Spaces:
Running
Running
Upload 7 files
Browse files- .gitattributes +1 -0
- Dockerfile +34 -0
- app.py +103 -0
- hs_gru.h5 +3 -0
- hs_gru.keras +3 -0
- index.html +194 -0
- requirements.txt +7 -0
- tokenizerpkl_gru.pkl +3 -0
.gitattributes
CHANGED
@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
|
|
33 |
*.zip filter=lfs diff=lfs merge=lfs -text
|
34 |
*.zst filter=lfs diff=lfs merge=lfs -text
|
35 |
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
|
|
|
33 |
*.zip filter=lfs diff=lfs merge=lfs -text
|
34 |
*.zst filter=lfs diff=lfs merge=lfs -text
|
35 |
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
36 |
+
hs_gru.keras filter=lfs diff=lfs merge=lfs -text
|
Dockerfile
ADDED
@@ -0,0 +1,34 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
FROM python:3.10-slim
|
2 |
+
|
3 |
+
# Set environment variables for Hugging Face cache
|
4 |
+
ENV HF_HOME=/code/hf_cache
|
5 |
+
ENV TRANSFORMERS_CACHE=/code/hf_cache
|
6 |
+
ENV PYTHONDONTWRITEBYTECODE=1
|
7 |
+
ENV PYTHONUNBUFFERED=1
|
8 |
+
|
9 |
+
# Create working directory
|
10 |
+
WORKDIR /code
|
11 |
+
|
12 |
+
# System dependencies
|
13 |
+
RUN apt-get update && apt-get install -y \
|
14 |
+
git \
|
15 |
+
wget \
|
16 |
+
build-essential \
|
17 |
+
&& rm -rf /var/lib/apt/lists/*
|
18 |
+
|
19 |
+
# Install Python packages
|
20 |
+
COPY requirements.txt .
|
21 |
+
RUN pip install --no-cache-dir --upgrade pip \
|
22 |
+
&& pip install --no-cache-dir -r requirements.txt
|
23 |
+
|
24 |
+
# Copy app code
|
25 |
+
COPY . .
|
26 |
+
|
27 |
+
# Create model cache directory
|
28 |
+
RUN mkdir -p /code/hf_cache
|
29 |
+
|
30 |
+
# Expose FastAPI port
|
31 |
+
EXPOSE 7860
|
32 |
+
|
33 |
+
# Start the app
|
34 |
+
CMD ["uvicorn", "app:app", "--host", "0.0.0.0", "--port", "7860"]
|
app.py
ADDED
@@ -0,0 +1,103 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from fastapi import FastAPI, Request, HTTPException
|
2 |
+
from fastapi.responses import HTMLResponse
|
3 |
+
from fastapi.staticfiles import StaticFiles
|
4 |
+
from fastapi.middleware.cors import CORSMiddleware
|
5 |
+
from pydantic import BaseModel
|
6 |
+
import tensorflow as tf
|
7 |
+
import pickle
|
8 |
+
from tensorflow.keras.preprocessing.sequence import pad_sequences
|
9 |
+
from transformers import AutoTokenizer, AutoModelForSequenceClassification
|
10 |
+
import torch
|
11 |
+
from fastapi.responses import JSONResponse
|
12 |
+
# Initialize FastAPI
|
13 |
+
app = FastAPI()
|
14 |
+
|
15 |
+
# Load GRU model and tokenizer
|
16 |
+
gru_model = tf.keras.models.load_model('hs_gru.h5')
|
17 |
+
with open('tokenizerpkl_gru.pkl', 'rb') as f:
|
18 |
+
gru_tokenizer = pickle.load(f)
|
19 |
+
gru_maxlen = 100
|
20 |
+
|
21 |
+
# Load RoBERTa model
|
22 |
+
# Load RoBERTa model
|
23 |
+
roberta_model_name = "facebook/roberta-hate-speech-dynabench-r4-target"
|
24 |
+
roberta_tokenizer = AutoTokenizer.from_pretrained(roberta_model_name)
|
25 |
+
if roberta_tokenizer.pad_token is None:
|
26 |
+
roberta_tokenizer.add_special_tokens({'pad_token': '[PAD]'})
|
27 |
+
roberta_model = AutoModelForSequenceClassification.from_pretrained(roberta_model_name)
|
28 |
+
roberta_model.resize_token_embeddings(len(roberta_tokenizer))
|
29 |
+
|
30 |
+
#load toxigen-hatebert model
|
31 |
+
toxigen_model_name = "tomh/toxigen_roberta"
|
32 |
+
toxigen_tokenizer = AutoTokenizer.from_pretrained(toxigen_model_name)
|
33 |
+
if toxigen_tokenizer.pad_token is None:
|
34 |
+
toxigen_tokenizer.add_special_tokens({'pad_token': '[PAD]'})
|
35 |
+
toxigen_model = AutoModelForSequenceClassification.from_pretrained(toxigen_model_name)
|
36 |
+
toxigen_model.resize_token_embeddings(len(toxigen_tokenizer))
|
37 |
+
|
38 |
+
# Enable CORS
|
39 |
+
app.add_middleware(
|
40 |
+
CORSMiddleware,
|
41 |
+
allow_origins=["*"],
|
42 |
+
allow_credentials=True,
|
43 |
+
allow_methods=["*"],
|
44 |
+
allow_headers=["*"],
|
45 |
+
)
|
46 |
+
|
47 |
+
# Mount static directory
|
48 |
+
# app.mount("/static", StaticFiles(directory="static"), name="static")
|
49 |
+
|
50 |
+
# Pydantic input model
|
51 |
+
class TextInput(BaseModel):
|
52 |
+
text: str
|
53 |
+
|
54 |
+
@app.get("/", response_class=HTMLResponse)
|
55 |
+
def read_root():
|
56 |
+
with open("index.html", "r") as f:
|
57 |
+
return f.read()
|
58 |
+
|
59 |
+
@app.get("/health")
|
60 |
+
def health_check():
|
61 |
+
return {"message": "Hate Speech Detection API is running!"}
|
62 |
+
|
63 |
+
@app.post("/predict")
|
64 |
+
def predict_ensemble(input: TextInput):
|
65 |
+
try:
|
66 |
+
text = input.text
|
67 |
+
# print(f"Received input: {input.text}")
|
68 |
+
|
69 |
+
# ----- GRU Prediction -----
|
70 |
+
seq = gru_tokenizer.texts_to_sequences([text])
|
71 |
+
padded = pad_sequences(seq, maxlen=gru_maxlen, padding='post')
|
72 |
+
gru_prob = float(gru_model.predict(padded)[0][0])
|
73 |
+
|
74 |
+
# ----- RoBERTa Prediction -----
|
75 |
+
inputs_roberta = roberta_tokenizer(text, return_tensors="pt", truncation=True, padding=True)
|
76 |
+
with torch.no_grad():
|
77 |
+
logits_roberta = roberta_model(**inputs_roberta).logits
|
78 |
+
probs_roberta = torch.nn.functional.softmax(logits_roberta, dim=1)
|
79 |
+
roberta_prob = float(probs_roberta[0][1].item())
|
80 |
+
|
81 |
+
# -----toxigen -hatebert Prediction -----
|
82 |
+
inputs_toxigen = toxigen_tokenizer(text, return_tensors="pt", truncation=True, padding=True)
|
83 |
+
with torch.no_grad():
|
84 |
+
logits_toxigen = toxigen_model(**inputs_toxigen).logits
|
85 |
+
probs_toxigen = torch.nn.functional.softmax(logits_toxigen, dim=1)
|
86 |
+
toxigen_prob = float(probs_toxigen[0][1].item())
|
87 |
+
|
88 |
+
# ----- Weighted Ensemble -----
|
89 |
+
final_score = (0.3 * gru_prob) + (0.4 * roberta_prob) + (0.3 * toxigen_prob)
|
90 |
+
label = "Hate Speech" if final_score > 0.5 else "Not Hate Speech"
|
91 |
+
|
92 |
+
return {
|
93 |
+
# "text": text,
|
94 |
+
"gru_prob": round(gru_prob, 4),
|
95 |
+
"roberta_prob": round(roberta_prob, 4),
|
96 |
+
"toxigen_prob": round(toxigen_prob, 4),
|
97 |
+
"final_score": round(final_score, 4),
|
98 |
+
"prediction": label
|
99 |
+
}
|
100 |
+
|
101 |
+
except Exception as e:
|
102 |
+
print(f"Error during prediction: {str(e)}")
|
103 |
+
return JSONResponse(status_code=500, content={"detail": str(e)})
|
hs_gru.h5
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:d64d7840832a93d75389d7e15450bcf1914b6b5be1b23dd55543deb131c11528
|
3 |
+
size 26368672
|
hs_gru.keras
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:7699a8f4fbe4194d72813727483e91cccb8b3ccdcf7d234a7811bff043c0d4e4
|
3 |
+
size 26367562
|
index.html
ADDED
@@ -0,0 +1,194 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
<!DOCTYPE html>
|
2 |
+
<html lang="en">
|
3 |
+
<head>
|
4 |
+
<meta charset="UTF-8">
|
5 |
+
<title>Hate Speech Detection</title>
|
6 |
+
<style>
|
7 |
+
body {
|
8 |
+
font-family: Arial, sans-serif;
|
9 |
+
background: #f9f9f9;
|
10 |
+
color: #000;
|
11 |
+
margin: 0;
|
12 |
+
padding: 20px;
|
13 |
+
display: flex;
|
14 |
+
flex-direction: column;
|
15 |
+
align-items: center;
|
16 |
+
}
|
17 |
+
|
18 |
+
body.dark-mode {
|
19 |
+
background: #272626;
|
20 |
+
color: #fff;
|
21 |
+
}
|
22 |
+
|
23 |
+
h1 {
|
24 |
+
font-size: 40px;
|
25 |
+
margin-bottom: 10px;
|
26 |
+
}
|
27 |
+
|
28 |
+
.subtext {
|
29 |
+
font-size: 16px;
|
30 |
+
margin-bottom: 40px;
|
31 |
+
text-align: center;
|
32 |
+
max-width: 800px;
|
33 |
+
}
|
34 |
+
|
35 |
+
.container {
|
36 |
+
display: flex;
|
37 |
+
justify-content: center;
|
38 |
+
gap: 40px;
|
39 |
+
width: 100%;
|
40 |
+
max-width: 1000px;
|
41 |
+
}
|
42 |
+
|
43 |
+
.left-panel, .right-panel {
|
44 |
+
flex: 1;
|
45 |
+
display: flex;
|
46 |
+
flex-direction: column;
|
47 |
+
}
|
48 |
+
|
49 |
+
textarea {
|
50 |
+
width: 100%;
|
51 |
+
height: 140px;
|
52 |
+
padding: 10px;
|
53 |
+
font-size: 16px;
|
54 |
+
}
|
55 |
+
|
56 |
+
.buttons {
|
57 |
+
display: flex;
|
58 |
+
justify-content: space-between;
|
59 |
+
width: 100%;
|
60 |
+
margin-top: 10px;
|
61 |
+
}
|
62 |
+
|
63 |
+
.buttons button {
|
64 |
+
width: 49.5%;
|
65 |
+
padding: 12px;
|
66 |
+
font-size: 16px;
|
67 |
+
border: none;
|
68 |
+
border-radius: 4px;
|
69 |
+
cursor: pointer;
|
70 |
+
}
|
71 |
+
|
72 |
+
.submit-btn {
|
73 |
+
background: #4CAF50;
|
74 |
+
color: white;
|
75 |
+
}
|
76 |
+
|
77 |
+
.clear-btn {
|
78 |
+
background: #45a049;
|
79 |
+
color: white;
|
80 |
+
}
|
81 |
+
|
82 |
+
.share-btn {
|
83 |
+
width: 100%;
|
84 |
+
padding: 14px;
|
85 |
+
font-size: 16px;
|
86 |
+
background: #007BFF;
|
87 |
+
color: white;
|
88 |
+
border: none;
|
89 |
+
border-radius: 4px;
|
90 |
+
margin-top: auto;
|
91 |
+
}
|
92 |
+
|
93 |
+
.output {
|
94 |
+
background: white;
|
95 |
+
padding: 15px;
|
96 |
+
border-radius: 8px;
|
97 |
+
box-shadow: 0 0 10px rgba(0,0,0,0.1);
|
98 |
+
font-size: 16px;
|
99 |
+
min-height: 120px;
|
100 |
+
margin-bottom: 20px;
|
101 |
+
}
|
102 |
+
|
103 |
+
.dark-mode .output {
|
104 |
+
background: #1e1e1e;
|
105 |
+
color: #fff;
|
106 |
+
}
|
107 |
+
|
108 |
+
.mode-toggle {
|
109 |
+
margin-top: 60px;
|
110 |
+
font-size: 14px;
|
111 |
+
}
|
112 |
+
</style>
|
113 |
+
</head>
|
114 |
+
<body>
|
115 |
+
<h1 id="page-title">Hate Speech Detection</h1>
|
116 |
+
<p class="subtext">
|
117 |
+
This tool uses 3 models (GloVe-based Deep Learning (GRU), RoBERTa (Transformer), ToxiGen-HateBERT (Transformer))
|
118 |
+
to classify hate speech using an ensemble method.
|
119 |
+
</p>
|
120 |
+
|
121 |
+
<div class="container">
|
122 |
+
<div class="left-panel">
|
123 |
+
<form id="hs-form">
|
124 |
+
<textarea id="text-input" placeholder="Enter text here..."></textarea>
|
125 |
+
<div class="buttons">
|
126 |
+
<button type="button" class="clear-btn" onclick="clearText()">Clear Text</button>
|
127 |
+
<button type="submit" class="submit-btn">Submit</button>
|
128 |
+
</div>
|
129 |
+
</form>
|
130 |
+
</div>
|
131 |
+
|
132 |
+
<div class="right-panel">
|
133 |
+
<div class="output" id="output" style="display: none;"></div>
|
134 |
+
<button class="share-btn" onclick="shareText()">Share via Link</button>
|
135 |
+
</div>
|
136 |
+
</div>
|
137 |
+
|
138 |
+
<div class="mode-toggle">
|
139 |
+
<label><input type="checkbox" id="mode-toggle"> Dark Mode</label>
|
140 |
+
</div>
|
141 |
+
|
142 |
+
<script>
|
143 |
+
const outputDiv = document.getElementById("output");
|
144 |
+
const pageTitle = document.getElementById("page-title");
|
145 |
+
|
146 |
+
document.getElementById("hs-form").addEventListener("submit", async function (e) {
|
147 |
+
e.preventDefault();
|
148 |
+
const text = document.getElementById("text-input").value;
|
149 |
+
|
150 |
+
const response = await fetch("/predict", {
|
151 |
+
method: "POST",
|
152 |
+
headers: { "Content-Type": "application/json" },
|
153 |
+
body: JSON.stringify({ text: text }),
|
154 |
+
});
|
155 |
+
|
156 |
+
const result = await response.json();
|
157 |
+
outputDiv.style.display = "block";
|
158 |
+
outputDiv.innerHTML = `
|
159 |
+
<strong>Model 1:</strong> ${result.gru_prob}<br>
|
160 |
+
<strong>Model 2:</strong> ${result.roberta_prob}<br>
|
161 |
+
<strong>Model 3:</strong> ${result.toxigen_prob}<br>
|
162 |
+
<strong>Prediction:</strong> ${result.prediction}
|
163 |
+
`;
|
164 |
+
});
|
165 |
+
|
166 |
+
function clearText() {
|
167 |
+
document.getElementById("text-input").value = "";
|
168 |
+
outputDiv.style.display = "none";
|
169 |
+
outputDiv.innerHTML = "";
|
170 |
+
}
|
171 |
+
|
172 |
+
function shareText() {
|
173 |
+
const text = document.getElementById("text-input").value;
|
174 |
+
const shareUrl = `${window.location.href}?text=${encodeURIComponent(text)}`;
|
175 |
+
navigator.clipboard.writeText(shareUrl);
|
176 |
+
alert("Sharable link copied to clipboard!");
|
177 |
+
}
|
178 |
+
|
179 |
+
document.getElementById("mode-toggle").addEventListener("change", function () {
|
180 |
+
document.body.classList.toggle("dark-mode", this.checked);
|
181 |
+
pageTitle.style.color = this.checked ? "white" : "black";
|
182 |
+
});
|
183 |
+
|
184 |
+
// On load: apply shared text if present
|
185 |
+
window.onload = function () {
|
186 |
+
const urlParams = new URLSearchParams(window.location.search);
|
187 |
+
const sharedText = urlParams.get("text");
|
188 |
+
if (sharedText) {
|
189 |
+
document.getElementById("text-input").value = decodeURIComponent(sharedText);
|
190 |
+
}
|
191 |
+
};
|
192 |
+
</script>
|
193 |
+
</body>
|
194 |
+
</html>
|
requirements.txt
ADDED
@@ -0,0 +1,7 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
fastapi
|
2 |
+
uvicorn
|
3 |
+
transformers
|
4 |
+
torch
|
5 |
+
tensorflow
|
6 |
+
pydantic
|
7 |
+
python-multipart
|
tokenizerpkl_gru.pkl
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:f933c6a038133f7d49d3cda5e5e7a1c930813dafdde2ddd0f6dfc0037d2c8108
|
3 |
+
size 5651170
|