import argparse
import time
from datetime import datetime

try:
    from apex import amp
    from apex.parallel import DistributedDataParallel as DDP
    has_apex = True
except ImportError:
    has_apex = False

from data import Dataset, create_loader, resolve_data_config, FastCollateMixup, mixup_target
from models import create_model, resume_checkpoint
from utils import *
from loss import LabelSmoothingCrossEntropy, SoftTargetCrossEntropy
from optim import create_optimizer
from scheduler import create_scheduler

import torch
import torch.nn as nn
import torch.distributed as dist
import torchvision.utils

torch.backends.cudnn.benchmark = True

parser = argparse.ArgumentParser(description='Training')
parser.add_argument('data', metavar='DIR',
                    help='path to dataset')
parser.add_argument('--model', default='resnet101', type=str, metavar='MODEL',
                    help='Name of model to train (default: "countception"')
parser.add_argument('--num-classes', type=int, default=1000, metavar='N',
                    help='number of label classes (default: 1000)')
parser.add_argument('--opt', default='sgd', type=str, metavar='OPTIMIZER',
                    help='Optimizer (default: "sgd"')
parser.add_argument('--opt-eps', default=1e-8, type=float, metavar='EPSILON',
                    help='Optimizer Epsilon (default: 1e-8)')
parser.add_argument('--gp', default='avg', type=str, metavar='POOL',
                    help='Type of global pool, "avg", "max", "avgmax", "avgmaxc" (default: "avg")')
parser.add_argument('--tta', type=int, default=0, metavar='N',
                    help='Test/inference time augmentation (oversampling) factor. 0=None (default: 0)')
parser.add_argument('--pretrained', action='store_true', default=False,
                    help='Start with pretrained version of specified network (if avail)')
parser.add_argument('--img-size', type=int, default=224, metavar='N',
                    help='Image patch size (default: 224)')
parser.add_argument('--mean', type=float, nargs='+', default=None, metavar='MEAN',
                    help='Override mean pixel value of dataset')
parser.add_argument('--std', type=float, nargs='+', default=None, metavar='STD',
                    help='Override std deviation of of dataset')
parser.add_argument('--interpolation', default='', type=str, metavar='NAME',
                    help='Image resize interpolation type (overrides model)')
parser.add_argument('-b', '--batch-size', type=int, default=32, metavar='N',
                    help='input batch size for training (default: 32)')
parser.add_argument('-s', '--initial-batch-size', type=int, default=0, metavar='N',
                    help='initial input batch size for training (default: 0)')
parser.add_argument('--epochs', type=int, default=200, metavar='N',
                    help='number of epochs to train (default: 2)')
parser.add_argument('--start-epoch', default=None, type=int, metavar='N',
                    help='manual epoch number (useful on restarts)')
parser.add_argument('--decay-epochs', type=int, default=30, metavar='N',
                    help='epoch interval to decay LR')
parser.add_argument('--warmup-epochs', type=int, default=3, metavar='N',
                    help='epochs to warmup LR, if scheduler supports')
parser.add_argument('--decay-rate', '--dr', type=float, default=0.1, metavar='RATE',
                    help='LR decay rate (default: 0.1)')
parser.add_argument('--sched', default='step', type=str, metavar='SCHEDULER',
                    help='LR scheduler (default: "step"')
parser.add_argument('--drop', type=float, default=0.0, metavar='DROP',
                    help='Dropout rate (default: 0.)')
parser.add_argument('--reprob', type=float, default=0., metavar='PCT',
                    help='Random erase prob (default: 0.)')
parser.add_argument('--remode', type=str, default='const',
                    help='Random erase mode (default: "const")')
parser.add_argument('--lr', type=float, default=0.01, metavar='LR',
                    help='learning rate (default: 0.01)')
parser.add_argument('--warmup-lr', type=float, default=0.0001, metavar='LR',
                    help='warmup learning rate (default: 0.0001)')
parser.add_argument('--momentum', type=float, default=0.9, metavar='M',
                    help='SGD momentum (default: 0.9)')
parser.add_argument('--weight-decay', type=float, default=0.0001,
                    help='weight decay (default: 0.0001)')
parser.add_argument('--mixup', type=float, default=0.0,
                    help='mixup alpha, mixup enabled if > 0. (default: 0.)')
parser.add_argument('--mixup-off-epoch', default=0, type=int, metavar='N',
                    help='turn off mixup after this epoch, disabled if 0 (default: 0)')
parser.add_argument('--smoothing', type=float, default=0.1,
                    help='label smoothing (default: 0.1)')
parser.add_argument('--bn-tf', action='store_true', default=False,
                    help='Use Tensorflow BatchNorm defaults for models that support it (default: False)')
parser.add_argument('--bn-momentum', type=float, default=None,
                    help='BatchNorm momentum override (if not None)')
parser.add_argument('--bn-eps', type=float, default=None,
                    help='BatchNorm epsilon override (if not None)')
parser.add_argument('--seed', type=int, default=42, metavar='S',
                    help='random seed (default: 42)')
parser.add_argument('--log-interval', type=int, default=50, metavar='N',
                    help='how many batches to wait before logging training status')
parser.add_argument('--recovery-interval', type=int, default=1000, metavar='N',
                    help='how many batches to wait before writing recovery checkpoint')
parser.add_argument('-j', '--workers', type=int, default=4, metavar='N',
                    help='how many training processes to use (default: 1)')
parser.add_argument('--num-gpu', type=int, default=1,
                    help='Number of GPUS to use')
parser.add_argument('--initial-checkpoint', default='', type=str, metavar='PATH',
                    help='path to init checkpoint (default: none)')
parser.add_argument('--resume', default='', type=str, metavar='PATH',
                    help='path to latest checkpoint (default: none)')
parser.add_argument('--save-images', action='store_true', default=False,
                    help='save images of input bathes every log interval for debugging')
parser.add_argument('--amp', action='store_true', default=False,
                    help='use NVIDIA amp for mixed precision training')
parser.add_argument('--no-prefetcher', action='store_true', default=False,
                    help='disable fast prefetcher')
parser.add_argument('--output', default='', type=str, metavar='PATH',
                    help='path to output folder (default: none, current dir)')
parser.add_argument('--eval-metric', default='prec1', type=str, metavar='EVAL_METRIC',
                    help='Best metric (default: "prec1"')
parser.add_argument("--local_rank", default=0, type=int)


def main():
    args = parser.parse_args()

    args.prefetcher = not args.no_prefetcher
    args.distributed = False
    if 'WORLD_SIZE' in os.environ:
        args.distributed = int(os.environ['WORLD_SIZE']) > 1
        if args.distributed and args.num_gpu > 1:
            print('Using more than one GPU per process in distributed mode is not allowed. Setting num_gpu to 1.')
            args.num_gpu = 1

    args.device = 'cuda:0'
    args.world_size = 1
    r = -1
    if args.distributed:
        args.num_gpu = 1
        args.device = 'cuda:%d' % args.local_rank
        torch.cuda.set_device(args.local_rank)
        torch.distributed.init_process_group(backend='nccl',
                                             init_method='env://')
        args.world_size = torch.distributed.get_world_size()
        r = torch.distributed.get_rank()

    if args.distributed:
        print('Training in distributed mode with multiple processes, 1 GPU per process. Process %d, total %d.'
              % (r, args.world_size))
    else:
        print('Training with a single process on %d GPUs.' % args.num_gpu)

    # FIXME seed handling for multi-process distributed?
    torch.manual_seed(args.seed)

    output_dir = ''
    if args.local_rank == 0:
        if args.output:
            output_base = args.output
        else:
            output_base = './output'
        exp_name = '-'.join([
            datetime.now().strftime("%Y%m%d-%H%M%S"),
            args.model,
            str(args.img_size)])
        output_dir = get_outdir(output_base, 'train', exp_name)

    model = create_model(
        args.model,
        pretrained=args.pretrained,
        num_classes=args.num_classes,
        drop_rate=args.drop,
        global_pool=args.gp,
        bn_tf=args.bn_tf,
        bn_momentum=args.bn_momentum,
        bn_eps=args.bn_eps,
        checkpoint_path=args.initial_checkpoint)

    print('Model %s created, param count: %d' %
          (args.model, sum([m.numel() for m in model.parameters()])))

    data_config = resolve_data_config(model, args, verbose=args.local_rank == 0)

    # optionally resume from a checkpoint
    start_epoch = 0
    optimizer_state = None
    if args.resume:
        optimizer_state, start_epoch = resume_checkpoint(model, args.resume, args.start_epoch)

    if args.num_gpu > 1:
        if args.amp:
            print('Warning: AMP does not work well with nn.DataParallel, disabling. '
                  'Use distributed mode for multi-GPU AMP.')
            args.amp = False
        model = nn.DataParallel(model, device_ids=list(range(args.num_gpu))).cuda()
    else:
        model.cuda()

    optimizer = create_optimizer(args, model)
    if optimizer_state is not None:
        optimizer.load_state_dict(optimizer_state)

    if has_apex and args.amp:
        model, optimizer = amp.initialize(model, optimizer, opt_level='O1')
        use_amp = True
        print('AMP enabled')
    else:
        use_amp = False
        print('AMP disabled')

    if args.distributed:
        model = DDP(model, delay_allreduce=True)

    lr_scheduler, num_epochs = create_scheduler(args, optimizer)
    if start_epoch > 0:
        lr_scheduler.step(start_epoch)
    if args.local_rank == 0:
        print('Scheduled epochs: ', num_epochs)

    train_dir = os.path.join(args.data, 'train')
    if not os.path.exists(train_dir):
        print('Error: training folder does not exist at: %s' % train_dir)
        exit(1)
    dataset_train = Dataset(train_dir)

    collate_fn = None
    if args.prefetcher and args.mixup > 0:
        collate_fn = FastCollateMixup(args.mixup, args.smoothing, args.num_classes)

    loader_train = create_loader(
        dataset_train,
        input_size=data_config['input_size'],
        batch_size=args.batch_size,
        is_training=True,
        use_prefetcher=args.prefetcher,
        rand_erase_prob=args.reprob,
        rand_erase_mode=args.remode,
        interpolation='random',  # FIXME cleanly resolve this? data_config['interpolation'],
        mean=data_config['mean'],
        std=data_config['std'],
        num_workers=args.workers,
        distributed=args.distributed,
        collate_fn=collate_fn,
    )

    eval_dir = os.path.join(args.data, 'validation')
    if not os.path.isdir(eval_dir):
        print('Error: validation folder does not exist at: %s' % eval_dir)
        exit(1)
    dataset_eval = Dataset(eval_dir)

    loader_eval = create_loader(
        dataset_eval,
        input_size=data_config['input_size'],
        batch_size=4 * args.batch_size,
        is_training=False,
        use_prefetcher=args.prefetcher,
        interpolation=data_config['interpolation'],
        mean=data_config['mean'],
        std=data_config['std'],
        num_workers=args.workers,
        distributed=args.distributed,
    )

    if args.mixup > 0.:
        # smoothing is handled with mixup label transform
        train_loss_fn = SoftTargetCrossEntropy().cuda()
        validate_loss_fn = nn.CrossEntropyLoss().cuda()
    elif args.smoothing:
        train_loss_fn = LabelSmoothingCrossEntropy(smoothing=args.smoothing).cuda()
        validate_loss_fn = nn.CrossEntropyLoss().cuda()
    else:
        train_loss_fn = nn.CrossEntropyLoss().cuda()
        validate_loss_fn = train_loss_fn

    eval_metric = args.eval_metric
    saver = None
    if output_dir:
        decreasing = True if eval_metric == 'loss' else False
        saver = CheckpointSaver(checkpoint_dir=output_dir, decreasing=decreasing)
    best_metric = None
    best_epoch = None
    try:
        for epoch in range(start_epoch, num_epochs):
            if args.distributed:
                loader_train.sampler.set_epoch(epoch)

            train_metrics = train_epoch(
                epoch, model, loader_train, optimizer, train_loss_fn, args,
                lr_scheduler=lr_scheduler, saver=saver, output_dir=output_dir, use_amp=use_amp)

            eval_metrics = validate(
                model, loader_eval, validate_loss_fn, args)

            if lr_scheduler is not None:
                lr_scheduler.step(epoch, eval_metrics[eval_metric])

            update_summary(
                epoch, train_metrics, eval_metrics, os.path.join(output_dir, 'summary.csv'),
                write_header=best_metric is None)

            if saver is not None:
                # save proper checkpoint with eval metric
                best_metric, best_epoch = saver.save_checkpoint({
                    'epoch': epoch + 1,
                    'arch': args.model,
                    'state_dict': model.state_dict(),
                    'optimizer': optimizer.state_dict(),
                    'args': args,
                    },
                    epoch=epoch + 1,
                    metric=eval_metrics[eval_metric])

    except KeyboardInterrupt:
        pass
    if best_metric is not None:
        print('*** Best metric: {0} (epoch {1})'.format(best_metric, best_epoch))


def train_epoch(
        epoch, model, loader, optimizer, loss_fn, args,
        lr_scheduler=None, saver=None, output_dir='', use_amp=False):

    if args.prefetcher and args.mixup > 0 and loader.mixup_enabled:
        if args.mixup_off_epoch and epoch >= args.mixup_off_epoch:
            loader.mixup_enabled = False

    batch_time_m = AverageMeter()
    data_time_m = AverageMeter()
    losses_m = AverageMeter()

    model.train()

    end = time.time()
    last_idx = len(loader) - 1
    num_updates = epoch * len(loader)
    for batch_idx, (input, target) in enumerate(loader):
        last_batch = batch_idx == last_idx
        data_time_m.update(time.time() - end)
        if not args.prefetcher:
            input = input.cuda()
            target = target.cuda()
            if args.mixup > 0.:
                lam = 1.
                if not args.mixup_off_epoch or epoch < args.mixup_off_epoch:
                    lam = np.random.beta(args.mixup, args.mixup)
                input.mul_(lam).add_(1 - lam, input.flip(0))
                target = mixup_target(target, args.num_classes, lam, args.smoothing)

        output = model(input)

        loss = loss_fn(output, target)
        if not args.distributed:
            losses_m.update(loss.item(), input.size(0))

        optimizer.zero_grad()
        if use_amp:
            with amp.scale_loss(loss, optimizer) as scaled_loss:
                scaled_loss.backward()
        else:
            loss.backward()
        optimizer.step()

        torch.cuda.synchronize()
        num_updates += 1

        batch_time_m.update(time.time() - end)
        if last_batch or batch_idx % args.log_interval == 0:
            lrl = [param_group['lr'] for param_group in optimizer.param_groups]
            lr = sum(lrl) / len(lrl)

            if args.distributed:
                reduced_loss = reduce_tensor(loss.data, args.world_size)
                losses_m.update(reduced_loss.item(), input.size(0))

            if args.local_rank == 0:
                print('Train: {} [{}/{} ({:.0f}%)]  '
                      'Loss: {loss.val:.6f} ({loss.avg:.4f})  '
                      'Time: {batch_time.val:.3f}s, {rate:.3f}/s  '
                      '({batch_time.avg:.3f}s, {rate_avg:.3f}/s)  '
                      'LR: {lr:.4f}  '
                      'Data: {data_time.val:.3f} ({data_time.avg:.3f})'.format(
                    epoch,
                    batch_idx, len(loader),
                    100. * batch_idx / last_idx,
                    loss=losses_m,
                    batch_time=batch_time_m,
                    rate=input.size(0) * args.world_size / batch_time_m.val,
                    rate_avg=input.size(0) * args.world_size / batch_time_m.avg,
                    lr=lr,
                    data_time=data_time_m))

                if args.save_images and output_dir:
                    torchvision.utils.save_image(
                        input,
                        os.path.join(output_dir, 'train-batch-%d.jpg' % batch_idx),
                        padding=0,
                        normalize=True)

        if args.local_rank == 0 and (
                saver is not None and last_batch or (batch_idx + 1) % args.recovery_interval == 0):
            save_epoch = epoch + 1 if last_batch else epoch
            saver.save_recovery({
                'epoch': save_epoch,
                'arch': args.model,
                'state_dict':  model.state_dict(),
                'optimizer': optimizer.state_dict(),
                'args': args,
                },
                epoch=save_epoch,
                batch_idx=batch_idx)

        if lr_scheduler is not None:
            lr_scheduler.step_update(num_updates=num_updates, metric=losses_m.avg)

        end = time.time()

    return OrderedDict([('loss', losses_m.avg)])


def validate(model, loader, loss_fn, args):
    batch_time_m = AverageMeter()
    losses_m = AverageMeter()
    prec1_m = AverageMeter()
    prec5_m = AverageMeter()

    model.eval()

    end = time.time()
    last_idx = len(loader) - 1
    with torch.no_grad():
        for batch_idx, (input, target) in enumerate(loader):
            last_batch = batch_idx == last_idx
            if not args.prefetcher:
                input = input.cuda()
                target = target.cuda()

            output = model(input)
            if isinstance(output, (tuple, list)):
                output = output[0]

            # augmentation reduction
            reduce_factor = args.tta
            if reduce_factor > 1:
                output = output.unfold(0, reduce_factor, reduce_factor).mean(dim=2)
                target = target[0:target.size(0):reduce_factor]

            loss = loss_fn(output, target)
            prec1, prec5 = accuracy(output, target, topk=(1, 5))

            if args.distributed:
                reduced_loss = reduce_tensor(loss.data, args.world_size)
                prec1 = reduce_tensor(prec1, args.world_size)
                prec5 = reduce_tensor(prec5, args.world_size)
            else:
                reduced_loss = loss.data

            torch.cuda.synchronize()

            losses_m.update(reduced_loss.item(), input.size(0))
            prec1_m.update(prec1.item(), output.size(0))
            prec5_m.update(prec5.item(), output.size(0))

            batch_time_m.update(time.time() - end)
            end = time.time()
            if args.local_rank == 0 and (last_batch or batch_idx % args.log_interval == 0):
                print('Test: [{0}/{1}]\t'
                      'Time {batch_time.val:.3f} ({batch_time.avg:.3f})  '
                      'Loss {loss.val:.4f} ({loss.avg:.4f})  '
                      'Prec@1 {top1.val:.4f} ({top1.avg:.4f})  '
                      'Prec@5 {top5.val:.4f} ({top5.avg:.4f})'.format(
                    batch_idx, last_idx,
                    batch_time=batch_time_m, loss=losses_m,
                    top1=prec1_m, top5=prec5_m))

    metrics = OrderedDict([('loss', losses_m.avg), ('prec1', prec1_m.avg), ('prec5', prec5_m.avg)])

    return metrics


def reduce_tensor(tensor, n):
    rt = tensor.clone()
    dist.all_reduce(rt, op=dist.ReduceOp.SUM)
    rt /= n
    return rt


if __name__ == '__main__':
    main()