xautodl/exps/algos/ENAS.py

##################################################
# Copyright (c) Xuanyi Dong [GitHub D-X-Y], 2020 #
##########################################################################
# Efficient Neural Architecture Search via Parameters Sharing, ICML 2018 #
##########################################################################
import os, sys, time, glob, random, argparse
import numpy as np
from copy import deepcopy
import torch
import torch.nn as nn
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 load_config, dict2config, configure2str
from datasets import get_datasets, get_nas_search_loaders
from procedures import prepare_seed, prepare_logger, save_checkpoint, copy_checkpoint, get_optim_scheduler
from utils import get_model_infos, obtain_accuracy
from log_utils import AverageMeter, time_string, convert_secs2time
from models import get_cell_based_tiny_net, get_search_spaces
from nas_201_api import NASBench201API as API


def train_shared_cnn(xloader, shared_cnn, controller, criterion, scheduler, optimizer, epoch_str, print_freq, logger):
    data_time, batch_time = AverageMeter(), AverageMeter()
    losses, top1s, top5s, xend = AverageMeter(), AverageMeter(), AverageMeter(), time.time()

    shared_cnn.train()
    controller.eval()

    for step, (inputs, targets) in enumerate(xloader):
        scheduler.update(None, 1.0 * step / len(xloader))
        targets = targets.cuda(non_blocking=True)
        # measure data loading time
        data_time.update(time.time() - xend)

        with torch.no_grad():
            _, _, sampled_arch = controller()

        optimizer.zero_grad()
        shared_cnn.module.update_arch(sampled_arch)
        _, logits = shared_cnn(inputs)
        loss = criterion(logits, targets)
        loss.backward()
        torch.nn.utils.clip_grad_norm_(shared_cnn.parameters(), 5)
        optimizer.step()
        # record
        prec1, prec5 = obtain_accuracy(logits.data, targets.data, topk=(1, 5))
        losses.update(loss.item(), inputs.size(0))
        top1s.update(prec1.item(), inputs.size(0))
        top5s.update(prec5.item(), inputs.size(0))

        # measure elapsed time
        batch_time.update(time.time() - xend)
        xend = time.time()

        if step % print_freq == 0 or step + 1 == len(xloader):
            Sstr = "*Train-Shared-CNN* " + time_string() + " [{:}][{:03d}/{:03d}]".format(epoch_str, step, len(xloader))
            Tstr = "Time {batch_time.val:.2f} ({batch_time.avg:.2f}) Data {data_time.val:.2f} ({data_time.avg:.2f})".format(
                batch_time=batch_time, data_time=data_time
            )
            Wstr = "[Loss {loss.val:.3f} ({loss.avg:.3f})  Prec@1 {top1.val:.2f} ({top1.avg:.2f}) Prec@5 {top5.val:.2f} ({top5.avg:.2f})]".format(
                loss=losses, top1=top1s, top5=top5s
            )
            logger.log(Sstr + " " + Tstr + " " + Wstr)
    return losses.avg, top1s.avg, top5s.avg


def train_controller(xloader, shared_cnn, controller, criterion, optimizer, config, epoch_str, print_freq, logger):
    # config. (containing some necessary arg)
    #   baseline: The baseline score (i.e. average val_acc) from the previous epoch
    data_time, batch_time = AverageMeter(), AverageMeter()
    GradnormMeter, LossMeter, ValAccMeter, EntropyMeter, BaselineMeter, RewardMeter, xend = (
        AverageMeter(),
        AverageMeter(),
        AverageMeter(),
        AverageMeter(),
        AverageMeter(),
        AverageMeter(),
        time.time(),
    )

    shared_cnn.eval()
    controller.train()
    controller.zero_grad()
    # for step, (inputs, targets) in enumerate(xloader):
    loader_iter = iter(xloader)
    for step in range(config.ctl_train_steps * config.ctl_num_aggre):
        try:
            inputs, targets = next(loader_iter)
        except:
            loader_iter = iter(xloader)
            inputs, targets = next(loader_iter)
        targets = targets.cuda(non_blocking=True)
        # measure data loading time
        data_time.update(time.time() - xend)

        log_prob, entropy, sampled_arch = controller()
        with torch.no_grad():
            shared_cnn.module.update_arch(sampled_arch)
            _, logits = shared_cnn(inputs)
            val_top1, val_top5 = obtain_accuracy(logits.data, targets.data, topk=(1, 5))
            val_top1 = val_top1.view(-1) / 100
        reward = val_top1 + config.ctl_entropy_w * entropy
        if config.baseline is None:
            baseline = val_top1
        else:
            baseline = config.baseline - (1 - config.ctl_bl_dec) * (config.baseline - reward)

        loss = -1 * log_prob * (reward - baseline)

        # account
        RewardMeter.update(reward.item())
        BaselineMeter.update(baseline.item())
        ValAccMeter.update(val_top1.item() * 100)
        LossMeter.update(loss.item())
        EntropyMeter.update(entropy.item())

        # Average gradient over controller_num_aggregate samples
        loss = loss / config.ctl_num_aggre
        loss.backward(retain_graph=True)

        # measure elapsed time
        batch_time.update(time.time() - xend)
        xend = time.time()
        if (step + 1) % config.ctl_num_aggre == 0:
            grad_norm = torch.nn.utils.clip_grad_norm_(controller.parameters(), 5.0)
            GradnormMeter.update(grad_norm)
            optimizer.step()
            controller.zero_grad()

        if step % print_freq == 0:
            Sstr = (
                "*Train-Controller* "
                + time_string()
                + " [{:}][{:03d}/{:03d}]".format(epoch_str, step, config.ctl_train_steps * config.ctl_num_aggre)
            )
            Tstr = "Time {batch_time.val:.2f} ({batch_time.avg:.2f}) Data {data_time.val:.2f} ({data_time.avg:.2f})".format(
                batch_time=batch_time, data_time=data_time
            )
            Wstr = "[Loss {loss.val:.3f} ({loss.avg:.3f})  Prec@1 {top1.val:.2f} ({top1.avg:.2f}) Reward {reward.val:.2f} ({reward.avg:.2f})] Baseline {basel.val:.2f} ({basel.avg:.2f})".format(
                loss=LossMeter, top1=ValAccMeter, reward=RewardMeter, basel=BaselineMeter
            )
            Estr = "Entropy={:.4f} ({:.4f})".format(EntropyMeter.val, EntropyMeter.avg)
            logger.log(Sstr + " " + Tstr + " " + Wstr + " " + Estr)

    return LossMeter.avg, ValAccMeter.avg, BaselineMeter.avg, RewardMeter.avg, baseline.item()


def get_best_arch(controller, shared_cnn, xloader, n_samples=10):
    with torch.no_grad():
        controller.eval()
        shared_cnn.eval()
        archs, valid_accs = [], []
        loader_iter = iter(xloader)
        for i in range(n_samples):
            try:
                inputs, targets = next(loader_iter)
            except:
                loader_iter = iter(xloader)
                inputs, targets = next(loader_iter)

            _, _, sampled_arch = controller()
            arch = shared_cnn.module.update_arch(sampled_arch)
            _, logits = shared_cnn(inputs)
            val_top1, val_top5 = obtain_accuracy(logits.cpu().data, targets.data, topk=(1, 5))

            archs.append(arch)
            valid_accs.append(val_top1.item())

        best_idx = np.argmax(valid_accs)
        best_arch, best_valid_acc = archs[best_idx], valid_accs[best_idx]
        return best_arch, best_valid_acc


def valid_func(xloader, network, criterion):
    data_time, batch_time = AverageMeter(), AverageMeter()
    arch_losses, arch_top1, arch_top5 = AverageMeter(), AverageMeter(), AverageMeter()
    network.eval()
    end = time.time()
    with torch.no_grad():
        for step, (arch_inputs, arch_targets) in enumerate(xloader):
            arch_targets = arch_targets.cuda(non_blocking=True)
            # measure data loading time
            data_time.update(time.time() - end)
            # prediction
            _, logits = network(arch_inputs)
            arch_loss = criterion(logits, arch_targets)
            # record
            arch_prec1, arch_prec5 = obtain_accuracy(logits.data, arch_targets.data, topk=(1, 5))
            arch_losses.update(arch_loss.item(), arch_inputs.size(0))
            arch_top1.update(arch_prec1.item(), arch_inputs.size(0))
            arch_top5.update(arch_prec5.item(), arch_inputs.size(0))
            # measure elapsed time
            batch_time.update(time.time() - end)
            end = time.time()
    return arch_losses.avg, arch_top1.avg, arch_top5.avg


def main(xargs):
    assert torch.cuda.is_available(), "CUDA is not available."
    torch.backends.cudnn.enabled = True
    torch.backends.cudnn.benchmark = False
    torch.backends.cudnn.deterministic = True
    torch.set_num_threads(xargs.workers)
    prepare_seed(xargs.rand_seed)
    logger = prepare_logger(args)

    train_data, test_data, xshape, class_num = get_datasets(xargs.dataset, xargs.data_path, -1)
    logger.log("use config from : {:}".format(xargs.config_path))
    config = load_config(xargs.config_path, {"class_num": class_num, "xshape": xshape}, logger)
    _, train_loader, valid_loader = get_nas_search_loaders(
        train_data, test_data, xargs.dataset, "configs/nas-benchmark/", config.batch_size, xargs.workers
    )
    # since ENAS will train the controller on valid-loader, we need to use train transformation for valid-loader
    valid_loader.dataset.transform = deepcopy(train_loader.dataset.transform)
    if hasattr(valid_loader.dataset, "transforms"):
        valid_loader.dataset.transforms = deepcopy(train_loader.dataset.transforms)
    # data loader
    logger.log(
        "||||||| {:10s} ||||||| Train-Loader-Num={:}, Valid-Loader-Num={:}, batch size={:}".format(
            xargs.dataset, len(train_loader), len(valid_loader), config.batch_size
        )
    )
    logger.log("||||||| {:10s} ||||||| Config={:}".format(xargs.dataset, config))

    search_space = get_search_spaces("cell", xargs.search_space_name)
    model_config = dict2config(
        {
            "name": "ENAS",
            "C": xargs.channel,
            "N": xargs.num_cells,
            "max_nodes": xargs.max_nodes,
            "num_classes": class_num,
            "space": search_space,
            "affine": False,
            "track_running_stats": bool(xargs.track_running_stats),
        },
        None,
    )
    shared_cnn = get_cell_based_tiny_net(model_config)
    controller = shared_cnn.create_controller()

    w_optimizer, w_scheduler, criterion = get_optim_scheduler(shared_cnn.parameters(), config)
    a_optimizer = torch.optim.Adam(
        controller.parameters(), lr=config.controller_lr, betas=config.controller_betas, eps=config.controller_eps
    )
    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(shared_cnn, xshape)
    # logger.log('{:}'.format(shared_cnn))
    # logger.log('FLOP = {:.2f} M, Params = {:.2f} MB'.format(flop, param))
    logger.log("search-space : {:}".format(search_space))
    if xargs.arch_nas_dataset is None:
        api = None
    else:
        api = API(xargs.arch_nas_dataset)
    logger.log("{:} create API = {:} done".format(time_string(), api))
    shared_cnn, controller, criterion = torch.nn.DataParallel(shared_cnn).cuda(), controller.cuda(), criterion.cuda()

    last_info, model_base_path, model_best_path = logger.path("info"), logger.path("model"), logger.path("best")

    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"]
        baseline = checkpoint["baseline"]
        valid_accuracies = checkpoint["valid_accuracies"]
        shared_cnn.load_state_dict(checkpoint["shared_cnn"])
        controller.load_state_dict(checkpoint["controller"])
        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, baseline = 0, {"best": -1}, {}, None

    # start training
    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)
        logger.log(
            "\n[Search the {:}-th epoch] {:}, LR={:}, baseline={:}".format(
                epoch_str, need_time, min(w_scheduler.get_lr()), baseline
            )
        )

        cnn_loss, cnn_top1, cnn_top5 = train_shared_cnn(
            train_loader,
            shared_cnn,
            controller,
            criterion,
            w_scheduler,
            w_optimizer,
            epoch_str,
            xargs.print_freq,
            logger,
        )
        logger.log(
            "[{:}] shared-cnn : loss={:.2f}, accuracy@1={:.2f}%, accuracy@5={:.2f}%".format(
                epoch_str, cnn_loss, cnn_top1, cnn_top5
            )
        )
        ctl_loss, ctl_acc, ctl_baseline, ctl_reward, baseline = train_controller(
            valid_loader,
            shared_cnn,
            controller,
            criterion,
            a_optimizer,
            dict2config(
                {
                    "baseline": baseline,
                    "ctl_train_steps": xargs.controller_train_steps,
                    "ctl_num_aggre": xargs.controller_num_aggregate,
                    "ctl_entropy_w": xargs.controller_entropy_weight,
                    "ctl_bl_dec": xargs.controller_bl_dec,
                },
                None,
            ),
            epoch_str,
            xargs.print_freq,
            logger,
        )
        search_time.update(time.time() - start_time)
        logger.log(
            "[{:}] controller : loss={:.2f}, accuracy={:.2f}%, baseline={:.2f}, reward={:.2f}, current-baseline={:.4f}, time-cost={:.1f} s".format(
                epoch_str, ctl_loss, ctl_acc, ctl_baseline, ctl_reward, baseline, search_time.sum
            )
        )
        best_arch, _ = get_best_arch(controller, shared_cnn, valid_loader)
        shared_cnn.module.update_arch(best_arch)
        _, best_valid_acc, _ = valid_func(valid_loader, shared_cnn, criterion)

        genotypes[epoch] = best_arch
        # check the best accuracy
        valid_accuracies[epoch] = best_valid_acc
        if best_valid_acc > valid_accuracies["best"]:
            valid_accuracies["best"] = best_valid_acc
            genotypes["best"] = best_arch
            find_best = True
        else:
            find_best = False

        logger.log("<<<--->>> The {:}-th epoch : {:}".format(epoch_str, genotypes[epoch]))
        # save checkpoint
        save_path = save_checkpoint(
            {
                "epoch": epoch + 1,
                "args": deepcopy(xargs),
                "baseline": baseline,
                "shared_cnn": shared_cnn.state_dict(),
                "controller": controller.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(args),
                "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, best_valid_acc
                )
            )
            copy_checkpoint(model_base_path, model_best_path, logger)
        if api is not None:
            logger.log("{:}".format(api.query_by_arch(genotypes[epoch], "200")))
        # measure elapsed time
        epoch_time.update(time.time() - start_time)
        start_time = time.time()

    logger.log("\n" + "-" * 100)
    logger.log("During searching, the best architecture is {:}".format(genotypes["best"]))
    logger.log("Its accuracy is {:.2f}%".format(valid_accuracies["best"]))
    logger.log("Randomly select {:} architectures and select the best.".format(xargs.controller_num_samples))
    start_time = time.time()
    final_arch, _ = get_best_arch(controller, shared_cnn, valid_loader, xargs.controller_num_samples)
    search_time.update(time.time() - start_time)
    shared_cnn.module.update_arch(final_arch)
    final_loss, final_top1, final_top5 = valid_func(valid_loader, shared_cnn, criterion)
    logger.log("The Selected Final Architecture : {:}".format(final_arch))
    logger.log("Loss={:.3f}, Accuracy@1={:.2f}%, Accuracy@5={:.2f}%".format(final_loss, final_top1, final_top5))
    logger.log(
        "ENAS : run {:} epochs, cost {:.1f} s, last-geno is {:}.".format(total_epoch, search_time.sum, final_arch)
    )
    if api is not None:
        logger.log("{:}".format(api.query_by_arch(final_arch)))
    logger.close()


if __name__ == "__main__":
    parser = argparse.ArgumentParser("ENAS")
    parser.add_argument("--data_path", type=str, help="Path to dataset")
    parser.add_argument(
        "--dataset",
        type=str,
        choices=["cifar10", "cifar100", "ImageNet16-120"],
        help="Choose between Cifar10/100 and ImageNet-16.",
    )
    # channels and number-of-cells
    parser.add_argument(
        "--track_running_stats",
        type=int,
        choices=[0, 1],
        help="Whether use track_running_stats or not in the BN layer.",
    )
    parser.add_argument("--search_space_name", type=str, help="The search space name.")
    parser.add_argument("--max_nodes", type=int, help="The maximum number of nodes.")
    parser.add_argument("--channel", type=int, help="The number of channels.")
    parser.add_argument("--num_cells", type=int, help="The number of cells in one stage.")
    parser.add_argument("--config_path", type=str, help="The config file to train ENAS.")
    parser.add_argument("--controller_train_steps", type=int, help=".")
    parser.add_argument("--controller_num_aggregate", type=int, help=".")
    parser.add_argument("--controller_entropy_weight", type=float, help="The weight for the entropy of the controller.")
    parser.add_argument("--controller_bl_dec", type=float, help=".")
    parser.add_argument("--controller_num_samples", type=int, help=".")
    # log
    parser.add_argument("--workers", type=int, default=2, help="number of data loading workers (default: 2)")
    parser.add_argument("--save_dir", type=str, help="Folder to save checkpoints and log.")
    parser.add_argument(
        "--arch_nas_dataset", type=str, help="The path to load the architecture dataset (nas-benchmark)."
    )
    parser.add_argument("--print_freq", type=int, help="print frequency (default: 200)")
    parser.add_argument("--rand_seed", type=int, help="manual seed")
    args = parser.parse_args()
    if args.rand_seed is None or args.rand_seed < 0:
        args.rand_seed = random.randint(1, 100000)
    main(args)