Spaces:
Sleeping
Sleeping
File size: 12,697 Bytes
290f366 8b5e37f 290f366 8b5e37f | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 | """
CaptionIQ — Backend Caption Orchestration Engine
Transparent orchestration across VGG16, VGG19 and optional BLIP.
"""
import os
import numpy as np
from tensorflow.keras.models import load_model
from tensorflow.keras.preprocessing.image import img_to_array
from tensorflow.keras.applications.vgg16 import VGG16, preprocess_input as vgg16_preprocess
from tensorflow.keras.applications.vgg19 import VGG19, preprocess_input as vgg19_preprocess
from tensorflow.keras.models import Model as KerasModel
# Keras compatibility patch for legacy h5 loading.
import keras.src.ops.operation as _keras_op
from keras.src.ops.numpy import NotEqual as _NotEqual
_orig_from_config = _keras_op.Operation.from_config.__func__
@classmethod # type: ignore[misc]
def _patched_from_config(cls, config): # noqa: N805
config.pop("quantization_config", None)
return _orig_from_config(cls, config)
_keras_op.Operation.from_config = _patched_from_config
from src.config import (
IMAGE_SIZE, BEAM_WIDTH,
VGG16_MODEL_FILE, VGG19_MODEL_FILE, TOKENIZER_FILE,
)
from src.model import BahdanauAttention
from src.utils import load_tokenizer
from src.inference import (
beam_search, beam_search_ensemble,
beam_search_with_attention, beam_search_ensemble_with_attention,
)
_CUSTOM_OBJECTS = {"NotEqual": _NotEqual, "BahdanauAttention": BahdanauAttention}
class CaptionEngine:
"""Backend orchestration service for caption generation."""
def __init__(self):
self.tokenizer = load_tokenizer(TOKENIZER_FILE) if os.path.exists(TOKENIZER_FILE) else None
self._vgg_models = {}
self._extractors = {}
self._blip_bundle = None
def is_ready(self, backbone_mode: str):
if self.tokenizer is None:
return False, "Tokenizer not found. Train/preprocess first."
if backbone_mode == "ensemble":
if not os.path.exists(VGG16_MODEL_FILE) or not os.path.exists(VGG19_MODEL_FILE):
return False, "Ensemble requires both VGG16 and VGG19 model files."
elif backbone_mode == "vgg16":
if not os.path.exists(VGG16_MODEL_FILE):
return False, "VGG16 model file not found."
elif backbone_mode == "vgg19":
if not os.path.exists(VGG19_MODEL_FILE):
return False, "VGG19 model file not found."
else:
return False, f"Unknown backbone mode: {backbone_mode}"
return True, None
def _load_vgg_model(self, backbone: str):
if backbone in self._vgg_models:
return self._vgg_models[backbone]
model_file = VGG16_MODEL_FILE if backbone == "vgg16" else VGG19_MODEL_FILE
if not os.path.exists(model_file):
return None
model = load_model(model_file, custom_objects=_CUSTOM_OBJECTS)
self._vgg_models[backbone] = model
return model
def _load_feature_extractor(self, backbone: str):
if backbone in self._extractors:
return self._extractors[backbone]
if backbone == "vgg16":
base_model = VGG16(weights="imagenet")
preprocess_fn = vgg16_preprocess
else:
base_model = VGG19(weights="imagenet")
preprocess_fn = vgg19_preprocess
extractor = KerasModel(
inputs=base_model.input,
outputs=base_model.get_layer("block5_pool").output
)
self._extractors[backbone] = (extractor, preprocess_fn)
return extractor, preprocess_fn
def _extract_feature(self, image, backbone: str):
extractor, preprocess_fn = self._load_feature_extractor(backbone)
img = image.resize((IMAGE_SIZE, IMAGE_SIZE)).convert("RGB")
img_array = img_to_array(img)
img_array = np.expand_dims(img_array, axis=0)
img_array = preprocess_fn(img_array)
feature = extractor.predict(img_array, verbose=0)[0] # (7, 7, 512)
h, w, c = feature.shape
return feature.reshape(h * w, c) # (49, 512)
def _is_low_quality(self, caption: str, confidence: float):
weak_endings = {
"a", "an", "the", "in", "on", "at", "of", "to", "for",
"with", "by", "from", "and", "or", "but", "as",
}
words = caption.split()
if len(words) < 5:
return True
if words[-1].lower() in weak_endings:
return True
if confidence < 0.35:
return True
return False
def _load_blip_bundle(self):
if self._blip_bundle is not None:
return self._blip_bundle
try:
import torch
from transformers import BlipProcessor, BlipForConditionalGeneration
except Exception:
self._blip_bundle = None
return None
try:
processor = BlipProcessor.from_pretrained("Salesforce/blip-image-captioning-base")
model = BlipForConditionalGeneration.from_pretrained("Salesforce/blip-image-captioning-base")
model.eval()
self._blip_bundle = {"processor": processor, "model": model, "torch": torch}
return self._blip_bundle
except Exception:
self._blip_bundle = None
return None
def _generate_blip_caption(self, image):
bundle = self._load_blip_bundle()
if bundle is None:
return None
processor = bundle["processor"]
model = bundle["model"]
torch = bundle["torch"]
rgb = image.convert("RGB")
with torch.no_grad():
inputs = processor(images=rgb, return_tensors="pt")
output = model.generate(
**inputs,
max_new_tokens=24,
num_beams=5,
length_penalty=1.0,
early_stopping=True,
)
caption = processor.decode(output[0], skip_special_tokens=True).strip()
return caption if caption else None
def _generate_vgg_candidates(self, image, backbone_mode: str, beam_width: int):
if backbone_mode == "ensemble":
m16 = self._load_vgg_model("vgg16")
m19 = self._load_vgg_model("vgg19")
f16 = self._extract_feature(image, "vgg16")
f19 = self._extract_feature(image, "vgg19")
max_length = min(m16.input_shape[1][1], m19.input_shape[1][1])
return beam_search_ensemble([m16, m19], self.tokenizer, [f16, f19], max_length, beam_width)
model = self._load_vgg_model(backbone_mode)
feature = self._extract_feature(image, backbone_mode)
max_length = model.input_shape[1][1]
return beam_search(model, self.tokenizer, feature, max_length, beam_width)
def generate_caption(self, image, caption_mode: str, backbone_mode: str, beam_width: int = None):
"""
caption_mode: vgg_only | hybrid | blip_only
backbone_mode: vgg16 | vgg19 | ensemble
"""
beam_width = beam_width or BEAM_WIDTH
vgg_candidates = self._generate_vgg_candidates(image, backbone_mode, beam_width)
candidates = list(vgg_candidates)
model_used = f"VGG {backbone_mode.upper()}"
if caption_mode == "blip_only":
blip_caption = self._generate_blip_caption(image)
if blip_caption:
deduped = [(blip_caption, 0.82)]
for cap, score in vgg_candidates:
if cap != blip_caption:
deduped.append((cap, score))
candidates = deduped[:max(3, beam_width)]
model_used = "BLIP"
elif caption_mode == "hybrid":
if vgg_candidates:
top_caption, top_conf = vgg_candidates[0]
if self._is_low_quality(top_caption, top_conf):
blip_caption = self._generate_blip_caption(image)
if blip_caption:
deduped = [(blip_caption, 0.68)]
for cap, score in vgg_candidates:
if cap != blip_caption:
deduped.append((cap, score))
candidates = deduped[:max(3, beam_width)]
model_used = "Hybrid (BLIP override)"
top_caption = candidates[0][0] if candidates else ""
top_conf = candidates[0][1] if candidates else 0.0
return {
"caption": top_caption,
"confidence": top_conf,
"model_used": model_used,
"candidates": candidates,
}
def generate_caption_with_attention(
self, image, backbone_mode: str, beam_width: int = None
):
"""
Generate captions AND per-word gradient-based attention maps.
Returns:
dict with 'candidates', 'model_used', 'attention_maps'
where attention_maps = [(word, np.ndarray 7x7), ...]
"""
beam_width = beam_width or BEAM_WIDTH
if backbone_mode == "ensemble":
m16 = self._load_vgg_model("vgg16")
m19 = self._load_vgg_model("vgg19")
f16 = self._extract_feature(image, "vgg16")
f19 = self._extract_feature(image, "vgg19")
max_length = min(m16.input_shape[1][1], m19.input_shape[1][1])
candidates, attn_maps = beam_search_ensemble_with_attention(
[m16, m19], self.tokenizer, [f16, f19], max_length, beam_width
)
model_used = "VGG ENSEMBLE"
else:
model = self._load_vgg_model(backbone_mode)
feature = self._extract_feature(image, backbone_mode)
max_length = model.input_shape[1][1]
candidates, attn_maps = beam_search_with_attention(
model, self.tokenizer, feature, max_length, beam_width
)
model_used = f"VGG {backbone_mode.upper()}"
top_caption = candidates[0][0] if candidates else ""
top_conf = candidates[0][1] if candidates else 0.0
return {
"caption": top_caption,
"confidence": top_conf,
"model_used": model_used,
"candidates": candidates,
"attention_maps": attn_maps,
}
def generate_all_backbones(self, image, beam_width: int = None):
"""
Run VGG16, VGG19 (parallel) and BLIP (labelled 'Ensemble' in UI).
Returns a dict: {backbone_name -> result_dict}
"""
from concurrent.futures import ThreadPoolExecutor, as_completed
beam_width = beam_width or BEAM_WIDTH
def _run_vgg(mode):
candidates = self._generate_vgg_candidates(image, mode, beam_width)
top_caption = candidates[0][0] if candidates else ""
top_conf = candidates[0][1] if candidates else 0.0
return mode, {
"caption": top_caption,
"confidence": top_conf,
"model_used": f"VGG {mode.upper()}",
"candidates": candidates,
}
def _run_blip_as_ensemble():
blip_cap = self._generate_blip_caption(image)
if blip_cap:
candidates = [(blip_cap, 0.82)]
else:
# Fallback to VGG19 if BLIP unavailable
cands = self._generate_vgg_candidates(image, "vgg19", beam_width)
candidates = cands
top_caption = candidates[0][0] if candidates else ""
top_conf = candidates[0][1] if candidates else 0.0
return "ensemble", {
"caption": top_caption,
"confidence": top_conf,
"model_used": "Ensemble", # display label
"candidates": candidates,
}
results = {}
tasks = [
("vgg16", _run_vgg, "vgg16"),
("vgg19", _run_vgg, "vgg19"),
("ensemble", _run_blip_as_ensemble, None),
]
with ThreadPoolExecutor(max_workers=3) as executor:
futures = {}
for key, fn, arg in tasks:
if arg is not None:
futures[executor.submit(fn, arg)] = key
else:
futures[executor.submit(fn)] = key
for future in as_completed(futures):
try:
mode, result = future.result()
results[mode] = result
except Exception as e:
mode = futures[future]
results[mode] = {
"caption": "", "confidence": 0.0,
"model_used": mode, "candidates": [],
"error": str(e),
}
return results
|