Gemma3n-challenge-demo / classifier.py
HMWCS
feat: add confidence score feature to model predictions
ddca90f
raw
history blame
17.5 kB
from transformers import AutoProcessor, AutoModelForImageTextToText
from PIL import Image
import torch
import logging
from typing import Union, Tuple
from config import Config
from knowledge_base import GarbageClassificationKnowledge
import re
class GarbageClassifier:
def __init__(self, config: Config = None):
self.config = config or Config()
self.knowledge = GarbageClassificationKnowledge()
self.processor = None
self.model = None
self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
# Setup logging
logging.basicConfig(level=logging.INFO)
self.logger = logging.getLogger(__name__)
def load_model(self):
"""Load the model and processor"""
try:
self.logger.info(f"Loading model: {self.config.MODEL_NAME}")
# Load processor
kwargs = {}
if self.config.HF_TOKEN:
kwargs["token"] = self.config.HF_TOKEN
self.processor = AutoProcessor.from_pretrained(
self.config.MODEL_NAME, **kwargs
)
# Load model
self.model = AutoModelForImageTextToText.from_pretrained(
self.config.MODEL_NAME,
torch_dtype=self.config.TORCH_DTYPE,
device_map=self.config.DEVICE_MAP,
)
self.logger.info("Model loaded successfully")
except Exception as e:
self.logger.error(f"Error loading model: {str(e)}")
raise
def preprocess_image(self, image: Image.Image) -> Image.Image:
"""
Preprocess image to meet Gemma3n requirements (512x512)
"""
# Convert to RGB if necessary
if image.mode != "RGB":
image = image.convert("RGB")
# Resize to 512x512 as required by Gemma3n
target_size = (512, 512)
# Calculate aspect ratio preserving resize
original_width, original_height = image.size
aspect_ratio = original_width / original_height
if aspect_ratio > 1:
# Width is larger
new_width = target_size[0]
new_height = int(target_size[0] / aspect_ratio)
else:
# Height is larger or equal
new_height = target_size[1]
new_width = int(target_size[1] * aspect_ratio)
# Resize image maintaining aspect ratio
image = image.resize((new_width, new_height), Image.Resampling.LANCZOS)
# Create a new image with target size and paste the resized image
processed_image = Image.new(
"RGB", target_size, (255, 255, 255)
) # White background
# Calculate position to center the image
x_offset = (target_size[0] - new_width) // 2
y_offset = (target_size[1] - new_height) // 2
processed_image.paste(image, (x_offset, y_offset))
return processed_image
def classify_image(self, image: Union[str, Image.Image]) -> Tuple[str, str, int]:
"""
Classify garbage in the image
Args:
image: PIL Image or path to image file
Returns:
Tuple of (classification_result, detailed_analysis, confidence_score)
"""
if self.model is None or self.processor is None:
raise RuntimeError("Model not loaded. Call load_model() first.")
try:
# Load and process image
if isinstance(image, str):
image = Image.open(image)
elif not isinstance(image, Image.Image):
raise ValueError("Image must be a PIL Image or file path")
# Preprocess image to meet Gemma3n requirements
processed_image = self.preprocess_image(image)
# Prepare messages with system prompt and user query
messages = [
{
"role": "system",
"content": [
{
"type": "text",
"text": self.knowledge.get_system_prompt(),
}
],
},
{
"role": "user",
"content": [
{"type": "image", "image": processed_image},
{
"type": "text",
"text": "Please classify what you see in this image. If it shows garbage/waste items, classify them according to the garbage classification standards. If it shows people, living things, or other non-waste items, classify it as 'Unable to classify' and explain why it's not garbage. Also provide a confidence score from 1-10 indicating how certain you are about your classification.",
},
],
},
]
# Apply chat template and tokenize
inputs = self.processor.apply_chat_template(
messages,
add_generation_prompt=True,
tokenize=True,
return_dict=True,
return_tensors="pt",
).to(self.model.device, dtype=self.model.dtype)
input_len = inputs["input_ids"].shape[-1]
outputs = self.model.generate(
**inputs,
max_new_tokens=self.config.MAX_NEW_TOKENS,
disable_compile=True,
)
response = self.processor.batch_decode(
outputs[:, input_len:],
skip_special_tokens=True,
)[0]
# Extract classification from response
classification = self._extract_classification(response)
# Extract reasoning from response
reasoning = self._extract_reasoning(response)
# Extract confidence score from response
confidence_score = self._extract_confidence_score(response, classification)
return classification, reasoning, confidence_score
except Exception as e:
self.logger.error(f"Error during classification: {str(e)}")
import traceback
traceback.print_exc()
return "Error", f"Classification failed: {str(e)}", 0
def _calculate_confidence_heuristic(self, response_lower: str, classification: str) -> int:
"""Calculate confidence based on response content and classification type"""
base_confidence = 5
# Confidence indicators (increase confidence)
high_confidence_words = ["clearly", "obviously", "definitely", "certainly", "exactly"]
medium_confidence_words = ["appears", "seems", "likely", "probably"]
# Uncertainty indicators (decrease confidence)
uncertainty_words = ["might", "could", "possibly", "maybe", "unclear", "difficult"]
# Adjust based on confidence words
for word in high_confidence_words:
if word in response_lower:
base_confidence += 2
break
for word in medium_confidence_words:
if word in response_lower:
base_confidence += 1
break
for word in uncertainty_words:
if word in response_lower:
base_confidence -= 2
break
# Classification-specific adjustments
if classification == "Unable to classify":
if any(indicator in response_lower for indicator in ["person", "people", "human", "living"]):
base_confidence += 1 # High confidence when clearly not waste
else:
base_confidence -= 1 # Lower confidence for unclear items
elif classification == "Error":
base_confidence = 1
else:
# Check for specific material mentions (increases confidence)
specific_materials = ["aluminum", "plastic", "glass", "metal", "cardboard", "paper"]
if any(material in response_lower for material in specific_materials):
base_confidence += 1
return min(max(base_confidence, 1), 10)
def _extract_confidence_score(self, response: str, classification: str) -> int:
"""Extract confidence score from response or calculate based on classification"""
response_lower = response.lower()
# Look for explicit confidence scores in the response
confidence_patterns = [
r'confidence[:\s]*(\d+)',
r'confident[:\s]*(\d+)',
r'certainty[:\s]*(\d+)',
r'score[:\s]*(\d+)',
r'(\d+)/10',
r'(\d+)\s*out\s*of\s*10'
]
for pattern in confidence_patterns:
match = re.search(pattern, response_lower)
if match:
score = int(match.group(1))
return min(max(score, 1), 10) # Clamp between 1-10
# If no explicit score found, calculate based on classification indicators
return self._calculate_confidence_heuristic(response_lower, classification)
def _extract_classification(self, response: str) -> str:
"""Extract the main classification from the response"""
response_lower = response.lower()
# First, look for positive waste category indicators
# Check exact category matches first
categories = self.knowledge.get_categories()
waste_categories = [cat for cat in categories if cat != "Unable to classify"]
for category in waste_categories:
if category.lower() in response_lower:
# Make sure it's not in a negative context
category_index = response_lower.find(category.lower())
context_before = response_lower[max(0, category_index-30):category_index]
# Only skip if there's a clear negation right before
if not any(neg in context_before[-10:] for neg in ["not", "cannot", "isn't", "doesn't"]):
return category
# Look for strong recyclable indicators
recyclable_indicators = [
"recyclable", "recycle", "aluminum", "plastic", "glass", "metal",
"foil", "can", "bottle", "cardboard", "paper", "tin", "steel", "iron"
]
if any(indicator in response_lower for indicator in recyclable_indicators):
# Check if it's explicitly said to be recyclable
recyclable_phrases = [
"recyclable", "can be recycled", "made of recyclable",
"recyclable material", "recyclable aluminum", "recyclable plastic"
]
if any(phrase in response_lower for phrase in recyclable_phrases):
return "Recyclable Waste"
# Check for specific materials
if any(material in response_lower for material in ["aluminum", "foil", "metal"]):
return "Recyclable Waste"
if any(material in response_lower for material in ["plastic", "bottle"]):
return "Recyclable Waste"
if any(material in response_lower for material in ["glass", "cardboard", "paper"]):
return "Recyclable Waste"
# Look for food waste indicators
food_indicators = [
"food", "fruit", "vegetable", "organic", "kitchen waste",
"peel", "core", "scraps", "leftovers"
]
if any(indicator in response_lower for indicator in food_indicators):
return "Food/Kitchen Waste"
# Look for hazardous waste indicators
hazardous_indicators = [
"battery", "chemical", "medicine", "paint", "toxic", "hazardous"
]
if any(indicator in response_lower for indicator in hazardous_indicators):
return "Hazardous Waste"
# Look for other waste indicators
other_waste_indicators = [
"cigarette", "ceramic", "dust", "diaper", "tissue", "other waste"
]
if any(indicator in response_lower for indicator in other_waste_indicators):
return "Other Waste"
# Only classify as "Unable to classify" if there are explicit indicators
unable_phrases = [
"unable to classify",
"cannot classify",
"cannot be classified as waste",
"not garbage", "not waste", "not trash"
]
if any(phrase in response_lower for phrase in unable_phrases):
return "Unable to classify"
# Check for non-garbage items (people, living things, etc.)
non_garbage_indicators = [
"person", "people", "human", "face", "man", "woman",
"living", "alive", "animal", "pet",
"portrait", "photo of a person"
]
if any(indicator in response_lower for indicator in non_garbage_indicators):
return "Unable to classify"
# If we found waste-related content but no clear category, try to infer
waste_related = any(word in response_lower for word in [
"waste", "trash", "garbage", "discard", "throw", "bin"
])
if waste_related:
# Default to Other Waste if it's clearly waste but unclear category
return "Other Waste"
# If no clear classification found and no clear non-waste indicators,
# default to "Unable to classify"
return "Unable to classify"
def _extract_reasoning(self, response: str) -> str:
"""Extract only the reasoning content, removing all formatting markers and classification info"""
import re
# Remove all formatting markers
cleaned_response = response.replace("**Classification**:", "")
cleaned_response = cleaned_response.replace("**Reasoning**:", "")
cleaned_response = re.sub(r'\*\*.*?\*\*:', '', cleaned_response) # Remove any **text**: patterns
cleaned_response = cleaned_response.replace("**", "") # Remove remaining ** markers
# Remove category names that might appear at the beginning
categories = self.knowledge.get_categories()
for category in categories:
if cleaned_response.strip().startswith(category):
cleaned_response = cleaned_response.replace(category, "", 1)
break
# Remove common material names that might appear at the beginning
material_names = [
"Glass", "Plastic", "Metal", "Paper", "Cardboard", "Aluminum",
"Steel", "Iron", "Tin", "Foil", "Wood", "Ceramic", "Fabric",
"Recyclable Waste", "Food/Kitchen Waste", "Hazardous Waste", "Other Waste"
]
# Clean the response
cleaned_response = cleaned_response.strip()
# Remove material names at the beginning
for material in material_names:
if cleaned_response.startswith(material):
# Remove the material name and any following punctuation/whitespace
cleaned_response = cleaned_response[len(material):].lstrip(" .,;:")
break
# Split into sentences and clean up
sentences = []
# Split by common sentence endings, but keep the endings
parts = re.split(r'([.!?])\s+', cleaned_response)
# Rejoin parts to maintain sentence structure
reconstructed_parts = []
for i in range(0, len(parts), 2):
if i < len(parts):
sentence = parts[i]
if i + 1 < len(parts):
sentence += parts[i + 1] # Add the punctuation back
reconstructed_parts.append(sentence)
for part in reconstructed_parts:
part = part.strip()
if not part:
continue
# Skip parts that are just category names or material names
if part in categories or part.rstrip(".,;:") in material_names:
continue
# Skip parts that start with category names or material names
is_category_line = False
for item in categories + material_names:
if part.startswith(item):
is_category_line = True
break
if is_category_line:
continue
# Clean up the sentence
part = re.sub(r'^[A-Za-z\s]+:', '', part).strip() # Remove "Category:" type prefixes
if part and len(part) > 3: # Only keep meaningful content
sentences.append(part)
# Join sentences
reasoning = ' '.join(sentences)
# Final cleanup - remove any remaining standalone material words at the beginning
reasoning_words = reasoning.split()
if reasoning_words and reasoning_words[0] in [m.lower() for m in material_names]:
reasoning_words = reasoning_words[1:]
reasoning = ' '.join(reasoning_words)
# Ensure proper capitalization
if reasoning:
reasoning = reasoning[0].upper() + reasoning[1:] if len(reasoning) > 1 else reasoning.upper()
# Ensure proper punctuation
if not reasoning.endswith(('.', '!', '?')):
reasoning += '.'
return reasoning if reasoning else "Analysis not available"
def get_categories_info(self):
"""Get information about all categories"""
return self.knowledge.get_category_descriptions()