upload evaluate
Browse files- evaluate.py +228 -0
evaluate.py
ADDED
@@ -0,0 +1,228 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import logging
|
3 |
+
import torch.nn.functional as F
|
4 |
+
from PIL import Image
|
5 |
+
from transformers import AutoTokenizer, AutoModel, Swinv2Model
|
6 |
+
from torchvision import transforms
|
7 |
+
from src.model.model import MisinformationDetectionModel
|
8 |
+
|
9 |
+
logger = logging.getLogger(__name__)
|
10 |
+
|
11 |
+
|
12 |
+
class MisinformationPredictor:
|
13 |
+
def __init__(
|
14 |
+
self,
|
15 |
+
model_path,
|
16 |
+
device="cuda" if torch.cuda.is_available() else "cpu",
|
17 |
+
embed_dim=256,
|
18 |
+
num_heads=8,
|
19 |
+
dropout=0.1,
|
20 |
+
hidden_dim=64,
|
21 |
+
num_classes=3,
|
22 |
+
mlp_ratio=4.0,
|
23 |
+
text_input_dim=384,
|
24 |
+
image_input_dim=1024,
|
25 |
+
fused_attn=False,
|
26 |
+
text_encoder="microsoft/deberta-v3-xsmall",
|
27 |
+
):
|
28 |
+
"""
|
29 |
+
Initialize the predictor with a trained model and required encoders.
|
30 |
+
|
31 |
+
Args:
|
32 |
+
model_path: Path to the saved model checkpoint
|
33 |
+
text_encoder: Name/path of the text encoder model
|
34 |
+
device: Device to run inference on
|
35 |
+
Other args: Model architecture parameters
|
36 |
+
"""
|
37 |
+
self.device = torch.device(device)
|
38 |
+
|
39 |
+
# Initialize tokenizer and encoders
|
40 |
+
logger.info("Loading encoders...")
|
41 |
+
self.tokenizer = AutoTokenizer.from_pretrained(text_encoder)
|
42 |
+
self.text_encoder = AutoModel.from_pretrained(text_encoder).to(self.device)
|
43 |
+
self.image_encoder = Swinv2Model.from_pretrained(
|
44 |
+
"microsoft/swinv2-base-patch4-window8-256"
|
45 |
+
).to(self.device)
|
46 |
+
|
47 |
+
# Set encoders to eval mode
|
48 |
+
self.text_encoder.eval()
|
49 |
+
self.image_encoder.eval()
|
50 |
+
|
51 |
+
# Initialize model
|
52 |
+
self.model = MisinformationDetectionModel(
|
53 |
+
text_input_dim=text_input_dim,
|
54 |
+
image_input_dim=image_input_dim,
|
55 |
+
embed_dim=embed_dim,
|
56 |
+
num_heads=num_heads,
|
57 |
+
dropout=dropout,
|
58 |
+
hidden_dim=hidden_dim,
|
59 |
+
num_classes=num_classes,
|
60 |
+
mlp_ratio=mlp_ratio,
|
61 |
+
fused_attn=fused_attn,
|
62 |
+
).to(self.device)
|
63 |
+
|
64 |
+
# Load model weights
|
65 |
+
logger.info(f"Loading model from {model_path}")
|
66 |
+
checkpoint = torch.load(model_path, map_location=self.device)
|
67 |
+
self.model.load_state_dict(checkpoint["model_state_dict"])
|
68 |
+
self.model.eval()
|
69 |
+
|
70 |
+
# Image preprocessing
|
71 |
+
self.image_transform = transforms.Compose(
|
72 |
+
[
|
73 |
+
transforms.Resize((256, 256)),
|
74 |
+
transforms.ToTensor(),
|
75 |
+
transforms.Normalize(
|
76 |
+
mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]
|
77 |
+
),
|
78 |
+
]
|
79 |
+
)
|
80 |
+
|
81 |
+
# Class mapping
|
82 |
+
self.idx_to_label = {0: "support", 1: "not_enough_information", 2: "refute"}
|
83 |
+
|
84 |
+
def process_image(self, image_path):
|
85 |
+
"""Process image from path to tensor."""
|
86 |
+
try:
|
87 |
+
image = Image.open(image_path).convert("RGB")
|
88 |
+
image = self.image_transform(image).unsqueeze(0) # Add batch dimension
|
89 |
+
return image.to(self.device)
|
90 |
+
except Exception as e:
|
91 |
+
logger.error(f"Error processing image {image_path}: {e}")
|
92 |
+
return None
|
93 |
+
|
94 |
+
@torch.no_grad()
|
95 |
+
def evaluate(
|
96 |
+
self, claim_text, claim_image_path, evidence_text, evidence_image_path
|
97 |
+
):
|
98 |
+
"""
|
99 |
+
Evaluate a single claim-evidence pair.
|
100 |
+
|
101 |
+
Args:
|
102 |
+
claim_text (str): The claim text
|
103 |
+
claim_image_path (str): Path to the claim image
|
104 |
+
evidence_text (str): The evidence text
|
105 |
+
evidence_image_path (str): Path to the evidence image
|
106 |
+
|
107 |
+
Returns:
|
108 |
+
dict: Dictionary containing predictions from all modality combinations
|
109 |
+
"""
|
110 |
+
try:
|
111 |
+
# Process text inputs
|
112 |
+
claim_text_inputs = self.tokenizer(
|
113 |
+
claim_text,
|
114 |
+
truncation=True,
|
115 |
+
padding="max_length",
|
116 |
+
max_length=512,
|
117 |
+
return_tensors="pt",
|
118 |
+
).to(self.device)
|
119 |
+
|
120 |
+
evidence_text_inputs = self.tokenizer(
|
121 |
+
evidence_text,
|
122 |
+
truncation=True,
|
123 |
+
padding="max_length",
|
124 |
+
max_length=512,
|
125 |
+
return_tensors="pt",
|
126 |
+
).to(self.device)
|
127 |
+
|
128 |
+
# Get text embeddings
|
129 |
+
claim_text_embeds = self.text_encoder(**claim_text_inputs).last_hidden_state
|
130 |
+
evidence_text_embeds = self.text_encoder(
|
131 |
+
**evidence_text_inputs
|
132 |
+
).last_hidden_state
|
133 |
+
|
134 |
+
# Process image inputs
|
135 |
+
claim_image = self.process_image(claim_image_path)
|
136 |
+
evidence_image = self.process_image(evidence_image_path)
|
137 |
+
|
138 |
+
# Process claim image
|
139 |
+
if claim_image is not None:
|
140 |
+
claim_image_embeds = self.image_encoder(claim_image).last_hidden_state
|
141 |
+
else:
|
142 |
+
logger.warning(
|
143 |
+
"Claim image processing failed, setting embedding to None"
|
144 |
+
)
|
145 |
+
claim_image_embeds = None
|
146 |
+
|
147 |
+
# Process evidence image
|
148 |
+
if evidence_image is not None:
|
149 |
+
evidence_image_embeds = self.image_encoder(
|
150 |
+
evidence_image
|
151 |
+
).last_hidden_state
|
152 |
+
else:
|
153 |
+
logger.warning(
|
154 |
+
"Evidence image processing failed, setting embedding to None"
|
155 |
+
)
|
156 |
+
evidence_image_embeds = None
|
157 |
+
|
158 |
+
# Get model predictions
|
159 |
+
(y_t_t, y_t_i), (y_i_t, y_i_i) = self.model(
|
160 |
+
X_t=claim_text_embeds,
|
161 |
+
X_i=claim_image_embeds,
|
162 |
+
E_t=evidence_text_embeds,
|
163 |
+
E_i=evidence_image_embeds,
|
164 |
+
)
|
165 |
+
|
166 |
+
# Process predictions with confidence scores
|
167 |
+
predictions = {}
|
168 |
+
|
169 |
+
def process_output(output, path_name):
|
170 |
+
if output is not None:
|
171 |
+
probs = F.softmax(output, dim=-1)
|
172 |
+
pred_idx = probs.argmax(dim=-1).item()
|
173 |
+
confidence = probs[0][pred_idx].item()
|
174 |
+
return {
|
175 |
+
"label": self.idx_to_label[pred_idx],
|
176 |
+
"confidence": confidence,
|
177 |
+
"probabilities": {
|
178 |
+
self.idx_to_label[i]: p.item()
|
179 |
+
for i, p in enumerate(probs[0])
|
180 |
+
},
|
181 |
+
}
|
182 |
+
return None
|
183 |
+
|
184 |
+
predictions["text_text"] = process_output(y_t_t, "text_text")
|
185 |
+
predictions["text_image"] = process_output(y_t_i, "text_image")
|
186 |
+
predictions["image_text"] = process_output(y_i_t, "image_text")
|
187 |
+
predictions["image_image"] = process_output(y_i_i, "image_image")
|
188 |
+
|
189 |
+
return {
|
190 |
+
path: pred["label"] if pred else None
|
191 |
+
for path, pred in predictions.items()
|
192 |
+
}
|
193 |
+
|
194 |
+
except Exception as e:
|
195 |
+
logger.error(f"Error during evaluation: {e}")
|
196 |
+
return None
|
197 |
+
|
198 |
+
|
199 |
+
if __name__ == "__main__":
|
200 |
+
# Example usage
|
201 |
+
logging.basicConfig(level=logging.INFO)
|
202 |
+
|
203 |
+
predictor = MisinformationPredictor(model_path="ckpts/model.pt", device="cpu")
|
204 |
+
|
205 |
+
# Example prediction
|
206 |
+
predictions = predictor.evaluate(
|
207 |
+
claim_text="Musician Kodak Black was shot outside of a nightclub in Florida in December 2016.",
|
208 |
+
claim_image_path="./data/raw/factify/extracted/images/test/0_claim.jpg",
|
209 |
+
evidence_text="On 26 December 2016, the web site Gummy Post published an article claiming \
|
210 |
+
that musician Kodak Black was shot outside a nightclub in Florida. \
|
211 |
+
This article is a hoax. While Gummy Post cited a 'police report', no records exist \
|
212 |
+
of any shooting involving Kodak Black (real name Dieuson Octave) in Florida during December 2016. \
|
213 |
+
Additionally, the video Gummy Post shared as evidence showed an unrelated crime scene.",
|
214 |
+
evidence_image_path="./data/raw/factify/extracted/images/test/0_evidence.jpg",
|
215 |
+
)
|
216 |
+
|
217 |
+
print(predictions)
|
218 |
+
# Print predictions
|
219 |
+
# if predictions:
|
220 |
+
# print("\nPredictions:")
|
221 |
+
# for path, pred in predictions.items():
|
222 |
+
# if pred:
|
223 |
+
# print(f"\n{path}:")
|
224 |
+
# print(f" Label: {pred['label']}")
|
225 |
+
# print(f" Confidence: {pred['confidence']:.4f}")
|
226 |
+
# print(" Probabilities:")
|
227 |
+
# for label, prob in pred["probabilities"].items():
|
228 |
+
# print(f" {label}: {prob:.4f}")
|