with {:}-th epoch." .format(last_info, start_epoch)) else: logger.log("=> do not find the last-info file : {:}".format(last_info)) start_epoch, valid_accuracies, genotypes = 0, {'best': -1}, \ {-1: network.genotype(xargs.save_dir + xargs.genotype_file)} start_time, epoch_time = time.time(), AverageMeter() total_epoch = opt_config.epochs for epoch in range(start_epoch, total_epoch): w_scheduler.step(epoch=epoch) need_time = 'Time Left: {:}'.format( convert_secs2time(epoch_time.avg * (total_epoch - epoch), True)) epoch_str = '{:03d}-{:03d}'.format(epoch, total_epoch) network.set_tau(xargs.tau_max - (xargs.tau_max - xargs.tau_min) * epoch / (total_epoch - 1)) logger.log('\n[Search the {:}-th epoch] tau={:.2f} {:}'.format( epoch_str, network.get_tau(), need_time)) epoch_start = time.time() base_losses, base_top1, base_top5, arch_losses, arch_top1, arch_top5 = \ search(arch_config, search_loader, network, criterion, w_optimizer, a_optimizer, xargs.print_frequency, epoch_str, logger) epoch_time.update(time.time() - epoch_start) logger.log( '[{:}] searching : loss={:.2f}, accuracy@1={:.2f}%, accuracy@5={:.2f}%' .format(epoch_str, base_losses, base_top1, base_top5)) logger.log( '[{:}] evaluate : loss={:.2f}, accuracy@1={:.2f}%, accuracy@5={:.2f}%' .format(epoch_str, arch_losses, arch_top1, arch_top5)) valid_accuracies[epoch] = arch_top1 if arch_top1 > valid_accuracies['best']: valid_accuracies['best'] = arch_top1 genotypes['best'] = network.genotype(xargs.save_dir + xargs.genotype_file) find_best = True else: find_best = False genotypes[epoch] = network.genotype(xargs.save_dir + xargs.genotype_file) logger.log('<<<--->>> The {:}-th epoch : {:}'.format( epoch_str, genotypes[epoch])) save_path = save_checkpoint( { 'epoch': epoch + 1, 'args': deepcopy(xargs), 'network': network.state_dict(), 'w_optimizer': w_optimizer.state_dict(), 'a_optimizer': a_optimizer.state_dict(), 'w_scheduler': w_scheduler.state_dict(), 'genotypes': genotypes, 'valid_accuracies': valid_accuracies }, model_base_path, logger) last_info = save_checkpoint( { 'epoch': epoch + 1, 'args': deepcopy(xargs), 'last_checkpoint': save_path, }, logger.path('info'), logger) if find_best: logger.log( '<<<--->>> The {:}-th epoch : find the highest validation accuracy : {:.2f}%.' .format(epoch_str, arch_top1)) copy_checkpoint(model_base_path, model_best_path, logger) with torch.no_grad(): logger.log('{:}'.format(network.show_alphas())) logger.log('\n' + '-' * 100) # check the performance from the architecture dataset logger.log('GDAS : run {:} epochs, use time : {:}.'.format( total_epoch, convert_secs2time(time.time() - start_time, True))) logger.log('best geno is {:}.'.format(genotypes['best']))
def train(xargs): lib_dir = (Path(__file__).parent / '..' / '..' / 'lib').resolve() if str(lib_dir) not in sys.path: sys.path.insert(0, str(lib_dir)) assert (torch.cuda.is_available(), 'CUDA is not available.') # start cudnn cudnn.enabled = True # make each conv is the same cudnn.benchmark = False # make sure the same seed has the same result cudnn.deterministic = True torch.set_num_threads(xargs.workers) prepare_seed(xargs.rand_seed) logger = prepare_logger(xargs) # get original data train_data, valid_data, xshape, class_num = get_dataset( xargs.dataset, xargs.data_path, -1) config = load_config(xargs.config_path, { 'class_num': class_num, 'xshape': xshape }, logger) search_loader, _, valid_loader = get_nas_search_loaders( train_data, valid_data, xargs.dataset, 'config/', config.batch_size, xargs.workers) logger.log( '||||||| {:10s} ||||||| Search-Loader-Num={:}, batch size={:}'.format( xargs.dataset, len(search_loader), config.batch_size)) logger.log('||||||| {:10s} ||||||| Config={:}'.format( xargs.dataset, config)) search_space = DARTS_SPACE model_config = load_config( xargs.model_config, { 'num_classes': class_num, 'space': search_space, 'affine': False, 'track_running_stats': bool(xargs.track_running_stats) }, None) search_model = get_cell_based_tiny_net(model_config) # logger.log('search-model :\n{:}'.format(search_model)) # logger.log('model-config : {:}'.format(model_config)) w_optimizer, w_scheduler, criterion = get_optim_scheduler( search_model.get_weights(), config) a_optimizer = torch.optim.Adam(search_model.get_alphas(), lr=xargs.arch_learning_rate, betas=(0.5, 0.999), weight_decay=xargs.arch_weight_decay) # logger.log('w-optimizer : {:}'.format(w_optimizer)) # logger.log('a-optimizer : {:}'.format(a_optimizer)) # logger.log('w-scheduler : {:}'.format(w_scheduler)) # logger.log('criterion : {:}'.format(criterion)) # flop, param = get_model_infos(search_model, xshape) # logger.log('FLOP = {:.2f} M, Params = {:.2f} MB'.format(flop, param)) logger.log('search-space [{:} ops] : {:}'.format(len(search_space), search_space)) if xargs.arch_nas_dataset is None: api = None else: pass # api = API(xargs.arch_nas_dataset) last_info, model_base_path, model_best_path = logger.path( 'info'), logger.path('model'), logger.path('best') # network, criterion = torch.nn.DataParallel(search_model).cuda(), criterion.cuda() network = search_model if last_info.exists(): # automatically resume from previous checkpoint logger.log("=> loading checkpoint of the last-info '{:}' start".format( last_info)) last_info = torch.load(last_info) start_epoch = last_info['epoch'] checkpoint = torch.load(last_info['last_checkpoint']) genotypes = checkpoint['genotypes'] valid_accuracies = checkpoint['valid_accuracies'] search_model.load_state_dict(checkpoint['search_model']) w_scheduler.load_state_dict(checkpoint['w_scheduler']) w_optimizer.load_state_dict(checkpoint['w_optimizer']) a_optimizer.load_state_dict(checkpoint['a_optimizer']) logger.log( "=> loading checkpoint of the last-info '{:}' start with {:}-th epoch." .format(last_info, start_epoch)) else: logger.log("=> do not find the last-info file : {:}".format(last_info)) start_epoch, valid_accuracies, genotypes = 0, { 'best': -1 }, { -1: search_model.genotype() } start_time, search_time, epoch_time, total_epoch = time.time( ), AverageMeter(), AverageMeter(), config.epochs + config.warmup for epoch in range(start_epoch, total_epoch): w_scheduler.update(epoch, 0.0) need_time = 'Time Left: {:}'.format( convert_secs2time(epoch_time.val * (total_epoch - epoch), True)) epoch_str = '{:03d}-{:03d}'.format(epoch, total_epoch) search_model.set_tau(xargs.tau_max - (xargs.tau_max - xargs.tau_min) * epoch / (total_epoch - 1)) logger.log('\n[Search the {:}-th epoch] {:}, tau={:}, LR={:}'.format( epoch_str, need_time, search_model.get_tau(), min(w_scheduler.get_lr()))) search_w_loss, search_w_top1, search_w_top5, valid_a_loss, valid_a_top1, valid_a_top5 \ = search_func(search_loader, network, criterion, w_scheduler, w_optimizer, a_optimizer, epoch_str, xargs.print_freq, logger) search_time.update(time.time() - start_time) logger.log( '[{:}] searching : loss={:.2f}, accuracy@1={:.2f}%, accuracy@5={:.2f}%, time-cost={:.1f} s' .format(epoch_str, search_w_loss, search_w_top1, search_w_top5, search_time.sum)) logger.log( '[{:}] evaluate : loss={:.2f}, accuracy@1={:.2f}%, accuracy@5={:.2f}%' .format(epoch_str, valid_a_loss, valid_a_top1, valid_a_top5)) # check the best accuracy valid_accuracies[epoch] = valid_a_top1 if valid_a_top1 > valid_accuracies['best']: valid_accuracies['best'] = valid_a_top1 genotypes['best'] = search_model.genotype() find_best = True else: find_best = False genotypes[epoch] = search_model.genotype() logger.log('<<<--->>> The {:}-th epoch : {:}'.format( epoch_str, genotypes[epoch])) # save checkpoint save_path = save_checkpoint( { 'epoch': epoch + 1, 'args': deepcopy(xargs), 'search_model': search_model.state_dict(), 'w_optimizer': w_optimizer.state_dict(), 'a_optimizer': a_optimizer.state_dict(), 'w_scheduler': w_scheduler.state_dict(), 'genotypes': genotypes, 'valid_accuracies': valid_accuracies }, model_base_path, logger) last_info = save_checkpoint( { 'epoch': epoch + 1, 'args': deepcopy(xargs), 'last_checkpoint': save_path, }, logger.path('info'), logger) if find_best: logger.log( '<<<--->>> The {:}-th epoch : find the highest validation accuracy : {:.2f}%.' .format(epoch_str, valid_a_top1)) copy_checkpoint(model_base_path, model_best_path, logger) with torch.no_grad(): logger.log('{:}'.format(search_model.show_alphas())) # if api is not None: logger.log('{:}'.format(api.query_by_arch(genotypes[epoch]))) # measure elapsed time epoch_time.update(time.time() - start_time) start_time = time.time() logger.log('\n' + '-' * 100) # check the performance from the architecture dataset logger.log( 'GDAS : run {:} epochs, cost {:.1f} s, last-geno is {:}.'.format( total_epoch, search_time.sum, genotypes[total_epoch - 1])) # if api is not None: logger.log('{:}'.format(api.query_by_arch(genotypes[total_epoch - 1]))) logger.close()