Create app.py
Browse files
app.py
ADDED
@@ -0,0 +1,154 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import gradio as gr
|
2 |
+
from transformers import AutoTokenizer, AutoModelForSequenceClassification
|
3 |
+
import torch
|
4 |
+
|
5 |
+
class TextDetectionApp:
|
6 |
+
def __init__(self):
|
7 |
+
# Load DeBERTa model and tokenizer
|
8 |
+
self.deberta_tokenizer = AutoTokenizer.from_pretrained("zeyadusf/deberta-DAIGT-MODELS")
|
9 |
+
self.deberta_model = AutoModelForSequenceClassification.from_pretrained("zeyadusf/deberta-DAIGT-MODELS")
|
10 |
+
|
11 |
+
# Load RoBERTa model and tokenizer
|
12 |
+
self.roberta_tokenizer = AutoTokenizer.from_pretrained("zeyadusf/roberta-DAIGT-kaggle")
|
13 |
+
self.roberta_model = AutoModelForSequenceClassification.from_pretrained("zeyadusf/roberta-DAIGT-kaggle")
|
14 |
+
|
15 |
+
# Load Feedforward model
|
16 |
+
self.ff_model = torch.jit.load("model_scripted.pt")
|
17 |
+
|
18 |
+
def api_huggingface(self, text):
|
19 |
+
"""
|
20 |
+
Generate predictions using the DeBERTa and RoBERTa models.
|
21 |
+
|
22 |
+
Args:
|
23 |
+
text (str): The input text to classify.
|
24 |
+
|
25 |
+
Returns:
|
26 |
+
tuple: Predictions from RoBERTa and DeBERTa models.
|
27 |
+
"""
|
28 |
+
# DeBERTa predictions
|
29 |
+
deberta_inputs = self.deberta_tokenizer(text, return_tensors="pt", truncation=True, padding=True)
|
30 |
+
deberta_outputs = self.deberta_model(**deberta_inputs)
|
31 |
+
deberta_logits = deberta_outputs.logits
|
32 |
+
deberta_scores = torch.softmax(deberta_logits, dim=1)
|
33 |
+
deberta_predictions = [
|
34 |
+
{"label": f"LABEL_{i}", "score": score.item()}
|
35 |
+
for i, score in enumerate(deberta_scores[0])
|
36 |
+
]
|
37 |
+
|
38 |
+
# RoBERTa predictions
|
39 |
+
roberta_inputs = self.roberta_tokenizer(text, return_tensors="pt", truncation=True, padding=True)
|
40 |
+
roberta_outputs = self.roberta_model(**roberta_inputs)
|
41 |
+
roberta_logits = roberta_outputs.logits
|
42 |
+
roberta_scores = torch.softmax(roberta_logits, dim=1)
|
43 |
+
roberta_predictions = [
|
44 |
+
{"label": f"LABEL_{i}", "score": score.item()}
|
45 |
+
for i, score in enumerate(roberta_scores[0])
|
46 |
+
]
|
47 |
+
|
48 |
+
return roberta_predictions, deberta_predictions
|
49 |
+
|
50 |
+
def generate_ff_input(self, models_results):
|
51 |
+
"""
|
52 |
+
Generates input features for the Feedforward model from the API output.
|
53 |
+
|
54 |
+
Parameters:
|
55 |
+
models_results (tuple): Tuple containing the results of DeBERTa and RoBERTa models.
|
56 |
+
|
57 |
+
Returns:
|
58 |
+
torch.Tensor: Feedforward model input features tensor.
|
59 |
+
"""
|
60 |
+
roberta, deberta = models_results
|
61 |
+
input_ff = []
|
62 |
+
try:
|
63 |
+
if roberta[0]['label'] == 'LABEL_0':
|
64 |
+
input_ff.append(roberta[0]['score'])
|
65 |
+
input_ff.append(roberta[1]['score'])
|
66 |
+
else:
|
67 |
+
input_ff.append(roberta[1]['score'])
|
68 |
+
input_ff.append(roberta[0]['score'])
|
69 |
+
|
70 |
+
if deberta[0]['label'] == 'LABEL_0':
|
71 |
+
input_ff.append(deberta[0]['score'])
|
72 |
+
input_ff.append(deberta[1]['score'])
|
73 |
+
else:
|
74 |
+
input_ff.append(deberta[1]['score'])
|
75 |
+
input_ff.append(deberta[0]['score'])
|
76 |
+
|
77 |
+
except Exception as e:
|
78 |
+
print(f"Error {e}: The text is long")
|
79 |
+
|
80 |
+
input_ff = torch.tensor(input_ff, dtype=torch.float32)
|
81 |
+
input_ff = input_ff.view(1, -1)
|
82 |
+
return input_ff
|
83 |
+
|
84 |
+
def detect_text(self, text):
|
85 |
+
"""
|
86 |
+
Detects whether the input text is generated or human-written using the Feedforward model.
|
87 |
+
|
88 |
+
Returns:
|
89 |
+
float: The detection result.
|
90 |
+
"""
|
91 |
+
with torch.no_grad():
|
92 |
+
self.output = self.ff_model(self.generate_ff_input(self.api_huggingface(text)))[0][0].item()
|
93 |
+
return self.output
|
94 |
+
|
95 |
+
def classify_text(self, text, model_choice):
|
96 |
+
"""
|
97 |
+
Classifies the input text using the selected model.
|
98 |
+
|
99 |
+
Args:
|
100 |
+
text (str): The input text to classify.
|
101 |
+
model_choice (str): The model to use ('DeBERTa', 'RoBERTa', or 'Feedforward').
|
102 |
+
|
103 |
+
Returns:
|
104 |
+
str: The classification result.
|
105 |
+
"""
|
106 |
+
if model_choice == 'DeBERTa':
|
107 |
+
# Tokenize input
|
108 |
+
inputs = self.deberta_tokenizer(text, return_tensors="pt", truncation=True, padding=True)
|
109 |
+
|
110 |
+
# Run model
|
111 |
+
outputs = self.deberta_model(**inputs)
|
112 |
+
|
113 |
+
# Get classification results
|
114 |
+
logits = outputs.logits
|
115 |
+
predicted_class_id = logits.argmax().item()
|
116 |
+
return f"DeBERTa Prediction: Class {predicted_class_id}"
|
117 |
+
|
118 |
+
elif model_choice == 'RoBERTa':
|
119 |
+
# Tokenize input
|
120 |
+
inputs = self.roberta_tokenizer(text, return_tensors="pt", truncation=True, padding=True)
|
121 |
+
|
122 |
+
# Run model
|
123 |
+
outputs = self.roberta_model(**inputs)
|
124 |
+
|
125 |
+
# Get classification results
|
126 |
+
logits = outputs.logits
|
127 |
+
predicted_class_id = logits.argmax().item()
|
128 |
+
return f"RoBERTa Prediction: Class {predicted_class_id}"
|
129 |
+
|
130 |
+
elif model_choice == 'Feedforward':
|
131 |
+
# Run feedforward detection
|
132 |
+
detection_score = self.detect_text(text)
|
133 |
+
return f"Feedforward Detection Score: {detection_score}"
|
134 |
+
|
135 |
+
else:
|
136 |
+
return "Invalid model selection."
|
137 |
+
|
138 |
+
|
139 |
+
# Initialize the app
|
140 |
+
app = TextDetectionApp()
|
141 |
+
|
142 |
+
# Gradio Interface
|
143 |
+
iface = gr.Interface(
|
144 |
+
fn=app.classify_text,
|
145 |
+
inputs=[
|
146 |
+
gr.Textbox(lines=2, placeholder="Enter your text here..."),
|
147 |
+
gr.Radio(choices=["DeBERTa", "RoBERTa", "Feedforward"], label="Model Choice")
|
148 |
+
],
|
149 |
+
outputs="text",
|
150 |
+
title="Text Classification with Multiple Models",
|
151 |
+
description="Classify text using DeBERTa, RoBERTa, or a custom Feedforward model."
|
152 |
+
)
|
153 |
+
|
154 |
+
iface.launch()
|