import math
from collections import OrderedDict
import warnings

import torch
import torch.nn as nn
from torch import Tensor
from torch.jit.annotations import Dict, List, Tuple

from ..utils import load_state_dict_from_url

from . import _utils as det_utils
from .anchor_utils import AnchorGenerator
from .transform import GeneralizedRCNNTransform
from .backbone_utils import resnet_fpn_backbone
from ...ops.feature_pyramid_network import LastLevelP6P7
from ...ops import sigmoid_focal_loss
from ...ops import boxes as box_ops


__all__ = [
    "RetinaNet", "retinanet_resnet50_fpn",
]


def _sum(x: List[Tensor]) -> Tensor:
    res = x[0]
    for i in x[1:]:
        res = res + i
    return res


class RetinaNetHead(nn.Module):
    """
    A regression and classification head for use in RetinaNet.

    Arguments:
        in_channels (int): number of channels of the input feature
        num_anchors (int): number of anchors to be predicted
        num_classes (int): number of classes to be predicted
    """

    def __init__(self, in_channels, num_anchors, num_classes):
        super().__init__()
        self.classification_head = RetinaNetClassificationHead(in_channels, num_anchors, num_classes)
        self.regression_head = RetinaNetRegressionHead(in_channels, num_anchors)

    def compute_loss(self, targets, head_outputs, anchors, matched_idxs):
        # type: (List[Dict[str, Tensor]], Dict[str, Tensor], List[Tensor], List[Tensor]) -> Dict[str, Tensor]
        return {
            'classification': self.classification_head.compute_loss(targets, head_outputs, matched_idxs),
            'bbox_regression': self.regression_head.compute_loss(targets, head_outputs, anchors, matched_idxs),
        }

    def forward(self, x):
        # type: (List[Tensor]) -> Dict[str, Tensor]
        return {
            'cls_logits': self.classification_head(x),
            'bbox_regression': self.regression_head(x)
        }


class RetinaNetClassificationHead(nn.Module):
    """
    A classification head for use in RetinaNet.

    Arguments:
        in_channels (int): number of channels of the input feature
        num_anchors (int): number of anchors to be predicted
        num_classes (int): number of classes to be predicted
    """

    def __init__(self, in_channels, num_anchors, num_classes, prior_probability=0.01):
        super().__init__()

        conv = []
        for _ in range(4):
            conv.append(nn.Conv2d(in_channels, in_channels, kernel_size=3, stride=1, padding=1))
            conv.append(nn.ReLU())
        self.conv = nn.Sequential(*conv)

        for layer in self.conv.children():
            if isinstance(layer, nn.Conv2d):
                torch.nn.init.normal_(layer.weight, std=0.01)
                torch.nn.init.constant_(layer.bias, 0)

        self.cls_logits = nn.Conv2d(in_channels, num_anchors * num_classes, kernel_size=3, stride=1, padding=1)
        torch.nn.init.normal_(self.cls_logits.weight, std=0.01)
        torch.nn.init.constant_(self.cls_logits.bias, -math.log((1 - prior_probability) / prior_probability))

        self.num_classes = num_classes
        self.num_anchors = num_anchors

        # This is to fix using det_utils.Matcher.BETWEEN_THRESHOLDS in TorchScript.
        # TorchScript doesn't support class attributes.
        # https://github.com/pytorch/vision/pull/1697#issuecomment-630255584
        self.BETWEEN_THRESHOLDS = det_utils.Matcher.BETWEEN_THRESHOLDS

    def compute_loss(self, targets, head_outputs, matched_idxs):
        # type: (List[Dict[str, Tensor]], Dict[str, Tensor], List[Tensor]) -> Tensor
        losses = []

        cls_logits = head_outputs['cls_logits']

        for targets_per_image, cls_logits_per_image, matched_idxs_per_image in zip(targets, cls_logits, matched_idxs):
            # determine only the foreground
            foreground_idxs_per_image = matched_idxs_per_image >= 0
            num_foreground = foreground_idxs_per_image.sum()
            # no matched_idxs means there were no annotations in this image
            # TODO: enable support for images without annotations that works on distributed
            if False:  # matched_idxs_per_image.numel() == 0:
                gt_classes_target = torch.zeros_like(cls_logits_per_image)
                valid_idxs_per_image = torch.arange(cls_logits_per_image.shape[0])
            else:
                # create the target classification
                gt_classes_target = torch.zeros_like(cls_logits_per_image)
                gt_classes_target[
                    foreground_idxs_per_image,
                    targets_per_image['labels'][matched_idxs_per_image[foreground_idxs_per_image]]
                ] = 1.0

                # find indices for which anchors should be ignored
                valid_idxs_per_image = matched_idxs_per_image != self.BETWEEN_THRESHOLDS

            # compute the classification loss
            losses.append(sigmoid_focal_loss(
                cls_logits_per_image[valid_idxs_per_image],
                gt_classes_target[valid_idxs_per_image],
                reduction='sum',
            ) / max(1, num_foreground))

        return _sum(losses) / len(targets)

    def forward(self, x):
        # type: (List[Tensor]) -> Tensor
        all_cls_logits = []

        for features in x:
            cls_logits = self.conv(features)
            cls_logits = self.cls_logits(cls_logits)

            # Permute classification output from (N, A * K, H, W) to (N, HWA, K).
            N, _, H, W = cls_logits.shape
            cls_logits = cls_logits.view(N, -1, self.num_classes, H, W)
            cls_logits = cls_logits.permute(0, 3, 4, 1, 2)
            cls_logits = cls_logits.reshape(N, -1, self.num_classes)  # Size=(N, HWA, 4)

            all_cls_logits.append(cls_logits)

        return torch.cat(all_cls_logits, dim=1)


class RetinaNetRegressionHead(nn.Module):
    """
    A regression head for use in RetinaNet.

    Arguments:
        in_channels (int): number of channels of the input feature
        num_anchors (int): number of anchors to be predicted
    """
    __annotations__ = {
        'box_coder': det_utils.BoxCoder,
    }

    def __init__(self, in_channels, num_anchors):
        super().__init__()

        conv = []
        for _ in range(4):
            conv.append(nn.Conv2d(in_channels, in_channels, kernel_size=3, stride=1, padding=1))
            conv.append(nn.ReLU())
        self.conv = nn.Sequential(*conv)

        self.bbox_reg = nn.Conv2d(in_channels, num_anchors * 4, kernel_size=3, stride=1, padding=1)
        torch.nn.init.normal_(self.bbox_reg.weight, std=0.01)
        torch.nn.init.zeros_(self.bbox_reg.bias)

        for layer in self.conv.children():
            if isinstance(layer, nn.Conv2d):
                torch.nn.init.normal_(layer.weight, std=0.01)
                torch.nn.init.zeros_(layer.bias)

        self.box_coder = det_utils.BoxCoder(weights=(1.0, 1.0, 1.0, 1.0))

    def compute_loss(self, targets, head_outputs, anchors, matched_idxs):
        # type: (List[Dict[str, Tensor]], Dict[str, Tensor], List[Tensor], List[Tensor]) -> Tensor
        losses = []

        bbox_regression = head_outputs['bbox_regression']

        for targets_per_image, bbox_regression_per_image, anchors_per_image, matched_idxs_per_image in \
                zip(targets, bbox_regression, anchors, matched_idxs):
            # no matched_idxs means there were no annotations in this image
            # TODO enable support for images without annotations with distributed support
            # if matched_idxs_per_image.numel() == 0:
            #     continue

            # get the targets corresponding GT for each proposal
            # NB: need to clamp the indices because we can have a single
            # GT in the image, and matched_idxs can be -2, which goes
            # out of bounds
            matched_gt_boxes_per_image = targets_per_image['boxes'][matched_idxs_per_image.clamp(min=0)]

            # determine only the foreground indices, ignore the rest
            foreground_idxs_per_image = matched_idxs_per_image >= 0
            num_foreground = foreground_idxs_per_image.sum()

            # select only the foreground boxes
            matched_gt_boxes_per_image = matched_gt_boxes_per_image[foreground_idxs_per_image, :]
            bbox_regression_per_image = bbox_regression_per_image[foreground_idxs_per_image, :]
            anchors_per_image = anchors_per_image[foreground_idxs_per_image, :]

            # compute the regression targets
            target_regression = self.box_coder.encode_single(matched_gt_boxes_per_image, anchors_per_image)

            # compute the loss
            losses.append(torch.nn.functional.l1_loss(
                bbox_regression_per_image,
                target_regression,
                size_average=False
            ) / max(1, num_foreground))

        return _sum(losses) / max(1, len(targets))

    def forward(self, x):
        # type: (List[Tensor]) -> Tensor
        all_bbox_regression = []

        for features in x:
            bbox_regression = self.conv(features)
            bbox_regression = self.bbox_reg(bbox_regression)

            # Permute bbox regression output from (N, 4 * A, H, W) to (N, HWA, 4).
            N, _, H, W = bbox_regression.shape
            bbox_regression = bbox_regression.view(N, -1, 4, H, W)
            bbox_regression = bbox_regression.permute(0, 3, 4, 1, 2)
            bbox_regression = bbox_regression.reshape(N, -1, 4)  # Size=(N, HWA, 4)

            all_bbox_regression.append(bbox_regression)

        return torch.cat(all_bbox_regression, dim=1)


class RetinaNet(nn.Module):
    """
    Implements RetinaNet.

    The input to the model is expected to be a list of tensors, each of shape [C, H, W], one for each
    image, and should be in 0-1 range. Different images can have different sizes.

    The behavior of the model changes depending if it is in training or evaluation mode.

    During training, the model expects both the input tensors, as well as a targets (list of dictionary),
    containing:
        - boxes (FloatTensor[N, 4]): the ground-truth boxes in [x1, y1, x2, y2] format, with values
          between 0 and H and 0 and W
        - labels (Int64Tensor[N]): the class label for each ground-truth box

    The model returns a Dict[Tensor] during training, containing the classification and regression
    losses.

    During inference, the model requires only the input tensors, and returns the post-processed
    predictions as a List[Dict[Tensor]], one for each input image. The fields of the Dict are as
    follows:
        - boxes (FloatTensor[N, 4]): the predicted boxes in [x1, y1, x2, y2] format, with values between
          0 and H and 0 and W
        - labels (Int64Tensor[N]): the predicted labels for each image
        - scores (Tensor[N]): the scores for each prediction

    Arguments:
        backbone (nn.Module): the network used to compute the features for the model.
            It should contain an out_channels attribute, which indicates the number of output
            channels that each feature map has (and it should be the same for all feature maps).
            The backbone should return a single Tensor or an OrderedDict[Tensor].
        num_classes (int): number of output classes of the model (excluding the background).
        min_size (int): minimum size of the image to be rescaled before feeding it to the backbone
        max_size (int): maximum size of the image to be rescaled before feeding it to the backbone
        image_mean (Tuple[float, float, float]): mean values used for input normalization.
            They are generally the mean values of the dataset on which the backbone has been trained
            on
        image_std (Tuple[float, float, float]): std values used for input normalization.
            They are generally the std values of the dataset on which the backbone has been trained on
        anchor_generator (AnchorGenerator): module that generates the anchors for a set of feature
            maps.
        head (nn.Module): Module run on top of the feature pyramid.
            Defaults to a module containing a classification and regression module.
        score_thresh (float): Score threshold used for postprocessing the detections.
        nms_thresh (float): NMS threshold used for postprocessing the detections.
        detections_per_img (int): Number of best detections to keep after NMS.
        fg_iou_thresh (float): minimum IoU between the anchor and the GT box so that they can be
            considered as positive during training.
        bg_iou_thresh (float): maximum IoU between the anchor and the GT box so that they can be
            considered as negative during training.

    Example:

        >>> import torch
        >>> import torchvision
        >>> from torchvision.models.detection import RetinaNet
        >>> from torchvision.models.detection.anchor_utils import AnchorGenerator
        >>> # load a pre-trained model for classification and return
        >>> # only the features
        >>> backbone = torchvision.models.mobilenet_v2(pretrained=True).features
        >>> # RetinaNet needs to know the number of
        >>> # output channels in a backbone. For mobilenet_v2, it's 1280
        >>> # so we need to add it here
        >>> backbone.out_channels = 1280
        >>>
        >>> # let's make the network generate 5 x 3 anchors per spatial
        >>> # location, with 5 different sizes and 3 different aspect
        >>> # ratios. We have a Tuple[Tuple[int]] because each feature
        >>> # map could potentially have different sizes and
        >>> # aspect ratios
        >>> anchor_generator = AnchorGenerator(
        >>>     sizes=((32, 64, 128, 256, 512),),
        >>>     aspect_ratios=((0.5, 1.0, 2.0),)
        >>> )
        >>>
        >>> # put the pieces together inside a RetinaNet model
        >>> model = RetinaNet(backbone,
        >>>                   num_classes=2,
        >>>                   anchor_generator=anchor_generator)
        >>> model.eval()
        >>> x = [torch.rand(3, 300, 400), torch.rand(3, 500, 400)]
        >>> predictions = model(x)
    """
    __annotations__ = {
        'box_coder': det_utils.BoxCoder,
        'proposal_matcher': det_utils.Matcher,
    }

    def __init__(self, backbone, num_classes,
                 # transform parameters
                 min_size=800, max_size=1333,
                 image_mean=None, image_std=None,
                 # Anchor parameters
                 anchor_generator=None, head=None,
                 proposal_matcher=None,
                 score_thresh=0.05,
                 nms_thresh=0.5,
                 detections_per_img=300,
                 fg_iou_thresh=0.5, bg_iou_thresh=0.4):
        super().__init__()

        if not hasattr(backbone, "out_channels"):