File size: 7,969 Bytes
c93bcc6 16b70ae c93bcc6 16b70ae 2d8a966 c93bcc6 2d8a966 16b70ae c93bcc6 16b70ae 7ebdef3 16b70ae 7ebdef3 16b70ae c93bcc6 16b70ae c93bcc6 16b70ae | 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 | from typing import Literal, Optional, Tuple, Union
import torch
import torch.nn as nn
import torch.nn.functional as F
from transformers import AutoConfig, AutoModel, PreTrainedModel
# Handle import for both local development and HuggingFace Hub loading
try:
from .configuration_spatial_embeddings import SpatialEmbeddingsConfig
except ImportError:
# When loaded from HuggingFace Hub, relative imports may not work
# Try absolute import instead
try:
from configuration_spatial_embeddings import SpatialEmbeddingsConfig
except ImportError:
# Last resort: import from the module directly (Hugging Face dynamic module loading)
from configuration_spatial_embeddings import SpatialEmbeddingsConfig
class EmbeddingProjector(nn.Module):
"""
Configurable MLP projection head for embedding transformation.
(Copied from embedding_pipeline/train/model.py for self-contained publishing)
"""
def __init__(
self,
input_dim: int = 768,
hidden_dim: int = 512,
output_dim: int = 256,
dropout: float = 0.1,
num_hidden_layers: int = 1,
hidden_dim_multiplier: float = 1.0,
activation: Literal["gelu", "relu", "silu"] = "gelu",
use_residual: bool = True,
use_layer_norm: bool = True,
):
super().__init__()
self.input_dim = input_dim
self.hidden_dim = hidden_dim
self.output_dim = output_dim
self.use_residual = use_residual
self.use_layer_norm = use_layer_norm
self.num_hidden_layers = num_hidden_layers
self.hidden_dim_multiplier = hidden_dim_multiplier
self.activation_name = activation
self.hidden_dims = self._compute_hidden_dims(
hidden_dim, num_hidden_layers, hidden_dim_multiplier
)
self.activation = self._resolve_activation(activation)
# First hidden block
self.input_layer = nn.Linear(input_dim, self.hidden_dims[0])
if use_layer_norm:
self.input_norm = nn.LayerNorm(self.hidden_dims[0])
self.input_dropout = nn.Dropout(dropout)
# Additional hidden blocks (if any)
self.hidden_layers = nn.ModuleList()
if use_layer_norm:
self.hidden_norms = nn.ModuleList()
else:
self.hidden_norms = None
self.hidden_dropouts = nn.ModuleList()
for idx in range(1, len(self.hidden_dims)):
layer = nn.Linear(self.hidden_dims[idx - 1], self.hidden_dims[idx])
self.hidden_layers.append(layer)
if use_layer_norm:
self.hidden_norms.append(nn.LayerNorm(self.hidden_dims[idx]))
self.hidden_dropouts.append(nn.Dropout(dropout))
# Output block
self.output_layer = nn.Linear(self.hidden_dims[-1], output_dim)
if use_layer_norm:
self.output_norm = nn.LayerNorm(output_dim)
self.output_dropout = nn.Dropout(dropout)
# Residual shortcut (projects input directly to output)
if use_residual:
self.residual_proj = nn.Linear(input_dim, output_dim)
@staticmethod
def _compute_hidden_dims(
base_hidden_dim: int, num_layers: int, multiplier: float
) -> list[int]:
dims: list[int] = []
current_dim = base_hidden_dim
for layer_idx in range(num_layers):
if layer_idx == 0:
dims.append(base_hidden_dim)
else:
current_dim = max(16, int(round(current_dim * multiplier)))
dims.append(current_dim)
return dims
@staticmethod
def _resolve_activation(name: str) -> nn.Module:
if name == "gelu":
return nn.GELU()
if name == "relu":
return nn.ReLU()
if name == "silu":
return nn.SiLU()
raise ValueError(f"Unsupported activation: {name}")
def forward(self, x: torch.Tensor) -> torch.Tensor:
# First hidden block
out = self.input_layer(x)
if self.use_layer_norm:
out = self.input_norm(out)
out = self.activation(out)
out = self.input_dropout(out)
# Additional hidden blocks
for idx, layer in enumerate(self.hidden_layers):
out = layer(out)
if self.use_layer_norm and self.hidden_norms is not None:
out = self.hidden_norms[idx](out)
out = self.activation(out)
out = self.hidden_dropouts[idx](out)
# Output block
out = self.output_layer(out)
if self.use_layer_norm:
out = self.output_norm(out)
out = self.output_dropout(out)
# Residual connection
if self.use_residual:
residual = self.residual_proj(x)
out = out + residual
# L2 normalization
out = F.normalize(out, p=2, dim=1)
return out
class SpatialEmbeddingsModel(PreTrainedModel):
config_class = SpatialEmbeddingsConfig
_tied_weights_keys = []
def __init__(self, config: SpatialEmbeddingsConfig):
super().__init__(config)
self.config = config
# transformers ≥5.6 accesses all_tied_weights_keys before tie_weights() is called
# for custom trust_remote_code models; set it here since we have no tied weights.
if not hasattr(self, "all_tied_weights_keys"):
self.all_tied_weights_keys = {}
# Initialize backbone from config rather than from_pretrained.
# Calling from_pretrained inside __init__ is an anti-pattern in transformers ≥5:
# the outer from_pretrained wraps __init__ in a meta-device context
# (low_cpu_mem_usage=True), and a nested from_pretrained inside that context
# raises an error. Using from_config is safe in any device context.
# Backbone weights are populated by the outer from_pretrained weight-loading
# step (inference) or explicitly by convert_to_hf.py (publishing).
backbone_config = AutoConfig.from_pretrained(
config.backbone_model_name, trust_remote_code=True
)
self.backbone = AutoModel.from_config(backbone_config, trust_remote_code=True)
# Initialize projector
self.projector = EmbeddingProjector(
input_dim=config.input_dim,
hidden_dim=config.hidden_dim,
output_dim=config.output_dim,
dropout=config.dropout,
num_hidden_layers=config.num_hidden_layers,
hidden_dim_multiplier=config.hidden_dim_multiplier,
activation=config.activation,
use_residual=config.use_residual,
use_layer_norm=config.use_layer_norm,
)
def forward(
self,
pixel_values: Optional[torch.Tensor] = None,
return_dict: Optional[bool] = None,
**kwargs,
) -> Union[Tuple, torch.Tensor]:
"""
Args:
pixel_values: Tensor of shape (batch_size, channels, height, width)
return_dict: Whether to return a dictionary or tuple
Returns:
If return_dict is True (default for HF), returns object with 'embeddings'.
Otherwise returns (embeddings,).
"""
# Pass through backbone
outputs = self.backbone(pixel_values=pixel_values, return_dict=True, **kwargs)
# Extract pooled output (CLS token or similar)
# DINOv2 outputs pooler_output in some versions, or last_hidden_state
if hasattr(outputs, "pooler_output") and outputs.pooler_output is not None:
backbone_emb = outputs.pooler_output
else:
# Fallback: Use CLS token from last hidden state
backbone_emb = outputs.last_hidden_state[:, 0]
# Project to specialized embedding
specialized_emb = self.projector(backbone_emb)
if return_dict:
return {"embeddings": specialized_emb, "backbone_outputs": outputs}
return (specialized_emb,)
|