CLIP(Contrastive Language-Image Pretraining)主体网络代码详解

CLIP是OpenAI于2021年发表的工作，其采用无监督学习中的对比学习的训练方法，使用了规模巨大的数据集（4亿个图片文本对）来进行训练，其在多个数据集上均得到了让人欣喜的结果，有效地证实了NLP与CV结合所具备的巨大的潜力，并基于此产生了许多有趣的工作。在这里分享一下我对于CLIP主体网络代码的理解，可能会存在诸多纰漏，请大家多多指教。

paper：http://proceedings.mlr.press/v139/radford21a/radford21a.pdf

code： https://github.com/openai/CLIP

from collections import OrderedDict
from typing import Tuple, Union
from torch import nn
import numpy as np
import torch
import torch.nn.functional

#----------------------------------------------#
# ModifiedResNet50中标准残差结构--Bottleneck的定义
#----------------------------------------------#
class Bottleneck(nn.Module):
    expansion = 4

    def __init__(self, inplanes, planes, stride=1):
        super(Bottleneck,self).__init__()
        #---------------------------------------------------#
        # 所有的卷积层的步长均为1,但是当步长大于1时,在第二次卷积之后
        # 将会有一个平均池化层
        #---------------------------------------------------#
        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)
        #---------------------------------#
        # 当步长大于1时,将会通过一个平均池化层,
        #  否则将会直接对其跳过
        #---------------------------------#
        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
        #--------------------------------------#
        # 执行该if语句时,"downsample layer"
        # 将会由二维平均池化,卷积以及BatchNorm2d组成
        #--------------------------------------#
        if stride > 1 or inplanes != planes * Bottleneck.expansion:
            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))
        #------------------------------------------#
        # 当downsample层不为空时,其将会对原始的输入张量
        # 执行三个序列操作
        #------------------------------------------#
        if self.downsample is not None:
            identity = self.downsample(x)

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

#-----------------------------------#
# 对于AttentionPool2d这个类的定义
# 在ModifiedResNet50中被使用
#-----------------------------------#
class AttentionPool2d(nn.Module):
    def __init__(self, spacial_dim: int, embed_dim: int, num_heads: int, output_dim: int = None):
        super(AttentionPool2d,self).__init__()
        #--------------------------------------------------#
        #  nn.Parameter()的作用为作为nn.Module中的可训练参数使用
        #--------------------------------------------------#
        self.positional_embedding = nn.Parameter(torch.randn(spacial_dim ** 2 + 1, embed_dim) / embed_dim ** 0.5)
        #-------------------------------#
        # 通过全连接层来获取以下四个映射量
        #-------------------------------#
        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):
        #---------------------------------------------------------------#
        # 首先进行张量shape的转变,由 batch_size,c,h,w -> (h*w),batch_size,c
        #---------------------------------------------------------------#
        x = x.reshape(x.shape[0], x.shape[1], x.shape[2] * x.shape[3]).permute(2, 0, 1)
        #-------------------------------------------#
        # (h*w),batch_size,c -> (h*w+1),batch_size,c
        #-------------------------------------------#
        x = torch.cat([x.mean(dim=0, keepdim=True), x], dim=0)
        #--------------------------------------------------------------------#
        # tensor的shape以及type均不发生改变,所做的只是将位置信息嵌入至原先的tensor中
        # shape:(h*w+1),batch_size,c
        #--------------------------------------------------------------------#
        x = x + self.positional_embedding[:, None, :].to(x.dtype)
        #---------------------------------------#
        # 将输入的张量pass through 多头注意力机制模块
        #---------------------------------------#
        x, _ = torch.nn.functional.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
        )

        return x[0]

#-------------------------------------#
# CLIP中所使用到的ModifiedResNet50的定义
#-------------------------------------#
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
    - 最后的平均池化层我们使用一个 QKV注意力池化层来进行替代
    """

    def __init__(self, layers, output_dim, heads, input_resolution=224, width=64):
        super(ModifiedResNet,self).__init__()
        self.output_dim       = output_dim
        self.input_resolution = input_resolution
        #----------------------------------#
        # the 3- "stem" convolution layers
        #----------------------------------#
        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.relu    = nn.ReLU(inplace=True)
        self.avgpool = nn.AvgPool2d(2)
        #------------------------------------#
        # residual layers in ModifiedResNet50
        # 共计四层
        #------------------------------------#
        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
        #--------------------------------------------------#
        # 对于最后的平均池化层我们使用一个QKV注意力池化层来进行替代
        #--------------------------------------------------#
        self.attnpool = AttentionPool2d(input_resolution // 32, embed_dim, heads, output_dim)
    #----------------------------------------#
    # ModifiedResNet50中的残差层的定义
    # 其中的blocks即为标准的残差结构--Bottleneck
    #----------------------------------------#
    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)
    #-------------------------------#
    # ModifiedResNet50的前向传播函数
    #-------------------------------#
    def forward(self, x):
        #-----------------------------------------#
        # As to the 3-"stem" convolution layers
        # 在这里我们将三个卷积层集成到一个函数中使用
        # 每一层均为 conv->bn->relu
        #-----------------------------------------#
        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
        #------------------------------------------#
        # 这行code的作用在于对输入的张量进行一个type的转换
        #------------------------------------------#
        x = x.type(self.conv1.weight.dtype)
        #---------------------#
        # 过三个卷积层
        #---------------------#
        x = stem(x)
        #----------------------------------#
        # 过ModifiedResNet50中的残差结构,共4层
        #----------------------------------#
        x = self.layer1(x)
        x = self.layer2(x)
        x = self.layer3(x)
        x = self.layer4(x)
        #--------------------------------------------------#
        # 对于最后的平均池化层我们使用一个QKV注意力池化层来进行替代
        #--------------------------------------------------#
        x = self.attnpool(x)

        return x

#-----------------------------------#
# transformer模块中所使用到的LayerNorm层
#-----------------------------------#
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)

#--------------------------------#
# QuickGELU激活函数的定义
# 在transformer结构中的MLP层中被使用
#--------------------------------#
class QuickGELU(nn.Module):
    def forward(self, x: torch.Tensor):
        return x * torch.sigmoid(1.702 * x)

#-------------------------------------------------#
# transformer模块的定义,将会在transformer结构中被使用
# 1.多头注意力层
# 2.LayerNorm层
# 3.MLP层
#-------------------------------------------------#
class ResidualAttentionBlock(nn.Module):
    def __init__(self, d_model: int, n_head: int, attn_mask: torch.Tensor = None):
        super(ResidualAttentionBlock,self).__init__()
        #----------------------#
        # 多头注意力机制
        #----------------------#
        self.attn = nn.MultiheadAttention(d_model, n_head)
        self.ln_1 = LayerNorm(d_model)
        #-------------------------------------------------------------------#
        # 在MLP层中首先是进行一次全连接,之后是过QuickGELU激活函数,最后是通过投影进行映射
        #-------------------------------------------------------------------#
        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
    #-------------------------------------#
    # 该函数的作用是对输入的张量使用多头注意力机制
    #-------------------------------------#
    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
        return self.attn(x, x, x, need_weights=False, attn_mask=self.attn_mask)[0]
    #---------------------------------------------------------------#
    # 在这个前向传播函数中,对于transformer模块进行了定义以及说明
    #---------------------------------------------------------------#
    def forward(self, x: torch.Tensor):
        x = x + self.attention(self.ln_1(x))
        x = x + self.mlp(self.ln_2(x))

        return x

#-------------------------------------#
# transformer结构的定义
# 即为多个transformer模块按照顺序进行堆叠
#-------------------------------------#
class Transformer(nn.Module):
    def __init__(self, width: int, layers: int, heads: int, attn_mask: torch.Tensor = None):
        super(Transformer,self).__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)

#---------------------------------#
# VisionTransformer结构的定义
# 输入图片的通道数为3
#---------------------------------#
class VisionTransformer(nn.Module):
    def __init__(self, input_resolution: int, patch_size: int, width: int, layers: int, heads: int, output_dim: int):
        super(VisionTransformer,self).__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
        #------------------------------------------#
        # 在这里我们可以用nn.Parameter()来将这
        # 个随机初始化的Tensor注册为可学习的参数Parameter
        #------------------------------------------#
        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          = Transformer(width, layers, heads)
        self.ln_post              = LayerNorm(width)
        self.proj                 = nn.Parameter(scale * torch.randn(width, output_dim))

    def forward(self, x: torch.Tensor):
        #-----------------------------------------------------------------------------------------#
        # 此处的卷积可以将张量的shape转变为batch_size,width,grid,grid(grid=input_resolution/patch_size)
        #-----------------------------------------------------------------------------------------#
        x = self.conv1(x)
        #---------------------------------------------#
        # reshape之后,shape=batch_size,width,grid ** 2
        #---------------------------------------------#
        x = x.reshape(x.shape[0], x.shape[1], -1)
        #----------------------------#
        # 转置之后,shape为
        # batch_size,grid ** 2,width
        #----------------------------#
        x = x.permute(0, 2, 1)
        #------------------------------------------------------#
        # pass这条语句之后,shape=batch_size,grid ** 2 + 1,width
        #------------------------------------------------------#
        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)
        #--------------------------------------------#
        # 加上其位置编码信息,并且pass through LayerNorm层
        #--------------------------------------------#
        x = x + self.positional_embedding.to(x.dtype)
        x = self.ln_pre(x)
        #-----------------------------------------------#
        # shape先转变为grid ** 2 + 1,batch_size,width
        # 之后经由transformer结构编码
        # 最后再进行转置,恢复为batch_size,grid ** 2 + 1,width
        # 再pass through LayerNorm层
        #-----------------------------------------------#
        x = x.permute(1, 0, 2)
        x = self.transformer(x)
        x = x.permute(1, 0, 2)
        x = self.ln_post(x[:, 0, :])
        #-------------------------#
        # 若成立则将会进行矩阵乘法运算
        #-------------------------#
        if self.proj is not None:
            x = x @ self.proj

        return x

#------------------------#
# CLIP模型的定义
#------------------------#
class CLIP(nn.Module):
    def __init__(self,embed_dim: int,
                 #-------------------#
                 # vision部分的函数定义
                 #-------------------#
                 image_resolution: int, vision_layers: Union[Tuple[int, int, int, int], int], vision_width: int, vision_patch_size: int,
                 #-------------------#
                 # text部分的函数定义
                 #-------------------#
                 context_length: int, vocab_size: int, transformer_width: int, transformer_heads: int, transformer_layers: int):
        super(CLIP,self).__init__()
        #------------------------#
        # 定义文本的长度
        #------------------------#
        self.context_length = context_length
        #--------------------------------------------#
        # image encoder
        # 对于image部分,可以使用ModifiedResNet50或者ViT
        #--------------------------------------------#
        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
            )
        #---------------------------------------#
        # text encoder
        # 对于文字部分则直接使用Text Transformer即可
        #---------------------------------------#
        self.transformer = Transformer(
            width        = transformer_width,
            layers       = transformer_layers,
            heads        = transformer_heads,
            attn_mask    = self.build_attention_mask()
        )

        self.vocab_size           = vocab_size
        #----------------------------------------------#
        # token嵌入以及位置嵌入还有对于LayerNorm的进一步的定义
        #----------------------------------------------#
        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)
        #--------------------------------------------------------------------------#
        # 在这里我们可以用nn.Parameter()来将这个随机初始化的Tensor注册为可学习的参数Parameter
        # torch.empty用于返回一个未初始化的tensor
        # torch.zeros用于将tensor中元素值全置为0
        # torch.ones用于将tensor中元素值全置为1
        # logit_scale与 cosine similarities有关
        #--------------------------------------------------------------------------#
        self.text_projection      = nn.Parameter(torch.empty(transformer_width, embed_dim))
        self.logit_scale          = nn.Parameter(torch.ones([]) * np.log(1 / 0.07))
        #-------------------#
        # 定义了此类中的一个函数
        #-------------------#
        self.initialize_parameters()
    #---------------------------#
    # 部分权值的初始化操作
    #---------------------------#
    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
    #-------------------#
    # image encoder函数
    #-------------------#
    def encode_image(self, image):
        return self.visual(image.type(self.dtype))
    #------------------------------------#
    # text encoder函数
    # 这里为一个单纯的 Text transformer结构
    #------------------------------------#
    def encode_text(self, text):
        x = self.token_embedding(text).type(self.dtype)  #shape=[batch_size, n_ctx, d_model]
        #----------------#
        # 嵌入位置信息
        #----------------#
        x = x + self.positional_embedding.type(self.dtype)
        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]
        # transformer.width为上面定义的d_model
        # 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
    #--------------------------------------------------------------#
    # CLIP这个类的前向传播函数的定义
    # 即为CLIP整体模型的定义
    #--------------------------------------------------------------#
    def forward(self, image, text):
        #---------------------------------------------#
        # 使用ModifiedResNet50或者ViT来完成图像信息的编码
        # 使用Text transformer来完成文本信息的编码
        #---------------------------------------------#
        image_features   = self.encode_image(image)
        text_features    = self.encode_text(text)
        #---------------------------------#
        # joint multimodal embedding
        # normalized features
        #---------------------------------#
        image_features   = image_features / image_features.norm(dim=-1, keepdim=True)
        text_features    = text_features / text_features.norm(dim=-1, keepdim=True)
        #--------------------------------#
        # 计算图像以及文本的相似度
        # 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()
        #-------------------------------------------------#
        # 所返回的张量的shape为
        # shape = [global_batch_size, global_batch_size]
        #-------------------------------------------------#
        return logits_per_image, logits_per_text

#-----------------------------#
# 为了训练加速使用到了混合精度运算
#-----------------------------#
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)

#-----------------------------------------------#
# CLIP模型的创建
#-----------------------------------------------#
def build_model(state_dict: dict):
    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]
    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 = CLIP(
        embed_dim,
        image_resolution, vision_layers, vision_width, vision_patch_size,
        context_length, vocab_size, transformer_width, transformer_heads, transformer_layers
    )

    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)
    return model.eval()