Spaces:
Runtime error
Runtime error
gabrielchua
commited on
Commit
•
073508d
1
Parent(s):
3df2176
add demo
Browse files- app.py +191 -0
- requirements.txt +8 -0
app.py
ADDED
@@ -0,0 +1,191 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
"""
|
2 |
+
app.py
|
3 |
+
"""
|
4 |
+
|
5 |
+
import json
|
6 |
+
from typing import Any, Dict, List, Optional
|
7 |
+
|
8 |
+
import gradio as gr
|
9 |
+
import numpy as np
|
10 |
+
import pandas as pd
|
11 |
+
import spaces
|
12 |
+
import torch
|
13 |
+
from huggingface_hub import hf_hub_download
|
14 |
+
from sklearn.linear_model import RidgeClassifier
|
15 |
+
from transformers import AutoModel, AutoTokenizer
|
16 |
+
|
17 |
+
# Define the list of available Lionguard models
|
18 |
+
Lionguard_models: List[str] = [
|
19 |
+
"dsaidgovsg/Lionguard-binary-v1.0",
|
20 |
+
"dsaidgovsg/Lionguard-harassment-v1.0",
|
21 |
+
"dsaidgovsg/Lionguard-hateful-v1.0",
|
22 |
+
"dsaidgovsg/Lionguard-public_harm-v1.0",
|
23 |
+
"dsaidgovsg/Lionguard-self_harm-v1.0",
|
24 |
+
"dsaidgovsg/Lionguard-sexual-v1.0",
|
25 |
+
"dsaidgovsg/Lionguard-toxic-v1.0",
|
26 |
+
"dsaidgovsg/Lionguard-violent-v1.0",
|
27 |
+
]
|
28 |
+
|
29 |
+
def load_config(model_repo: str) -> Dict[str, Any]:
|
30 |
+
"""
|
31 |
+
Load the configuration for a given model repository.
|
32 |
+
|
33 |
+
Args:
|
34 |
+
model_repo (str): The model repository name.
|
35 |
+
|
36 |
+
Returns:
|
37 |
+
Dict[str, Any]: The configuration dictionary.
|
38 |
+
"""
|
39 |
+
config_path = hf_hub_download(repo_id=model_repo, filename="config.json")
|
40 |
+
with open(config_path, 'r') as f:
|
41 |
+
return json.load(f)
|
42 |
+
|
43 |
+
def load_all_configs() -> Dict[str, Dict[str, Any]]:
|
44 |
+
"""
|
45 |
+
Load configurations for all Lionguard models.
|
46 |
+
|
47 |
+
Returns:
|
48 |
+
Dict[str, Dict[str, Any]]: A dictionary of model configurations.
|
49 |
+
"""
|
50 |
+
model_configs = {}
|
51 |
+
for model_repo in Lionguard_models:
|
52 |
+
model_configs[model_repo] = load_config(model_repo)
|
53 |
+
print("All model configurations loaded.")
|
54 |
+
return model_configs
|
55 |
+
|
56 |
+
@spaces.GPU
|
57 |
+
def get_embeddings(device: str, data: List[str], config: Dict[str, Any]) -> np.ndarray:
|
58 |
+
"""
|
59 |
+
Generate embeddings for the input data using the specified model configuration.
|
60 |
+
|
61 |
+
Args:
|
62 |
+
device (str): The device to use for computations.
|
63 |
+
data (List[str]): The input text data.
|
64 |
+
config (Dict[str, Any]): The model configuration.
|
65 |
+
|
66 |
+
Returns:
|
67 |
+
np.ndarray: The generated embeddings.
|
68 |
+
"""
|
69 |
+
tokenizer = AutoTokenizer.from_pretrained(config['tokenizer'])
|
70 |
+
model = AutoModel.from_pretrained(config['embedding_model'])
|
71 |
+
model.eval()
|
72 |
+
model.to(device)
|
73 |
+
|
74 |
+
batch_size = config['batch_size']
|
75 |
+
num_batches = int(np.ceil(len(data)/batch_size))
|
76 |
+
output = []
|
77 |
+
for i in range(num_batches):
|
78 |
+
sentences = data[i*batch_size:(i+1)*batch_size]
|
79 |
+
encoded_input = tokenizer(sentences, max_length=config['max_length'], padding=True, truncation=True, return_tensors='pt')
|
80 |
+
encoded_input.to(device)
|
81 |
+
with torch.no_grad():
|
82 |
+
model_output = model(**encoded_input)
|
83 |
+
sentence_embeddings = model_output[0][:, 0]
|
84 |
+
sentence_embeddings = torch.nn.functional.normalize(sentence_embeddings, p=2, dim=1)
|
85 |
+
output.extend(sentence_embeddings.cpu().numpy())
|
86 |
+
return np.array(output)
|
87 |
+
|
88 |
+
def set_model_atttributes(model: RidgeClassifier, attributes: Dict[str, Any]) -> RidgeClassifier:
|
89 |
+
"""
|
90 |
+
Set the attributes for the Ridge Classifier model.
|
91 |
+
|
92 |
+
Args:
|
93 |
+
model (RidgeClassifier): The Ridge Classifier model.
|
94 |
+
attributes (Dict[str, Any]): The attributes to set.
|
95 |
+
|
96 |
+
Returns:
|
97 |
+
RidgeClassifier: The updated Ridge Classifier model.
|
98 |
+
"""
|
99 |
+
model.coef_ = np.array(attributes['coef_'])
|
100 |
+
model.intercept_ = np.array(attributes['intercept_'])
|
101 |
+
model.n_features_in_ = np.array(attributes['n_features_in_'])
|
102 |
+
return model
|
103 |
+
|
104 |
+
def convert_decision_to_proba(d: np.ndarray) -> np.ndarray:
|
105 |
+
"""
|
106 |
+
Convert decision function values to probabilities.
|
107 |
+
|
108 |
+
Args:
|
109 |
+
d (np.ndarray): The decision function values.
|
110 |
+
|
111 |
+
Returns:
|
112 |
+
np.ndarray: The converted probabilities.
|
113 |
+
"""
|
114 |
+
d = np.c_[-d, d]
|
115 |
+
probs = np.exp(d) / np.sum(np.exp(d), axis=1, keepdims=True)
|
116 |
+
return probs
|
117 |
+
|
118 |
+
def predict_all(text: str, model_configs: Dict[str, Dict[str, Any]]) -> Optional[pd.DataFrame]:
|
119 |
+
"""
|
120 |
+
Predict probabilities for all Lionguard models given an input text.
|
121 |
+
|
122 |
+
Args:
|
123 |
+
text (str): The input text to predict on.
|
124 |
+
model_configs (Dict[str, Dict[str, Any]]): The configurations for all models.
|
125 |
+
|
126 |
+
Returns:
|
127 |
+
pd.DataFrame: A DataFrame containing prediction probabilities for each category.
|
128 |
+
"""
|
129 |
+
if not text.strip():
|
130 |
+
return None
|
131 |
+
|
132 |
+
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
133 |
+
|
134 |
+
first_model = next(iter(model_configs))
|
135 |
+
config = model_configs[first_model]
|
136 |
+
embeddings = get_embeddings(device, [text], config)
|
137 |
+
embeddings_df = pd.DataFrame(embeddings)
|
138 |
+
|
139 |
+
results = []
|
140 |
+
for model_repo in Lionguard_models:
|
141 |
+
if model_repo not in model_configs:
|
142 |
+
print(f"Configuration for {model_repo} not found. Skipping...")
|
143 |
+
continue
|
144 |
+
|
145 |
+
config = model_configs[model_repo]
|
146 |
+
|
147 |
+
model_fp = hf_hub_download(repo_id=model_repo, filename=config['model_name'])
|
148 |
+
with open(model_fp, 'r') as json_file:
|
149 |
+
model_params = json.load(json_file)
|
150 |
+
|
151 |
+
model = RidgeClassifier()
|
152 |
+
model_attributes = model_params["attributes"]
|
153 |
+
model_params.pop("attributes", None)
|
154 |
+
model.set_params(**model_params)
|
155 |
+
model = set_model_atttributes(model, model_attributes)
|
156 |
+
|
157 |
+
preds = convert_decision_to_proba(model.decision_function(embeddings_df))[:,1]
|
158 |
+
|
159 |
+
model_name = model_repo.split('/')[-1].split('-')[1]
|
160 |
+
results.append({"Category": model_name, "Probability": float(preds[0])})
|
161 |
+
|
162 |
+
result_df = pd.DataFrame(results).sort_values("Probability", ascending=False)
|
163 |
+
|
164 |
+
if result_df.shape[0] > 0:
|
165 |
+
return result_df
|
166 |
+
else:
|
167 |
+
return None
|
168 |
+
|
169 |
+
def create_interface(model_configs: Dict[str, Dict[str, Any]]) -> gr.Interface:
|
170 |
+
"""
|
171 |
+
Create the Gradio interface for the Lionguard demo.
|
172 |
+
|
173 |
+
Args:
|
174 |
+
model_configs (Dict[str, Dict[str, Any]]): The configurations for all models.
|
175 |
+
|
176 |
+
Returns:
|
177 |
+
gr.Interface: The Gradio interface object.
|
178 |
+
"""
|
179 |
+
return gr.Interface(
|
180 |
+
fn=lambda text: predict_all(text, model_configs),
|
181 |
+
inputs=gr.Textbox(lines=3, placeholder="Enter text here..."),
|
182 |
+
outputs=gr.DataFrame(label="Prediction Probabilities"),
|
183 |
+
title="🦁 Lionguard Demo",
|
184 |
+
description="Lionguard is a Singapore-contextualized moderation classifier that can serve against unsafe LLM outputs.",
|
185 |
+
allow_flagging="never"
|
186 |
+
)
|
187 |
+
|
188 |
+
if __name__ == "__main__":
|
189 |
+
model_configs = load_all_configs()
|
190 |
+
iface = create_interface(model_configs)
|
191 |
+
iface.launch()
|
requirements.txt
ADDED
@@ -0,0 +1,8 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
gradio==4.22.0
|
2 |
+
huggingface-hub==0.20.2
|
3 |
+
numpy==1.24.4
|
4 |
+
pandas==2.2.1
|
5 |
+
torch==2.1.2
|
6 |
+
scikit-learn==1.3.0
|
7 |
+
spaces==0.28.3
|
8 |
+
transformers==4.37.2
|