File size: 10,781 Bytes
844428c | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 | """Model classes for RSP models compatible with transformers"""
import sys
import os
from pathlib import Path
import torch
import torch.nn as nn
from transformers import PreTrainedModel
from transformers.modeling_outputs import ImageClassifierOutputWithNoAttention
from safetensors.torch import load_file
# Import local modular model
from modular_resnet import ResNet, Bottleneck
# Import other models from sibling directories if needed
_parent_dir = Path(__file__).parent.parent
import importlib.util
# Import SwinTransformer from RSP-Swin-T
_swin_path = _parent_dir / "RSP-Swin-T" / "modular_swin.py"
if _swin_path.exists():
spec = importlib.util.spec_from_file_location("modular_swin_swin", _swin_path)
swin_module = importlib.util.module_from_spec(spec)
spec.loader.exec_module(swin_module)
SwinTransformer = swin_module.SwinTransformer
else:
SwinTransformer = None
# Import ViTAE from RSP-ViTAEv2-S
_vitae_path = _parent_dir / "RSP-ViTAEv2-S" / "modular_vitae_window_noshift.py"
if _vitae_path.exists():
spec = importlib.util.spec_from_file_location("modular_vitae_window_noshift_vitae", _vitae_path)
vitae_module = importlib.util.module_from_spec(spec)
spec.loader.exec_module(vitae_module)
ViTAE_Window_NoShift_12_basic_stages4_14 = vitae_module.ViTAE_Window_NoShift_12_basic_stages4_14
else:
ViTAE_Window_NoShift_12_basic_stages4_14 = None
# Import configuration - handle both relative and absolute imports
try:
from configuration_rsp import RSPResNetConfig, RSPSwinConfig, RSPViTAEConfig
except ImportError:
# Fallback: import from same directory
import importlib.util
config_path = Path(__file__).parent / "configuration_rsp.py"
spec = importlib.util.spec_from_file_location("configuration_rsp", config_path)
config_module = importlib.util.module_from_spec(spec)
spec.loader.exec_module(config_module)
RSPResNetConfig = config_module.RSPResNetConfig
RSPSwinConfig = config_module.RSPSwinConfig
RSPViTAEConfig = config_module.RSPViTAEConfig
class RSPResNetForImageClassification(PreTrainedModel):
"""RSP ResNet model for image classification"""
config_class = RSPResNetConfig
def __init__(self, config):
super().__init__(config)
# Build ResNet model from config
block = Bottleneck if config.block == "Bottleneck" else None
if block is None:
raise ValueError(f"Unsupported block type: {config.block}")
self.model = ResNet(
block=block,
layers=config.layers,
num_classes=config.num_labels
)
def forward(self, pixel_values=None, labels=None, return_dict=None, **kwargs):
"""
Args:
pixel_values: Input images (B, C, H, W)
labels: Optional labels for loss computation
return_dict: Whether to return a ModelOutput instead of a plain tuple
"""
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
if pixel_values is None:
raise ValueError("pixel_values must be provided")
logits = self.model(pixel_values)
loss = None
if labels is not None:
loss_fct = nn.CrossEntropyLoss()
loss = loss_fct(logits.view(-1, self.config.num_labels), labels.view(-1))
if not return_dict:
output = (logits,)
return (loss,) + output if loss is not None else output
return ImageClassifierOutputWithNoAttention(
loss=loss,
logits=logits,
hidden_states=None,
)
@classmethod
def from_pretrained(cls, pretrained_model_name_or_path, *model_args, **kwargs):
"""Load model from pretrained checkpoint"""
config = kwargs.pop("config", None)
if config is None:
config = RSPResNetConfig.from_pretrained(pretrained_model_name_or_path)
model = cls(config)
# Load weights from safetensors
model_path = Path(pretrained_model_name_or_path)
safetensors_path = model_path / "model.safetensors"
if safetensors_path.exists():
state_dict = load_file(str(safetensors_path))
# Remove 'model.' prefix if present
state_dict_clean = {}
for k, v in state_dict.items():
if k.startswith("model."):
state_dict_clean[k[6:]] = v
else:
state_dict_clean[k] = v
model.model.load_state_dict(state_dict_clean, strict=False)
else:
raise FileNotFoundError(f"Model weights not found at {safetensors_path}")
return model
class RSPSwinForImageClassification(PreTrainedModel):
"""RSP Swin Transformer model for image classification"""
config_class = RSPSwinConfig
def __init__(self, config):
super().__init__(config)
# Build SwinTransformer model from config
self.model = SwinTransformer(
img_size=config.image_size,
patch_size=config.patch_size,
in_chans=config.num_channels,
num_classes=config.num_labels,
embed_dim=config.embed_dim,
depths=config.depths,
num_heads=config.num_heads,
window_size=config.window_size,
mlp_ratio=config.mlp_ratio,
qkv_bias=config.qkv_bias,
ape=config.ape,
patch_norm=config.patch_norm,
)
def forward(self, pixel_values=None, labels=None, return_dict=None, **kwargs):
"""
Args:
pixel_values: Input images (B, C, H, W)
labels: Optional labels for loss computation
return_dict: Whether to return a ModelOutput instead of a plain tuple
"""
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
if pixel_values is None:
raise ValueError("pixel_values must be provided")
logits = self.model(pixel_values)
loss = None
if labels is not None:
loss_fct = nn.CrossEntropyLoss()
loss = loss_fct(logits.view(-1, self.config.num_labels), labels.view(-1))
if not return_dict:
output = (logits,)
return (loss,) + output if loss is not None else output
return ImageClassifierOutputWithNoAttention(
loss=loss,
logits=logits,
hidden_states=None,
)
@classmethod
def from_pretrained(cls, pretrained_model_name_or_path, *model_args, **kwargs):
"""Load model from pretrained checkpoint"""
config = kwargs.pop("config", None)
if config is None:
config = RSPSwinConfig.from_pretrained(pretrained_model_name_or_path)
model = cls(config)
# Load weights from safetensors
model_path = Path(pretrained_model_name_or_path)
safetensors_path = model_path / "model.safetensors"
if safetensors_path.exists():
state_dict = load_file(str(safetensors_path))
# Remove 'model.' prefix if present
state_dict_clean = {}
for k, v in state_dict.items():
if k.startswith("model."):
state_dict_clean[k[6:]] = v
else:
state_dict_clean[k] = v
model.model.load_state_dict(state_dict_clean, strict=False)
else:
raise FileNotFoundError(f"Model weights not found at {safetensors_path}")
return model
class RSPViTAEForImageClassification(PreTrainedModel):
"""RSP ViTAE model for image classification"""
config_class = RSPViTAEConfig
def __init__(self, config):
super().__init__(config)
# Build ViTAE model from config
# Note: ViTAE_Window_NoShift_12_basic_stages4_14 already sets most parameters as defaults:
# - stages=4, embed_dims=[64, 64, 128, 256], token_dims=[64, 128, 256, 512]
# - downsample_ratios=[4, 2, 2, 2], NC_depth=[2, 2, 8, 2], etc.
# We only pass parameters that need to be overridden (img_size, num_classes)
# The function accepts **kwargs, so we can pass window_size if needed
self.model = ViTAE_Window_NoShift_12_basic_stages4_14(
pretrained=False,
img_size=config.image_size,
num_classes=config.num_labels,
window_size=7,
)
def forward(self, pixel_values=None, labels=None, return_dict=None, **kwargs):
"""
Args:
pixel_values: Input images (B, C, H, W)
labels: Optional labels for loss computation
return_dict: Whether to return a ModelOutput instead of a plain tuple
"""
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
if pixel_values is None:
raise ValueError("pixel_values must be provided")
logits = self.model(pixel_values)
loss = None
if labels is not None:
loss_fct = nn.CrossEntropyLoss()
loss = loss_fct(logits.view(-1, self.config.num_labels), labels.view(-1))
if not return_dict:
output = (logits,)
return (loss,) + output if loss is not None else output
return ImageClassifierOutputWithNoAttention(
loss=loss,
logits=logits,
hidden_states=None,
)
@classmethod
def from_pretrained(cls, pretrained_model_name_or_path, *model_args, **kwargs):
"""Load model from pretrained checkpoint"""
config = kwargs.pop("config", None)
if config is None:
config = RSPViTAEConfig.from_pretrained(pretrained_model_name_or_path)
model = cls(config)
# Load weights from safetensors
model_path = Path(pretrained_model_name_or_path)
safetensors_path = model_path / "model.safetensors"
if safetensors_path.exists():
state_dict = load_file(str(safetensors_path))
# Remove 'model.' prefix if present
state_dict_clean = {}
for k, v in state_dict.items():
if k.startswith("model."):
state_dict_clean[k[6:]] = v
else:
state_dict_clean[k] = v
model.model.load_state_dict(state_dict_clean, strict=False)
else:
raise FileNotFoundError(f"Model weights not found at {safetensors_path}")
return model
|