cisen/model/clip.py

from collections import OrderedDict
from typing import Tuple, Union
from transformers import AutoTokenizer, AutoModelForSequenceClassification
import numpy as np
import torch
import torch.nn.functional as F
from torch import nn
from ..utils.dataset import tokenize
from ..utils.simple_tokenizer import SimpleTokenizer as _Tokenizer
_tokenizer = _Tokenizer()


class Bottleneck(nn.Module):
    expansion = 4

    def __init__(self, inplanes, planes, stride=1):
        super().__init__()

        # all conv layers have stride 1. an avgpool is performed after the second convolution when stride > 1
        self.conv1 = nn.Conv2d(inplanes, planes, 1, bias=False)
        self.bn1 = nn.BatchNorm2d(planes)

        self.conv2 = nn.Conv2d(planes, planes, 3, padding=1, bias=False)
        self.bn2 = nn.BatchNorm2d(planes)

        self.avgpool = nn.AvgPool2d(stride) if stride > 1 else nn.Identity()

        self.conv3 = nn.Conv2d(planes, planes * self.expansion, 1, bias=False)
        self.bn3 = nn.BatchNorm2d(planes * self.expansion)

        self.relu = nn.ReLU(inplace=True)
        self.downsample = None
        self.stride = stride

        if stride > 1 or inplanes != planes * Bottleneck.expansion:
            # downsampling layer is prepended with an avgpool, and the subsequent convolution has stride 1
            self.downsample = nn.Sequential(
                OrderedDict([("-1", nn.AvgPool2d(stride)),
                             ("0",
                              nn.Conv2d(inplanes,
                                        planes * self.expansion,
                                        1,
                                        stride=1,
                                        bias=False)),
                             ("1", nn.BatchNorm2d(planes * self.expansion))]))

    def forward(self, x: torch.Tensor):
        identity = x

        out = self.relu(self.bn1(self.conv1(x)))
        out = self.relu(self.bn2(self.conv2(out)))
        out = self.avgpool(out)
        out = self.bn3(self.conv3(out))

        if self.downsample is not None:
            identity = self.downsample(x)

        out += identity
        out = self.relu(out)
        return out


"""
    attenpool used in CRIS (output: C1/C2/C3  3 deiffent feature maps)
"""
class ModifiedAttentionPool2d(nn.Module):
    def __init__(self,
                 spacial_dim: int,
                 embed_dim: int,
                 num_heads: int,
                 output_dim: int = None):
        super().__init__()
        self.spacial_dim = spacial_dim
        self.positional_embedding = nn.Parameter(
            torch.randn(spacial_dim**2 + 1, embed_dim) / embed_dim**0.5)
        self.k_proj = nn.Linear(embed_dim, embed_dim)
        self.q_proj = nn.Linear(embed_dim, embed_dim)
        self.v_proj = nn.Linear(embed_dim, embed_dim)
        self.c_proj = nn.Linear(embed_dim, output_dim or embed_dim)
        self.num_heads = num_heads
        # residual
        self.connect = nn.Sequential(
            nn.Conv2d(embed_dim, output_dim, 1, stride=1, bias=False),
            nn.BatchNorm2d(output_dim))

    def resize_pos_embed(self, pos_embed, input_shpae):
        """Resize pos_embed weights.
        Resize pos_embed using bicubic interpolate method.
        Args:
            pos_embed (torch.Tensor): Position embedding weights.
            input_shpae (tuple): Tuple for (downsampled input image height,
                downsampled input image width).
            pos_shape (tuple): The resolution of downsampled origin training
                image.
            mode (str): Algorithm used for upsampling:
                ``'nearest'`` | ``'linear'`` | ``'bilinear'`` | ``'bicubic'`` |
                ``'trilinear'``. Default: ``'nearest'``
        Return:
            torch.Tensor: The resized pos_embed of shape [B, C, L_new]
        """
        assert pos_embed.ndim == 3, 'shape of pos_embed must be [B, L, C]'
        pos_h = pos_w = self.spacial_dim
        cls_token_weight = pos_embed[:, 0]
        pos_embed_weight = pos_embed[:, (-1 * pos_h * pos_w):]
        pos_embed_weight = pos_embed_weight.reshape(
            1, pos_h, pos_w, pos_embed.shape[2]).permute(0, 3, 1, 2)
        pos_embed_weight = F.interpolate(pos_embed_weight,
                                         size=input_shpae,
                                         align_corners=False,
                                         mode='bicubic')
        cls_token_weight = cls_token_weight.unsqueeze(1)
        pos_embed_weight = torch.flatten(pos_embed_weight, 2).transpose(1, 2)
        # pos_embed = torch.cat((cls_token_weight, pos_embed_weight), dim=1)
        return pos_embed_weight.transpose(-2, -1)

    def forward(self, x):
        B, C, H, W = x.size()
        res = self.connect(x)
        x = x.reshape(B, C, -1)  # NC(HW)
        # x = torch.cat([x.mean(dim=-1, keepdim=True), x], dim=-1)  # NC(1+HW)
        pos_embed = self.positional_embedding.unsqueeze(0)
        pos_embed = self.resize_pos_embed(pos_embed, (H, W))  # NC(HW)
        x = x + pos_embed.to(x.dtype)  # NC(HW)
        x = x.permute(2, 0, 1)  # (HW)NC
        x, _ = F.multi_head_attention_forward(
            query=x,
            key=x,
            value=x,
            embed_dim_to_check=x.shape[-1],
            num_heads=self.num_heads,
            q_proj_weight=self.q_proj.weight,
            k_proj_weight=self.k_proj.weight,
            v_proj_weight=self.v_proj.weight,
            in_proj_weight=None,
            in_proj_bias=torch.cat(
                [self.q_proj.bias, self.k_proj.bias, self.v_proj.bias]),
            bias_k=None,
            bias_v=None,
            add_zero_attn=False,
            dropout_p=0,
            out_proj_weight=self.c_proj.weight,
            out_proj_bias=self.c_proj.bias,
            use_separate_proj_weight=True,
            training=self.training,
            need_weights=False)
        xt = x[0]
        x = x.permute(1, 2, 0).reshape(B, -1, H, W)
        x = x + res
        x = F.relu(x, True)

        return x, xt


"""
    attenpool used in Clip (output: a tensor (b, dim) image encoding)
"""
class AttentionPool2d(nn.Module):
    def __init__(self, spacial_dim: int, embed_dim: int, num_heads: int, output_dim: int = None):
        super().__init__()
        self.positional_embedding = nn.Parameter(torch.randn(spacial_dim ** 2 + 1, embed_dim) / embed_dim ** 0.5)
        self.k_proj = nn.Linear(embed_dim, embed_dim)
        self.q_proj = nn.Linear(embed_dim, embed_dim)
        self.v_proj = nn.Linear(embed_dim, embed_dim)
        self.c_proj = nn.Linear(embed_dim, output_dim or embed_dim)
        self.num_heads = num_heads

    def forward(self, x):
        x = x.flatten(start_dim=2).permute(2, 0, 1)  # NCHW -> (HW)NC
        x = torch.cat([x.mean(dim=0, keepdim=True), x], dim=0)  # (HW+1)NC
        x = x + self.positional_embedding[:, None, :].to(x.dtype)  # (HW+1)NC
        x, _ = F.multi_head_attention_forward(
            query=x[:1], key=x, value=x,
            embed_dim_to_check=x.shape[-1],
            num_heads=self.num_heads,
            q_proj_weight=self.q_proj.weight,
            k_proj_weight=self.k_proj.weight,
            v_proj_weight=self.v_proj.weight,
            in_proj_weight=None,
            in_proj_bias=torch.cat([self.q_proj.bias, self.k_proj.bias, self.v_proj.bias]),
            bias_k=None,
            bias_v=None,
            add_zero_attn=False,
            dropout_p=0,
            out_proj_weight=self.c_proj.weight,
            out_proj_bias=self.c_proj.bias,
            use_separate_proj_weight=True,
            training=self.training,
            need_weights=False
        )
        return x.squeeze(0)


class ModifiedResNet(nn.Module):
    """
    A ResNet class that is similar to torchvision's but contains the following changes:
    - There are now 3 "stem" convolutions as opposed to 1, with an average pool instead of a max pool.
    - Performs anti-aliasing strided convolutions, where an avgpool is prepended to convolutions with stride > 1
    - The final pooling layer is a QKV attention instead of an average pool
    """
    def __init__(self,
                 layers,
                 output_dim,
                 heads,
                 input_resolution=224,
                 width=64):
        super().__init__()
        self.output_dim = output_dim
        self.input_resolution = input_resolution

        # the 3-layer stem
        self.conv1 = nn.Conv2d(3,
                               width // 2,
                               kernel_size=3,
                               stride=2,
                               padding=1,
                               bias=False)
        self.bn1 = nn.BatchNorm2d(width // 2)
        self.conv2 = nn.Conv2d(width // 2,
                               width // 2,
                               kernel_size=3,
                               padding=1,
                               bias=False)
        self.bn2 = nn.BatchNorm2d(width // 2)
        self.conv3 = nn.Conv2d(width // 2,
                               width,
                               kernel_size=3,
                               padding=1,
                               bias=False)
        self.bn3 = nn.BatchNorm2d(width)
        self.avgpool = nn.AvgPool2d(2)
        self.relu = nn.ReLU(inplace=True)

        # residual layers
        self._inplanes = width  # this is a *mutable* variable used during construction
        self.layer1 = self._make_layer(width, layers[0])
        self.layer2 = self._make_layer(width * 2, layers[1], stride=2)
        self.layer3 = self._make_layer(width * 4, layers[2], stride=2)
        self.layer4 = self._make_layer(width * 8, layers[3], stride=2)

        embed_dim = width * 32  # the ResNet feature dimension

        self.attnpool = AttentionPool2d(input_resolution // 32, embed_dim,
                                        heads, output_dim)
        # self.modifiedattnpool = ModifiedAttentionPool2d(input_resolution // 32, embed_dim,
        #                                 heads, output_dim)

    def _make_layer(self, planes, blocks, stride=1):
        layers = [Bottleneck(self._inplanes, planes, stride)]

        self._inplanes = planes * Bottleneck.expansion
        for _ in range(1, blocks):
            layers.append(Bottleneck(self._inplanes, planes))

        return nn.Sequential(*layers)

    def forward(self, x):
        def stem(x):
            for conv, bn in [(self.conv1, self.bn1), (self.conv2, self.bn2),
                             (self.conv3, self.bn3)]:

                x = self.relu(bn(conv(x)))

            x = self.avgpool(x)
            return x

        x = x.type(self.conv1.weight.dtype)
        x = stem(x)

        x = self.layer1(x)

        x2 = self.layer2(x)

        x3 = self.layer3(x2)
        x4 = self.layer4(x3)
        x5 = self.attnpool(x4)
        # x4 = self.modifiedattnpool(x4)

        return (x2, x3, x4), x5


class LayerNorm(nn.LayerNorm):
    """Subclass torch's LayerNorm to handle fp16."""
    def forward(self, x: torch.Tensor):
        orig_type = x.dtype
        ret = super().forward(x.type(torch.float32))
        return ret.type(orig_type)


class QuickGELU(nn.Module):
    def forward(self, x: torch.Tensor):
        return x * torch.sigmoid(1.702 * x)


class ResidualAttentionBlock(nn.Module):
    def __init__(self,
                 d_model: int,
                 n_head: int,
                 attn_mask: torch.Tensor = None):
        super().__init__()
        # print(n_head)
        self.attn = nn.MultiheadAttention(d_model, n_head)
        self.ln_1 = LayerNorm(d_model)
        self.mlp = nn.Sequential(
            OrderedDict([("c_fc", nn.Linear(d_model, d_model * 4)),
                         ("gelu", QuickGELU()),
                         ("c_proj", nn.Linear(d_model * 4, d_model))]))
        self.ln_2 = LayerNorm(d_model)
        self.attn_mask = attn_mask

    def attention(self, x: torch.Tensor):
        self.attn_mask = self.attn_mask.to(
            dtype=x.dtype,
            device=x.device) if self.attn_mask is not None else None
        res = self.attn(x, x, x, need_weights=False,
                         attn_mask=self.attn_mask)[0]
        # print(res)
        return res

    def forward(self, x: torch.Tensor):
        # a = self.attention(self.ln_1(x))
        x = x + self.attention(self.ln_1(x))

        x = x + self.mlp(self.ln_2(x))
        return x

class Transformer(nn.Module):
    def __init__(self, width: int, layers: int, heads: int, attn_mask: torch.Tensor = None):
        super().__init__()
        self.width = width
        self.layers = layers
        self.resblocks = nn.Sequential(*[ResidualAttentionBlock(width, heads, attn_mask) for _ in range(layers)])

    def forward(self, x: torch.Tensor):
        return self.resblocks(x)

class ViTTransformer(nn.Module):
    def __init__(self, width: int, layers: int, heads: int, attn_mask: torch.Tensor = None):
        super().__init__()
        self.width = width
        self.layers = layers
        self.resblocks = nn.Sequential(*[ResidualAttentionBlock(width, heads, attn_mask) for _ in range(layers)])

    def forward(self, x: torch.Tensor):
        outputs = []
        i = 1
        for block in self.resblocks:
            x = block(x)
            if i > 7:
                outputs.append(x)
            i = i + 1
        return outputs


class VisionTransformer(nn.Module):
    def __init__(self, input_resolution: int, patch_size: int, width: int,
                 layers: int, heads: int, output_dim: int):
        super().__init__()
        self.input_resolution = input_resolution
        self.output_dim = output_dim
        self.conv1 = nn.Conv2d(in_channels=3,
                               out_channels=width,
                               kernel_size=patch_size,
                               stride=patch_size,
                               bias=False)

        scale = width ** -0.5
        self.class_embedding = nn.Parameter(scale * torch.randn(width))
        self.positional_embedding = nn.Parameter(scale * torch.randn(
            (input_resolution // patch_size) ** 2 + 1, width))
        self.ln_pre = LayerNorm(width)

        self.transformer = ViTTransformer(width, layers, heads)

        self.ln_post = LayerNorm(width)
        self.proj = nn.Parameter(scale * torch.randn(width, output_dim))

    def forward(self, x: torch.Tensor):
        # input: batch, 3, 224, 224

        # batch, 1024, 16, 16
        x = self.conv1(x)  # shape = [*, width, grid, grid]
        # batch, 1024, 256
        x = x.reshape(x.shape[0], x.shape[1],
                      -1)  # shape = [*, width, grid ** 2]
        # batch, 256, 1024
        x = x.permute(0, 2, 1)  # shape = [*, grid ** 2, width]
        # batch, 257, 1024
        x = torch.cat([
            self.class_embedding.to(x.dtype) + torch.zeros(
                x.shape[0], 1, x.shape[-1], dtype=x.dtype, device=x.device), x
        ],
            dim=1)  # shape = [*, grid ** 2 + 1, width]

        x = x + self.positional_embedding.to(x.dtype)

        x = self.ln_pre(x)
        # 257, batch, 1024
        x = x.permute(1, 0, 2)  # NLD -> LND

        out = self.transformer(x)
        # batch, 257, 1024
        x1, x2 ,x3, x4 = out[0], out[1], out[2], out[3]
        x1 = x1.permute(1, 0, 2)
        x2 = x2.permute(1, 0, 2)
        x3 = x3.permute(1, 0, 2)
        x4 = x4.permute(1, 0, 2)  # LND -> NLD

        # 用于分类
        x = self.ln_post(x4[:, 0, :])
        #feature
        # x_f = self.ln_post(x[:, 1:, :])


        if self.proj is not None:
            x = x @ self.proj

        return (x1[:, 1:, :], x2[:, 1:, :], x3[:, 1:, :], x4[:, 1:, :]), x

class ModifiedVisionTransformer(nn.Module):
    def __init__(self, input_resolution: int, patch_size: int, width: int,
                 layers: int, heads: int, output_dim: int):
        super().__init__()
        self.input_resolution = input_resolution
        self.output_dim = output_dim
        self.conv1 = nn.Conv2d(in_channels=3,
                               out_channels=width,
                               kernel_size=patch_size,
                               stride=patch_size,
                               bias=False)

        self.conv2 = nn.Conv2d(in_channels=3,
                               out_channels=width // 2,
                               kernel_size=patch_size // 2,
                               stride=patch_size // 2,
                               bias=False)

        self.conv3 = nn.Conv2d(in_channels=3,
                               out_channels=width,
                               kernel_size=patch_size * 2,
                               stride=patch_size * 2,
                               bias=False)
        self.conv_layers = [self.conv1, self.conv2]
        scale = width**-0.5

        self.class_embedding1 = nn.Parameter(scale * torch.randn(width))
        self.class_embedding2 = nn.Parameter(scale * torch.randn(width // 2))
        self.cls_layers = [self.class_embedding1, self.class_embedding2]

        self.positional_embedding1 = nn.Parameter(scale * torch.randn(
            (input_resolution // patch_size)**2 + 1, width))
        self.positional_embedding2 = nn.Parameter(scale * torch.randn(
            (input_resolution // (patch_size // 2)) ** 2 + 1, width // 2))
        self.pos_layers = [self.positional_embedding1, self.positional_embedding2]

        self.ln_pre1 = LayerNorm(width)
        self.ln_pre2 = LayerNorm(width // 2)
        self.pre_layers = [self.ln_pre1, self.ln_pre2]

        self.transformer1 = Transformer(width, layers, heads)
        self.transformer2 = Transformer(width // 2, layers, heads)
        self.tran_layers = [self.transformer1, self.transformer2]

        self.ln_post1 = LayerNorm(width)
        self.ln_post2 = LayerNorm(width // 2)
        self.post_layers = [self.ln_post1, self.ln_post2]

        self.proj1 = nn.Parameter(scale * torch.randn(width, output_dim * 2))
        self.proj2 = nn.Parameter(scale * torch.randn(width // 2, output_dim))
        self.proj_layers = [self.proj1, self.proj2]


    def forward(self, x: torch.Tensor):
        # input: batch, 3, 224, 224
        input = x
        # batch, 1024, 16, 16
        out = []
        f = []
        cl = []
        for i in range(2):
            x = self.conv_layers[i](input)  # shape = [*, width, grid, grid]

            b, c, w, h = x.shape
            # batch, 1024, 256
            x = x.reshape(x.shape[0], x.shape[1],
                          -1)  # shape = [*, width, grid ** 2]
            # batch, 256, 1024
            x = x.permute(0, 2, 1)  # shape = [*, grid ** 2, width]
            # batch, 257, 1024
            x = torch.cat([
                self.cls_layers[i].to(x.dtype) + torch.zeros(
                    x.shape[0], 1, x.shape[-1], dtype=x.dtype, device=x.device), x
            ],
                          dim=1)  # shape = [*, grid ** 2 + 1, width]

            x = x + self.pos_layers[i].to(x.dtype)

            x = self.pre_layers[i](x)
            # 257, batch, 1024
            x = x.permute(1, 0, 2)  # NLD -> LND

            x, cls = self.tran_layers[i](x)
            # batch, 257, 1024
            x = x.permute(1, 0, 2)  # LND -> NLD

            # 用于分类
            # x = self.ln_post(x[:, 0, :])
            # feature
            x = self.post_layers[i](x[:, 1:, :])



            if self.proj_layers[i] is not None:
                x = x @ self.proj_layers[i]
                cls = [j @ self.proj_layers[i] for j in cls]

            feat = x.permute(0,2,1).reshape(b, x.shape[2] , w, h)
            out.append(x)
            f.append(feat)
            cl.append(cls)
        return out, f, cl

"""
    Long CLIP
"""
class LCLIP(nn.Module):
    def __init__(self,
                 embed_dim: int,
                 # vision
                 image_resolution: int,
                 vision_layers: Union[Tuple[int, int, int, int], int],
                 vision_width: int,
                 vision_patch_size: int,
                 # text
                 context_length: int,
                 vocab_size: int,
                 transformer_width: int,
                 transformer_heads: int,
                 transformer_layers: int, 
                 load_from_clip: bool
                 ):
        super().__init__()
        self.context_length = 248

        if isinstance(vision_layers, (tuple, list)):
            vision_heads = vision_width * 32 // 64
            self.visual = ModifiedResNet(
                layers=vision_layers,
                output_dim=embed_dim,
                heads=vision_heads,
                input_resolution=image_resolution,
                width=vision_width
            )
        else:
            vision_heads = vision_width // 64
            self.visual = VisionTransformer(
                input_resolution=image_resolution,
                patch_size=vision_patch_size,
                width=vision_width,
                layers=vision_layers,
                heads=vision_heads,
                output_dim=embed_dim
            )

        self.transformer = Transformer(
            width=transformer_width,
            layers=transformer_layers,
            heads=transformer_heads,
            attn_mask=self.build_attention_mask()
        )

        self.vocab_size = vocab_size
        self.token_embedding = nn.Embedding(vocab_size, transformer_width)
        # self.positional_embedding = nn.Parameter(torch.empty(248, transformer_width))

        if load_from_clip == False:
            self.positional_embedding = nn.Parameter(torch.empty(248, transformer_width))
            self.positional_embedding_res = nn.Parameter(torch.empty(248, transformer_width))

        else:
            self.positional_embedding = nn.Parameter(torch.empty(248, transformer_width))

        self.ln_final = LayerNorm(transformer_width)

        self.text_projection = nn.Parameter(torch.empty(transformer_width, embed_dim))
        self.logit_scale = nn.Parameter(torch.ones([]) * np.log(1 / 0.07))

        self.initialize_parameters()
        self.mask1 = torch.zeros([248, 1])
        self.mask1[:20, :] = 1
        self.mask2 = torch.zeros([248, 1])
        self.mask2[20:, :] = 1

    
    def initialize_parameters(self):
        nn.init.normal_(self.token_embedding.weight, std=0.02)
        nn.init.normal_(self.positional_embedding, std=0.01)

        if isinstance(self.visual, ModifiedResNet):
            if self.visual.attnpool is not None:
                std = self.visual.attnpool.c_proj.in_features ** -0.5
                nn.init.normal_(self.visual.attnpool.q_proj.weight, std=std)
                nn.init.normal_(self.visual.attnpool.k_proj.weight, std=std)
                nn.init.normal_(self.visual.attnpool.v_proj.weight, std=std)
                nn.init.normal_(self.visual.attnpool.c_proj.weight, std=std)

            for resnet_block in [self.visual.layer1, self.visual.layer2, self.visual.layer3, self.visual.layer4]:
                for name, param in resnet_block.named_parameters():
                    if name.endswith("bn3.weight"):
                        nn.init.zeros_(param)

        proj_std = (self.transformer.width ** -0.5) * ((2 * self.transformer.layers) ** -0.5)
        attn_std = self.transformer.width ** -0.5
        fc_std = (2 * self.transformer.width) ** -0.5
        for block in self.transformer.resblocks:
            nn.init.normal_(block.attn.in_proj_weight, std=attn_std)
            nn.init.normal_(block.attn.out_proj.weight, std=proj_std)
            nn.init.normal_(block.mlp.c_fc.weight, std=fc_std)
            nn.init.normal_(block.mlp.c_proj.weight, std=proj_std)

        if self.text_projection is not None:
            nn.init.normal_(self.text_projection, std=self.transformer.width ** -0.5)

    def build_attention_mask(self):
        # lazily create causal attention mask, with full attention between the vision tokens
        # pytorch uses additive attention mask; fill with -inf
        mask = torch.empty(self.context_length, self.context_length)
        mask.fill_(float("-inf"))
        mask.triu_(1)  # zero out the lower diagonal
        return mask

    @property
    def dtype(self):
        return self.visual.conv1.weight.dtype

    def encode_image(self, image):
        return self.visual(image.type(self.dtype))

    def encode_text(self, text): 
        x = self.token_embedding(text).type(self.dtype)  # [batch_size, n_ctx, d_model]
        
        # x = x + (self.positional_embedding.to(x.device) * self.mask1.to(x.device)).type(self.dtype).to(x.device) + (self.positional_embedding_res.to(x.device) * self.mask2.to(x.device)).type(self.dtype).to(x.device) 
        x = x + (self.positional_embedding.to(x.device) * self.mask1.to(x.device)).type(self.dtype).to(x.device)
        x = x.permute(1, 0, 2)  # NLD -> LND
        x = self.transformer(x)
        x = x.permute(1, 0, 2)  # LND -> NLD
        x = self.ln_final(x).type(self.dtype)

        # x.shape = [batch_size, n_ctx, transformer.width]
        # take features from the eot embedding (eot_token is the highest number in each sequence)
        x = x[torch.arange(x.shape[0]), text.argmax(dim=-1)] @ self.text_projection

        return x

    def encode_text_full(self, text): 
        x = self.token_embedding(text).type(self.dtype)  # [batch_size, n_ctx, d_model]
        
        x = x + (self.positional_embedding.to(x.device) * self.mask1.to(x.device)).type(self.dtype).to(x.device) + (self.positional_embedding_res.to(x.device) * self.mask2.to(x.device)).type(self.dtype).to(x.device) 
        
        x = x.permute(1, 0, 2)  # NLD -> LND
        x = self.transformer(x)
        x = x.permute(1, 0, 2)  # LND -> NLD
        x = self.ln_final(x).type(self.dtype)

        # x.shape = [batch_size, n_ctx, transformer.width]
        # take features from the eot embedding (eot_token is the highest number in each sequence)
        #x = x[torch.arange(x.shape[0]), text.argmax(dim=-1)] @ self.text_projection

        return x


    def forward(self, image, text):
        image_features = self.encode_image(image)
        text_features, _ = self.encode_text(text)

        # normalized features
        image_features = image_features / image_features.norm(dim=1, keepdim=True)
        text_features = text_features / text_features.norm(dim=1, keepdim=True)

        # cosine similarity as logits
        logit_scale = self.logit_scale.exp()
        logits_per_image = logit_scale * image_features @ text_features.t()
        logits_per_text = logits_per_image.t()

        # shape = [global_batch_size, global_batch_size]
        return logits_per_image, logits_per_text
"""
    original CLIP
"""
class CLIP(nn.Module):
    def __init__(
            self,
            embed_dim: int,
            # vision
            image_resolution: int,
            vision_layers: Union[Tuple[int, int, int, int], int],
            vision_width: int,
            vision_patch_size: int,
            # text
            context_length: int,
            txt_length: int,
            vocab_size: int,
            transformer_width: int,
            transformer_heads: int,
            transformer_layers: int):
        super().__init__()

        self.context_length = context_length

        if isinstance(vision_layers, (tuple, list)):
            vision_heads = vision_width * 32 // 64
            self.visual = ModifiedResNet(layers=vision_layers,
                                         output_dim=embed_dim,
                                         heads=vision_heads,
                                         input_resolution=image_resolution,
                                         width=vision_width)
            # self.fq_attnpool = AttentionPool2d(image_resolution // 32, vision_width* 32,
            #                                    vision_heads, embed_dim)
        else:
            vision_heads = vision_width // 64
            self.visual = VisionTransformer(input_resolution=image_resolution,
                                            patch_size=vision_patch_size,
                                            width=vision_width,
                                            layers=vision_layers,
                                            heads=vision_heads,
                                            output_dim=embed_dim)

        self.transformer = Transformer(
            width=transformer_width,
            layers=transformer_layers,
            heads=transformer_heads,
            attn_mask=self.build_attention_mask(txt_length))

        self.vocab_size = vocab_size
        self.token_embedding = nn.Embedding(vocab_size, transformer_width)
        self.positional_embedding = nn.Parameter(
            torch.empty(self.context_length, transformer_width))
        self.ln_final = LayerNorm(transformer_width)

        self.text_projection = nn.Parameter(
            torch.empty(transformer_width, embed_dim))
        self.logit_scale = nn.Parameter(torch.ones([]) * np.log(1 / 0.07))

        self.token_embedding.requires_grad_ = False
        self.initialize_parameters()

    def initialize_parameters(self):
        nn.init.normal_(self.token_embedding.weight, std=0.02)
        nn.init.normal_(self.positional_embedding, std=0.01)

        if isinstance(self.visual, ModifiedResNet):
            if self.visual.attnpool is not None:
                std = self.visual.attnpool.c_proj.in_features**-0.5
                nn.init.normal_(self.visual.attnpool.q_proj.weight, std=std)
                nn.init.normal_(self.visual.attnpool.k_proj.weight, std=std)
                nn.init.normal_(self.visual.attnpool.v_proj.weight, std=std)
                nn.init.normal_(self.visual.attnpool.c_proj.weight, std=std)

            for resnet_block in [
                    self.visual.layer1, self.visual.layer2, self.visual.layer3,
                    self.visual.layer4
            ]:
                for name, param in resnet_block.named_parameters():
                    if name.endswith("bn3.weight"):
                        nn.init.zeros_(param)

        proj_std = (self.transformer.width**-0.5) * (
            (2 * self.transformer.layers)**-0.5)
        attn_std = self.transformer.width**-0.5
        fc_std = (2 * self.transformer.width)**-0.5
        for block in self.transformer.resblocks:
            nn.init.normal_(block.attn.in_proj_weight, std=attn_std)
            nn.init.normal_(block.attn.out_proj.weight, std=proj_std)
            nn.init.normal_(block.mlp.c_fc.weight, std=fc_std)
            nn.init.normal_(block.mlp.c_proj.weight, std=proj_std)

        if self.text_projection is not None:
            nn.init.normal_(self.text_projection,
                            std=self.transformer.width**-0.5)

    def build_attention_mask(self, context_length):
        # lazily create causal attention mask, with full attention between the vision tokens
        # pytorch uses additive attention mask; fill with -inf
        mask = torch.empty(context_length, context_length)
        mask.fill_(float("-inf"))
        mask.triu_(1)  # zero out the lower diagonal
        return mask

    @property
    def dtype(self):
        return self.visual.conv1.weight.dtype

    def encode_image(self, image):
        return self.visual(image.type(self.dtype))

    def encode_fq(self, image):
        return self.fq_attnpool(image.type(self.dtype))

    def encode_text(self, text):
        a = self.token_embedding
        x = self.token_embedding(text).type(
            self.dtype)  # [batch_size, n_ctx, d_model]

        x = x + self.positional_embedding.type(self.dtype)[:x.size(1)]
        # print(x.shape)
        # print(x)
        
        x = x.permute(1, 0, 2)  # NLD -> LND
        x = self.transformer(x)
        x = x.permute(1, 0, 2)  # LND -> NLD
        x = self.ln_final(x).type(self.dtype)
        # print(text[0])
        # x.shape = [batch_size, n_ctx, transformer.width]
        # take features from the eot embedding (eot_token is the highest number in each sequence)
        state = x[torch.arange(x.shape[0]),
                  text.argmax(dim=-1)] @ self.text_projection
        # x = x @ self.text_projection
        # state = x[torch.arange(x.shape[0]), text.argmax(dim=-1)]

        return x, state

    def forward(self, image, text):
        image_features = self.encode_image(image)
        text_features = self.encode_text(text)

        # normalized features
        image_features = image_features / image_features.norm(dim=-1,
                                                              keepdim=True)
        text_features = text_features / text_features.norm(dim=-1,
                                                           keepdim=True)

        # cosine similarity as logits
        logit_scale = self.logit_scale.exp()
        logits_per_image = logit_scale * image_features @ text_features.t()
        logits_per_text = logits_per_image.t()

        # shape = [global_batch_size, global_batch_size]
        return logits_per_image, logits_per_text

"""
    modified CLIP : without text encoder
"""

class zhCLIP(nn.Module):
    def __init__(self,
            embed_dim,
            # vision
            image_resolution: int,
            vision_layers: Union[Tuple[int, int, int, int], int],
            vision_width: int,
            vision_patch_size: int):
        super().__init__()



        if isinstance(vision_layers, (tuple, list)):
            vision_heads = vision_width * 32 // 64
            self.visual = ModifiedResNet(layers=vision_layers,
                                         output_dim=embed_dim,
                                         heads=vision_heads,
                                         input_resolution=image_resolution,
                                         width=vision_width)
            self.fq_attnpool = AttentionPool2d(image_resolution // 32, vision_width* 32,
                                               vision_heads, embed_dim)
        else:
            vision_heads = vision_width // 64
            self.visual = ModifiedVisionTransformer(input_resolution=image_resolution,
                                            patch_size=vision_patch_size,
                                            width=vision_width,
                                            layers=vision_layers,
                                            heads=vision_heads,
                                            output_dim=embed_dim)

        self.logit_scale = nn.Parameter(torch.ones([]) * np.log(1 / 0.07))
        self.initialize_parameters()

    def initialize_parameters(self):

        if isinstance(self.visual, ModifiedResNet):
            if self.visual.attnpool is not None:
                std = self.visual.attnpool.c_proj.in_features**-0.5
                nn.init.normal_(self.visual.attnpool.q_proj.weight, std=std)
                nn.init.normal_(self.visual.attnpool.k_proj.weight, std=std)
                nn.init.normal_(self.visual.attnpool.v_proj.weight, std=std)
                nn.init.normal_(self.visual.attnpool.c_proj.weight, std=std)

            for resnet_block in [
                    self.visual.layer1, self.visual.layer2, self.visual.layer3,
                    self.visual.layer4
            ]:
                for name, param in resnet_block.named_parameters():
                    if name.endswith("bn3.weight"):
                        nn.init.zeros_(param)


    def build_attention_mask(self, context_length):
        # lazily create causal attention mask, with full attention between the vision tokens
        # pytorch uses additive attention mask; fill with -inf
        mask = torch.empty(context_length, context_length)
        mask.fill_(float("-inf"))
        mask.triu_(1)  # zero out the lower diagonal
        return mask

    @property
    def dtype(self):
        return self.visual.conv1.weight.dtype

    def encode_image(self, image):
        return self.visual(image.type(self.dtype))

    def encode_fq(self, image):
        return self.fq_attnpool(image.type(self.dtype))

    def forward(self, image, text):
        image_features = self.encode_image(image)
        text_features = self.encode_text(text)

        # normalized features
        image_features = image_features / image_features.norm(dim=-1,
                                                              keepdim=True)
        text_features = text_features / text_features.norm(dim=-1,
                                                           keepdim=True)

        # cosine similarity as logits
        logit_scale = self.logit_scale.exp()
        logits_per_image = logit_scale * image_features @ text_features.t()
        logits_per_text = logits_per_image.t()

        # shape = [global_batch_size, global_batch_size]
        return logits_per_image, logits_per_text


def convert_weights(model: nn.Module):
    """Convert applicable model parameters to fp16"""
    def _convert_weights_to_fp16(l):
        if isinstance(l, (nn.Conv1d, nn.Conv2d, nn.Linear)):
            l.weight.data = l.weight.data.half()
            if l.bias is not None:
                l.bias.data = l.bias.data.half()

        if isinstance(l, nn.MultiheadAttention):
            for attr in [
                    *[f"{s}_proj_weight" for s in ["in", "q", "k", "v"]],
                    "in_proj_bias", "bias_k", "bias_v"
            ]:
                tensor = getattr(l, attr)
                if tensor is not None:
                    tensor.data = tensor.data.half()

        for name in ["text_projection", "proj"]:
            if hasattr(l, name):
                attr = getattr(l, name)
                if attr is not None:
                    attr.data = attr.data.half()

    model.apply(_convert_weights_to_fp16)

class PromptLearner(nn.Module):

    def __init__(self, transformer_width, context_length, vocab_size,
                 transformer_layers, transformer_heads, bert_embed_dim):
        super().__init__()

        self.transformer_width = transformer_width
        self.context_length = context_length
        self.vocab_size = vocab_size
        self.token_embedding = nn.Embedding(self.vocab_size, self.transformer_width)

        self.transformer = Transformer(
            width=transformer_width,
            layers=transformer_layers,
            heads=transformer_heads,
            attn_mask=self.build_attention_mask()
        )

        self.positional_embedding = nn.Parameter(torch.empty(self.context_length, transformer_width))
        self.ln_final = LayerNorm(transformer_width)

        self.text_projection = nn.Parameter(torch.empty(transformer_width, bert_embed_dim))


        # self.load_from_openai_model(pretrained_model=clip_pretrain)

    def build_attention_mask(self):
        # lazily create causal attention mask, with full attention between the vision tokens
        # pytorch uses additive attention mask; fill with -inf
        mask = torch.empty(self.context_length, self.context_length)
        mask.fill_(float("-inf"))
        mask.triu_(1)  # zero out the lower diagonal
        return mask

    def init_label_emb(self, labels_path):

        label = open(labels_path, 'r').readlines()
        # label81 = open(unseen_labels_path, 'r').readlines()
        # label1006 = label925 + label81
        self.name_lens = [len(_tokenizer.encode(name)) for name in label]
        self.label_token = torch.zeros((len(self.name_lens), self.context_length), dtype=torch.long)
        for i, c in enumerate(label):
            self.label_token[i] = tokenize(f"There is a {c.strip()} in the scene")
        self.label_emb = torch.zeros((len(self.name_lens), max(self.name_lens), self.transformer_width))
        for i, embed in enumerate(self.token_embedding(self.label_token)):
            self.label_emb[i][:self.name_lens[i]] = embed[4:4 + self.name_lens[i]].clone().detach()

    # def load_from_openai_model(self, pretrained_model):
    #     state_dict = clip.load(pretrained_model, jit=False)[0].state_dict()
    #     load_dict = {}
    #     for k, v in state_dict.items():
    #         if not k.startswith("visual") and (
    #                 k not in ["logit_scale", "input_resolution", "context_length", "vocab_size"]):
    #             load_dict[k] = v
    #     msg = self.load_state_dict(load_dict)

    def load_label_emb(self, label=None):
        self.name_lens = [len(_tokenizer.encode(name.split("\t")[-1])) for name in label]
        self.label_token = torch.zeros((len(self.name_lens), self.context_length), dtype=torch.long).cuda()
        for i, c in enumerate(label):
            name = c.split("\t")[-1]
            self.label_token[i] = tokenize(f"There is a {name.strip()} in the scene")
        self.label_emb = torch.zeros((len(self.name_lens), max(self.name_lens), self.transformer_width)).cuda()
        for i, embed in enumerate(self.token_embedding(self.label_token)):
            self.label_emb[i][:self.name_lens[i]] = embed[4:4 + self.name_lens[i]].clone().detach()

    def forward(self, device):

        label_embeds = self.token_embedding(self.label_token.to(device))

        for i in range(label_embeds.shape[0]):
            label_embeds[i, 4:4 + self.name_lens[i], :] = self.label_emb[i][:self.name_lens[i]]

        x = label_embeds + self.positional_embedding
        x = x.permute(1, 0, 2)  # NLD -> LND

        x = self.transformer(x)
        x = x.permute(1, 0, 2)  # LND -> NLD
        x = self.ln_final(x)

        res = x[torch.arange(x.shape[0]), self.label_token.argmax(dim=-1)] @ self.text_projection

        return res

def build_promptlearner(state_dict: dict):
    embed_dim = state_dict["text_projection"].shape[1]
    context_length = state_dict["positional_embedding"].shape[0]
    vocab_size = state_dict["token_embedding.weight"].shape[0]
    transformer_width = state_dict["ln_final.weight"].shape[0]
    transformer_heads = transformer_width // 64
    transformer_layers = len(
        set(
            k.split(".")[2] for k in state_dict
            if k.startswith(f"transformer.resblocks")))
    model = PromptLearner(transformer_width, context_length, vocab_size,
                          transformer_layers, transformer_heads, embed_dim)
    # model = PromptLearner(embed_dim, vision_patch_size, context_length, txt_length, vocab_size,
    #              transformer_width, transformer_heads, transformer_layers)
    load_dict = {}
    for k, v in state_dict.items():
        if not k.startswith("visual") and (
                k not in ["logit_scale", "input_resolution", "context_length", "vocab_size"]):
            load_dict[k] = v

    convert_weights(model)
    model.load_state_dict(load_dict, False)

    return model
 
def build_model(state_dict: dict, txt_length: int):
    vit = "visual.proj" in state_dict

    if vit:
        vision_width = state_dict["visual.conv1.weight"].shape[0]
        vision_layers = len([
            k for k in state_dict.keys()
            if k.startswith("visual.") and k.endswith(".attn.in_proj_weight")
        ])
        vision_patch_size = state_dict["visual.conv1.weight"].shape[-1]
        grid_size = round(
            (state_dict["visual.positional_embedding"].shape[0] - 1)**0.5)
        image_resolution = vision_patch_size * grid_size
    else:
        counts: list = [
            len(
                set(
                    k.split(".")[2] for k in state_dict
                    if k.startswith(f"visual.layer{b}")))
            for b in [1, 2, 3, 4]
        ]
        vision_layers = tuple(counts)
        vision_width = state_dict["visual.layer1.0.conv1.weight"].shape[0]
        output_width = round(
            (state_dict["visual.attnpool.positional_embedding"].shape[0] -
             1)**0.5)
        vision_patch_size = None
        assert output_width**2 + 1 == state_dict[
            "visual.attnpool.positional_embedding"].shape[0]
        image_resolution = output_width * 32

    vision_heads = vision_width * 32 // 64
    embed_dim = state_dict["text_projection"].shape[1]
    # context_length = state_dict["positional_embedding"].shape[0]
    context_length = txt_length
    vocab_size = state_dict["token_embedding.weight"].shape[0]
    transformer_width = state_dict["ln_final.weight"].shape[0]
    transformer_heads = transformer_width // 64
    transformer_layers = len(
        set(
            k.split(".")[2] for k in state_dict
            if k.startswith(f"transformer.resblocks")))

    model = CLIP(embed_dim, image_resolution, vision_layers, vision_width,
                 vision_patch_size, context_length, txt_length, vocab_size,
                 transformer_width, transformer_heads, transformer_layers)

    for key in ["input_resolution", "context_length", "vocab_size", 'positional_embedding']:
        if key in state_dict:
            del state_dict[key]

    convert_weights(model)
    model.load_state_dict(state_dict, False)
    return model.eval(), image_resolution, vision_heads, embed_dim, vision_width, vision_patch_size

def build_lclip_model(state_dict: dict, load_from_clip: bool):
    vit = "visual.proj" in state_dict

    if vit:
        vision_width = state_dict["visual.conv1.weight"].shape[0]
        vision_layers = len([k for k in state_dict.keys() if k.startswith("visual.") and k.endswith(".attn.in_proj_weight")])
        vision_patch_size = state_dict["visual.conv1.weight"].shape[-1]
        grid_size = round((state_dict["visual.positional_embedding"].shape[0] - 1) ** 0.5)
        image_resolution = vision_patch_size * grid_size

    else:
        counts: list = [len(set(k.split(".")[2] for k in state_dict if k.startswith(f"visual.layer{b}"))) for b in [1, 2, 3, 4]]
        vision_layers = tuple(counts)
        vision_width = state_dict["visual.layer1.0.conv1.weight"].shape[0]
        output_width = round((state_dict["visual.attnpool.positional_embedding"].shape[0] - 1) ** 0.5)
        vision_patch_size = None
        assert output_width ** 2 + 1 == state_dict["visual.attnpool.positional_embedding"].shape[0]
        image_resolution = output_width * 32

    embed_dim = state_dict["text_projection"].shape[1]
    # print(embed_dim)
    context_length = state_dict["positional_embedding"].shape[0]
    vocab_size = state_dict["token_embedding.weight"].shape[0]
    transformer_width = state_dict["ln_final.weight"].shape[0]
    transformer_heads = transformer_width // 64
    transformer_layers = len(set(k.split(".")[2] for k in state_dict if k.startswith("transformer.resblocks")))

    model = LCLIP(
        embed_dim,
        image_resolution, vision_layers, vision_width, vision_patch_size,
        context_length, vocab_size, transformer_width, transformer_heads, transformer_layers, load_from_clip
    )

    for key in ["input_resolution", "context_length", "vocab_size"]:
        if key in state_dict:
            del state_dict[key]

    convert_weights(model)
    # model.load_state_dict(state_dict)
    model.load_state_dict(state_dict, strict=False)
    vision_heads = vision_width // 64
    # print(vision_heads)
    return model.eval(), image_resolution, vision_heads, embed_dim, vision_width, vision_patch_size

def build_modified_model(state_dict: dict, txt_length: int):
    vit = "visual.proj" in state_dict

    if vit:
        vision_width = state_dict["visual.conv1.weight"].shape[0]
        vision_layers = len([
            k for k in state_dict.keys()
            if k.startswith("visual.") and k.endswith(".attn.in_proj_weight")
        ])
        vision_patch_size = state_dict["visual.conv1.weight"].shape[-1]
        grid_size = round(
            (state_dict["visual.positional_embedding"].shape[0] - 1)**0.5)
        image_resolution = vision_patch_size * grid_size
    else:
        counts: list = [
            len(
                set(
                    k.split(".")[2] for k in state_dict
                    if k.startswith(f"visual.layer{b}")))
            for b in [1, 2, 3, 4]
        ]
        vision_layers = tuple(counts)
        vision_width = state_dict["visual.layer1.0.conv1.weight"].shape[0]
        
        output_width = round(
            (state_dict["visual.attnpool.positional_embedding"].shape[0] -
             1)**0.5)
        vision_patch_size = None
        assert output_width**2 + 1 == state_dict[
            "visual.attnpool.positional_embedding"].shape[0]
        image_resolution = output_width * 32
    embed_dim = state_dict["text_projection"].shape[1]

    model = zhCLIP(embed_dim, image_resolution, vision_layers, vision_width,
                 vision_patch_size)

    for key in ["input_resolution", "context_length", "vocab_size"]:
        if key in state_dict:
            del state_dict[key]

    convert_weights(model)
    model.load_state_dict(state_dict, False)
    return model.eval()