even more device handling
Browse files- custom_st.py +121 -5
custom_st.py
CHANGED
|
@@ -4,6 +4,7 @@ import os
|
|
| 4 |
import math
|
| 5 |
from io import BytesIO
|
| 6 |
from typing import Any, Dict, List, Literal, Optional, Union
|
|
|
|
| 7 |
|
| 8 |
import requests
|
| 9 |
import torch
|
|
@@ -121,27 +122,142 @@ class Transformer(nn.Module):
|
|
| 121 |
image_data = base64.b64decode(data)
|
| 122 |
return Image.open(BytesIO(image_data))
|
| 123 |
|
| 124 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 125 |
processed_texts = []
|
| 126 |
processed_images = []
|
| 127 |
dummy_image = Image.new('RGB', (56, 56))
|
| 128 |
|
| 129 |
for sample in texts:
|
| 130 |
if isinstance(sample, str):
|
| 131 |
-
|
| 132 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 133 |
elif isinstance(sample, Image.Image):
|
| 134 |
processed_texts.append(self.document_prompt)
|
| 135 |
processed_images.append(self._resize_image(sample))
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 136 |
|
| 137 |
return processed_texts, processed_images
|
| 138 |
|
| 139 |
def forward(self, features: Dict[str, torch.Tensor]) -> Dict[str, torch.Tensor]:
|
| 140 |
-
cache_position = torch.arange(0, features['input_ids'].shape[
|
| 141 |
inputs = self.model.prepare_inputs_for_generation(
|
| 142 |
**features, cache_position=cache_position, use_cache=False
|
| 143 |
)
|
| 144 |
|
|
|
|
|
|
|
|
|
|
|
|
|
| 145 |
with torch.no_grad():
|
| 146 |
output = self.model(
|
| 147 |
**inputs,
|
|
@@ -155,7 +271,7 @@ class Transformer(nn.Module):
|
|
| 155 |
)
|
| 156 |
return features
|
| 157 |
|
| 158 |
-
def tokenize(self, texts: List[Union[str, Image.Image]], padding: str = 'longest') -> Dict[str, torch.Tensor]:
|
| 159 |
processed_texts, processed_images = self._process_input(texts)
|
| 160 |
|
| 161 |
return self.processor(
|
|
|
|
| 4 |
import math
|
| 5 |
from io import BytesIO
|
| 6 |
from typing import Any, Dict, List, Literal, Optional, Union
|
| 7 |
+
from urllib.parse import urlparse
|
| 8 |
|
| 9 |
import requests
|
| 10 |
import torch
|
|
|
|
| 122 |
image_data = base64.b64decode(data)
|
| 123 |
return Image.open(BytesIO(image_data))
|
| 124 |
|
| 125 |
+
@staticmethod
|
| 126 |
+
def _is_valid_url(url: str) -> bool:
|
| 127 |
+
try:
|
| 128 |
+
result = urlparse(url)
|
| 129 |
+
# Check if scheme and netloc are present and scheme is http/https
|
| 130 |
+
return all([result.scheme in ('http', 'https'), result.netloc])
|
| 131 |
+
except Exception:
|
| 132 |
+
return False
|
| 133 |
+
|
| 134 |
+
@staticmethod
|
| 135 |
+
def _is_safe_path(path: str) -> bool:
|
| 136 |
+
try:
|
| 137 |
+
# Convert to absolute path and normalize
|
| 138 |
+
abs_path = os.path.abspath(os.path.normpath(path))
|
| 139 |
+
# Check if file exists and is a regular file (not a directory or special file)
|
| 140 |
+
return os.path.isfile(abs_path)
|
| 141 |
+
except Exception:
|
| 142 |
+
return False
|
| 143 |
+
|
| 144 |
+
@staticmethod
|
| 145 |
+
def _load_image_from_url(url: str) -> Image.Image:
|
| 146 |
+
try:
|
| 147 |
+
response = requests.get(
|
| 148 |
+
url,
|
| 149 |
+
stream=True,
|
| 150 |
+
timeout=10, # Add timeout
|
| 151 |
+
headers={'User-Agent': 'Mozilla/5.0'} # Add user agent
|
| 152 |
+
)
|
| 153 |
+
response.raise_for_status()
|
| 154 |
+
|
| 155 |
+
# Check content type
|
| 156 |
+
content_type = response.headers.get('content-type', '')
|
| 157 |
+
if not content_type.startswith('image/'):
|
| 158 |
+
raise ValueError(f"Invalid content type: {content_type}")
|
| 159 |
+
|
| 160 |
+
# Limit file size (e.g., 10MB)
|
| 161 |
+
content = BytesIO()
|
| 162 |
+
size = 0
|
| 163 |
+
max_size = 10 * 1024 * 1024 # 10MB
|
| 164 |
+
|
| 165 |
+
for chunk in response.iter_content(chunk_size=8192):
|
| 166 |
+
size += len(chunk)
|
| 167 |
+
if size > max_size:
|
| 168 |
+
raise ValueError("File too large")
|
| 169 |
+
content.write(chunk)
|
| 170 |
+
|
| 171 |
+
content.seek(0)
|
| 172 |
+
return Image.open(content)
|
| 173 |
+
except Exception as e:
|
| 174 |
+
raise ValueError(f"Failed to load image from URL: {str(e)}")
|
| 175 |
+
|
| 176 |
+
@staticmethod
|
| 177 |
+
def _load_image_from_path(image_path: str) -> Image.Image:
|
| 178 |
+
try:
|
| 179 |
+
# Convert to absolute path and normalize
|
| 180 |
+
abs_path = os.path.abspath(os.path.normpath(image_path))
|
| 181 |
+
|
| 182 |
+
# Check file size before loading
|
| 183 |
+
file_size = os.path.getsize(abs_path)
|
| 184 |
+
max_size = 10 * 1024 * 1024 # 10MB
|
| 185 |
+
if file_size > max_size:
|
| 186 |
+
raise ValueError("File too large")
|
| 187 |
+
|
| 188 |
+
with Image.open(abs_path) as img:
|
| 189 |
+
# Make a copy to ensure file handle is closed
|
| 190 |
+
return img.copy()
|
| 191 |
+
except Exception as e:
|
| 192 |
+
raise ValueError(f"Failed to load image from path: {str(e)}")
|
| 193 |
+
|
| 194 |
+
@staticmethod
|
| 195 |
+
def _load_image_from_bytes(image_bytes: bytes) -> Image.Image:
|
| 196 |
+
try:
|
| 197 |
+
# Check size
|
| 198 |
+
if len(image_bytes) > 10 * 1024 * 1024: # 10MB
|
| 199 |
+
raise ValueError("Image data too large")
|
| 200 |
+
|
| 201 |
+
return Image.open(BytesIO(image_bytes))
|
| 202 |
+
except Exception as e:
|
| 203 |
+
raise ValueError(f"Failed to load image from bytes: {str(e)}")
|
| 204 |
+
|
| 205 |
+
def _process_input(self, texts: List[Union[str, Image.Image, bytes]]) -> tuple[List[str], List[Image.Image]]:
|
| 206 |
processed_texts = []
|
| 207 |
processed_images = []
|
| 208 |
dummy_image = Image.new('RGB', (56, 56))
|
| 209 |
|
| 210 |
for sample in texts:
|
| 211 |
if isinstance(sample, str):
|
| 212 |
+
# Check if the string is a valid URL
|
| 213 |
+
if self._is_valid_url(sample):
|
| 214 |
+
try:
|
| 215 |
+
img = self._load_image_from_url(sample)
|
| 216 |
+
processed_texts.append(self.document_prompt)
|
| 217 |
+
processed_images.append(self._resize_image(img))
|
| 218 |
+
except Exception as e:
|
| 219 |
+
# If URL loading fails, treat as regular text
|
| 220 |
+
processed_texts.append(self.query_prompt % sample)
|
| 221 |
+
processed_images.append(dummy_image)
|
| 222 |
+
# Check if the string is a valid file path
|
| 223 |
+
elif self._is_safe_path(sample):
|
| 224 |
+
try:
|
| 225 |
+
img = self._load_image_from_path(sample)
|
| 226 |
+
processed_texts.append(self.document_prompt)
|
| 227 |
+
processed_images.append(self._resize_image(img))
|
| 228 |
+
except Exception as e:
|
| 229 |
+
# If image loading fails, treat as regular text
|
| 230 |
+
processed_texts.append(self.query_prompt % sample)
|
| 231 |
+
processed_images.append(dummy_image)
|
| 232 |
+
else:
|
| 233 |
+
# Regular text query
|
| 234 |
+
processed_texts.append(self.query_prompt % sample)
|
| 235 |
+
processed_images.append(dummy_image)
|
| 236 |
elif isinstance(sample, Image.Image):
|
| 237 |
processed_texts.append(self.document_prompt)
|
| 238 |
processed_images.append(self._resize_image(sample))
|
| 239 |
+
elif isinstance(sample, bytes):
|
| 240 |
+
try:
|
| 241 |
+
img = self._load_image_from_bytes(sample)
|
| 242 |
+
processed_texts.append(self.document_prompt)
|
| 243 |
+
processed_images.append(self._resize_image(img))
|
| 244 |
+
except Exception as e:
|
| 245 |
+
# If bytes can't be converted to image, use dummy
|
| 246 |
+
processed_texts.append(self.document_prompt)
|
| 247 |
+
processed_images.append(dummy_image)
|
| 248 |
|
| 249 |
return processed_texts, processed_images
|
| 250 |
|
| 251 |
def forward(self, features: Dict[str, torch.Tensor]) -> Dict[str, torch.Tensor]:
|
| 252 |
+
cache_position = torch.arange(0, features['input_ids'].shape[1])
|
| 253 |
inputs = self.model.prepare_inputs_for_generation(
|
| 254 |
**features, cache_position=cache_position, use_cache=False
|
| 255 |
)
|
| 256 |
|
| 257 |
+
# ensure inputs are on the same device as the model
|
| 258 |
+
device = next(self.model.parameters()).device
|
| 259 |
+
inputs = {k: v.to(device) for k, v in inputs.items() if isinstance(v, torch.Tensor)}
|
| 260 |
+
|
| 261 |
with torch.no_grad():
|
| 262 |
output = self.model(
|
| 263 |
**inputs,
|
|
|
|
| 271 |
)
|
| 272 |
return features
|
| 273 |
|
| 274 |
+
def tokenize(self, texts: List[Union[str, Image.Image, bytes]], padding: str = 'longest') -> Dict[str, torch.Tensor]:
|
| 275 |
processed_texts, processed_images = self._process_input(texts)
|
| 276 |
|
| 277 |
return self.processor(
|