# Copyright (c) Meta Platforms, Inc. and affiliates.
#
# This source code is licensed under the Apache License, Version 2.0
# found in the LICENSE file in the root directory of this source tree.

# References:
#   https://github.com/facebookresearch/dino/blob/master/vision_transformer.py
#   https://github.com/rwightman/pytorch-image-models/tree/master/timm/models/vision_transformer.py

import logging
import os
import warnings

from torch import unbind

#from xformers.ops import memory_efficient_attention


from torch import Tensor
from torch import nn
import torch.nn.functional as F

XFORMERS_AVAILABLE = False


class Attention(nn.Module):
    def __init__(
        self,
        dim: int,
        num_heads: int = 8,
        qkv_bias: bool = True,
        proj_bias: bool = True,
        attn_drop: float = 0.0,
        proj_drop: float = 0.0,
        norm_layer: nn.Module = nn.LayerNorm,
        qk_norm: bool = False,
        fused_attn: bool = True,  # use F.scaled_dot_product_attention or not
        rope=None,
    ) -> None:
        super().__init__()
        assert dim % num_heads == 0, "dim should be divisible by num_heads"
        self.num_heads = num_heads
        self.head_dim = dim // num_heads
        self.scale = self.head_dim**-0.5
        self.fused_attn = fused_attn

        self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
        self.q_norm = norm_layer(self.head_dim) if qk_norm else nn.Identity()
        self.k_norm = norm_layer(self.head_dim) if qk_norm else nn.Identity()
        self.attn_drop = nn.Dropout(attn_drop)
        self.proj = nn.Linear(dim, dim, bias=proj_bias)
        self.proj_drop = nn.Dropout(proj_drop)
        self.rope = rope

    def forward(self, x: Tensor, pos=None) -> Tensor:
        B, N, C = x.shape
        qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, self.head_dim).permute(2, 0, 3, 1, 4)
        q, k, v = qkv.unbind(0)
        q, k = self.q_norm(q), self.k_norm(k)

        if self.rope is not None:
            q = self.rope(q, pos)
            k = self.rope(k, pos)

        if self.fused_attn:
            x = F.scaled_dot_product_attention(
                q,
                k,
                v,
                dropout_p=self.attn_drop.p if self.training else 0.0,
            )
        else:
            q = q * self.scale
            attn = q @ k.transpose(-2, -1)
            attn = attn.softmax(dim=-1)
            attn = self.attn_drop(attn)
            x = attn @ v
        # attn_vis = ((q * self.scale) @ k.transpose(-2, -1)).softmax(dim=-1)
        # import matplotlib.pyplot as plt
        # for i in range (attn_vis.shape[0]):
        #     for j in range (attn_vis.shape[1]):
        #         for k in range(7):
        #             _ = attn_vis[i][j][k][7:].reshape(1,1,32,32)
        #             _ = F.interpolate(_, size=(256, 256), mode='nearest')
        #             _ = _.squeeze().cpu().detach().numpy()
        #             os.makedirs("/share/project/cwm/houyuan.chen/UPS_Lightning/plot/results/Net2_v15_epoch5_vis_for_goblet_attn_map/", exist_ok=True)
        #             plt.imsave(f"/share/project/cwm/houyuan.chen/UPS_Lightning/plot/results/Net2_v15_epoch5_vis_for_goblet_attn_map/attn_vis_img_{i}_head_{j}_token_{k}.png", _)


        x = x.transpose(1, 2).reshape(B, N, C)
        x = self.proj(x)
        x = self.proj_drop(x)
        return x


class MemEffAttention(Attention):
    def forward(self, x: Tensor, attn_bias=None, pos=None) -> Tensor:
        assert pos is None
        if not XFORMERS_AVAILABLE:
            if attn_bias is not None:
                raise AssertionError("xFormers is required for using nested tensors")
            return super().forward(x)

        B, N, C = x.shape
        qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads)

        q, k, v = unbind(qkv, 2)

        x = memory_efficient_attention(q, k, v, attn_bias=attn_bias)
        x = x.reshape([B, N, C])

        x = self.proj(x)
        x = self.proj_drop(x)
        return x