nvidia/C-RADIOv4-H

1# Copyright (c) 2023-2024, NVIDIA CORPORATION.  All rights reserved.2#3# NVIDIA CORPORATION and its licensors retain all intellectual property4# and proprietary rights in and to this software, related documentation5# and any modifications thereto.  Any use, reproduction, disclosure or6# distribution of this software and related documentation without an express7# license agreement from NVIDIA CORPORATION is strictly prohibited.8 9import math10from typing import Union, Tuple, Optional11 12import torch13import torch.nn.functional as F14from torch import nn15from einops import rearrange16 17from .cls_token import ClsToken18 19input_dim_t = Union[int, Tuple[int, int]]20 21try:22    # raise ImportError()23    from indirect_grid_sample import indirect_grid_sample24except ImportError:25    indirect_grid_sample = None26 27class ViTPatchGenerator(nn.Module):28    def __init__(self,29                 patch_size: int,30                 embed_dim: int,31                 input_dims: input_dim_t,32                 abs_pos: bool = True,33                 normalize_patches: bool = False,34                 cls_token: bool = False,35                 max_input_dims: Optional[input_dim_t] = None,36                 pos_dropout: float = 0.0,37                 return_pos_enc: bool = False,38                 num_cls_tokens: int = 1,39                 register_multiple: Optional[int] = None,40                 num_registers: Optional[int] = None,41                 patch_bias: bool = False,42                 device=None, dtype=None,43    ):44        super().__init__()45 46        if isinstance(input_dims, int):47            input_dims = (input_dims, input_dims)48 49        if max_input_dims is None:50            max_input_dims = input_dims51        if isinstance(max_input_dims, int):52            max_input_dims = (max_input_dims, max_input_dims)53 54        max_input_dims = tuple(55            int(math.ceil(d / patch_size) * patch_size)56            for d in max_input_dims57        )58 59        self.cpe_mode = max_input_dims != input_dims60        self.pos_dropout = pos_dropout61        self.return_pos_enc = return_pos_enc62 63        factory = dict(device=device, dtype=dtype)64 65        self.patch_size = patch_size66        self.abs_pos = abs_pos67        self.embed_dim = embed_dim68 69        self.num_rows = max_input_dims[0] // patch_size70        self.num_cols = max_input_dims[1] // patch_size71        self.input_dims = tuple(d // patch_size for d in input_dims)72        self.num_patches = self.num_rows * self.num_cols73        self.max_input_dims = max_input_dims74 75        self.im_to_patches = Im2Patches(patch_size)76        self.embedder = ViTPatchLinear(patch_size, embed_dim, bias=patch_bias, **factory)77 78        if abs_pos:79            scale = embed_dim ** -0.580            self.pos_embed = nn.Parameter(torch.randn(1, self.num_patches, embed_dim, **factory) * scale)81 82        self.cls_token = ClsToken(83            embed_dim,84            num_tokens=num_cls_tokens,85            enabled=cls_token,86            register_multiple=register_multiple,87            num_registers=num_registers,88        )89 90        self.patch_normalizer = nn.LayerNorm(embed_dim) if normalize_patches else nn.Identity()91 92        self.num_video_frames = None93 94    def forward(self, x: torch.Tensor) -> torch.Tensor:95        patches = self.embed_patches(x)96        patches, pos_enc = self.apply_pos_enc(patches, input_size=x.shape[2:])97        patches = self.cls_token(patches)98        patches = self.patch_normalizer(patches)99        if self.return_pos_enc:100            return patches, pos_enc101        return patches102 103    @property104    def apply_cls_token(self):105        return self.cls_token.enabled106 107    @property108    def num_cls_tokens(self):109        return self.cls_token.num_tokens110 111    @property112    def num_cls_patches(self):113        return self.cls_token.num_patches114 115    @property116    def num_registers(self):117        return self.cls_token.num_registers118 119    @property120    def num_skip(self):121        return self.num_cls_tokens + self.num_registers122 123    def no_weight_decay(self):124        return [125            'pos_embed',126        ]127 128    def _load_embed(self, src_embed: torch.Tensor, targ_embed: nn.Parameter):129        if src_embed.shape != targ_embed.shape:130            src_size = int(math.sqrt(src_embed.shape[1]))131 132            assert src_size ** 2 == src_embed.shape[1], 'Unable to interpolate non-square embedding'133 134            src_embed = rearrange(src_embed, 'b (h w) c -> b c h w', h=src_size, w=src_size)135            src_embed = F.interpolate(src_embed, size=(self.num_rows, self.num_cols), mode='bicubic', align_corners=True, antialias=False)136            src_embed = rearrange(src_embed, 'b c h w -> b (h w) c')137        targ_embed.data.copy_(src_embed)138 139    def _load_projection(self, src_proj_weight: torch.Tensor, targ_proj_weight: torch.Tensor):140        if src_proj_weight.shape != targ_proj_weight.shape:141            src_patch_size = int(math.sqrt(src_proj_weight.shape[1] // 3))142 143            assert (src_patch_size ** 2) * 3 == src_proj_weight.shape[1], 'Unable to interpolate non-square patch size'144 145            src_proj_weight = rearrange(src_proj_weight, 'b (c h w) -> b c h w', c=3, h=src_patch_size, w=src_patch_size)146            src_proj_weight = F.interpolate(src_proj_weight, size=(self.patch_size, self.patch_size), mode='bicubic', align_corners=True, antialias=False)147            src_proj_weight = rearrange(src_proj_weight, 'b c h w -> b (c h w)')148        targ_proj_weight.data.copy_(src_proj_weight)149 150    def embed_patches(self, x: torch.Tensor) -> torch.Tensor:151        patches = self.im_to_patches(x)152        patches = self.embedder(patches)153        return patches154 155    def apply_pos_enc(self,156                      patches: torch.Tensor,157                      patch_idxs: Optional[torch.Tensor] = None,158                      input_size: Optional[Tuple[int, int]] = None,159    ) -> torch.Tensor:160        if not self.abs_pos:161            return patches162 163        pos_enc = self.get_pos_enc(patches.shape[0], patch_idxs, input_size)164 165        if self.training and self.pos_dropout > 0:166            keeps = torch.rand(patches.shape[0], 1, 1, dtype=pos_enc.dtype, device=pos_enc.device) > self.pos_dropout167            pos_enc_drop = torch.where(keeps, pos_enc, 0)168        else:169            pos_enc_drop = pos_enc170 171        return patches + pos_enc_drop, pos_enc172 173    def get_pos_enc(self,174                    batch_size: int,175                    patch_idxs: Optional[torch.Tensor] = None,176                    input_size: Optional[Tuple[int, int]] = None,177    ) -> torch.Tensor:178        if input_size is None:179            input_dims = self.input_dims180        else:181            input_dims = tuple(d // self.patch_size for d in input_size)182 183        pos_embed = self._get_pos_embeddings(batch_size, input_dims)184 185        if patch_idxs is None:186            return pos_embed187 188        exp_patch_idxs = patch_idxs.unsqueeze(-1).expand(-1, -1, pos_embed.shape[-1])189 190        pos_embed = torch.gather(pos_embed.expand(patch_idxs.shape[0], -1, -1), dim=1, index=exp_patch_idxs)191        return pos_embed192 193 194    def _get_pos_embeddings(self, batch_size: int, input_dims: Tuple[int, int]):195        if (self.num_rows, self.num_cols) == input_dims:196            return self.pos_embed197 198        pos_embed = self.pos_embed.reshape(1, self.num_rows, self.num_cols, -1).permute(0, 3, 1, 2)199 200        def window_select(pos_embed):201            if input_dims[0] < pos_embed.shape[-2]:202                pos_embed = pos_embed[..., :input_dims[0], :]203            if input_dims[1] < pos_embed.shape[-1]:204                pos_embed = pos_embed[..., :, :input_dims[1]]205            return pos_embed206 207        if self.cpe_mode:208            if self.training:209                if self.num_video_frames is not None:210                    if batch_size % self.num_video_frames != 0:211                        raise ValueError(f'Batch size {batch_size} must be divisible by num_video_frames {self.num_video_frames} for CPE mode.')212 213                    batch_size //= self.num_video_frames214 215                min_scale = math.sqrt(0.1)216                scale = torch.rand(batch_size, 1, 1, device=pos_embed.device) * (1 - min_scale) + min_scale217                aspect_min = math.log(3 / 4)218                aspect_max = -aspect_min219                aspect = torch.exp(torch.rand(batch_size, 1, 1, device=pos_embed.device) * (aspect_max - aspect_min) + aspect_min)220 221                scale_x = scale * aspect222                scale_y = scale * (1 / aspect)223                scale_xy = torch.stack([scale_x, scale_y], dim=-1).clamp_(0, 1)224 225                pos_xy = torch.rand(batch_size, 1, 1, 2, device=pos_embed.device) * (1 - scale_xy)226 227                lin_x = torch.linspace(0, 1, steps=input_dims[1], device=pos_embed.device)[None, None].expand(batch_size, input_dims[0], -1)228                lin_y = torch.linspace(0, 1, steps=input_dims[0], device=pos_embed.device)[None, :, None].expand(batch_size, -1, input_dims[1])229 230                lin_xy = torch.stack([lin_x, lin_y], dim=-1)231 232                grid_xy = lin_xy * scale_xy + pos_xy233 234                # Convert to [-1, 1] range235                grid_xy.mul_(2).sub_(1)236 237                pos_embed = F.grid_sample(238                    pos_embed.float().expand(batch_size, -1, -1, -1),239                    grid=grid_xy,240                    mode='bilinear',241                    padding_mode='zeros',242                    align_corners=True,243                ).to(pos_embed.dtype)244 245                if self.num_video_frames is not None:246                    pos_embed = torch.repeat_interleave(pos_embed, self.num_video_frames, dim=0)247            else:248                # i_rows, i_cols = input_dims249                # p_rows, p_cols = pos_embed.shape[2:]250                # if i_rows <= p_rows and i_cols <= p_cols:251                #     left = (p_cols - i_cols) // 2252                #     top = (p_rows - i_rows) // 2253                #     pos_embed = pos_embed[..., top:top+i_rows, left:left+i_cols]254                # else:255                max_dim = max(input_dims)256                pos_embed = F.interpolate(pos_embed.float(), size=(max_dim, max_dim), align_corners=False, mode='bilinear').to(pos_embed.dtype)257 258                pos_embed = window_select(pos_embed)259        else:260            pos_embed = window_select(pos_embed)261 262        if pos_embed.shape[-2:] != input_dims:263            pos_embed = F.interpolate(pos_embed.float(), size=input_dims, align_corners=False, mode='bilinear').to(pos_embed.dtype)264 265        pos_embed = pos_embed.flatten(2).permute(0, 2, 1)266 267        return pos_embed268 269 270class Im2Patches(nn.Module):271    def __init__(self, patch_size: int):272        super().__init__()273        self.patch_size = patch_size274 275    def forward(self, x: torch.Tensor) -> torch.Tensor:276        if self.patch_size == 1:277            patches = x.flatten(2)278            patches = patches.permute(0, 2, 1)279            return patches280 281        py = x.shape[-2] // self.patch_size282        px = x.shape[-1] // self.patch_size283        patches = rearrange(x, 'b c (py yy) (px xx) -> b (py px) (c yy xx)',284                            py=py, yy=self.patch_size,285                            px=px, xx=self.patch_size,286        )287        return patches288 289 290class ViTPatchLinear(nn.Linear):291    def __init__(self, patch_size: int, embed_dim: int, bias: bool = False, **factory):292        super().__init__(293            3 * (patch_size ** 2),294            embed_dim,295            bias=bias,296            **factory297        )298        self.patch_size = patch_size299