# Copyright 2021 Huawei Technologies Co., Ltd
#
# Licensed under the BSD 3-Clause License  (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# https://opensource.org/licenses/BSD-3-Clause
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# 2021.06.15-Changed for implementation of TNT model
#            Huawei Technologies Co., Ltd. 
""" Loader Factory, Fast Collate, CUDA Prefetcher

Prefetcher and Fast Collate inspired by NVIDIA APEX example at
https://github.com/NVIDIA/apex/commit/d5e2bb4bdeedd27b1dfaf5bb2b24d6c000dee9be#diff-cf86c282ff7fba81fad27a559379d5bf

Hacked together by / Copyright 2020 Ross Wightman
"""

import torch.utils.data
import torch.distributed as dist
import numpy as np
import math
import random

from timm.data.transforms_factory import create_transform
from timm.data.constants import IMAGENET_DEFAULT_MEAN, IMAGENET_DEFAULT_STD
from timm.data.distributed_sampler import OrderedDistributedSampler
# from timm.data.random_erasing import RandomErasing
from timm.data.mixup import FastCollateMixup
from timm.data.loader import fast_collate, MultiEpochsDataLoader

from .rasampler import RASampler

def _get_pixels(per_pixel, rand_color, patch_size, dtype=torch.float32, device='npu'):
    # NOTE I've seen CUDA illegal memory access errors being caused by the normal_()
    # paths, flip the order so normal is run on CPU if this becomes a problem
    # Issue has been fixed in master https://github.com/pytorch/pytorch/issues/19508
    # This is for npu, modified from timm
    if per_pixel:
        return torch.empty(patch_size, dtype=dtype, device=device).normal_()
    elif rand_color:
        return torch.empty((patch_size[0], 1, 1), dtype=dtype, device=device).normal_()
    else:
        return torch.zeros((patch_size[0], 1, 1), dtype=dtype, device=device)

class RandomErasing:
    """ Randomly selects a rectangle region in an image and erases its pixels.
        'Random Erasing Data Augmentation' by Zhong et al.
        See https://arxiv.org/pdf/1708.04896.pdf
        This variant of RandomErasing is intended to be applied to either a batch
        or single image tensor after it has been normalized by dataset mean and std.
    Args:
         probability: Probability that the Random Erasing operation will be performed.
         min_area: Minimum percentage of erased area wrt input image area.
         max_area: Maximum percentage of erased area wrt input image area.
         min_aspect: Minimum aspect ratio of erased area.
         mode: pixel color mode, one of 'const', 'rand', or 'pixel'
            'const' - erase block is constant color of 0 for all channels
            'rand'  - erase block is same per-channel random (normal) color
            'pixel' - erase block is per-pixel random (normal) color
        max_count: maximum number of erasing blocks per image, area per box is scaled by count.
            per-image count is randomly chosen between 1 and this value.
    """

    def __init__(
            self,
            probability=0.5, min_area=0.02, max_area=1/3, min_aspect=0.3, max_aspect=None,
            mode='const', min_count=1, max_count=None, num_splits=0, device='npu'):
        self.probability = probability
        self.min_area = min_area
        self.max_area = max_area
        max_aspect = max_aspect or 1 / min_aspect
        self.log_aspect_ratio = (math.log(min_aspect), math.log(max_aspect))
        self.min_count = min_count
        self.max_count = max_count or min_count
        self.num_splits = num_splits
        mode = mode.lower()
        self.rand_color = False
        self.per_pixel = False
        if mode == 'rand':
            self.rand_color = True  # per block random normal
        elif mode == 'pixel':
            self.per_pixel = True  # per pixel random normal
        else:
            assert not mode or mode == 'const'
        self.device = device

    def _erase(self, img, chan, img_h, img_w, dtype):
        if random.random() > self.probability:
            return
        area = img_h * img_w
        count = self.min_count if self.min_count == self.max_count else \
            random.randint(self.min_count, self.max_count)
        for _ in range(count):
            for attempt in range(10):
                target_area = random.uniform(self.min_area, self.max_area) * area / count
                aspect_ratio = math.exp(random.uniform(*self.log_aspect_ratio))
                h = int(round(math.sqrt(target_area * aspect_ratio)))
                w = int(round(math.sqrt(target_area / aspect_ratio)))
                if w < img_w and h < img_h:
                    top = random.randint(0, img_h - h)
                    left = random.randint(0, img_w - w)
                    img[:, top:top + h, left:left + w] = _get_pixels(
                        self.per_pixel, self.rand_color, (chan, h, w),
                        dtype=dtype, device=self.device)
                    break

    def __call__(self, input):
        if len(input.size()) == 3:
            self._erase(input, *input.size(), input.dtype)
        else:
            batch_size, chan, img_h, img_w = input.size()
            # skip first slice of batch if num_splits is set (for clean portion of samples)
            batch_start = batch_size // self.num_splits if self.num_splits > 1 else 0
            for i in range(batch_start, batch_size):
                self._erase(input[i], chan, img_h, img_w, input.dtype)
        return input

class PrefetchLoader:

    def __init__(self,
                 loader,
                 mean=IMAGENET_DEFAULT_MEAN,
                 std=IMAGENET_DEFAULT_STD,
                 fp16=False,
                 re_prob=0.,
                 re_mode='const',
                 re_count=1,
                 re_num_splits=0):
        self.loader = loader
        self.mean = torch.tensor([x * 255 for x in mean]).npu().view(1, 3, 1, 1)
        self.std = torch.tensor([x * 255 for x in std]).npu().view(1, 3, 1, 1)
        self.fp16 = fp16
        if fp16:
            self.mean = self.mean.half()
            self.std = self.std.half()
        if re_prob > 0.:
            self.random_erasing = RandomErasing(
                probability=re_prob, mode=re_mode, max_count=re_count, num_splits=re_num_splits)
        else:
            self.random_erasing = None

    def __iter__(self):
        stream = torch.npu.Stream()
        first = True

        for next_input, next_target in self.loader:
            with torch.npu.stream(stream):
                next_input = next_input.npu(non_blocking=True)
                next_target = next_target.npu(non_blocking=True)
                if self.fp16:
                    next_input = next_input.half().sub_(self.mean).div_(self.std)
                else:
                    next_input = next_input.float().sub_(self.mean).div_(self.std)
                if self.random_erasing is not None:
                    next_input = self.random_erasing(next_input)

            if not first:
                yield input, target
            else:
                first = False

            torch.npu.current_stream().wait_stream(stream)
            input = next_input
            target = next_target

        yield input, target

    def __len__(self):
        return len(self.loader)

    @property
    def sampler(self):
        return self.loader.sampler

    @property
    def dataset(self):
        return self.loader.dataset

    @property
    def mixup_enabled(self):
        if isinstance(self.loader.collate_fn, FastCollateMixup):
            return self.loader.collate_fn.mixup_enabled
        else:
            return False

    @mixup_enabled.setter
    def mixup_enabled(self, x):
        if isinstance(self.loader.collate_fn, FastCollateMixup):
            self.loader.collate_fn.mixup_enabled = x

def is_dist_avail_and_initialized():
    if not dist.is_available():
        return False
    if not dist.is_initialized():
        return False
    return True


def get_world_size():
    if not is_dist_avail_and_initialized():
        return 1
    return dist.get_world_size()


def get_rank():
    if not is_dist_avail_and_initialized():
        return 0
    return dist.get_rank()


def create_loader(
        dataset,
        input_size,
        batch_size,
        is_training=False,
        use_prefetcher=True,
        no_aug=False,
        re_prob=0.,
        re_mode='const',
        re_count=1,
        re_split=False,
        scale=None,
        ratio=None,
        hflip=0.5,
        vflip=0.,
        color_jitter=0.4,
        auto_augment=None,
        num_aug_splits=0,
        interpolation='bilinear',
        mean=IMAGENET_DEFAULT_MEAN,
        std=IMAGENET_DEFAULT_STD,
        num_workers=1,
        distributed=False,
        crop_pct=None,
        collate_fn=None,
        pin_memory=False,
        fp16=False,
        tf_preprocessing=False,
        use_multi_epochs_loader=False,
        repeated_aug=False
):
    re_num_splits = 0
    if re_split:
        # apply RE to second half of batch if no aug split otherwise line up with aug split
        re_num_splits = num_aug_splits or 2
    dataset.transform = create_transform(
        input_size,
        is_training=is_training,
        use_prefetcher=use_prefetcher,
        no_aug=no_aug,
        scale=scale,
        ratio=ratio,
        hflip=hflip,
        vflip=vflip,
        color_jitter=color_jitter,
        auto_augment=auto_augment,
        interpolation=interpolation,
        mean=mean,
        std=std,
        crop_pct=crop_pct,
        tf_preprocessing=tf_preprocessing,
        re_prob=re_prob,
        re_mode=re_mode,
        re_count=re_count,
        re_num_splits=re_num_splits,
        separate=num_aug_splits > 0,
    )

    sampler = None
    if distributed:
        if is_training:
            if repeated_aug:
                print('using repeated_aug')
                num_tasks = get_world_size()
                global_rank = get_rank()
                sampler = RASampler(
                    dataset, num_replicas=num_tasks, rank=global_rank, shuffle=True
                )
            else:
                sampler = torch.utils.data.distributed.DistributedSampler(dataset)
        else:
            # This will add extra duplicate entries to result in equal num
            # of samples per-process, will slightly alter validation results
            sampler = OrderedDistributedSampler(dataset)
    else:
        if is_training and repeated_aug:
            print('using repeated_aug')
            num_tasks = get_world_size()
            global_rank = get_rank()
            sampler = RASampler(
                    dataset, num_replicas=num_tasks, rank=global_rank, shuffle=True
                )

    if collate_fn is None:
        collate_fn = fast_collate if use_prefetcher else torch.utils.data.dataloader.default_collate

    loader_class = torch.utils.data.DataLoader

    if use_multi_epochs_loader:
        loader_class = MultiEpochsDataLoader

    loader = loader_class(
        dataset,
        batch_size=batch_size,
        shuffle=sampler is None and is_training,
        num_workers=num_workers,
        sampler=sampler,
        collate_fn=collate_fn,
        pin_memory=pin_memory,
        drop_last=is_training,
    )
    if use_prefetcher:
        prefetch_re_prob = re_prob if is_training and not no_aug else 0.
        loader = PrefetchLoader(
            loader,
            mean=mean,
            std=std,
            fp16=fp16,
            re_prob=prefetch_re_prob,
            re_mode=re_mode,
            re_count=re_count,
            re_num_splits=re_num_splits
        )

    return loader