builder.py

import os
from siglip_encoder import SigLipVisionTower
import torch
import torch.nn as nn
import re


def build_vision_tower(vision_tower_cfg, **kwargs):
    vision_tower = getattr(vision_tower_cfg, "mm_vision_tower", getattr(vision_tower_cfg, "vision_tower", None))
    is_absolute_path_exists = os.path.exists(vision_tower)
    use_s2 = getattr(vision_tower_cfg, "s2", False)
    if "siglip" in vision_tower:
        return SigLipVisionTower(vision_tower, vision_tower_cfg=vision_tower_cfg, **kwargs)
    raise ValueError(f"Unknown vision tower: {vision_tower}")

def build_vision_resampler(confg, **kwargs):
    '''
        act as place holder, useless in our model
    '''
    print(f'No use of vision_resampler')

class IdentityMap(nn.Module):
    def __init__(self):
        super().__init__()

    def forward(self, x, *args, **kwargs):
        return x

    @property
    def config(self):
        return {"mm_projector_type": "identity"}


class SimpleResBlock(nn.Module):
    def __init__(self, channels):
        super().__init__()
        self.pre_norm = nn.LayerNorm(channels)

        self.proj = nn.Sequential(nn.Linear(channels, channels), nn.GELU(), nn.Linear(channels, channels))

    def forward(self, x):
        x = self.pre_norm(x)
        return x + self.proj(x)


def build_vision_projector(config, delay_load=False, **kwargs):
    projector_type = getattr(config, "mm_projector_type", "linear")

    if projector_type == "linear": ### this is default 
        return nn.Linear(config.mm_hidden_size, config.hidden_size)

    if projector_type == "pooler":
        return PoolerProjector(config, kwargs["vision_cfg"])

    mlp_gelu_match = re.match(r"^mlp(\d+)x_gelu$", projector_type)
    if mlp_gelu_match: ###this is default 
        mlp_depth = int(mlp_gelu_match.group(1)) ###mlx2x_gelu ----> 2 projector
        modules = [nn.Linear(config.mm_hidden_size, config.hidden_size)] ### 4096  - 4096 - gelu - 4096
        for _ in range(1, mlp_depth):
            modules.append(nn.GELU())
            modules.append(nn.Linear(config.hidden_size, config.hidden_size))
        return nn.Sequential(*modules)

    mlp_gelu_resnet_match = re.match(r"^mlp(\d+)x_res(\d+)x_gelu$", projector_type)
    if mlp_gelu_resnet_match:
        mlp_depth = int(mlp_gelu_resnet_match.group(1))
        res_depth = int(mlp_gelu_resnet_match.group(2))
        modules = [nn.Linear(config.mm_hidden_size, config.hidden_size)]
        for _ in range(1, mlp_depth):
            modules.append(nn.GELU())
            modules.append(nn.Linear(config.hidden_size, config.hidden_size))
        for _ in range(res_depth):
            modules.append(SimpleResBlock(config.hidden_size))
        return nn.Sequential(*modules)

    if projector_type == "identity":
        return IdentityMap()

    raise ValueError(f"Unknown projector type: {projector_type}")

class PoolerProjector(nn.Module):
    def __init__(self, config, vision_cfg):
        super().__init__()
        self._config = config
        self.hw = vision_cfg.image_size // vision_cfg.patch_size

        self.conv_pool = nn.Conv2d(config.mm_hidden_size, config.hidden_size, kernel_size=2, stride=2)

        self.proj = nn.Sequential(
            nn.GELU(),
            nn.Linear(config.hidden_size, config.hidden_size),
        )

    def forward(self, x, *args, **kwargs):
        height = width = self.hw
        assert height * width == x.shape[1]
        x = x.view(x.shape[0], height, width, -1).permute(0, 3, 1, 2)
        x = self.conv_pool(x)
        x = x.flatten(2).transpose(1, 2)
        x = self.proj(x)
        return x

    @property
    def config(self):
        return {"mm_projector_type": "pooler"}