cover/models/xclip_backbone.py

import copy
import math
from collections import OrderedDict
from typing import Tuple, Union

import clip
import numpy as np
import torch
import torch.nn.functional as F
from einops import rearrange
from timm.models.layers import trunc_normal_
from torch import nn
from torch.utils.checkpoint import checkpoint_sequential


def drop_path(x, drop_prob: float = 0.0, training: bool = False):
    """Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).
    This is the same as the DropConnect impl I created for EfficientNet, etc networks, however,
    the original name is misleading as 'Drop Connect' is a different form of dropout in a separate paper...
    See discussion: https://github.com/tensorflow/tpu/issues/494#issuecomment-532968956 ... I've opted for
    changing the layer and argument names to 'drop path' rather than mix DropConnect as a layer name and use
    'survival rate' as the argument.
    """
    if drop_prob == 0.0 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 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)
        return super().forward(x)


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__()

        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
        )
        return self.attn(x, x, x, need_weights=False, attn_mask=self.attn_mask)[0]

    def forward(self, x: torch.Tensor):
        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 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 = 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):
        x = self.conv1(x)  # shape = [*, width, grid, grid]
        x = x.reshape(x.shape[0], x.shape[1], -1)  # shape = [*, width, grid ** 2]
        x = x.permute(0, 2, 1)  # shape = [*, grid ** 2, 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,
        )  # shape = [*, grid ** 2 + 1, width]
        x = x + self.positional_embedding.to(x.dtype)
        x = self.ln_pre(x)

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

        x = self.ln_post(x[:, 0, :])

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


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,
        vocab_size: int,
        transformer_width: int,
        transformer_heads: int,
        transformer_layers: int,
    ):
        super().__init__()

        self.context_length = context_length

        # 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(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.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)

        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.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]
        # 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


class CrossFramelAttentionBlock(nn.Module):
    def __init__(
        self,
        d_model: int,
        n_head: int,
        attn_mask: torch.Tensor = None,
        droppath=0.0,
        T=0,
    ):
        super().__init__()
        self.T = T

        self.message_fc = nn.Linear(d_model, d_model)
        self.message_ln = LayerNorm(d_model)
        self.message_attn = nn.MultiheadAttention(d_model, n_head,)

        self.attn = nn.MultiheadAttention(d_model, n_head,)
        self.ln_1 = LayerNorm(d_model)

        self.drop_path = DropPath(droppath) if droppath > 0.0 else nn.Identity()
        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
        )
        return self.attn(x, x, x, need_weights=False, attn_mask=self.attn_mask)[0]

    def forward(self, x):
        l, bt, d = x.size()
        b = bt // self.T
        x = x.view(l, b, self.T, d)

        msg_token = self.message_fc(x[0, :, :, :])
        msg_token = msg_token.view(b, self.T, 1, d)

        msg_token = msg_token.permute(1, 2, 0, 3).view(self.T, b, d)
        msg_token = msg_token + self.drop_path(
            self.message_attn(
                self.message_ln(msg_token),
                self.message_ln(msg_token),
                self.message_ln(msg_token),
                need_weights=False,
            )[0]
        )
        msg_token = msg_token.view(self.T, 1, b, d).permute(1, 2, 0, 3)

        x = torch.cat([x, msg_token], dim=0)

        x = x.view(l + 1, -1, d)
        x = x + self.drop_path(self.attention(self.ln_1(x)))
        x = x[:l, :, :]
        x = x + self.drop_path(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,
        droppath=None,
        use_checkpoint=False,
        T=8,
    ):
        super().__init__()
        self.use_checkpoint = use_checkpoint
        if droppath is None:
            droppath = [0.0 for i in range(layers)]
        self.width = width
        self.layers = layers

        self.resblocks = nn.Sequential(
            *[
                CrossFramelAttentionBlock(width, heads, attn_mask, droppath[i], T)
                for i in range(layers)
            ]
        )

    def forward(self, x: torch.Tensor):
        if not self.use_checkpoint:
            return self.resblocks(x)
        else:
            return checkpoint_sequential(self.resblocks, 3, x)


class CrossFrameCommunicationTransformer(nn.Module):
    def __init__(
        self,
        input_resolution: int,
        patch_size: int,
        width: int,
        layers: int,
        heads: int,
        output_dim: int,
        droppath=None,
        T=8,
        use_checkpoint=False,
    ):
        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)

        ## Attention Blocks
        self.transformer = Transformer(
            width, layers, heads, droppath=droppath, use_checkpoint=use_checkpoint, T=T,
        )
        self.ln_post = LayerNorm(width)
        self.proj = nn.Parameter(scale * torch.randn(width, output_dim))

    def init_weights(self):
        self.apply(self._init_weights)

    def _init_weights(self, m):
        if isinstance(m, nn.Linear):
            trunc_normal_(m.weight, std=0.02)
            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 forward(self, x: torch.Tensor):
        x = self.conv1(x)  # shape = [*, width, grid, grid]
        x = x.reshape(x.shape[0], x.shape[1], -1)  # shape = [*, width, grid ** 2]
        x = x.permute(0, 2, 1)  # shape = [*, grid ** 2, 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,
        )  # shape = [*, grid ** 2 + 1, width]
        x = x + self.positional_embedding.to(x.dtype)

        x = self.ln_pre(x)

        x = x.permute(1, 0, 2)
        x = self.transformer(x)
        x = x.permute(1, 0, 2)

        cls_x = self.ln_post(x[:, 0, :])

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

        return cls_x, x[:, 1:, :]


class MulitHeadAttention(nn.Module):
    def __init__(
        self,
        dim,
        num_heads=8,
        qkv_bias=False,
        qk_scale=None,
        attn_drop=0.0,
        proj_drop=0.0,
    ):
        super().__init__()
        self.num_heads = num_heads
        head_dim = dim // num_heads

        self.scale = qk_scale or head_dim ** -0.5

        self.q_proj = nn.Linear(dim, dim, bias=qkv_bias)
        self.k_proj = nn.Linear(dim, dim, bias=qkv_bias)
        self.v_proj = nn.Linear(dim, dim, 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, q, k, v):
        B, N, C = q.shape
        B, M, C = k.shape
        q = (
            self.q_proj(q)
            .reshape(B, N, self.num_heads, C // self.num_heads)
            .permute(0, 2, 1, 3)
        )
        k = (
            self.k_proj(k)
            .reshape(B, M, self.num_heads, C // self.num_heads)
            .permute(0, 2, 1, 3)
        )
        v = (
            self.v_proj(v)
            .reshape(B, M, self.num_heads, C // self.num_heads)
            .permute(0, 2, 1, 3)
        )

        attn = (q @ k.transpose(-2, -1)) * self.scale
        attn = attn.softmax(dim=-1)
        attn = self.attn_drop(attn)

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


class PromptGeneratorLayer(nn.Module):
    def __init__(
        self, d_model, nhead, dropout=0.0,
    ):
        super().__init__()
        self.cross_attn = MulitHeadAttention(d_model, nhead, proj_drop=dropout)

        self.norm1 = nn.LayerNorm(d_model)
        self.norm3 = nn.LayerNorm(d_model)

        self.dropout = nn.Dropout(dropout)

        self.mlp = nn.Sequential(
            nn.Linear(d_model, d_model * 4),
            QuickGELU(),
            nn.Dropout(dropout),
            nn.Linear(d_model * 4, d_model),
        )

    def forward(self, x, visual):
        q = k = v = self.norm1(x)
        x = x + self.cross_attn(q, visual, visual)
        x = x + self.dropout(self.mlp(self.norm3(x)))
        return x


class VideoSpecificPrompt(nn.Module):
    def __init__(
        self, layers=2, embed_dim=512, alpha=0.1,
    ):
        super().__init__()
        self.norm = nn.LayerNorm(embed_dim)
        self.decoder = nn.ModuleList(
            [PromptGeneratorLayer(embed_dim, embed_dim // 64) for _ in range(layers)]
        )
        self.alpha = nn.Parameter(torch.ones(embed_dim) * alpha)
        self.apply(self._init_weights)

    def _init_weights(self, m):
        if isinstance(m, nn.Linear):
            trunc_normal_(m.weight, std=0.02)
            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 forward(self, text, visual):
        B, N, C = visual.shape
        visual = self.norm(visual)
        for layer in self.decoder:
            text = layer(text, visual)


from collections import OrderedDict

from timm.models.layers import trunc_normal_


class ResidualAttentionBlock(nn.Module):
    def __init__(self, d_model: int, n_head: int, attn_mask: torch.Tensor = None):
        super().__init__()

        self.attn = nn.MultiheadAttention(d_model, n_head)
        self.ln_1 = nn.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 = nn.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
        )
        return self.attn(x, x, x, need_weights=False, attn_mask=self.attn_mask)[0]

    def forward(self, x: torch.Tensor):
        x = x + self.attention(self.ln_1(x))
        x = x + self.mlp(self.ln_2(x))
        return x


class MultiframeIntegrationTransformer(nn.Module):
    def __init__(
        self, T, embed_dim=512, layers=1,
    ):
        super().__init__()
        self.T = T
        transformer_heads = embed_dim // 64
        self.positional_embedding = nn.Parameter(torch.empty(1, T, embed_dim))
        trunc_normal_(self.positional_embedding, std=0.02)
        self.resblocks = nn.Sequential(
            *[
                ResidualAttentionBlock(d_model=embed_dim, n_head=transformer_heads)
                for _ in range(layers)
            ]
        )

        self.apply(self._init_weights)

    def _init_weights(self, m):
        if isinstance(m, (nn.Linear,)):
            trunc_normal_(m.weight, std=0.02)
            if m.bias is not None:
                nn.init.zeros_(m.bias)
        elif isinstance(m, nn.LayerNorm):
            nn.init.zeros_(m.bias)
            nn.init.ones_(m.weight)

    def forward(self, x):
        ori_x = x
        x = x + self.positional_embedding
        x = x.permute(1, 0, 2)
        x = self.resblocks(x)
        x = x.permute(1, 0, 2)
        x = x.type(ori_x.dtype) + ori_x

        return x.mean(dim=1, keepdim=False)


class XCLIP(CLIP):
    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,
        # video
        T=8,
        droppath=0.0,
        mit_layers=1,
        # prompt
        prompts_alpha=1e-4,
        prompts_layers=1,
        # other
        use_cache=True,
        use_checkpoint=False,
    ):
        super().__init__(
            embed_dim,
            image_resolution,
            vision_layers,
            vision_width,
            vision_patch_size,
            context_length,
            vocab_size,
            transformer_width,
            transformer_heads,
            transformer_layers,
        )

        self.prompts_generator = VideoSpecificPrompt(
            layers=prompts_layers, embed_dim=embed_dim, alpha=prompts_alpha,
        )
        self.use_cache = use_cache
        self.mit = MultiframeIntegrationTransformer(
            T=T, embed_dim=embed_dim, layers=mit_layers,
        )

        dpr = (
            [x.item() for x in torch.linspace(0, droppath, vision_layers)]
            if droppath > 0.0
            else None
        )

        vision_heads = vision_width // 64
        self.visual = CrossFrameCommunicationTransformer(
            input_resolution=image_resolution,
            patch_size=vision_patch_size,
            width=vision_width,
            layers=vision_layers,
            heads=vision_heads,
            output_dim=embed_dim,
            droppath=dpr,
            T=T,
            use_checkpoint=use_checkpoint,
        )

        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(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.cache_text_features = None
        self.prompts_visual_ln = LayerNorm(vision_width)
        self.prompts_visual_proj = nn.Parameter(torch.randn(vision_width, embed_dim))

        self.initialize_parameters()

    @torch.jit.ignore
    def no_weight_decay_keywords(self):
        return {"positional_embedding"}

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

    def encode_text(self, text):
        x = self.token_embedding(text)
        eos_indx = text.argmax(dim=-1)
        K, N1, C = x.shape

        x = x + 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)
        # 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]), eos_indx] @ self.text_projection
        x = x.reshape(K, -1)
        return x

    def encode_video(self, image):
        b, t, c, h, w = image.size()
        image = image.reshape(-1, c, h, w)

        cls_features, img_features = self.encode_image(image)
        img_features = self.prompts_visual_ln(img_features)
        img_features = img_features @ self.prompts_visual_proj

        cls_features = cls_features.view(b, t, -1)
        img_features = img_features.view(b, t, -1, cls_features.shape[-1])

        video_features = self.mit(cls_features)

        return video_features, img_features

    def forward(self, image, **kwargs):
        image = rearrange(image, "b c t h w -> b t c h w")
        video_features, _ = self.encode_video(image)
        return video_features.reshape(*video_features.shape, 1, 1, 1)

    def cache_text(self, text):
        self.eval()
        with torch.no_grad():
            if self.cache_text_features is None:
                self.cache_text_features = self.encode_text(text)
        self.train()
        return self.cache_text_features

    def forward_original(self, image, text):
        b = image.shape[0]
        video_features, img_features = self.encode_video(image)
        img_features = img_features.mean(dim=1, keepdim=False)

        if self.use_cache:
            text_features = self.cache_text(text)
        else:
            text_features = self.encode_text(text)

        text_features = text_features.unsqueeze(0).expand(b, -1, -1)
        text_features = text_features + self.prompts_generator(
            text_features, img_features
        )

        video_features = video_features / video_features.norm(dim=-1, keepdim=True)
        text_features = text_features / text_features.norm(dim=-1, keepdim=True)
        logit_scale = self.logit_scale.exp()
        logits = torch.einsum("bd,bkd->bk", video_features, logit_scale * text_features)

        return logits


def build_x_clip_model(
    pretrained_path="./pretrained_weights/k400_32_8.pth",
    droppath=0.0,
    use_checkpoint=False,
    logger=None,
    prompts_alpha=1e-1,
    prompts_layers=2,
    use_cache=True,
    mit_layers=4,
    **kwargs,
):
    state_dict = torch.load(pretrained_path, map_location="cpu")["model"]
    T = int(pretrained_path.split("_")[-1].split(".")[0])
    print(T)
    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 = XCLIP(
        embed_dim,
        image_resolution,
        vision_layers,
        vision_width,
        vision_patch_size,
        context_length,
        vocab_size,
        transformer_width,
        transformer_heads,
        transformer_layers,
        T=T,
        droppath=droppath,
        mit_layers=mit_layers,
        prompts_alpha=prompts_alpha,
        prompts_layers=prompts_layers,
        use_checkpoint=use_checkpoint,
        use_cache=use_cache,
    )

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

    msg = model.load_state_dict(state_dict, strict=False)

    return model.eval()