import torch
from torch import nn
from einops import rearrange, repeat
from einops.layers.torch import Rearrange
 
def pair(t):
    return t if isinstance(t, tuple) else (t, t)
 
# classes
 
class PreNorm(nn.Module):
    def __init__(self, dim, fn):
        super().__init__()
        self.norm = nn.LayerNorm(dim)
        self.fn = fn
    def forward(self, x, **kwargs):
        return self.fn(self.norm(x), **kwargs)
 
class FeedForward(nn.Module):    # MLP
    def __init__(self, dim, hidden_dim, dropout = 0.):
        super().__init__()
        self.net = nn.Sequential(
            nn.Linear(dim, hidden_dim),
            nn.GELU(),
            nn.Dropout(dropout),
            nn.Linear(hidden_dim, dim),
            nn.Dropout(dropout)
        )
    def forward(self, x):
        return self.net(x)
 
class Attention(nn.Module):
    def __init__(self, dim, heads, dropout = 0.):
        super().__init__()
        # inner_dim = dim_head * heads
        dim_head = dim // heads
        project_out = not (heads == 1 and dim_head == dim)
 
        self.heads = heads
        self.scale = dim_head ** -0.5
 
        self.attend = nn.Softmax(dim = -1)
        self.to_qkv = nn.Linear(dim, dim * 3, bias = False)
 
        self.to_out = nn.Sequential(
            nn.Linear(dim, dim),
            nn.Dropout(dropout)
        ) if project_out else nn.Identity()
 
    def forward(self, x):
        B, N, C = x.shape
        # qkv(): -> [batch_size, num_patches + 1, 3 * total_embed_dim]
        # reshape: -> [batch_size, num_patches + 1, 3, num_heads, embed_dim_per_head]
        # permute: -> [3, batch_size, num_heads, num_patches + 1, embed_dim_per_head]796/8
        qkv = self.to_qkv(x).reshape(B, N, 3, self.heads, C // self.heads).permute(2, 0, 3, 1, 4)
        # [batch_size, num_heads, num_patches + 1, embed_dim_per_head]
        q, k, v = qkv[0], qkv[1], qkv[2]  # make torchscript happy (cannot use tensor as tuple)
 
        dots = torch.matmul(q, k.transpose(-1, -2)) * self.scale
 
        attn = self.attend(dots)
 
        out = torch.matmul(attn, v).transpose(1, 2).reshape(B, N, C)
        # out = rearrange(out, 'b h n d -> b n (h d)')
        return self.to_out(out)
 
class Transformer(nn.Module):
    def __init__(self, dim, depth, heads, mlp_dim, dropout = 0.):
        super().__init__()
        self.layers = nn.ModuleList([])
        for _ in range(depth):
            self.layers.append(nn.ModuleList([
                PreNorm(dim, Attention(dim, heads = heads, dropout = dropout)),
                PreNorm(dim, FeedForward(dim, mlp_dim, dropout = dropout))
            ]))
    def forward(self, x):
        for attn, ff in self.layers:
            x = attn(x) + x
            x = ff(x) + x
        return x
 
class ViT(nn.Module):
    def __init__(self, *, image_size, patch_size, num_classes, dim, depth, heads, mlp_dim, pool = 'cls', channels = 3, dropout = 0., emb_dropout = 0.):
        super().__init__()
        image_height, image_width = pair(image_size) ## 224*224
        patch_height, patch_width = pair(patch_size)## 16 * 16
 
        assert image_height % patch_height == 0 and image_width % patch_width == 0, 'Image dimensions must be divisible by the patch size.'
 
        num_patches = (image_height // patch_height) * (image_width // patch_width)
        patch_dim = channels * patch_height * patch_width
        assert pool in {'cls', 'mean'}, 'pool type must be either cls (cls token) or mean (mean pooling)'
 
        self.to_patch_embedding = nn.Sequential(
            nn.Conv2d(channels, dim, kernel_size=patch_height, stride=patch_height),
        )
        self.pos_embedding = nn.Parameter(torch.randn(1, num_patches + 1, dim))
        self.cls_token = nn.Parameter(torch.randn(1, 1, dim))
        self.dropout = nn.Dropout(emb_dropout)
 
        self.transformer = Transformer(dim, depth, heads, mlp_dim, dropout)
 
        self.pool = pool
        self.to_latent = nn.Identity()
 
        self.mlp_head = nn.Sequential(
            nn.LayerNorm(dim),
            nn.Linear(dim, num_classes)
        )
 
    def forward(self, img):
        x = self.to_patch_embedding(img).flatten(2).transpose(1,2) ## img 1 3 224 224  输出形状x : 1 196 1024
        b, n, _ = x.shape ##
        cls_tokens = self.cls_token.expand(x.shape[0], -1, -1)
        # cls_tokens = repeat(self.cls_token, '() n d -> b n d', b = b)
        x = torch.cat((cls_tokens, x), dim=1)
        x += self.pos_embedding[:, :(n + 1)]
        x = self.dropout(x)
 
        x = self.transformer(x)
 
        x = x.mean(dim = 1) if self.pool == 'mean' else x[:, 0]
 
        x = self.to_latent(x)
        return self.mlp_head(x)
 
v = ViT(
    image_size = 224,
    patch_size = 16,
    num_classes = 1000,
    dim = 768,
    depth = 6,
    heads = 8,
    mlp_dim = 768*4,
    dropout = 0.1,
    emb_dropout = 0.1
)
 
img = torch.randn(1, 3, 224, 224)
preds = v(img) # (1, 1000)