# ์ฐธ์กฐ `ultralytics/models/sam/modules/tiny_encoder.py`

์ฐธ๊ณ

์ด ํ์ผ์ https://github.com/ultralytics/ ultralytics/blob/main/ ultralytics/models/ sam/modules/tiny_encoder .py์์ ํ์ธํ  ์ ์์ต๋๋ค. ๋ฌธ์ ๋ฅผ ๋ฐ๊ฒฌํ๋ฉด ํ ๋ฆฌํ์คํธ ๐ ๏ธ ์ ๊ธฐ์ฌํ์ฌ ๋ฌธ์ ๋ฅผ ํด๊ฒฐํ๋๋ก ๋์์ฃผ์ธ์. ๊ฐ์ฌํฉ๋๋ค ๐!

## `ultralytics.models.sam.modules.tiny_encoder.Conv2d_BN`

๊ธฐ์ง: `Sequential`

2D ์ปจ๋ณผ๋ฃจ์์ ์ํํ ํ ์ผ๊ด ์ ๊ทํ๋ฅผ ์ํํ๋ ์์ฐจ ์ปจํ์ด๋์๋๋ค.

์ ์์ค ์ฝ๋ `ultralytics/models/sam/modules/tiny_encoder.py`
 ```23 24 25 26 27 28 29 30 31 32 33 34 35``` ``````class Conv2d_BN(torch.nn.Sequential): """A sequential container that performs 2D convolution followed by batch normalization.""" def __init__(self, a, b, ks=1, stride=1, pad=0, dilation=1, groups=1, bn_weight_init=1): """Initializes the MBConv model with given input channels, output channels, expansion ratio, activation, and drop path. """ super().__init__() self.add_module("c", torch.nn.Conv2d(a, b, ks, stride, pad, dilation, groups, bias=False)) bn = torch.nn.BatchNorm2d(b) torch.nn.init.constant_(bn.weight, bn_weight_init) torch.nn.init.constant_(bn.bias, 0) self.add_module("bn", bn) ``````

### `__init__(a, b, ks=1, stride=1, pad=0, dilation=1, groups=1, bn_weight_init=1)`

์ฃผ์ด์ง ์๋ ฅ ์ฑ๋, ์ถ๋ ฅ ์ฑ๋, ํ์ฅ ๋น์จ, ํ์ฑํ, ๋๋กญ ๊ฒฝ๋ก๋ก MBConv ๋ชจ๋ธ์ ์ด๊ธฐํํฉ๋๋ค. ๋๋กญ ๊ฒฝ๋ก๋ก ์ด๊ธฐํํฉ๋๋ค.

์ ์์ค ์ฝ๋ `ultralytics/models/sam/modules/tiny_encoder.py`
 ```26 27 28 29 30 31 32 33 34 35``` ``````def __init__(self, a, b, ks=1, stride=1, pad=0, dilation=1, groups=1, bn_weight_init=1): """Initializes the MBConv model with given input channels, output channels, expansion ratio, activation, and drop path. """ super().__init__() self.add_module("c", torch.nn.Conv2d(a, b, ks, stride, pad, dilation, groups, bias=False)) bn = torch.nn.BatchNorm2d(b) torch.nn.init.constant_(bn.weight, bn_weight_init) torch.nn.init.constant_(bn.bias, 0) self.add_module("bn", bn) ``````

## `ultralytics.models.sam.modules.tiny_encoder.PatchEmbed`

๊ธฐ์ง: `Module`

์ด๋ฏธ์ง๋ฅผ ํจ์น์ ์๋ฒ ๋ํ๊ณ  ์ง์ ๋ ์๋ฒ ๋ ํฌ๊ธฐ๋ก ํฌ์ฌํฉ๋๋ค.

์ ์์ค ์ฝ๋ `ultralytics/models/sam/modules/tiny_encoder.py`
 ```38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60``` ``````class PatchEmbed(nn.Module): """Embeds images into patches and projects them into a specified embedding dimension.""" def __init__(self, in_chans, embed_dim, resolution, activation): """Initialize the PatchMerging class with specified input, output dimensions, resolution and activation function. """ super().__init__() img_size: Tuple[int, int] = to_2tuple(resolution) self.patches_resolution = (img_size[0] // 4, img_size[1] // 4) self.num_patches = self.patches_resolution[0] * self.patches_resolution[1] self.in_chans = in_chans self.embed_dim = embed_dim n = embed_dim self.seq = nn.Sequential( Conv2d_BN(in_chans, n // 2, 3, 2, 1), activation(), Conv2d_BN(n // 2, n, 3, 2, 1), ) def forward(self, x): """Runs input tensor 'x' through the PatchMerging model's sequence of operations.""" return self.seq(x) ``````

### `__init__(in_chans, embed_dim, resolution, activation)`

์ง์ ๋ ์๋ ฅ, ์ถ๋ ฅ ํฌ๊ธฐ, ํด์๋, ํ์ฑํ ํจ์๋ฅผ ์ฌ์ฉํ์ฌ PatchMerging ํด๋์ค๋ฅผ ์ด๊ธฐํํฉ๋๋ค. ํจ์๋ฅผ ํธ์ถํฉ๋๋ค.

์ ์์ค ์ฝ๋ `ultralytics/models/sam/modules/tiny_encoder.py`
 ```41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56``` ``````def __init__(self, in_chans, embed_dim, resolution, activation): """Initialize the PatchMerging class with specified input, output dimensions, resolution and activation function. """ super().__init__() img_size: Tuple[int, int] = to_2tuple(resolution) self.patches_resolution = (img_size[0] // 4, img_size[1] // 4) self.num_patches = self.patches_resolution[0] * self.patches_resolution[1] self.in_chans = in_chans self.embed_dim = embed_dim n = embed_dim self.seq = nn.Sequential( Conv2d_BN(in_chans, n // 2, 3, 2, 1), activation(), Conv2d_BN(n // 2, n, 3, 2, 1), ) ``````

### `forward(x)`

์๋ ฅ tensor 'x'๋ฅผ ํจ์น๋ณํฉ ๋ชจ๋ธ์ ์์ ์์์ ๋ฐ๋ผ ์คํํฉ๋๋ค.

์ ์์ค ์ฝ๋ `ultralytics/models/sam/modules/tiny_encoder.py`
 ```58 59 60``` ``````def forward(self, x): """Runs input tensor 'x' through the PatchMerging model's sequence of operations.""" return self.seq(x) ``````

## `ultralytics.models.sam.modules.tiny_encoder.MBConv`

๊ธฐ์ง: `Module`

๋ชจ๋ฐ์ผ ์ญ๋ณ๋ชฉํ์(MBConv) ๋ ์ด์ด, EfficientNet ์ํคํ์ฒ์ ์ผ๋ถ์๋๋ค.

์ ์์ค ์ฝ๋ `ultralytics/models/sam/modules/tiny_encoder.py`
 ```63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98``` ``````class MBConv(nn.Module): """Mobile Inverted Bottleneck Conv (MBConv) layer, part of the EfficientNet architecture.""" def __init__(self, in_chans, out_chans, expand_ratio, activation, drop_path): """Initializes a convolutional layer with specified dimensions, input resolution, depth, and activation function. """ super().__init__() self.in_chans = in_chans self.hidden_chans = int(in_chans * expand_ratio) self.out_chans = out_chans self.conv1 = Conv2d_BN(in_chans, self.hidden_chans, ks=1) self.act1 = activation() self.conv2 = Conv2d_BN(self.hidden_chans, self.hidden_chans, ks=3, stride=1, pad=1, groups=self.hidden_chans) self.act2 = activation() self.conv3 = Conv2d_BN(self.hidden_chans, out_chans, ks=1, bn_weight_init=0.0) self.act3 = activation() # NOTE: `DropPath` is needed only for training. # self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity() self.drop_path = nn.Identity() def forward(self, x): """Implements the forward pass for the model architecture.""" shortcut = x x = self.conv1(x) x = self.act1(x) x = self.conv2(x) x = self.act2(x) x = self.conv3(x) x = self.drop_path(x) x += shortcut return self.act3(x) ``````

### `__init__(in_chans, out_chans, expand_ratio, activation, drop_path)`

์ง์ ๋ ์น์, ์๋ ฅ ํด์๋, ๊น์ด, ํ์ฑํ ํจ์๋ฅผ ์ฌ์ฉํ์ฌ ์ปจ๋ณผ๋ฃจ์ ๋ ์ด์ด๋ฅผ ์ด๊ธฐํํฉ๋๋ค. ํจ์๋ฅผ ํธ์ถํฉ๋๋ค.

์ ์์ค ์ฝ๋ `ultralytics/models/sam/modules/tiny_encoder.py`
 ```66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86``` ``````def __init__(self, in_chans, out_chans, expand_ratio, activation, drop_path): """Initializes a convolutional layer with specified dimensions, input resolution, depth, and activation function. """ super().__init__() self.in_chans = in_chans self.hidden_chans = int(in_chans * expand_ratio) self.out_chans = out_chans self.conv1 = Conv2d_BN(in_chans, self.hidden_chans, ks=1) self.act1 = activation() self.conv2 = Conv2d_BN(self.hidden_chans, self.hidden_chans, ks=3, stride=1, pad=1, groups=self.hidden_chans) self.act2 = activation() self.conv3 = Conv2d_BN(self.hidden_chans, out_chans, ks=1, bn_weight_init=0.0) self.act3 = activation() # NOTE: `DropPath` is needed only for training. # self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity() self.drop_path = nn.Identity() ``````

### `forward(x)`

๋ชจ๋ธ ์ํคํ์ฒ์ ๋ํ ํฌ์๋ ํจ์ค๋ฅผ ๊ตฌํํฉ๋๋ค.

์ ์์ค ์ฝ๋ `ultralytics/models/sam/modules/tiny_encoder.py`
 ```88 89 90 91 92 93 94 95 96 97 98``` ``````def forward(self, x): """Implements the forward pass for the model architecture.""" shortcut = x x = self.conv1(x) x = self.act1(x) x = self.conv2(x) x = self.act2(x) x = self.conv3(x) x = self.drop_path(x) x += shortcut return self.act3(x) ``````

## `ultralytics.models.sam.modules.tiny_encoder.PatchMerging`

๊ธฐ์ง: `Module`

๊ธฐ๋ฅ ๋งต์์ ์ธ์ ํ ํจ์น๋ฅผ ๋ณํฉํ๊ณ  ์๋ก์ด ์ฐจ์์ผ๋ก ํฌ์ํฉ๋๋ค.

์ ์์ค ์ฝ๋ `ultralytics/models/sam/modules/tiny_encoder.py`
 ```101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133``` ``````class PatchMerging(nn.Module): """Merges neighboring patches in the feature map and projects to a new dimension.""" def __init__(self, input_resolution, dim, out_dim, activation): """Initializes the ConvLayer with specific dimension, input resolution, depth, activation, drop path, and other optional parameters. """ super().__init__() self.input_resolution = input_resolution self.dim = dim self.out_dim = out_dim self.act = activation() self.conv1 = Conv2d_BN(dim, out_dim, 1, 1, 0) stride_c = 1 if out_dim in [320, 448, 576] else 2 self.conv2 = Conv2d_BN(out_dim, out_dim, 3, stride_c, 1, groups=out_dim) self.conv3 = Conv2d_BN(out_dim, out_dim, 1, 1, 0) def forward(self, x): """Applies forward pass on the input utilizing convolution and activation layers, and returns the result.""" if x.ndim == 3: H, W = self.input_resolution B = len(x) # (B, C, H, W) x = x.view(B, H, W, -1).permute(0, 3, 1, 2) x = self.conv1(x) x = self.act(x) x = self.conv2(x) x = self.act(x) x = self.conv3(x) return x.flatten(2).transpose(1, 2) ``````

### `__init__(input_resolution, dim, out_dim, activation)`

ํน์  ์น์, ์๋ ฅ ํด์๋, ๊น์ด, ํ์ฑํ, ๋๋กญ ๊ฒฝ๋ก ๋ฐ ๊ธฐํ ์ต์ ํ๋ผ๋ฏธํฐ๋ก ์ด๊ธฐํํฉ๋๋ค.

์ ์์ค ์ฝ๋ `ultralytics/models/sam/modules/tiny_encoder.py`
 ```104 105 106 107 108 109 110 111 112 113 114 115 116 117``` ``````def __init__(self, input_resolution, dim, out_dim, activation): """Initializes the ConvLayer with specific dimension, input resolution, depth, activation, drop path, and other optional parameters. """ super().__init__() self.input_resolution = input_resolution self.dim = dim self.out_dim = out_dim self.act = activation() self.conv1 = Conv2d_BN(dim, out_dim, 1, 1, 0) stride_c = 1 if out_dim in [320, 448, 576] else 2 self.conv2 = Conv2d_BN(out_dim, out_dim, 3, stride_c, 1, groups=out_dim) self.conv3 = Conv2d_BN(out_dim, out_dim, 1, 1, 0) ``````

### `forward(x)`

์ปจ๋ณผ๋ฃจ์ ๋ฐ ํ์ฑํ ๋ ์ด์ด๋ฅผ ํ์ฉํ์ฌ ์๋ ฅ์ ํฌ์๋ ํจ์ค๋ฅผ ์ ์ฉํ๊ณ  ๊ฒฐ๊ณผ๋ฅผ ๋ฐํํฉ๋๋ค.

์ ์์ค ์ฝ๋ `ultralytics/models/sam/modules/tiny_encoder.py`
 ```119 120 121 122 123 124 125 126 127 128 129 130 131 132 133``` ``````def forward(self, x): """Applies forward pass on the input utilizing convolution and activation layers, and returns the result.""" if x.ndim == 3: H, W = self.input_resolution B = len(x) # (B, C, H, W) x = x.view(B, H, W, -1).permute(0, 3, 1, 2) x = self.conv1(x) x = self.act(x) x = self.conv2(x) x = self.act(x) x = self.conv3(x) return x.flatten(2).transpose(1, 2) ``````

## `ultralytics.models.sam.modules.tiny_encoder.ConvLayer`

๊ธฐ์ง: `Module`

์ปจ๋ณผ๋ฃจ์ ๋ ์ด์ด๋ ์ฌ๋ฌ ๊ฐ์ MobileNetV3 ์คํ์ผ ์ญ๋ณ๋ชฉ ์ปจ๋ณผ๋ฃจ์(MBConv)์ ํน์ง์ผ๋ก ํฉ๋๋ค.

์ ํ์ ์ผ๋ก ์ถ๋ ฅ์ ๋ค์ด์ํ ์ฐ์ฐ์ ์ ์ฉํ๊ณ  ๊ทธ๋ผ๋ฐ์ด์ ์ฒดํฌํฌ์ธํธ๋ฅผ ์ง์ํฉ๋๋ค.

์ ์์ค ์ฝ๋ `ultralytics/models/sam/modules/tiny_encoder.py`
 ```136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200``` ``````class ConvLayer(nn.Module): """ Convolutional Layer featuring multiple MobileNetV3-style inverted bottleneck convolutions (MBConv). Optionally applies downsample operations to the output, and provides support for gradient checkpointing. """ def __init__( self, dim, input_resolution, depth, activation, drop_path=0.0, downsample=None, use_checkpoint=False, out_dim=None, conv_expand_ratio=4.0, ): """ Initializes the ConvLayer with the given dimensions and settings. Args: dim (int): The dimensionality of the input and output. input_resolution (Tuple[int, int]): The resolution of the input image. depth (int): The number of MBConv layers in the block. activation (Callable): Activation function applied after each convolution. drop_path (Union[float, List[float]]): Drop path rate. Single float or a list of floats for each MBConv. downsample (Optional[Callable]): Function for downsampling the output. None to skip downsampling. use_checkpoint (bool): Whether to use gradient checkpointing to save memory. out_dim (Optional[int]): The dimensionality of the output. None means it will be the same as `dim`. conv_expand_ratio (float): Expansion ratio for the MBConv layers. """ super().__init__() self.dim = dim self.input_resolution = input_resolution self.depth = depth self.use_checkpoint = use_checkpoint # Build blocks self.blocks = nn.ModuleList( [ MBConv( dim, dim, conv_expand_ratio, activation, drop_path[i] if isinstance(drop_path, list) else drop_path, ) for i in range(depth) ] ) # Patch merging layer self.downsample = ( None if downsample is None else downsample(input_resolution, dim=dim, out_dim=out_dim, activation=activation) ) def forward(self, x): """Processes the input through a series of convolutional layers and returns the activated output.""" for blk in self.blocks: x = checkpoint.checkpoint(blk, x) if self.use_checkpoint else blk(x) return x if self.downsample is None else self.downsample(x) ``````

### `__init__(dim, input_resolution, depth, activation, drop_path=0.0, downsample=None, use_checkpoint=False, out_dim=None, conv_expand_ratio=4.0)`

์ฃผ์ด์ง ์น์์ ์ค์ ์ผ๋ก ConvLayer๋ฅผ ์ด๊ธฐํํฉ๋๋ค.

๋งค๊ฐ๋ณ์:

์ด๋ฆ ์ ํ ์ค๋ช ๊ธฐ๋ณธ๊ฐ
`dim` `int`

์๋ ฅ ๋ฐ ์ถ๋ ฅ์ ์ฐจ์์๋๋ค.

ํ์
`input_resolution` `Tuple[int, int]`

์๋ ฅ ์ด๋ฏธ์ง์ ํด์๋์๋๋ค.

ํ์
`depth` `int`

๋ธ๋ก์ ์๋ MBConv ๋ ์ด์ด ์์๋๋ค.

ํ์
`activation` `Callable`

๊ฐ ์ปจ๋ณผ๋ฃจ์ ํ์ ์ ์ฉ๋๋ ํ์ฑํ ํจ์์๋๋ค.

ํ์
`drop_path` `Union[float, List[float]]`

๋๋กญ ๊ฒฝ๋ก ์๋. ๋จ์ผ ํ๋กํธ ๋๋ ๊ฐ MBConv์ ๋ํ ํ๋กํธ ๋ชฉ๋ก์๋๋ค.

`0.0`
`downsample` `Optional[Callable]`

์ถ๋ ฅ์ ๋ค์ด์ํ๋งํ๋ ๊ธฐ๋ฅ์๋๋ค. ์์์ผ๋ก ์ค์ ํ๋ฉด ๋ค์ด์ํ๋ง์ ๊ฑด๋๋๋๋ค.

`None`
`use_checkpoint` `bool`

๋ฉ๋ชจ๋ฆฌ ์ ์ฝ์ ์ํด ๊ทธ๋ผ๋ฐ์ด์ ์ฒดํฌํฌ์ธํธ๋ฅผ ์ฌ์ฉํ ์ง ์ฌ๋ถ์๋๋ค.

`False`
`out_dim` `Optional[int]`

์ถ๋ ฅ์ ์ฐจ์์๋๋ค. ์์์ ๋ค์๊ณผ ๊ฐ์์ ์๋ฏธํฉ๋๋ค. `dim`.

`None`
`conv_expand_ratio` `float`

MBConv ๋ ์ด์ด์ ํ์ฅ ๋น์จ์๋๋ค.

`4.0`
์ ์์ค ์ฝ๋ `ultralytics/models/sam/modules/tiny_encoder.py`
 ```143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194``` ``````def __init__( self, dim, input_resolution, depth, activation, drop_path=0.0, downsample=None, use_checkpoint=False, out_dim=None, conv_expand_ratio=4.0, ): """ Initializes the ConvLayer with the given dimensions and settings. Args: dim (int): The dimensionality of the input and output. input_resolution (Tuple[int, int]): The resolution of the input image. depth (int): The number of MBConv layers in the block. activation (Callable): Activation function applied after each convolution. drop_path (Union[float, List[float]]): Drop path rate. Single float or a list of floats for each MBConv. downsample (Optional[Callable]): Function for downsampling the output. None to skip downsampling. use_checkpoint (bool): Whether to use gradient checkpointing to save memory. out_dim (Optional[int]): The dimensionality of the output. None means it will be the same as `dim`. conv_expand_ratio (float): Expansion ratio for the MBConv layers. """ super().__init__() self.dim = dim self.input_resolution = input_resolution self.depth = depth self.use_checkpoint = use_checkpoint # Build blocks self.blocks = nn.ModuleList( [ MBConv( dim, dim, conv_expand_ratio, activation, drop_path[i] if isinstance(drop_path, list) else drop_path, ) for i in range(depth) ] ) # Patch merging layer self.downsample = ( None if downsample is None else downsample(input_resolution, dim=dim, out_dim=out_dim, activation=activation) ) ``````

### `forward(x)`

์ผ๋ จ์ ์ปจ๋ณผ๋ฃจ์ ๋ ์ด์ด๋ฅผ ํตํด ์๋ ฅ์ ์ฒ๋ฆฌํ๊ณ  ํ์ฑํ๋ ์ถ๋ ฅ์ ๋ฐํํฉ๋๋ค.

์ ์์ค ์ฝ๋ `ultralytics/models/sam/modules/tiny_encoder.py`
 ```196 197 198 199 200``` ``````def forward(self, x): """Processes the input through a series of convolutional layers and returns the activated output.""" for blk in self.blocks: x = checkpoint.checkpoint(blk, x) if self.use_checkpoint else blk(x) return x if self.downsample is None else self.downsample(x) ``````

## `ultralytics.models.sam.modules.tiny_encoder.Mlp`

๊ธฐ์ง: `Module`

๋ณ์๊ธฐ ์ํคํ์ฒ๋ฅผ ์ํ ๋ฉํฐ๋ ์ด์ด ํผ์ํธ๋ก (MLP).

์ด ๋ ์ด์ด๋ in_features๋ก ์๋ ฅ์ ๋ฐ๊ณ , ๋ ์ด์ด ์ ๊ทํ์ ์์ ํ ์ฐ๊ฒฐ๋ ๋ ๊ฐ์ ๋ ์ด์ด๋ฅผ ์ ์ฉํฉ๋๋ค.

์ ์์ค ์ฝ๋ `ultralytics/models/sam/modules/tiny_encoder.py`
 ```203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228``` ``````class Mlp(nn.Module): """ Multi-layer Perceptron (MLP) for transformer architectures. This layer takes an input with in_features, applies layer normalization and two fully-connected layers. """ def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.0): """Initializes Attention module with the given parameters including dimension, key_dim, number of heads, etc.""" super().__init__() out_features = out_features or in_features hidden_features = hidden_features or in_features self.norm = nn.LayerNorm(in_features) self.fc1 = nn.Linear(in_features, hidden_features) self.fc2 = nn.Linear(hidden_features, out_features) self.act = act_layer() self.drop = nn.Dropout(drop) def forward(self, x): """Applies operations on input x and returns modified x, runs downsample if not None.""" x = self.norm(x) x = self.fc1(x) x = self.act(x) x = self.drop(x) x = self.fc2(x) return self.drop(x) ``````

### `__init__(in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.0)`

์ฐจ์, ํค_๋ค, ํค๋ ์ ๋ฑ ์ฃผ์ด์ง ๋งค๊ฐ๋ณ์๋ฅผ ์ฌ์ฉํ์ฌ ์ฃผ์ ๋ชจ๋์ ์ด๊ธฐํํฉ๋๋ค.

์ ์์ค ์ฝ๋ `ultralytics/models/sam/modules/tiny_encoder.py`
 ```210 211 212 213 214 215 216 217 218 219``` ``````def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.0): """Initializes Attention module with the given parameters including dimension, key_dim, number of heads, etc.""" super().__init__() out_features = out_features or in_features hidden_features = hidden_features or in_features self.norm = nn.LayerNorm(in_features) self.fc1 = nn.Linear(in_features, hidden_features) self.fc2 = nn.Linear(hidden_features, out_features) self.act = act_layer() self.drop = nn.Dropout(drop) ``````

### `forward(x)`

์๋ ฅ x์ ์ฐ์ฐ์ ์ ์ฉํ๊ณ  ์์ ๋ x๋ฅผ ๋ฐํํ๋ฉฐ, ์์์ด ์๋ ๊ฒฝ์ฐ ๋ค์ด์ํ๋ง์ ์คํํฉ๋๋ค.

์ ์์ค ์ฝ๋ `ultralytics/models/sam/modules/tiny_encoder.py`
 ```221 222 223 224 225 226 227 228``` ``````def forward(self, x): """Applies operations on input x and returns modified x, runs downsample if not None.""" x = self.norm(x) x = self.fc1(x) x = self.act(x) x = self.drop(x) x = self.fc2(x) return self.drop(x) ``````

## `ultralytics.models.sam.modules.tiny_encoder.Attention`

๊ธฐ์ง: `Module`

๊ณต๊ฐ ์ธ์์ ์ง์ํ๋ ๋ฉํฐ ํค๋ ์ฃผ์๋ ฅ ๋ชจ๋๋ก ๊ณต๊ฐ ํด์๋์ ๋ฐ๋ผ ์ฃผ์๋ ฅ ํธํฅ์ ์ ์ฉํฉ๋๋ค. ํด์๋. ํด์๋ ๊ทธ๋ฆฌ๋์์ ๊ณต๊ฐ ์์น ๊ฐ์ ๊ฐ ๊ณ ์  ์คํ์์ ๋ํด ํ๋ จ ๊ฐ๋ฅํ ์ฃผ์ ํธํฅ ๊ตฌํ ๊ทธ๋ฆฌ๋.

์์ฑ:

์ด๋ฆ ์ ํ ์ค๋ช
`ab` `Tensor`

์ถ๋ก ์ ์ํด ์บ์๋ ์ฃผ์ ํธํฅ์ ํ๋ จ ์ค์ ์ญ์ ๋ฉ๋๋ค.

์ ์์ค ์ฝ๋ `ultralytics/models/sam/modules/tiny_encoder.py`
 ```231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321``` ``````class Attention(torch.nn.Module): """ Multi-head attention module with support for spatial awareness, applying attention biases based on spatial resolution. Implements trainable attention biases for each unique offset between spatial positions in the resolution grid. Attributes: ab (Tensor, optional): Cached attention biases for inference, deleted during training. """ def __init__( self, dim, key_dim, num_heads=8, attn_ratio=4, resolution=(14, 14), ): """ Initializes the Attention module. Args: dim (int): The dimensionality of the input and output. key_dim (int): The dimensionality of the keys and queries. num_heads (int, optional): Number of attention heads. Default is 8. attn_ratio (float, optional): Attention ratio, affecting the dimensions of the value vectors. Default is 4. resolution (Tuple[int, int], optional): Spatial resolution of the input feature map. Default is (14, 14). Raises: AssertionError: If `resolution` is not a tuple of length 2. """ super().__init__() assert isinstance(resolution, tuple) and len(resolution) == 2 self.num_heads = num_heads self.scale = key_dim**-0.5 self.key_dim = key_dim self.nh_kd = nh_kd = key_dim * num_heads self.d = int(attn_ratio * key_dim) self.dh = int(attn_ratio * key_dim) * num_heads self.attn_ratio = attn_ratio h = self.dh + nh_kd * 2 self.norm = nn.LayerNorm(dim) self.qkv = nn.Linear(dim, h) self.proj = nn.Linear(self.dh, dim) points = list(itertools.product(range(resolution[0]), range(resolution[1]))) N = len(points) attention_offsets = {} idxs = [] for p1 in points: for p2 in points: offset = (abs(p1[0] - p2[0]), abs(p1[1] - p2[1])) if offset not in attention_offsets: attention_offsets[offset] = len(attention_offsets) idxs.append(attention_offsets[offset]) self.attention_biases = torch.nn.Parameter(torch.zeros(num_heads, len(attention_offsets))) self.register_buffer("attention_bias_idxs", torch.LongTensor(idxs).view(N, N), persistent=False) @torch.no_grad() def train(self, mode=True): """Sets the module in training mode and handles attribute 'ab' based on the mode.""" super().train(mode) if mode and hasattr(self, "ab"): del self.ab else: self.ab = self.attention_biases[:, self.attention_bias_idxs] def forward(self, x): # x """Performs forward pass over the input tensor 'x' by applying normalization and querying keys/values.""" B, N, _ = x.shape # B, N, C # Normalization x = self.norm(x) qkv = self.qkv(x) # (B, N, num_heads, d) q, k, v = qkv.view(B, N, self.num_heads, -1).split([self.key_dim, self.key_dim, self.d], dim=3) # (B, num_heads, N, d) q = q.permute(0, 2, 1, 3) k = k.permute(0, 2, 1, 3) v = v.permute(0, 2, 1, 3) self.ab = self.ab.to(self.attention_biases.device) attn = (q @ k.transpose(-2, -1)) * self.scale + ( self.attention_biases[:, self.attention_bias_idxs] if self.training else self.ab ) attn = attn.softmax(dim=-1) x = (attn @ v).transpose(1, 2).reshape(B, N, self.dh) return self.proj(x) ``````

### `__init__(dim, key_dim, num_heads=8, attn_ratio=4, resolution=(14, 14))`

์ฃผ์ ๋ชจ๋์ ์ด๊ธฐํํฉ๋๋ค.

๋งค๊ฐ๋ณ์:

์ด๋ฆ ์ ํ ์ค๋ช ๊ธฐ๋ณธ๊ฐ
`dim` `int`

์๋ ฅ ๋ฐ ์ถ๋ ฅ์ ์ฐจ์์๋๋ค.

ํ์
`key_dim` `int`

ํค ๋ฐ ์ฟผ๋ฆฌ์ ์ฐจ์์๋๋ค.

ํ์
`num_heads` `int`

์ฃผ์ ํค๋ ์์๋๋ค. ๊ธฐ๋ณธ๊ฐ์ 8๊ฐ์๋๋ค.

`8`
`attn_ratio` `float`

์ฃผ์ ๋น์จ๋ก, ๊ฐ ๋ฒกํฐ์ ํฌ๊ธฐ์ ์ํฅ์ ์ค๋๋ค. ๊ธฐ๋ณธ๊ฐ์ 4์๋๋ค.

`4`
`resolution` `Tuple[int, int]`

์๋ ฅ ํผ์ฒ ๋งต์ ๊ณต๊ฐ ํด์๋์๋๋ค. ๊ธฐ๋ณธ๊ฐ์ (14, 14)์๋๋ค.

`(14, 14)`

์ฌ๋ฆฌ๊ธฐ:

์ ํ ์ค๋ช
`AssertionError`

๋ง์ฝ `resolution` ๋ ๊ธธ์ด 2์ ํํ์ด ์๋๋๋ค.

์ ์์ค ์ฝ๋ `ultralytics/models/sam/modules/tiny_encoder.py`
 ```241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289``` ``````def __init__( self, dim, key_dim, num_heads=8, attn_ratio=4, resolution=(14, 14), ): """ Initializes the Attention module. Args: dim (int): The dimensionality of the input and output. key_dim (int): The dimensionality of the keys and queries. num_heads (int, optional): Number of attention heads. Default is 8. attn_ratio (float, optional): Attention ratio, affecting the dimensions of the value vectors. Default is 4. resolution (Tuple[int, int], optional): Spatial resolution of the input feature map. Default is (14, 14). Raises: AssertionError: If `resolution` is not a tuple of length 2. """ super().__init__() assert isinstance(resolution, tuple) and len(resolution) == 2 self.num_heads = num_heads self.scale = key_dim**-0.5 self.key_dim = key_dim self.nh_kd = nh_kd = key_dim * num_heads self.d = int(attn_ratio * key_dim) self.dh = int(attn_ratio * key_dim) * num_heads self.attn_ratio = attn_ratio h = self.dh + nh_kd * 2 self.norm = nn.LayerNorm(dim) self.qkv = nn.Linear(dim, h) self.proj = nn.Linear(self.dh, dim) points = list(itertools.product(range(resolution[0]), range(resolution[1]))) N = len(points) attention_offsets = {} idxs = [] for p1 in points: for p2 in points: offset = (abs(p1[0] - p2[0]), abs(p1[1] - p2[1])) if offset not in attention_offsets: attention_offsets[offset] = len(attention_offsets) idxs.append(attention_offsets[offset]) self.attention_biases = torch.nn.Parameter(torch.zeros(num_heads, len(attention_offsets))) self.register_buffer("attention_bias_idxs", torch.LongTensor(idxs).view(N, N), persistent=False) ``````

### `forward(x)`

์ ๊ทํ๋ฅผ ์ ์ฉํ๊ณ  ํค/๊ฐ์ ์ฟผ๋ฆฌํ์ฌ tensor 'x' ์๋ ฅ์ ๋ํ ์ ๋ฐฉํฅ ์ ๋ฌ์ ์ํํฉ๋๋ค.

์ ์์ค ์ฝ๋ `ultralytics/models/sam/modules/tiny_encoder.py`
 ```300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321``` ``````def forward(self, x): # x """Performs forward pass over the input tensor 'x' by applying normalization and querying keys/values.""" B, N, _ = x.shape # B, N, C # Normalization x = self.norm(x) qkv = self.qkv(x) # (B, N, num_heads, d) q, k, v = qkv.view(B, N, self.num_heads, -1).split([self.key_dim, self.key_dim, self.d], dim=3) # (B, num_heads, N, d) q = q.permute(0, 2, 1, 3) k = k.permute(0, 2, 1, 3) v = v.permute(0, 2, 1, 3) self.ab = self.ab.to(self.attention_biases.device) attn = (q @ k.transpose(-2, -1)) * self.scale + ( self.attention_biases[:, self.attention_bias_idxs] if self.training else self.ab ) attn = attn.softmax(dim=-1) x = (attn @ v).transpose(1, 2).reshape(B, N, self.dh) return self.proj(x) ``````

### `train(mode=True)`

๋ชจ๋์ ํ๋ จ ๋ชจ๋๋ก ์ค์ ํ๊ณ  ๋ชจ๋์ ๋ฐ๋ผ 'ab' ์์ฑ์ ์ฒ๋ฆฌํฉ๋๋ค.

์ ์์ค ์ฝ๋ `ultralytics/models/sam/modules/tiny_encoder.py`
 ```291 292 293 294 295 296 297 298``` ``````@torch.no_grad() def train(self, mode=True): """Sets the module in training mode and handles attribute 'ab' based on the mode.""" super().train(mode) if mode and hasattr(self, "ab"): del self.ab else: self.ab = self.attention_biases[:, self.attention_bias_idxs] ``````

## `ultralytics.models.sam.modules.tiny_encoder.TinyViTBlock`

๊ธฐ์ง: `Module`

์๋ ฅ์ ์์ฒด ์ฃผ์์ ๋ก์ปฌ ์ปจ๋ณผ๋ฃจ์์ ์ ์ฉํ๋ TinyViT ๋ธ๋ก์๋๋ค.

์ ์์ค ์ฝ๋ `ultralytics/models/sam/modules/tiny_encoder.py`
 ```324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434``` ``````class TinyViTBlock(nn.Module): """TinyViT Block that applies self-attention and a local convolution to the input.""" def __init__( self, dim, input_resolution, num_heads, window_size=7, mlp_ratio=4.0, drop=0.0, drop_path=0.0, local_conv_size=3, activation=nn.GELU, ): """ Initializes the TinyViTBlock. Args: dim (int): The dimensionality of the input and output. input_resolution (Tuple[int, int]): Spatial resolution of the input feature map. num_heads (int): Number of attention heads. window_size (int, optional): Window size for attention. Default is 7. mlp_ratio (float, optional): Ratio of mlp hidden dim to embedding dim. Default is 4. drop (float, optional): Dropout rate. Default is 0. drop_path (float, optional): Stochastic depth rate. Default is 0. local_conv_size (int, optional): The kernel size of the local convolution. Default is 3. activation (torch.nn, optional): Activation function for MLP. Default is nn.GELU. Raises: AssertionError: If `window_size` is not greater than 0. AssertionError: If `dim` is not divisible by `num_heads`. """ super().__init__() self.dim = dim self.input_resolution = input_resolution self.num_heads = num_heads assert window_size > 0, "window_size must be greater than 0" self.window_size = window_size self.mlp_ratio = mlp_ratio # NOTE: `DropPath` is needed only for training. # self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity() self.drop_path = nn.Identity() assert dim % num_heads == 0, "dim must be divisible by num_heads" head_dim = dim // num_heads window_resolution = (window_size, window_size) self.attn = Attention(dim, head_dim, num_heads, attn_ratio=1, resolution=window_resolution) mlp_hidden_dim = int(dim * mlp_ratio) mlp_activation = activation self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=mlp_activation, drop=drop) pad = local_conv_size // 2 self.local_conv = Conv2d_BN(dim, dim, ks=local_conv_size, stride=1, pad=pad, groups=dim) def forward(self, x): """Applies attention-based transformation or padding to input 'x' before passing it through a local convolution. """ H, W = self.input_resolution B, L, C = x.shape assert L == H * W, "input feature has wrong size" res_x = x if H == self.window_size and W == self.window_size: x = self.attn(x) else: x = x.view(B, H, W, C) pad_b = (self.window_size - H % self.window_size) % self.window_size pad_r = (self.window_size - W % self.window_size) % self.window_size padding = pad_b > 0 or pad_r > 0 if padding: x = F.pad(x, (0, 0, 0, pad_r, 0, pad_b)) pH, pW = H + pad_b, W + pad_r nH = pH // self.window_size nW = pW // self.window_size # Window partition x = ( x.view(B, nH, self.window_size, nW, self.window_size, C) .transpose(2, 3) .reshape(B * nH * nW, self.window_size * self.window_size, C) ) x = self.attn(x) # Window reverse x = x.view(B, nH, nW, self.window_size, self.window_size, C).transpose(2, 3).reshape(B, pH, pW, C) if padding: x = x[:, :H, :W].contiguous() x = x.view(B, L, C) x = res_x + self.drop_path(x) x = x.transpose(1, 2).reshape(B, C, H, W) x = self.local_conv(x) x = x.view(B, C, L).transpose(1, 2) return x + self.drop_path(self.mlp(x)) def extra_repr(self) -> str: """Returns a formatted string representing the TinyViTBlock's parameters: dimension, input resolution, number of attentions heads, window size, and MLP ratio. """ return ( f"dim={self.dim}, input_resolution={self.input_resolution}, num_heads={self.num_heads}, " f"window_size={self.window_size}, mlp_ratio={self.mlp_ratio}" ) ``````

### `__init__(dim, input_resolution, num_heads, window_size=7, mlp_ratio=4.0, drop=0.0, drop_path=0.0, local_conv_size=3, activation=nn.GELU)`

TinyViTBlock์ ์ด๊ธฐํํฉ๋๋ค.

๋งค๊ฐ๋ณ์:

์ด๋ฆ ์ ํ ์ค๋ช ๊ธฐ๋ณธ๊ฐ
`dim` `int`

์๋ ฅ ๋ฐ ์ถ๋ ฅ์ ์ฐจ์์๋๋ค.

ํ์
`input_resolution` `Tuple[int, int]`

์๋ ฅ ๊ธฐ๋ฅ ๋งต์ ๊ณต๊ฐ ํด์๋์๋๋ค.

ํ์
`num_heads` `int`

๊ด์ฌ ํค๋ ์์๋๋ค.

ํ์
`window_size` `int`

์ฃผ์๋ฅผ ๋๊ธฐ ์ํ ์ฐฝ ํฌ๊ธฐ. ๊ธฐ๋ณธ๊ฐ์ 7์๋๋ค.

`7`
`mlp_ratio` `float`

MLP ์จ๊น ์ด๋ก๊ฒ์ ์๋ฒ ๋ฉ ์ด๋ก๊ฒ์ ๋น์จ์๋๋ค. ๊ธฐ๋ณธ๊ฐ์ 4์๋๋ค.

`4.0`
`drop` `float`

ํ๋ฝ๋ฅ . ๊ธฐ๋ณธ๊ฐ์ 0์๋๋ค.

`0.0`
`drop_path` `float`

ํ๋ฅ ์  ๊น์ด ๋น์จ. ๊ธฐ๋ณธ๊ฐ์ 0์๋๋ค.

`0.0`
`local_conv_size` `int`

๋ก์ปฌ ์ปจ๋ณผ๋ฃจ์์ ์ปค๋ ํฌ๊ธฐ์๋๋ค. ๊ธฐ๋ณธ๊ฐ์ 3์๋๋ค.

`3`
`activation` `nn`

MLP์ฉ ํ์ฑํ ๊ธฐ๋ฅ. ๊ธฐ๋ณธ๊ฐ์ nn.GELU์๋๋ค.

`GELU`

์ฌ๋ฆฌ๊ธฐ:

์ ํ ์ค๋ช
`AssertionError`

๋ง์ฝ `window_size` ๋ 0๋ณด๋ค ํฌ์ง ์์ต๋๋ค.

`AssertionError`

๋ง์ฝ `dim` ๋ก ๋๋ ์ ์์ต๋๋ค. `num_heads`.

์ ์์ค ์ฝ๋ `ultralytics/models/sam/modules/tiny_encoder.py`
 ```327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380``` ``````def __init__( self, dim, input_resolution, num_heads, window_size=7, mlp_ratio=4.0, drop=0.0, drop_path=0.0, local_conv_size=3, activation=nn.GELU, ): """ Initializes the TinyViTBlock. Args: dim (int): The dimensionality of the input and output. input_resolution (Tuple[int, int]): Spatial resolution of the input feature map. num_heads (int): Number of attention heads. window_size (int, optional): Window size for attention. Default is 7. mlp_ratio (float, optional): Ratio of mlp hidden dim to embedding dim. Default is 4. drop (float, optional): Dropout rate. Default is 0. drop_path (float, optional): Stochastic depth rate. Default is 0. local_conv_size (int, optional): The kernel size of the local convolution. Default is 3. activation (torch.nn, optional): Activation function for MLP. Default is nn.GELU. Raises: AssertionError: If `window_size` is not greater than 0. AssertionError: If `dim` is not divisible by `num_heads`. """ super().__init__() self.dim = dim self.input_resolution = input_resolution self.num_heads = num_heads assert window_size > 0, "window_size must be greater than 0" self.window_size = window_size self.mlp_ratio = mlp_ratio # NOTE: `DropPath` is needed only for training. # self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity() self.drop_path = nn.Identity() assert dim % num_heads == 0, "dim must be divisible by num_heads" head_dim = dim // num_heads window_resolution = (window_size, window_size) self.attn = Attention(dim, head_dim, num_heads, attn_ratio=1, resolution=window_resolution) mlp_hidden_dim = int(dim * mlp_ratio) mlp_activation = activation self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=mlp_activation, drop=drop) pad = local_conv_size // 2 self.local_conv = Conv2d_BN(dim, dim, ks=local_conv_size, stride=1, pad=pad, groups=dim) ``````

### `extra_repr()`

ํฌ๊ธฐ, ์๋ ฅ ํด์๋, ์ดํ์ ํค๋ ์, ์ฐฝ ๋น์จ, ML ๋น์จ ๋ฑ TinyViTBlock์ ํ๋ผ๋ฏธํฐ๋ฅผ ๋ํ๋ด๋ ํ์ํ๋ ๋ฌธ์์ด์ ๋ฐํํฉ๋๋ค. ์ดํ์ ํค๋ ์, ์ฐฝ ํฌ๊ธฐ, MLP ๋น์จ์ ๋ฐํํฉ๋๋ค.

์ ์์ค ์ฝ๋ `ultralytics/models/sam/modules/tiny_encoder.py`
 ```427 428 429 430 431 432 433 434``` ``````def extra_repr(self) -> str: """Returns a formatted string representing the TinyViTBlock's parameters: dimension, input resolution, number of attentions heads, window size, and MLP ratio. """ return ( f"dim={self.dim}, input_resolution={self.input_resolution}, num_heads={self.num_heads}, " f"window_size={self.window_size}, mlp_ratio={self.mlp_ratio}" ) ``````

### `forward(x)`

๋ก์ปฌ ์ปจ๋ณผ๋ฃจ์์ ํต๊ณผํ๊ธฐ ์ ์ ์๋ ฅ 'x'์ ์ฃผ์ ๊ธฐ๋ฐ ๋ณํ ๋๋ ํจ๋ฉ์ ์ ์ฉํฉ๋๋ค. ์ปจ๋ณผ๋ฃจ์์ ๊ฑฐ์นฉ๋๋ค.

์ ์์ค ์ฝ๋ `ultralytics/models/sam/modules/tiny_encoder.py`
 ```382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425``` ``````def forward(self, x): """Applies attention-based transformation or padding to input 'x' before passing it through a local convolution. """ H, W = self.input_resolution B, L, C = x.shape assert L == H * W, "input feature has wrong size" res_x = x if H == self.window_size and W == self.window_size: x = self.attn(x) else: x = x.view(B, H, W, C) pad_b = (self.window_size - H % self.window_size) % self.window_size pad_r = (self.window_size - W % self.window_size) % self.window_size padding = pad_b > 0 or pad_r > 0 if padding: x = F.pad(x, (0, 0, 0, pad_r, 0, pad_b)) pH, pW = H + pad_b, W + pad_r nH = pH // self.window_size nW = pW // self.window_size # Window partition x = ( x.view(B, nH, self.window_size, nW, self.window_size, C) .transpose(2, 3) .reshape(B * nH * nW, self.window_size * self.window_size, C) ) x = self.attn(x) # Window reverse x = x.view(B, nH, nW, self.window_size, self.window_size, C).transpose(2, 3).reshape(B, pH, pW, C) if padding: x = x[:, :H, :W].contiguous() x = x.view(B, L, C) x = res_x + self.drop_path(x) x = x.transpose(1, 2).reshape(B, C, H, W) x = self.local_conv(x) x = x.view(B, C, L).transpose(1, 2) return x + self.drop_path(self.mlp(x)) ``````

## `ultralytics.models.sam.modules.tiny_encoder.BasicLayer`

๊ธฐ์ง: `Module`

TinyViT ์ํคํ์ฒ์ ํ ๋จ๊ณ๋ฅผ ์ํ ๊ธฐ๋ณธ TinyViT ๋ ์ด์ด์๋๋ค.

์ ์์ค ์ฝ๋ `ultralytics/models/sam/modules/tiny_encoder.py`
 ```437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516``` ``````class BasicLayer(nn.Module): """A basic TinyViT layer for one stage in a TinyViT architecture.""" def __init__( self, dim, input_resolution, depth, num_heads, window_size, mlp_ratio=4.0, drop=0.0, drop_path=0.0, downsample=None, use_checkpoint=False, local_conv_size=3, activation=nn.GELU, out_dim=None, ): """ Initializes the BasicLayer. Args: dim (int): The dimensionality of the input and output. input_resolution (Tuple[int, int]): Spatial resolution of the input feature map. depth (int): Number of TinyViT blocks. num_heads (int): Number of attention heads. window_size (int): Local window size. mlp_ratio (float, optional): Ratio of mlp hidden dim to embedding dim. Default is 4. drop (float, optional): Dropout rate. Default is 0. drop_path (float | tuple[float], optional): Stochastic depth rate. Default is 0. downsample (nn.Module | None, optional): Downsample layer at the end of the layer. Default is None. use_checkpoint (bool, optional): Whether to use checkpointing to save memory. Default is False. local_conv_size (int, optional): Kernel size of the local convolution. Default is 3. activation (torch.nn, optional): Activation function for MLP. Default is nn.GELU. out_dim (int | None, optional): The output dimension of the layer. Default is None. Raises: ValueError: If `drop_path` is a list of float but its length doesn't match `depth`. """ super().__init__() self.dim = dim self.input_resolution = input_resolution self.depth = depth self.use_checkpoint = use_checkpoint # Build blocks self.blocks = nn.ModuleList( [ TinyViTBlock( dim=dim, input_resolution=input_resolution, num_heads=num_heads, window_size=window_size, mlp_ratio=mlp_ratio, drop=drop, drop_path=drop_path[i] if isinstance(drop_path, list) else drop_path, local_conv_size=local_conv_size, activation=activation, ) for i in range(depth) ] ) # Patch merging layer self.downsample = ( None if downsample is None else downsample(input_resolution, dim=dim, out_dim=out_dim, activation=activation) ) def forward(self, x): """Performs forward propagation on the input tensor and returns a normalized tensor.""" for blk in self.blocks: x = checkpoint.checkpoint(blk, x) if self.use_checkpoint else blk(x) return x if self.downsample is None else self.downsample(x) def extra_repr(self) -> str: """Returns a string representation of the extra_repr function with the layer's parameters.""" return f"dim={self.dim}, input_resolution={self.input_resolution}, depth={self.depth}" ``````

### `__init__(dim, input_resolution, depth, num_heads, window_size, mlp_ratio=4.0, drop=0.0, drop_path=0.0, downsample=None, use_checkpoint=False, local_conv_size=3, activation=nn.GELU, out_dim=None)`

๋ฒ ์ด์ง ๋ ์ด์ด๋ฅผ ์ด๊ธฐํํฉ๋๋ค.

๋งค๊ฐ๋ณ์:

์ด๋ฆ ์ ํ ์ค๋ช ๊ธฐ๋ณธ๊ฐ
`dim` `int`

์๋ ฅ ๋ฐ ์ถ๋ ฅ์ ์ฐจ์์๋๋ค.

ํ์
`input_resolution` `Tuple[int, int]`

์๋ ฅ ๊ธฐ๋ฅ ๋งต์ ๊ณต๊ฐ ํด์๋์๋๋ค.

ํ์
`depth` `int`

TinyViT ๋ธ๋ก ์์๋๋ค.

ํ์
`num_heads` `int`

๊ด์ฌ ํค๋ ์์๋๋ค.

ํ์
`window_size` `int`

๋ก์ปฌ ์ฐฝ ํฌ๊ธฐ.

ํ์
`mlp_ratio` `float`

MLP ์จ๊น ์ด๋ก๊ฒ์ ์๋ฒ ๋ฉ ์ด๋ก๊ฒ์ ๋น์จ์๋๋ค. ๊ธฐ๋ณธ๊ฐ์ 4์๋๋ค.

`4.0`
`drop` `float`

ํ๋ฝ๋ฅ . ๊ธฐ๋ณธ๊ฐ์ 0์๋๋ค.

`0.0`
`drop_path` `float | tuple[float]`

ํ๋ฅ ์  ๊น์ด ๋น์จ. ๊ธฐ๋ณธ๊ฐ์ 0์๋๋ค.

`0.0`
`downsample` `Module | None`

๋ ์ด์ด ๋์์ ๋ ์ด์ด๋ฅผ ๋ค์ด์ํ๋งํฉ๋๋ค. ๊ธฐ๋ณธ๊ฐ์ ์์์๋๋ค.

`None`
`use_checkpoint` `bool`

๋ฉ๋ชจ๋ฆฌ ์ ์ฝ์ ์ํด ์ฒดํฌํฌ์ธํธ๋ฅผ ์ฌ์ฉํ ์ง ์ฌ๋ถ์๋๋ค. ๊ธฐ๋ณธ๊ฐ์ False์๋๋ค.

`False`
`local_conv_size` `int`

๋ก์ปฌ ์ปจ๋ณผ๋ฃจ์์ ์ปค๋ ํฌ๊ธฐ์๋๋ค. ๊ธฐ๋ณธ๊ฐ์ 3์๋๋ค.

`3`
`activation` `nn`

MLP์ฉ ํ์ฑํ ๊ธฐ๋ฅ. ๊ธฐ๋ณธ๊ฐ์ nn.GELU์๋๋ค.

`GELU`
`out_dim` `int | None`

๋ ์ด์ด์ ์ถ๋ ฅ ์น์์๋๋ค. ๊ธฐ๋ณธ๊ฐ์ ์์์๋๋ค.

`None`

์ฌ๋ฆฌ๊ธฐ:

์ ํ ์ค๋ช
`ValueError`

๋ง์ฝ `drop_path` ๋ ํ๋กํธ ๋ชฉ๋ก์ด์ง๋ง ๊ธธ์ด๊ฐ ์ผ์นํ์ง ์์ต๋๋ค. `depth`.

์ ์์ค ์ฝ๋ `ultralytics/models/sam/modules/tiny_encoder.py`
 ```440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506``` ``````def __init__( self, dim, input_resolution, depth, num_heads, window_size, mlp_ratio=4.0, drop=0.0, drop_path=0.0, downsample=None, use_checkpoint=False, local_conv_size=3, activation=nn.GELU, out_dim=None, ): """ Initializes the BasicLayer. Args: dim (int): The dimensionality of the input and output. input_resolution (Tuple[int, int]): Spatial resolution of the input feature map. depth (int): Number of TinyViT blocks. num_heads (int): Number of attention heads. window_size (int): Local window size. mlp_ratio (float, optional): Ratio of mlp hidden dim to embedding dim. Default is 4. drop (float, optional): Dropout rate. Default is 0. drop_path (float | tuple[float], optional): Stochastic depth rate. Default is 0. downsample (nn.Module | None, optional): Downsample layer at the end of the layer. Default is None. use_checkpoint (bool, optional): Whether to use checkpointing to save memory. Default is False. local_conv_size (int, optional): Kernel size of the local convolution. Default is 3. activation (torch.nn, optional): Activation function for MLP. Default is nn.GELU. out_dim (int | None, optional): The output dimension of the layer. Default is None. Raises: ValueError: If `drop_path` is a list of float but its length doesn't match `depth`. """ super().__init__() self.dim = dim self.input_resolution = input_resolution self.depth = depth self.use_checkpoint = use_checkpoint # Build blocks self.blocks = nn.ModuleList( [ TinyViTBlock( dim=dim, input_resolution=input_resolution, num_heads=num_heads, window_size=window_size, mlp_ratio=mlp_ratio, drop=drop, drop_path=drop_path[i] if isinstance(drop_path, list) else drop_path, local_conv_size=local_conv_size, activation=activation, ) for i in range(depth) ] ) # Patch merging layer self.downsample = ( None if downsample is None else downsample(input_resolution, dim=dim, out_dim=out_dim, activation=activation) ) ``````

### `extra_repr()`

๋ ์ด์ด์ ๋งค๊ฐ๋ณ์์ ํจ๊ป extra_repr ํจ์์ ๋ฌธ์์ด ํํ์ ๋ฐํํฉ๋๋ค.

์ ์์ค ์ฝ๋ `ultralytics/models/sam/modules/tiny_encoder.py`
 ```514 515 516``` ``````def extra_repr(self) -> str: """Returns a string representation of the extra_repr function with the layer's parameters.""" return f"dim={self.dim}, input_resolution={self.input_resolution}, depth={self.depth}" ``````

### `forward(x)`

์๋ ฅ tensor ์ ์๋ฐฉํฅ ์ ํ๋ฅผ ์ํํ์ฌ ์ ๊ทํ๋ tensor ์ ๋ฐํํฉ๋๋ค.

์ ์์ค ์ฝ๋ `ultralytics/models/sam/modules/tiny_encoder.py`
 ```508 509 510 511 512``` ``````def forward(self, x): """Performs forward propagation on the input tensor and returns a normalized tensor.""" for blk in self.blocks: x = checkpoint.checkpoint(blk, x) if self.use_checkpoint else blk(x) return x if self.downsample is None else self.downsample(x) ``````

## `ultralytics.models.sam.modules.tiny_encoder.LayerNorm2d`

๊ธฐ์ง: `Module`

2D์์ ๋ ์ด์ด ๋ธ๋ฉ๋ผ์ด์ ์ด์์ PyTorch ๊ตฌํ์๋๋ค.

์ ์์ค ์ฝ๋ `ultralytics/models/sam/modules/tiny_encoder.py`
 ```519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534``` ``````class LayerNorm2d(nn.Module): """A PyTorch implementation of Layer Normalization in 2D.""" def __init__(self, num_channels: int, eps: float = 1e-6) -> None: """Initialize LayerNorm2d with the number of channels and an optional epsilon.""" super().__init__() self.weight = nn.Parameter(torch.ones(num_channels)) self.bias = nn.Parameter(torch.zeros(num_channels)) self.eps = eps def forward(self, x: torch.Tensor) -> torch.Tensor: """Perform a forward pass, normalizing the input tensor.""" u = x.mean(1, keepdim=True) s = (x - u).pow(2).mean(1, keepdim=True) x = (x - u) / torch.sqrt(s + self.eps) return self.weight[:, None, None] * x + self.bias[:, None, None] ``````

### `__init__(num_channels, eps=1e-06)`

์ฑ๋ ์์ ์ก์ค๋ก (์ ํ ์ฌํญ)์ผ๋ก LayerNorm2d๋ฅผ ์ด๊ธฐํํฉ๋๋ค.

์ ์์ค ์ฝ๋ `ultralytics/models/sam/modules/tiny_encoder.py`
 ```522 523 524 525 526 527``` ``````def __init__(self, num_channels: int, eps: float = 1e-6) -> None: """Initialize LayerNorm2d with the number of channels and an optional epsilon.""" super().__init__() self.weight = nn.Parameter(torch.ones(num_channels)) self.bias = nn.Parameter(torch.zeros(num_channels)) self.eps = eps ``````

### `forward(x)`

ํฌ์๋ ํจ์ค๋ฅผ ์ํํ์ฌ ์๋ ฅ์ ์ ๊ทํํฉ๋๋ค tensor.

์ ์์ค ์ฝ๋ `ultralytics/models/sam/modules/tiny_encoder.py`
 ```529 530 531 532 533 534``` ``````def forward(self, x: torch.Tensor) -> torch.Tensor: """Perform a forward pass, normalizing the input tensor.""" u = x.mean(1, keepdim=True) s = (x - u).pow(2).mean(1, keepdim=True) x = (x - u) / torch.sqrt(s + self.eps) return self.weight[:, None, None] * x + self.bias[:, None, None] ``````

## `ultralytics.models.sam.modules.tiny_encoder.TinyViT`

๊ธฐ์ง: `Module`

๋น์  ์์์ ์ํ TinyViT ์ํคํ์ฒ.

์์ฑ:

์ด๋ฆ ์ ํ ์ค๋ช
`img_size` `int`

์ด๋ฏธ์ง ํฌ๊ธฐ๋ฅผ ์๋ ฅํฉ๋๋ค.

`in_chans` `int`

์๋ ฅ ์ฑ๋ ์์๋๋ค.

`num_classes` `int`

๋ถ๋ฅ ํด๋์ค ์์๋๋ค.

`embed_dims` `List[int]`

๊ฐ ๋ ์ด์ด์ ๋ํ ์๋ฒ ๋ฉ ์น์ ๋ชฉ๋ก์๋๋ค.

`depths` `List[int]`

๊ฐ ๋ ์ด์ด์ ๊น์ด ๋ชฉ๋ก์๋๋ค.

`num_heads` `List[int]`

๊ฐ ๋ ์ด์ด์ ๋ํ ๊ด์ฌ ํค๋ ์ ๋ชฉ๋ก์๋๋ค.

`window_sizes` `List[int]`

๊ฐ ๋ ์ด์ด์ ์ฐฝ ํฌ๊ธฐ ๋ชฉ๋ก์๋๋ค.

`mlp_ratio` `float`

์๋ฒ ๋ฉ ์ฐจ์์ ๋ํ MLP ์จ๊ฒจ์ง ์ฐจ์์ ๋น์จ์๋๋ค.

`drop_rate` `float`

๋๋กญ ๋ ์ด์ด์ ๋๋กญ๋ฅ ์๋๋ค.

`drop_path_rate` `float`

ํ๋ฅ ์  ๊น์ด์ ๋ํ ๋๋กญ ๊ฒฝ๋ก ์๋์๋๋ค.

`use_checkpoint` `bool`

ํจ์จ์ ์ธ ๋ฉ๋ชจ๋ฆฌ ์ฌ์ฉ์ ์ํด ์ฒดํฌํฌ์ธํธ๋ฅผ ์ฌ์ฉํ์ธ์.

`mbconv_expand_ratio` `float`

MBConv ๋ ์ด์ด์ ํ์ฅ ๋น์จ์๋๋ค.

`local_conv_size` `int`

๋ก์ปฌ ์ปจ๋ณผ๋ฃจ์ ์ปค๋ ํฌ๊ธฐ์๋๋ค.

`layer_lr_decay` `float`

๋ ์ด์ด๋ณ ํ์ต๋ฅ  ๊ฐ์ .

์ฐธ๊ณ

์ด ๊ตฌํ์ ๊น์ด, ๊ด์ฌ ํค๋ ๋ชฉ๋ก์ ๋ฐ์๋ค์ด๋๋ก ์ผ๋ฐํ๋์์ต๋๋ค, ์๋ฒ ๋ฉ ์น์ ๋ฐ ์ฐฝ ํฌ๊ธฐ ๋ชฉ๋ก์ ์์ฉํ๋๋ก ์ผ๋ฐํํ์ฌ ๋ค์ํ ๊ตฌ์ฑ์ TinyViT ๋ชจ๋ธ์ "์คํ"์ ๋ง๋ค ์ ์์ต๋๋ค.

์ ์์ค ์ฝ๋ `ultralytics/models/sam/modules/tiny_encoder.py`
 ```537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742``` ``````class TinyViT(nn.Module): """ The TinyViT architecture for vision tasks. Attributes: img_size (int): Input image size. in_chans (int): Number of input channels. num_classes (int): Number of classification classes. embed_dims (List[int]): List of embedding dimensions for each layer. depths (List[int]): List of depths for each layer. num_heads (List[int]): List of number of attention heads for each layer. window_sizes (List[int]): List of window sizes for each layer. mlp_ratio (float): Ratio of MLP hidden dimension to embedding dimension. drop_rate (float): Dropout rate for drop layers. drop_path_rate (float): Drop path rate for stochastic depth. use_checkpoint (bool): Use checkpointing for efficient memory usage. mbconv_expand_ratio (float): Expansion ratio for MBConv layer. local_conv_size (int): Local convolution kernel size. layer_lr_decay (float): Layer-wise learning rate decay. Note: This implementation is generalized to accept a list of depths, attention heads, embedding dimensions and window sizes, which allows you to create a "stack" of TinyViT models of varying configurations. """ def __init__( self, img_size=224, in_chans=3, num_classes=1000, embed_dims=[96, 192, 384, 768], depths=[2, 2, 6, 2], num_heads=[3, 6, 12, 24], window_sizes=[7, 7, 14, 7], mlp_ratio=4.0, drop_rate=0.0, drop_path_rate=0.1, use_checkpoint=False, mbconv_expand_ratio=4.0, local_conv_size=3, layer_lr_decay=1.0, ): """ Initializes the TinyViT model. Args: img_size (int, optional): The input image size. Defaults to 224. in_chans (int, optional): Number of input channels. Defaults to 3. num_classes (int, optional): Number of classification classes. Defaults to 1000. embed_dims (List[int], optional): List of embedding dimensions for each layer. Defaults to [96, 192, 384, 768]. depths (List[int], optional): List of depths for each layer. Defaults to [2, 2, 6, 2]. num_heads (List[int], optional): List of number of attention heads for each layer. Defaults to [3, 6, 12, 24]. window_sizes (List[int], optional): List of window sizes for each layer. Defaults to [7, 7, 14, 7]. mlp_ratio (float, optional): Ratio of MLP hidden dimension to embedding dimension. Defaults to 4. drop_rate (float, optional): Dropout rate. Defaults to 0. drop_path_rate (float, optional): Drop path rate for stochastic depth. Defaults to 0.1. use_checkpoint (bool, optional): Whether to use checkpointing for efficient memory usage. Defaults to False. mbconv_expand_ratio (float, optional): Expansion ratio for MBConv layer. Defaults to 4.0. local_conv_size (int, optional): Local convolution kernel size. Defaults to 3. layer_lr_decay (float, optional): Layer-wise learning rate decay. Defaults to 1.0. """ super().__init__() self.img_size = img_size self.num_classes = num_classes self.depths = depths self.num_layers = len(depths) self.mlp_ratio = mlp_ratio activation = nn.GELU self.patch_embed = PatchEmbed( in_chans=in_chans, embed_dim=embed_dims[0], resolution=img_size, activation=activation ) patches_resolution = self.patch_embed.patches_resolution self.patches_resolution = patches_resolution # Stochastic depth dpr = [x.item() for x in torch.linspace(0, drop_path_rate, sum(depths))] # stochastic depth decay rule # Build layers self.layers = nn.ModuleList() for i_layer in range(self.num_layers): kwargs = dict( dim=embed_dims[i_layer], input_resolution=( patches_resolution[0] // (2 ** (i_layer - 1 if i_layer == 3 else i_layer)), patches_resolution[1] // (2 ** (i_layer - 1 if i_layer == 3 else i_layer)), ), # input_resolution=(patches_resolution[0] // (2 ** i_layer), # patches_resolution[1] // (2 ** i_layer)), depth=depths[i_layer], drop_path=dpr[sum(depths[:i_layer]) : sum(depths[: i_layer + 1])], downsample=PatchMerging if (i_layer < self.num_layers - 1) else None, use_checkpoint=use_checkpoint, out_dim=embed_dims[min(i_layer + 1, len(embed_dims) - 1)], activation=activation, ) if i_layer == 0: layer = ConvLayer(conv_expand_ratio=mbconv_expand_ratio, **kwargs) else: layer = BasicLayer( num_heads=num_heads[i_layer], window_size=window_sizes[i_layer], mlp_ratio=self.mlp_ratio, drop=drop_rate, local_conv_size=local_conv_size, **kwargs, ) self.layers.append(layer) # Classifier head self.norm_head = nn.LayerNorm(embed_dims[-1]) self.head = nn.Linear(embed_dims[-1], num_classes) if num_classes > 0 else torch.nn.Identity() # Init weights self.apply(self._init_weights) self.set_layer_lr_decay(layer_lr_decay) self.neck = nn.Sequential( nn.Conv2d( embed_dims[-1], 256, kernel_size=1, bias=False, ), LayerNorm2d(256), nn.Conv2d( 256, 256, kernel_size=3, padding=1, bias=False, ), LayerNorm2d(256), ) def set_layer_lr_decay(self, layer_lr_decay): """Sets the learning rate decay for each layer in the TinyViT model.""" decay_rate = layer_lr_decay # Layers -> blocks (depth) depth = sum(self.depths) lr_scales = [decay_rate ** (depth - i - 1) for i in range(depth)] def _set_lr_scale(m, scale): """Sets the learning rate scale for each layer in the model based on the layer's depth.""" for p in m.parameters(): p.lr_scale = scale self.patch_embed.apply(lambda x: _set_lr_scale(x, lr_scales[0])) i = 0 for layer in self.layers: for block in layer.blocks: block.apply(lambda x: _set_lr_scale(x, lr_scales[i])) i += 1 if layer.downsample is not None: layer.downsample.apply(lambda x: _set_lr_scale(x, lr_scales[i - 1])) assert i == depth for m in [self.norm_head, self.head]: m.apply(lambda x: _set_lr_scale(x, lr_scales[-1])) for k, p in self.named_parameters(): p.param_name = k def _check_lr_scale(m): """Checks if the learning rate scale attribute is present in module's parameters.""" for p in m.parameters(): assert hasattr(p, "lr_scale"), p.param_name self.apply(_check_lr_scale) def _init_weights(self, m): """Initializes weights for linear layers and layer normalization in the given module.""" if isinstance(m, nn.Linear): # NOTE: This initialization is needed only for training. # trunc_normal_(m.weight, std=.02) if 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) @torch.jit.ignore def no_weight_decay_keywords(self): """Returns a dictionary of parameter names where weight decay should not be applied.""" return {"attention_biases"} def forward_features(self, x): """Runs the input through the model layers and returns the transformed output.""" x = self.patch_embed(x) # x input is (N, C, H, W) x = self.layers[0](x) start_i = 1 for i in range(start_i, len(self.layers)): layer = self.layers[i] x = layer(x) B, _, C = x.shape x = x.view(B, 64, 64, C) x = x.permute(0, 3, 1, 2) return self.neck(x) def forward(self, x): """Executes a forward pass on the input tensor through the constructed model layers.""" return self.forward_features(x) ``````

### `__init__(img_size=224, in_chans=3, num_classes=1000, embed_dims=[96, 192, 384, 768], depths=[2, 2, 6, 2], num_heads=[3, 6, 12, 24], window_sizes=[7, 7, 14, 7], mlp_ratio=4.0, drop_rate=0.0, drop_path_rate=0.1, use_checkpoint=False, mbconv_expand_ratio=4.0, local_conv_size=3, layer_lr_decay=1.0)`

TinyViT ๋ชจ๋ธ์ ์ด๊ธฐํํฉ๋๋ค.

๋งค๊ฐ๋ณ์:

์ด๋ฆ ์ ํ ์ค๋ช ๊ธฐ๋ณธ๊ฐ
`img_size` `int`

์๋ ฅ ์ด๋ฏธ์ง ํฌ๊ธฐ์๋๋ค. ๊ธฐ๋ณธ๊ฐ์ 224์๋๋ค.

`224`
`in_chans` `int`

์๋ ฅ ์ฑ๋ ์์๋๋ค. ๊ธฐ๋ณธ๊ฐ์ 3์๋๋ค.

`3`
`num_classes` `int`

๋ถ๋ฅ ํด๋์ค ์์๋๋ค. ๊ธฐ๋ณธ๊ฐ์ 1000์๋๋ค.

`1000`
`embed_dims` `List[int]`

๊ฐ ๋ ์ด์ด์ ์๋ฒ ๋ฉ ์น์ ๋ชฉ๋ก์๋๋ค. ๊ธฐ๋ณธ๊ฐ์ [96, 192, 384, 768]์๋๋ค.

`[96, 192, 384, 768]`
`depths` `List[int]`

๊ฐ ๋ ์ด์ด์ ๊น์ด ๋ชฉ๋ก์๋๋ค. ๊ธฐ๋ณธ๊ฐ์ [2, 2, 6, 2]์๋๋ค.

`[2, 2, 6, 2]`
`num_heads` `List[int]`

๊ฐ ๋ ์ด์ด์ ๋ํ ๊ด์ฌ ํค๋ ์ ๋ชฉ๋ก์๋๋ค. ๊ธฐ๋ณธ๊ฐ์ [3, 6, 12, 24]์๋๋ค.

`[3, 6, 12, 24]`
`window_sizes` `List[int]`

๊ฐ ๋ ์ด์ด์ ์ฐฝ ํฌ๊ธฐ ๋ชฉ๋ก์๋๋ค. ๊ธฐ๋ณธ๊ฐ์ [7, 7, 14, 7]์๋๋ค.

`[7, 7, 14, 7]`
`mlp_ratio` `float`

์๋ฒ ๋ฉ ์ฐจ์์ ๋ํ MLP ์จ๊ฒจ์ง ์ฐจ์์ ๋น์จ์๋๋ค. ๊ธฐ๋ณธ๊ฐ์ 4์๋๋ค.

`4.0`
`drop_rate` `float`

ํ๋ฝ๋ฅ . ๊ธฐ๋ณธ๊ฐ์ 0์๋๋ค.

`0.0`
`drop_path_rate` `float`

ํ๋ฅ ์  ๊น์ด์ ๋ํ ๋๋กญ ๊ฒฝ๋ก ์๋์๋๋ค. ๊ธฐ๋ณธ๊ฐ์ 0.1์๋๋ค.

`0.1`
`use_checkpoint` `bool`

ํจ์จ์ ์ธ ๋ฉ๋ชจ๋ฆฌ ์ฌ์ฉ์ ์ํด ์ฒดํฌํฌ์ธํธ๋ฅผ ์ฌ์ฉํ ์ง ์ฌ๋ถ์๋๋ค. ๊ธฐ๋ณธ๊ฐ์ False์๋๋ค.

`False`
`mbconv_expand_ratio` `float`

MBConv ๋ ์ด์ด์ ํ์ฅ ๋น์จ์๋๋ค. ๊ธฐ๋ณธ๊ฐ์ 4.0์๋๋ค.

`4.0`
`local_conv_size` `int`

๋ก์ปฌ ์ปจ๋ณผ๋ฃจ์ ์ปค๋ ํฌ๊ธฐ. ๊ธฐ๋ณธ๊ฐ์ 3์๋๋ค.

`3`
`layer_lr_decay` `float`

๋ ์ด์ด๋ณ ํ์ต๋ฅ  ๊ฐ์ . ๊ธฐ๋ณธ๊ฐ์ 1.0์๋๋ค.

`1.0`
์ ์์ค ์ฝ๋ `ultralytics/models/sam/modules/tiny_encoder.py`
 ```563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672``` ``````def __init__( self, img_size=224, in_chans=3, num_classes=1000, embed_dims=[96, 192, 384, 768], depths=[2, 2, 6, 2], num_heads=[3, 6, 12, 24], window_sizes=[7, 7, 14, 7], mlp_ratio=4.0, drop_rate=0.0, drop_path_rate=0.1, use_checkpoint=False, mbconv_expand_ratio=4.0, local_conv_size=3, layer_lr_decay=1.0, ): """ Initializes the TinyViT model. Args: img_size (int, optional): The input image size. Defaults to 224. in_chans (int, optional): Number of input channels. Defaults to 3. num_classes (int, optional): Number of classification classes. Defaults to 1000. embed_dims (List[int], optional): List of embedding dimensions for each layer. Defaults to [96, 192, 384, 768]. depths (List[int], optional): List of depths for each layer. Defaults to [2, 2, 6, 2]. num_heads (List[int], optional): List of number of attention heads for each layer. Defaults to [3, 6, 12, 24]. window_sizes (List[int], optional): List of window sizes for each layer. Defaults to [7, 7, 14, 7]. mlp_ratio (float, optional): Ratio of MLP hidden dimension to embedding dimension. Defaults to 4. drop_rate (float, optional): Dropout rate. Defaults to 0. drop_path_rate (float, optional): Drop path rate for stochastic depth. Defaults to 0.1. use_checkpoint (bool, optional): Whether to use checkpointing for efficient memory usage. Defaults to False. mbconv_expand_ratio (float, optional): Expansion ratio for MBConv layer. Defaults to 4.0. local_conv_size (int, optional): Local convolution kernel size. Defaults to 3. layer_lr_decay (float, optional): Layer-wise learning rate decay. Defaults to 1.0. """ super().__init__() self.img_size = img_size self.num_classes = num_classes self.depths = depths self.num_layers = len(depths) self.mlp_ratio = mlp_ratio activation = nn.GELU self.patch_embed = PatchEmbed( in_chans=in_chans, embed_dim=embed_dims[0], resolution=img_size, activation=activation ) patches_resolution = self.patch_embed.patches_resolution self.patches_resolution = patches_resolution # Stochastic depth dpr = [x.item() for x in torch.linspace(0, drop_path_rate, sum(depths))] # stochastic depth decay rule # Build layers self.layers = nn.ModuleList() for i_layer in range(self.num_layers): kwargs = dict( dim=embed_dims[i_layer], input_resolution=( patches_resolution[0] // (2 ** (i_layer - 1 if i_layer == 3 else i_layer)), patches_resolution[1] // (2 ** (i_layer - 1 if i_layer == 3 else i_layer)), ), # input_resolution=(patches_resolution[0] // (2 ** i_layer), # patches_resolution[1] // (2 ** i_layer)), depth=depths[i_layer], drop_path=dpr[sum(depths[:i_layer]) : sum(depths[: i_layer + 1])], downsample=PatchMerging if (i_layer < self.num_layers - 1) else None, use_checkpoint=use_checkpoint, out_dim=embed_dims[min(i_layer + 1, len(embed_dims) - 1)], activation=activation, ) if i_layer == 0: layer = ConvLayer(conv_expand_ratio=mbconv_expand_ratio, **kwargs) else: layer = BasicLayer( num_heads=num_heads[i_layer], window_size=window_sizes[i_layer], mlp_ratio=self.mlp_ratio, drop=drop_rate, local_conv_size=local_conv_size, **kwargs, ) self.layers.append(layer) # Classifier head self.norm_head = nn.LayerNorm(embed_dims[-1]) self.head = nn.Linear(embed_dims[-1], num_classes) if num_classes > 0 else torch.nn.Identity() # Init weights self.apply(self._init_weights) self.set_layer_lr_decay(layer_lr_decay) self.neck = nn.Sequential( nn.Conv2d( embed_dims[-1], 256, kernel_size=1, bias=False, ), LayerNorm2d(256), nn.Conv2d( 256, 256, kernel_size=3, padding=1, bias=False, ), LayerNorm2d(256), ) ``````

### `forward(x)`

์๋ ฅ tensor ์ ๋ํด ๊ตฌ์ฑ๋ ๋ชจ๋ธ ๋ ์ด์ด๋ฅผ ํตํด ํฌ์๋ ํจ์ค๋ฅผ ์คํํฉ๋๋ค.

์ ์์ค ์ฝ๋ `ultralytics/models/sam/modules/tiny_encoder.py`
 ```740 741 742``` ``````def forward(self, x): """Executes a forward pass on the input tensor through the constructed model layers.""" return self.forward_features(x) ``````

### `forward_features(x)`

๋ชจ๋ธ ๋ ์ด์ด๋ฅผ ํตํด ์๋ ฅ์ ์คํํ๊ณ  ๋ณํ๋ ์ถ๋ ฅ์ ๋ฐํํฉ๋๋ค.

์ ์์ค ์ฝ๋ `ultralytics/models/sam/modules/tiny_encoder.py`
 ```725 726 727 728 729 730 731 732 733 734 735 736 737 738``` ``````def forward_features(self, x): """Runs the input through the model layers and returns the transformed output.""" x = self.patch_embed(x) # x input is (N, C, H, W) x = self.layers[0](x) start_i = 1 for i in range(start_i, len(self.layers)): layer = self.layers[i] x = layer(x) B, _, C = x.shape x = x.view(B, 64, 64, C) x = x.permute(0, 3, 1, 2) return self.neck(x) ``````

### `no_weight_decay_keywords()`

๊ฐ์ค์น ๊ฐ์ ๊ฐ ์ ์ฉ๋์ง ์์์ผ ํ๋ ํ๋ผ๋ฏธํฐ ์ด๋ฆ์ ๋์๋๋ฆฌ๋ฅผ ๋ฐํํฉ๋๋ค.

์ ์์ค ์ฝ๋ `ultralytics/models/sam/modules/tiny_encoder.py`
 ```720 721 722 723``` ``````@torch.jit.ignore def no_weight_decay_keywords(self): """Returns a dictionary of parameter names where weight decay should not be applied.""" return {"attention_biases"} ``````

### `set_layer_lr_decay(layer_lr_decay)`

TinyViT ๋ชจ๋ธ์ ๊ฐ ๋ ์ด์ด์ ๋ํ ํ์ต๋ฅ  ๊ฐ์ ๋ฅผ ์ค์ ํฉ๋๋ค.

์ ์์ค ์ฝ๋ `ultralytics/models/sam/modules/tiny_encoder.py`
 ```674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707``` ``````def set_layer_lr_decay(self, layer_lr_decay): """Sets the learning rate decay for each layer in the TinyViT model.""" decay_rate = layer_lr_decay # Layers -> blocks (depth) depth = sum(self.depths) lr_scales = [decay_rate ** (depth - i - 1) for i in range(depth)] def _set_lr_scale(m, scale): """Sets the learning rate scale for each layer in the model based on the layer's depth.""" for p in m.parameters(): p.lr_scale = scale self.patch_embed.apply(lambda x: _set_lr_scale(x, lr_scales[0])) i = 0 for layer in self.layers: for block in layer.blocks: block.apply(lambda x: _set_lr_scale(x, lr_scales[i])) i += 1 if layer.downsample is not None: layer.downsample.apply(lambda x: _set_lr_scale(x, lr_scales[i - 1])) assert i == depth for m in [self.norm_head, self.head]: m.apply(lambda x: _set_lr_scale(x, lr_scales[-1])) for k, p in self.named_parameters(): p.param_name = k def _check_lr_scale(m): """Checks if the learning rate scale attribute is present in module's parameters.""" for p in m.parameters(): assert hasattr(p, "lr_scale"), p.param_name self.apply(_check_lr_scale) ``````

2023-11-12 ์์ฑ, 2023-11-25 ์๋ฐ์ดํธ๋จ
์์ฑ์: glenn-jocher (3), Laughing-q (1)