File size: 10,830 Bytes
685ecb2 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 |
from typing import Optional, Tuple
import torch
import torch.nn as nn
class SiglipVisionConfig:
def __init__(
self,
hidden_size=768,
intermediate_size=3072,
num_hidden_layers=12,
num_attention_heads=12,
num_channels=3,
image_size=224,
patch_size=16,
layer_norm_eps=1e-6,
attention_dropout=0.0,
num_image_tokens: int = None,
**kwargs,
):
super().__init__()
self.hidden_size = hidden_size
self.intermediate_size = intermediate_size
self.num_hidden_layers = num_hidden_layers
self.num_attention_heads = num_attention_heads
self.num_channels = num_channels
self.patch_size = patch_size
self.image_size = image_size
self.attention_dropout = attention_dropout
self.layer_norm_eps = layer_norm_eps
self.num_image_tokens = num_image_tokens
class SiglipVisionEmbeddings(nn.Module):
def __init__(self, config: SiglipVisionConfig):
super().__init__()
self.config = config
self.embed_dim = config.hidden_size
self.image_size = config.image_size
self.patch_size = config.patch_size
self.patch_embedding = nn.Conv2d(
in_channels=config.num_channels,
out_channels=self.embed_dim,
kernel_size=self.patch_size,
stride=self.patch_size,
padding="valid", # This indicates no padding is added
)
self.num_patches = (self.image_size // self.patch_size) ** 2
self.num_positions = self.num_patches
self.position_embedding = nn.Embedding(self.num_positions, self.embed_dim)
self.register_buffer(
"position_ids",
torch.arange(self.num_positions).expand((1, -1)),
persistent=False,
)
def forward(self, pixel_values: torch.FloatTensor) -> torch.Tensor:
_, _, height, width = (
pixel_values.shape
) # [Batch_Size, Channels, Height, Width]
# Convolve the `patch_size` kernel over the image, with no overlapping patches since the stride is equal to the kernel size
# The output of the convolution will have shape [Batch_Size, Embed_Dim, Num_Patches_H, Num_Patches_W]
# where Num_Patches_H = height // patch_size and Num_Patches_W = width // patch_size
patch_embeds = self.patch_embedding(pixel_values)
# [Batch_Size, Embed_Dim, Num_Patches_H, Num_Patches_W] -> [Batch_Size, Embed_Dim, Num_Patches]
# where Num_Patches = Num_Patches_H * Num_Patches_W
embeddings = patch_embeds.flatten(2)
# [Batch_Size, Embed_Dim, Num_Patches] -> [Batch_Size, Num_Patches, Embed_Dim]
embeddings = embeddings.transpose(1, 2)
# Add position embeddings to each patch. Each positional encoding is a vector of size [Embed_Dim]
embeddings = embeddings + self.position_embedding(self.position_ids)
# [Batch_Size, Num_Patches, Embed_Dim]
return embeddings
class SiglipAttention(nn.Module):
"""Multi-headed attention from 'Attention Is All You Need' paper"""
def __init__(self, config):
super().__init__()
self.config = config
self.embed_dim = config.hidden_size
self.num_heads = config.num_attention_heads
self.head_dim = self.embed_dim // self.num_heads
self.scale = self.head_dim**-0.5 # Equivalent to 1 / sqrt(self.head_dim)
self.dropout = config.attention_dropout
self.k_proj = nn.Linear(self.embed_dim, self.embed_dim)
self.v_proj = nn.Linear(self.embed_dim, self.embed_dim)
self.q_proj = nn.Linear(self.embed_dim, self.embed_dim)
self.out_proj = nn.Linear(self.embed_dim, self.embed_dim)
def forward(
self,
hidden_states: torch.Tensor,
) -> Tuple[torch.Tensor, Optional[torch.Tensor]]:
# hidden_states: [Batch_Size, Num_Patches, Embed_Dim]
batch_size, seq_len, _ = hidden_states.size()
# query_states: [Batch_Size, Num_Patches, Embed_Dim]
query_states = self.q_proj(hidden_states)
# key_states: [Batch_Size, Num_Patches, Embed_Dim]
key_states = self.k_proj(hidden_states)
# value_states: [Batch_Size, Num_Patches, Embed_Dim]
value_states = self.v_proj(hidden_states)
# query_states: [Batch_Size, Num_Heads, Num_Patches, Head_Dim]
query_states = query_states.view(
batch_size, seq_len, self.num_heads, self.head_dim
).transpose(1, 2)
key_states = key_states.view(
batch_size, seq_len, self.num_heads, self.head_dim
).transpose(1, 2)
value_states = value_states.view(
batch_size, seq_len, self.num_heads, self.head_dim
).transpose(1, 2)
# Calculate the attention using the formula Q * K^T / sqrt(d_k). attn_weights: [Batch_Size, Num_Heads, Num_Patches, Num_Patches]
attn_weights = (
torch.matmul(query_states, key_states.transpose(2, 3)) * self.scale
)
if attn_weights.size() != (batch_size, self.num_heads, seq_len, seq_len):
raise ValueError(
f"Attention weights should be of size {(batch_size, self.num_heads, seq_len, seq_len)}, but is"
f" {attn_weights.size()}"
)
# Apply the softmax row-wise. attn_weights: [Batch_Size, Num_Heads, Num_Patches, Num_Patches]
attn_weights = nn.functional.softmax(
attn_weights, dim=-1, dtype=torch.float32
).to(query_states.dtype)
# Apply dropout only during training
attn_weights = nn.functional.dropout(
attn_weights, p=self.dropout, training=self.training
)
# Multiply the attention weights by the value states. attn_output: [Batch_Size, Num_Heads, Num_Patches, Head_Dim]
attn_output = torch.matmul(attn_weights, value_states)
if attn_output.size() != (batch_size, self.num_heads, seq_len, self.head_dim):
raise ValueError(
f"`attn_output` should be of size {(batch_size, self.num_heads, seq_len, self.head_dim)}, but is"
f" {attn_output.size()}"
)
# [Batch_Size, Num_Heads, Num_Patches, Head_Dim] -> [Batch_Size, Num_Patches, Num_Heads, Head_Dim]
attn_output = attn_output.transpose(1, 2).contiguous()
# [Batch_Size, Num_Patches, Num_Heads, Head_Dim] -> [Batch_Size, Num_Patches, Embed_Dim]
attn_output = attn_output.reshape(batch_size, seq_len, self.embed_dim)
# [Batch_Size, Num_Patches, Embed_Dim]
attn_output = self.out_proj(attn_output)
return attn_output, attn_weights
class SiglipMLP(nn.Module):
def __init__(self, config):
super().__init__()
self.config = config
self.fc1 = nn.Linear(config.hidden_size, config.intermediate_size)
self.fc2 = nn.Linear(config.intermediate_size, config.hidden_size)
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
# [Batch_Size, Num_Patches, Embed_Dim] -> [Batch_Size, Num_Patches, Intermediate_Size]
hidden_states = self.fc1(hidden_states)
# hidden_states: [Batch_Size, Num_Patches, Intermediate_Size]
hidden_states = nn.functional.gelu(hidden_states, approximate="tanh")
# [Batch_Size, Num_Patches, Intermediate_Size] -> [Batch_Size, Num_Patches, Embed_Dim]
hidden_states = self.fc2(hidden_states)
return hidden_states
class SiglipEncoderLayer(nn.Module):
def __init__(self, config: SiglipVisionConfig):
super().__init__()
self.embed_dim = config.hidden_size
self.self_attn = SiglipAttention(config)
self.layer_norm1 = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_eps)
self.mlp = SiglipMLP(config)
self.layer_norm2 = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_eps)
# Ignore copy
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
# residual: [Batch_Size, Num_Patches, Embed_Dim]
residual = hidden_states
# [Batch_Size, Num_Patches, Embed_Dim] -> [Batch_Size, Num_Patches, Embed_Dim]
hidden_states = self.layer_norm1(hidden_states)
# [Batch_Size, Num_Patches, Embed_Dim] -> [Batch_Size, Num_Patches, Embed_Dim]
hidden_states, _ = self.self_attn(hidden_states=hidden_states)
# [Batch_Size, Num_Patches, Embed_Dim]
hidden_states = residual + hidden_states
# residual: [Batch_Size, Num_Patches, Embed_Dim]
residual = hidden_states
# [Batch_Size, Num_Patches, Embed_Dim] -> [Batch_Size, Num_Patches, Embed_Dim]
hidden_states = self.layer_norm2(hidden_states)
# [Batch_Size, Num_Patches, Embed_Dim] -> [Batch_Size, Num_Patches, Embed_Dim]
hidden_states = self.mlp(hidden_states)
# [Batch_Size, Num_Patches, Embed_Dim]
hidden_states = residual + hidden_states
return hidden_states
class SiglipEncoder(nn.Module):
def __init__(self, config: SiglipVisionConfig):
super().__init__()
self.config = config
self.layers = nn.ModuleList(
[SiglipEncoderLayer(config) for _ in range(config.num_hidden_layers)]
)
# Ignore copy
def forward(self, inputs_embeds: torch.Tensor) -> torch.Tensor:
# inputs_embeds: [Batch_Size, Num_Patches, Embed_Dim]
hidden_states = inputs_embeds
for encoder_layer in self.layers:
# [Batch_Size, Num_Patches, Embed_Dim] -> [Batch_Size, Num_Patches, Embed_Dim]
hidden_states = encoder_layer(hidden_states)
return hidden_states
class SiglipVisionTransformer(nn.Module):
def __init__(self, config: SiglipVisionConfig):
super().__init__()
self.config = config
embed_dim = config.hidden_size
self.embeddings = SiglipVisionEmbeddings(config)
self.encoder = SiglipEncoder(config)
self.post_layernorm = nn.LayerNorm(embed_dim, eps=config.layer_norm_eps)
def forward(self, pixel_values: torch.Tensor) -> torch.Tensor:
# pixel_values: [Batch_Size, Channels, Height, Width] -> [Batch_Size, Num_Patches, Embed_Dim]
hidden_states = self.embeddings(pixel_values)
last_hidden_state = self.encoder(inputs_embeds=hidden_states)
last_hidden_state = self.post_layernorm(last_hidden_state)
return last_hidden_state
class SiglipVisionModel(nn.Module):
def __init__(self, config: SiglipVisionConfig):
super().__init__()
self.config = config
self.vision_model = SiglipVisionTransformer(config)
def forward(self, pixel_values) -> Tuple:
# [Batch_Size, Channels, Height, Width] -> [Batch_Size, Num_Patches, Embed_Dim]
return self.vision_model(pixel_values=pixel_values)
|