|
| 1 | +from typing import Any, Dict |
| 2 | + |
| 3 | +import torch |
| 4 | +import torch.nn as nn |
| 5 | +import torch.nn.functional as F |
| 6 | +from timm.models.layers import DropPath |
| 7 | + |
| 8 | +from .base_modules import Identity |
| 9 | +from .misc_modules import LayerScale |
| 10 | +from .mlp import MlpBlock |
| 11 | +from .patch_embeddings import ContiguousEmbed |
| 12 | +from .token_mixers import RESHAPE_LOOKUP, TokenMixerBlock |
| 13 | + |
| 14 | + |
| 15 | +class MetaFormer(nn.Module): |
| 16 | + def __init__( |
| 17 | + self, |
| 18 | + in_channels: int, |
| 19 | + embed_kwargs: Dict[str, Any], |
| 20 | + mixer_kwargs: Dict[str, Any], |
| 21 | + mlp_kwargs: Dict[str, Any], |
| 22 | + out_channels: int = None, |
| 23 | + layer_scale: bool = False, |
| 24 | + dropout: float = 0.0, |
| 25 | + **kwargs |
| 26 | + ) -> None: |
| 27 | + """Create a generic Metaformer block with any token-mixer available. |
| 28 | +
|
| 29 | + Input shape: (B, in_channels, H, W) |
| 30 | + Output shape: (B, out_channels, H, W) |
| 31 | +
|
| 32 | + Parameters |
| 33 | + ---------- |
| 34 | + in_channels : int |
| 35 | + Number of input channels. |
| 36 | + embed_kwargs : Dict[str, Any] |
| 37 | + Key-word arguments for the patch embedding block. |
| 38 | + mixer_kwargs : Dict[str, Any] |
| 39 | + Key-word arguments for the token-mixer block. |
| 40 | + mlp_kwargs : Dict[str, Any] |
| 41 | + Key-word arguments for the final Mlp-block. |
| 42 | + out_channels : int, optional |
| 43 | + Number of output channels. |
| 44 | + layer_scale : bool, default=False |
| 45 | + Flag, whether to use layer-scaling. |
| 46 | + dropout : float, default=0.0 |
| 47 | + Drop-path probaility. |
| 48 | +
|
| 49 | + Examples |
| 50 | + -------- |
| 51 | + MetaFormer with exact memory-efficient self-attention: |
| 52 | + >>> import torch |
| 53 | + >>> import torch.nn as nn |
| 54 | +
|
| 55 | + >>> in_channels = 3 |
| 56 | + >>> head_dim = 64 |
| 57 | + >>> num_heads = 8 |
| 58 | + >>> query_dim = head_dim*num_heads |
| 59 | +
|
| 60 | + >>> # patch embedding kwargs |
| 61 | + >>> embed_kwargs = { |
| 62 | + "in_channels": 3, |
| 63 | + "kernel_size": 7, |
| 64 | + "stride": 4, |
| 65 | + "pad": 2, |
| 66 | + "head_dim": head_dim, |
| 67 | + "num_heads": num_heads, |
| 68 | + } |
| 69 | +
|
| 70 | + >>> # token-mixer kwargs |
| 71 | + >>> mixer_kwargs = { |
| 72 | + "token_mixer": "self-attention", |
| 73 | + "normalization": "ln", |
| 74 | + "residual": True, |
| 75 | + "norm_kwargs": { |
| 76 | + "normalized_shape": query_dim |
| 77 | + }, |
| 78 | + "mixer_kwargs": { |
| 79 | + "query_dim": query_dim, |
| 80 | + "name": "exact", |
| 81 | + "how": "memeff", |
| 82 | + "cross_attention_dim": None, |
| 83 | + } |
| 84 | + } |
| 85 | +
|
| 86 | + >>> # mlp-kwargs |
| 87 | + >>> mlp_kwargs = { |
| 88 | + "in_channels": query_dim, |
| 89 | + "norm_kwargs": {"normalized_shape": query_dim} |
| 90 | + } |
| 91 | +
|
| 92 | + >>> # init metaformer |
| 93 | + >>> metaformer = MetaFormer( |
| 94 | + in_channels=in_channels, |
| 95 | + embed_kwargs=embed_kwargs, |
| 96 | + mixer_kwargs=mixer_kwargs, |
| 97 | + mlp_kwargs=mlp_kwargs, |
| 98 | + layer_scale=True, |
| 99 | + dropout=0.1 |
| 100 | + ) |
| 101 | +
|
| 102 | + >>> x = torch.rand([8, 3, 256, 256]) |
| 103 | + >>> print(metaformer(x).shape) |
| 104 | + >>> # torch.Size([8, 4096, 512]) |
| 105 | +
|
| 106 | +
|
| 107 | + MetaFormer with multi-scale convolutional attention.: |
| 108 | + >>> import torch |
| 109 | + >>> import torch.nn as nn |
| 110 | +
|
| 111 | + >>> in_channels = 3 |
| 112 | + >>> head_dim = 64 |
| 113 | + >>> num_heads = 8 |
| 114 | + >>> query_dim = head_dim*num_heads |
| 115 | + >>> out_channels = 128 |
| 116 | +
|
| 117 | + >>> # patch embedding kwargs |
| 118 | + >>> embed_kwargs = { |
| 119 | + "in_channels": 3, |
| 120 | + "kernel_size": 7, |
| 121 | + "stride": 4, |
| 122 | + "pad": 2, |
| 123 | + "head_dim": head_dim, |
| 124 | + "num_heads": num_heads, |
| 125 | + } |
| 126 | +
|
| 127 | + >>> # token-mixer kwargs |
| 128 | + >>> mixer_kwargs = { |
| 129 | + "token_mixer": "mscan", |
| 130 | + "normalization": "bn", |
| 131 | + "norm_kwargs": { |
| 132 | + "num_features": query_dim, |
| 133 | + }, |
| 134 | + "mixer_kwargs":{ |
| 135 | + "in_channels": query_dim, |
| 136 | + } |
| 137 | + } |
| 138 | +
|
| 139 | + >>> # mlp-kwargs |
| 140 | + >>> mlp_kwargs = { |
| 141 | + "in_channels": query_dim, |
| 142 | + "norm_kwargs": {"normalized_shape": query_dim} |
| 143 | + } |
| 144 | +
|
| 145 | + >>> # init metaformer |
| 146 | + >>> metaformer = MetaFormer( |
| 147 | + in_channels=in_channels, |
| 148 | + out_channels=out_channels, |
| 149 | + embed_kwargs=embed_kwargs, |
| 150 | + mixer_kwargs=mixer_kwargs, |
| 151 | + mlp_kwargs=mlp_kwargs, |
| 152 | + layer_scale=True, |
| 153 | + dropout=0.1 |
| 154 | + ) |
| 155 | +
|
| 156 | + >>> x = torch.rand([8, 3, 256, 256]) |
| 157 | + >>> print(metaformer(x).shape) |
| 158 | + >>> # torch.Size([8, 128, 256, 256]) |
| 159 | + """ |
| 160 | + super().__init__() |
| 161 | + self.out_channels = out_channels if out_channels is not None else in_channels |
| 162 | + mixer_name = mixer_kwargs["token_mixer"] |
| 163 | + |
| 164 | + self.patch_embed = ContiguousEmbed( |
| 165 | + **embed_kwargs, flatten=not RESHAPE_LOOKUP[mixer_name] |
| 166 | + ) |
| 167 | + self.proj_dim = self.patch_embed.proj_dim |
| 168 | + |
| 169 | + self.mixer = TokenMixerBlock(**mixer_kwargs) |
| 170 | + self.mlp = MlpBlock(**mlp_kwargs) |
| 171 | + self.ls1 = ( |
| 172 | + LayerScale(dim=mlp_kwargs["in_channels"]) if layer_scale else Identity() |
| 173 | + ) |
| 174 | + self.ls2 = ( |
| 175 | + LayerScale(dim=mlp_kwargs["in_channels"]) if layer_scale else Identity() |
| 176 | + ) |
| 177 | + self.drop_path1 = DropPath() if dropout else Identity() |
| 178 | + self.drop_path2 = DropPath() if dropout else Identity() |
| 179 | + |
| 180 | + self.proj_out = nn.Conv2d( |
| 181 | + self.proj_dim, self.out_channels, kernel_size=1, stride=1, padding=0 |
| 182 | + ) |
| 183 | + |
| 184 | + self.downsample = Identity() |
| 185 | + if self.out_channels is not None: |
| 186 | + self.downsample = nn.Conv2d(in_channels, out_channels, 1) |
| 187 | + |
| 188 | + def forward(self, x: torch.Tensor, **kwargs) -> torch.Tensor: |
| 189 | + """Forward pass of the token mixer module.""" |
| 190 | + B, _, H, W = x.shape |
| 191 | + residual = self.downsample(x) |
| 192 | + |
| 193 | + # 1. embed and project |
| 194 | + x = self.patch_embed(x) |
| 195 | + |
| 196 | + # 2. token-mixing |
| 197 | + x = self.drop_path1(self.ls1(self.mixer(x, **kwargs))) |
| 198 | + |
| 199 | + # 3. mlp |
| 200 | + x = self.drop_path2(self.ls2(self.mlp(x))) |
| 201 | + |
| 202 | + # 4. Reshape back to image-like shape. |
| 203 | + p_H = self.patch_embed.get_patch_size(H) |
| 204 | + p_W = self.patch_embed.get_patch_size(W) |
| 205 | + x = x.reshape(B, p_H, p_W, self.proj_dim).permute(0, 3, 1, 2) |
| 206 | + |
| 207 | + # Upsample to input dims if patch size less than orig inp size |
| 208 | + # assumes that the input is square mat. |
| 209 | + # NOTE: the kernel_size, pad, & stride has to be set correctly for this to work |
| 210 | + if p_H < H: |
| 211 | + scale_factor = H // p_H |
| 212 | + x = F.interpolate(x, scale_factor=int(scale_factor), mode="bilinear") |
| 213 | + |
| 214 | + # 5. project to original input channels |
| 215 | + x = self.proj_out(x) |
| 216 | + |
| 217 | + # 6. residual |
| 218 | + return x + residual |
0 commit comments