# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
"""
Mostly copy-paste from timm library.
https://github.com/rwightman/pytorch-image-models/blob/master/timm/models/vision_transformer.py
"""
from typing import Optional
import math
from functools import partial

import torch
import torch.nn as nn


def _no_grad_trunc_normal_(tensor, mean, std, a, b):
    # Cut & paste from PyTorch official master until it's in a few official releases - RW
    # Method based on https://people.sc.fsu.edu/~jburkardt/presentations/truncated_normal.pdf
    def norm_cdf(x):
        # Computes standard normal cumulative distribution function
        return (1. + math.erf(x / math.sqrt(2.))) / 2.

    if (mean < a - 2 * std) or (mean > b + 2 * std):
        warnings.warn(
            "mean is more than 2 std from [a, b] in nn.init.trunc_normal_. The distribution of values may be incorrect.",
            stacklevel=2
        )

    with torch.no_grad():
        # Values are generated by using a truncated uniform distribution and
        # then using the inverse CDF for the normal distribution.
        # Get upper and lower cdf values
        l = norm_cdf((a - mean) / std)
        u = norm_cdf((b - mean) / std)

        # Uniformly fill tensor with values from [l, u], then translate to
        # [2l-1, 2u-1].
        tensor.uniform_(2 * l - 1, 2 * u - 1)

        # Use inverse cdf transform for normal distribution to get truncated
        # standard normal
        tensor.erfinv_()

        # Transform to proper mean, std
        tensor.mul_(std * math.sqrt(2.))
        tensor.add_(mean)

        # Clamp to ensure it's in the proper range
        tensor.clamp_(min=a, max=b)
        return tensor


def trunc_normal_(tensor, mean=0., std=1., a=-2., b=2.):
    # type: (Tensor, float, float, float, float) -> Tensor
    return _no_grad_trunc_normal_(tensor, mean, std, a, b)


def drop_path(x, drop_prob: float = 0., training: bool = False):
    if drop_prob == 0. or not training:
        return x
    keep_prob = 1 - drop_prob
    shape = (x.shape[0],) + (1,) * (x.ndim - 1)  # work with diff dim tensors, not just 2D ConvNets
    random_tensor = keep_prob + torch.rand(shape, dtype=x.dtype, device=x.device)
    random_tensor.floor_()  # binarize
    output = x.div(keep_prob) * random_tensor
    return output


class DropPath(nn.Module):
    """Drop paths (Stochastic Depth) per sample  (when applied in main path of residual blocks).
    """
    def __init__(self, drop_prob=None):
        super(DropPath, self).__init__()
        self.drop_prob = drop_prob

    def forward(self, x):
        return drop_path(x, self.drop_prob, self.training)


class Mlp(nn.Module):
    def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.):
        super().__init__()
        out_features = out_features or in_features
        hidden_features = hidden_features or in_features
        self.fc1 = nn.Linear(in_features, hidden_features)
        self.act = act_layer()
        self.fc2 = nn.Linear(hidden_features, out_features)
        self.drop = nn.Dropout(drop)

    def forward(self, x):
        x = self.fc1(x)
        x = self.act(x)
        x = self.drop(x)
        x = self.fc2(x)
        x = self.drop(x)
        return x


class Attention(nn.Module):
    def __init__(self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0., proj_drop=0.):
        super().__init__()
        self.num_heads = num_heads
        head_dim = dim // num_heads
        self.scale = qk_scale or head_dim ** -0.5  # square root of dimension for normalisation

        self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
        self.attn_drop = nn.Dropout(attn_drop)

        self.proj = nn.Linear(dim, dim)
        self.proj_drop = nn.Dropout(proj_drop)

    def forward(self, x):
        B, N, C = x.shape  # B x (cls token + # patch tokens) x dim

        qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
        # qkv: 3 x B x Nh x (cls token + # patch tokens) x (dim // Nh)

        q, k, v = qkv[0], qkv[1], qkv[2]
        # q, k, v: B x Nh x (cls token + # patch tokens) x (dim // Nh)

        # q: B x Nh x (cls token + # patch tokens) x (dim // Nh)
        # k.transpose(-2, -1) = B x Nh x (dim // Nh) x (cls token + # patch tokens)
        # attn: B x Nh x (cls token + # patch tokens) x (cls token + # patch tokens)
        attn = (q @ k.transpose(-2, -1)) * self.scale  # @ operator is for matrix multiplication
        attn = attn.softmax(dim=-1)  # B x Nh x (cls token + # patch tokens) x (cls token + # patch tokens)
        attn = self.attn_drop(attn)

        # attn = B x Nh x (cls token + # patch tokens) x (cls token + # patch tokens)
        # v = B x Nh x (cls token + # patch tokens) x (dim // Nh)
        # attn @ v = B x Nh x (cls token + # patch tokens) x (dim // Nh)
        # (attn @ v).transpose(1, 2) = B x (cls token + # patch tokens) x Nh x (dim // Nh)
        x = (attn @ v).transpose(1, 2).reshape(B, N, C)  # B x (cls token + # patch tokens) x dim
        x = self.proj(x)  # B x (cls token + # patch tokens) x dim
        x = self.proj_drop(x)
        return x, attn


class Block(nn.Module):
    def __init__(self,
                 dim, num_heads,
                 mlp_ratio=4.,
                 qkv_bias=False,
                 qk_scale=None,
                 drop=0.,
                 attn_drop=0.,
                 drop_path=0.,
                 act_layer=nn.GELU,
                 norm_layer=nn.LayerNorm):
        super().__init__()
        self.norm1 = norm_layer(dim)
        self.attn = Attention(
            dim,
            num_heads=num_heads,
            qkv_bias=qkv_bias,
            qk_scale=qk_scale,
            attn_drop=attn_drop,
            proj_drop=drop
        )

        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()

        self.norm2 = norm_layer(dim)
        mlp_hidden_dim = int(dim * mlp_ratio)
        self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)

    def forward(self, x, return_attention=False):
        y, attn = self.attn(self.norm1(x))
        if return_attention:
            return attn
        x = x + self.drop_path(y)
        x = x + self.drop_path(self.mlp(self.norm2(x)))
        return x


class PatchEmbed(nn.Module):
    """ Image to Patch Embedding"""
    def __init__(self, img_size=(224, 224), patch_size=16, in_chans=3, embed_dim=768):
        super().__init__()
        num_patches = (img_size[0] // patch_size) * (img_size[1] // patch_size)
        self.img_size = img_size
        self.patch_size = patch_size
        self.num_patches = num_patches

        self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=patch_size)

    def forward(self, x):
        B, C, H, W = x.shape
        x = self.proj(x)
        x = x.flatten(2).transpose(1, 2)  # B x (P_H * P_W) x C
        return x


class VisionTransformer(nn.Module):
    """ Vision Transformer """
    def __init__(self,
                 img_size=(224, 224),
                 patch_size=16,
                 in_chans=3,
                 num_classes=0,
                 embed_dim=768,
                 depth=12,
                 num_heads=12,
                 mlp_ratio=4.,
                 qkv_bias=False,
                 qk_scale=None,
                 drop_rate=0.,
                 attn_drop_rate=0.,
                 drop_path_rate=0.,
                 norm_layer=nn.LayerNorm):
        super().__init__()
        self.num_features = self.embed_dim = embed_dim

        self.patch_embed = PatchEmbed(
            img_size=(224, 224),  # noel: this is to load pretrained model.
            patch_size=patch_size,
            in_chans=in_chans,
            embed_dim=embed_dim
        )
        num_patches = self.patch_embed.num_patches

        self.cls_token = nn.Parameter(torch.zeros(1, 1, embed_dim))
        self.pos_embed = nn.Parameter(torch.zeros(1, num_patches + 1, embed_dim))
        self.pos_drop = nn.Dropout(p=drop_rate)

        dpr = [x.item() for x in torch.linspace(0, drop_path_rate, depth)]  # stochastic depth decay rule
        self.blocks = nn.ModuleList([
            Block(
                dim=embed_dim,
                num_heads=num_heads,
                mlp_ratio=mlp_ratio,
                qkv_bias=qkv_bias,
                qk_scale=qk_scale,
                drop=drop_rate,
                attn_drop=attn_drop_rate,
                drop_path=dpr[i],
                norm_layer=norm_layer
            ) for i in range(depth)])
        self.norm = norm_layer(embed_dim)

        # Classifier head
        self.head = nn.Linear(embed_dim, num_classes) if num_classes > 0 else nn.Identity()

        trunc_normal_(self.pos_embed, std=.02)
        trunc_normal_(self.cls_token, std=.02)
        self.apply(self._init_weights)

        self.depth = depth
        self.embed_dim = self.n_embs = embed_dim
        self.mlp_ratio = mlp_ratio
        self.n_heads = num_heads
        self.patch_size = patch_size

    def _init_weights(self, m):
        if isinstance(m, nn.Linear):
            trunc_normal_(m.weight, std=.02)
            if isinstance(m, nn.Linear) and m.bias is not None:
                nn.init.constant_(m.bias, 0)
        elif isinstance(m, nn.LayerNorm):
            nn.init.constant_(m.bias, 0)
            nn.init.constant_(m.weight, 1.0)

    def make_input_divisible(self, x: torch.Tensor) -> torch.Tensor:
        """Pad some pixels to make the input size divisible by the patch size."""
        B, _, H_0, W_0 = x.shape
        pad_w = (self.patch_size - W_0 % self.patch_size) % self.patch_size
        pad_h = (self.patch_size - H_0 % self.patch_size) % self.patch_size

        x = nn.functional.pad(x, (0, pad_w, 0, pad_h), value=0)
        return x

    def prepare_tokens(self, x):
        B, nc, h, w = x.shape
        x: torch.Tensor = self.make_input_divisible(x)
        patch_embed_h, patch_embed_w = x.shape[-2] // self.patch_size, x.shape[-1] // self.patch_size

        x = self.patch_embed(x)  # patch linear embedding

        # add positional encoding to each token
        # add the [CLS] token to the embed patch tokens
        cls_tokens = self.cls_token.expand(B, -1, -1)
        x = torch.cat((cls_tokens, x), dim=1)
        x = x + self.interpolate_pos_encoding(x, self.pos_embed, size=(patch_embed_h, patch_embed_w))
        return self.pos_drop(x)

    @staticmethod
    def split_token(x, token_type: str):
        if token_type == "cls":
            return x[:, 0, :]
        elif token_type == "patch":
            return x[:, 1:, :]
        else:
            return x

    # noel
    def forward(self, x, layer: Optional[str] = None):
        x: torch.Tensor = self.prepare_tokens(x)

        features: dict = {}
        for i, blk in enumerate(self.blocks):
            x = blk(x)
            features[f"layer{i + 1}"] = self.norm(x)

        if layer is not None:
            return features[layer]
        else:
            return features

    # noel - for DINO's visual
    def get_last_selfattention(self, x):
        x = self.prepare_tokens(x)
        for i, blk in enumerate(self.blocks):
            if i < len(self.blocks) - 1:
                x = blk(x)
            else:
                # return attention of the last block
                return blk(x, return_attention=True)

    def get_tokens(
            self,
            x,
            layers: list,
            patch_tokens: bool = False,
            norm: bool = True,
            input_tokens: bool = False,
            post_pe: bool = False
    ):
        """Return intermediate tokens."""
        list_tokens: list = []

        B = x.shape[0]
        x = self.patch_embed(x)

        cls_tokens = self.cls_token.expand(B, -1, -1)

        x = torch.cat((cls_tokens, x), dim=1)

        if input_tokens:
            list_tokens.append(x)

        pos_embed = self.interpolate_pos_encoding(x, self.pos_embed)
        x = x + pos_embed

        if post_pe:
            list_tokens.append(x)

        x = self.pos_drop(x)

        for i, blk in enumerate(self.blocks):
            x = blk(x)  # B x # patches x dim
            if layers is None or i in layers:
                list_tokens.append(self.norm(x) if norm else x)

        tokens = torch.stack(list_tokens, dim=1)  # B x n_layers x (1 + # patches) x dim

        if not patch_tokens:
            return tokens[:, :, 0, :]  # index [CLS] tokens only, B x n_layers x dim

        else:
            return tokens

    def forward_features(self, x):
        B = x.shape[0]
        x = self.patch_embed(x)

        cls_tokens = self.cls_token.expand(B, -1, -1)
        x = torch.cat((cls_tokens, x), dim=1)
        pos_embed = self.interpolate_pos_encoding(x, self.pos_embed)
        x = x + pos_embed
        x = self.pos_drop(x)

        for blk in self.blocks:
            x = blk(x)

        if self.norm is not None:
            x = self.norm(x)

        return x[:, 0]

    def interpolate_pos_encoding(self, x, pos_embed, size):
        """Interpolate the learnable positional encoding to match the number of patches.

        x: B x (1 + N patches) x dim_embedding
        pos_embed: B x (1 + N patches) x dim_embedding

        return interpolated positional embedding
        """
        npatch = x.shape[1] - 1  # (H // patch_size * W // patch_size)
        N = pos_embed.shape[1] - 1  # 784 (= 28 x 28)
        if npatch == N:
            return pos_embed
        class_emb, pos_embed = pos_embed[:, 0], pos_embed[:, 1:]  # a learnable CLS token, learnable position embeddings

        dim = x.shape[-1]  # dimension of embeddings
        pos_embed = nn.functional.interpolate(
            pos_embed.reshape(1, int(math.sqrt(N)), int(math.sqrt(N)), dim).permute(0, 3, 1, 2),  # B x dim x 28 x 28
            size=size,
            mode='bicubic',
            align_corners=False
        )

        pos_embed = pos_embed.permute(0, 2, 3, 1).view(1, -1, dim)
        pos_embed = torch.cat((class_emb.unsqueeze(0), pos_embed), dim=1)
        return pos_embed

    def forward_selfattention(self, x, return_interm_attn=False):
        B, nc, w, h = x.shape
        N = self.pos_embed.shape[1] - 1
        x = self.patch_embed(x)

        # interpolate patch embeddings
        dim = x.shape[-1]
        w0 = w // self.patch_embed.patch_size
        h0 = h // self.patch_embed.patch_size
        class_pos_embed = self.pos_embed[:, 0]
        patch_pos_embed = self.pos_embed[:, 1:]
        patch_pos_embed = nn.functional.interpolate(
            patch_pos_embed.reshape(1, int(math.sqrt(N)), int(math.sqrt(N)), dim).permute(0, 3, 1, 2),
            scale_factor=(w0 / math.sqrt(N), h0 / math.sqrt(N)),
            mode='bicubic'
        )
        if w0 != patch_pos_embed.shape[-2]:
            helper = torch.zeros(h0)[None, None, None, :].repeat(1, dim, w0 - patch_pos_embed.shape[-2], 1).to(x.device)
            patch_pos_embed = torch.cat((patch_pos_embed, helper), dim=-2)
        if h0 != patch_pos_embed.shape[-1]:
            helper = torch.zeros(w0)[None, None, :, None].repeat(1, dim, 1, h0 - patch_pos_embed.shape[-1]).to(x.device)
            pos_embed = torch.cat((patch_pos_embed, helper), dim=-1)
        patch_pos_embed = patch_pos_embed.permute(0, 2, 3, 1).view(1, -1, dim)
        pos_embed = torch.cat((class_pos_embed.unsqueeze(0), patch_pos_embed), dim=1)

        cls_tokens = self.cls_token.expand(B, -1, -1)  # self.cls_token: 1 x 1 x emb_dim -> ?
        x = torch.cat((cls_tokens, x), dim=1)
        x = x + pos_embed
        x = self.pos_drop(x)

        if return_interm_attn:
            list_attn = []
            for i, blk in enumerate(self.blocks):
                attn = blk(x, return_attention=True)
                x = blk(x)
                list_attn.append(attn)
            return torch.cat(list_attn, dim=0)

        else:
            for i, blk in enumerate(self.blocks):
                if i < len(self.blocks) - 1:
                    x = blk(x)
                else:
                    return blk(x, return_attention=True)

    def forward_return_n_last_blocks(self, x, n=1, return_patch_avgpool=False):
        B = x.shape[0]
        x = self.patch_embed(x)

        cls_tokens = self.cls_token.expand(B, -1, -1)

        x = torch.cat((cls_tokens, x), dim=1)
        pos_embed = self.interpolate_pos_encoding(x, self.pos_embed)
        x = x + pos_embed
        x = self.pos_drop(x)

        # we will return the [CLS] tokens from the `n` last blocks
        output = []
        for i, blk in enumerate(self.blocks):
            x = blk(x)
            if len(self.blocks) - i <= n:
                # get only CLS token (B x dim)
                output.append(self.norm(x)[:, 0])
        if return_patch_avgpool:
            x = self.norm(x)
            # In addition to the [CLS] tokens from the `n` last blocks, we also return 
            # the patch tokens from the last block. This is useful for linear eval.
            output.append(torch.mean(x[:, 1:], dim=1))
        return torch.cat(output, dim=-1)

    def return_patch_emb_from_n_last_blocks(self, x, n=1, return_patch_avgpool=False):
        """Return intermediate patch embeddings, rather than CLS token, from the last n blocks."""
        B = x.shape[0]
        x = self.patch_embed(x)

        cls_tokens = self.cls_token.expand(B, -1, -1)

        x = torch.cat((cls_tokens, x), dim=1)
        pos_embed = self.interpolate_pos_encoding(x, self.pos_embed)
        x = x + pos_embed
        x = self.pos_drop(x)

        # we will return the [CLS] tokens from the `n` last blocks
        output = []
        for i, blk in enumerate(self.blocks):
            x = blk(x)
            if len(self.blocks) - i <= n:
                output.append(self.norm(x)[:, 1:])  # get only CLS token (B x dim)

        if return_patch_avgpool:
            x = self.norm(x)
            # In addition to the [CLS] tokens from the `n` last blocks, we also return
            # the patch tokens from the last block. This is useful for linear eval.
            output.append(torch.mean(x[:, 1:], dim=1))
        return torch.stack(output, dim=-1)  # B x n_patches x dim x n


def deit_tiny(patch_size=16, **kwargs):
    model = VisionTransformer(
        patch_size=patch_size,
        embed_dim=192,
        depth=12,
        num_heads=3,
        mlp_ratio=4,
        qkv_bias=True,
        norm_layer=partial(nn.LayerNorm, eps=1e-6),
        **kwargs)
    return model


def deit_small(patch_size=16, **kwargs):
    depth = kwargs.pop("depth") if "depth" in kwargs else 12
    model = VisionTransformer(
        patch_size=patch_size,
        embed_dim=384,
        depth=depth,
        num_heads=6,
        mlp_ratio=4,
        qkv_bias=True,
        norm_layer=partial(nn.LayerNorm, eps=1e-6),
        **kwargs
    )
    return model


def vit_base(patch_size=16, **kwargs):
    model = VisionTransformer(
        patch_size=patch_size, embed_dim=768, depth=12, num_heads=12, mlp_ratio=4,
        qkv_bias=True, norm_layer=partial(nn.LayerNorm, eps=1e-6), **kwargs)
    return model


class DINOHead(nn.Module):
    def __init__(self, in_dim, out_dim, use_bn=False, norm_last_layer=True, nlayers=3, hidden_dim=2048, bottleneck_dim=256):
        super().__init__()
        nlayers = max(nlayers, 1)
        if nlayers == 1:
            self.mlp = nn.Linear(in_dim, bottleneck_dim)
        else:
            layers = [nn.Linear(in_dim, hidden_dim)]
            if use_bn:
                layers.append(nn.BatchNorm1d(hidden_dim))
            layers.append(nn.GELU())
            for _ in range(nlayers - 2):
                layers.append(nn.Linear(hidden_dim, hidden_dim))
                if use_bn:
                    layers.append(nn.BatchNorm1d(hidden_dim))
                layers.append(nn.GELU())
            layers.append(nn.Linear(hidden_dim, bottleneck_dim))
            self.mlp = nn.Sequential(*layers)
        self.apply(self._init_weights)
        self.last_layer = nn.utils.weight_norm(nn.Linear(bottleneck_dim, out_dim, bias=False))
        self.last_layer.weight_g.data.fill_(1)
        if norm_last_layer:
            self.last_layer.weight_g.requires_grad = False

    def _init_weights(self, m):
        if isinstance(m, nn.Linear):
            trunc_normal_(m.weight, std=.02)
            if isinstance(m, nn.Linear) and m.bias is not None:
                nn.init.constant_(m.bias, 0)

    def forward(self, x):
        x = self.mlp(x)
        x = nn.functional.normalize(x, dim=-1, p=2)
        x = self.last_layer(x)
        return x