Spaces:
Running
Running
Bayesian model
Browse files
app.py
CHANGED
@@ -1,19 +1,66 @@
|
|
1 |
import gradio as gr
|
|
|
|
|
|
|
|
|
2 |
|
3 |
-
|
4 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
5 |
|
6 |
demo = gr.Interface(
|
7 |
fn=predict,
|
8 |
inputs=[
|
9 |
-
gr.Dropdown(choices=
|
|
|
|
|
|
|
10 |
gr.TextArea(label="Email"),
|
11 |
],
|
12 |
outputs=[gr.Number(label="Spam probability")],
|
13 |
title="Bayes or Spam?",
|
14 |
-
description="Choose your model, and predict if your email is a spam! 📨<br>COMING SOON:
|
15 |
examples=[
|
16 |
-
[
|
|
|
17 |
],
|
18 |
article="This is a demo of the models in the [Bayes or Spam?](https://github.com/tbitai/bayes-or-spam) project.",
|
19 |
)
|
|
|
1 |
import gradio as gr
|
2 |
+
from huggingface_hub import hf_hub_download
|
3 |
+
import json
|
4 |
+
import tensorflow as tf
|
5 |
+
import numpy as np
|
6 |
|
7 |
+
model_probs_path = hf_hub_download(repo_id="tbitai/bayes-enron1-spam", filename="probs.json")
|
8 |
+
with open(model_probs_path) as f:
|
9 |
+
model_probs = json.load(f)
|
10 |
+
|
11 |
+
UNK = '[UNK]'
|
12 |
+
|
13 |
+
def tokenize(text):
|
14 |
+
return tf.keras.preprocessing.text.text_to_word_sequence(text)
|
15 |
+
|
16 |
+
def combine(probs):
|
17 |
+
if any(p == 0 for p in probs):
|
18 |
+
return 0
|
19 |
+
prod = np.prod(probs)
|
20 |
+
neg_prod = np.prod([1 - p for p in probs])
|
21 |
+
if prod + neg_prod == 0: # Still possible due to floating point arithmetic
|
22 |
+
return 0.5 # Assume that prod and neg_prod are equally small
|
23 |
+
return prod / (prod + neg_prod)
|
24 |
+
|
25 |
+
def get_interesting_probs(probs, intr_threshold):
|
26 |
+
return sorted(probs,
|
27 |
+
key=lambda p: abs(p - 0.5),
|
28 |
+
reverse=True)[:intr_threshold]
|
29 |
+
|
30 |
+
def unbias(p):
|
31 |
+
return (2 * p) / (p + 1)
|
32 |
+
|
33 |
+
def predict_bayes(text, intr_threshold, unbiased=False):
|
34 |
+
words = tokenize(text)
|
35 |
+
probs = [model_probs.get(w, model_probs[UNK]) for w in words]
|
36 |
+
if unbiased:
|
37 |
+
probs = [unbias(p) for p in probs]
|
38 |
+
interesting_probs = get_interesting_probs(probs, intr_threshold)
|
39 |
+
return combine(interesting_probs)
|
40 |
+
|
41 |
+
MODELS = [
|
42 |
+
BAYES := "Bayes Enron1 spam",
|
43 |
+
]
|
44 |
+
|
45 |
+
def predict(model, unbiased, intr_threshold, input_txt):
|
46 |
+
if model == BAYES:
|
47 |
+
return predict_bayes(input_txt, unbiased=unbiased, intr_threshold=intr_threshold)
|
48 |
|
49 |
demo = gr.Interface(
|
50 |
fn=predict,
|
51 |
inputs=[
|
52 |
+
gr.Dropdown(choices=MODELS, value=BAYES, label="Model"),
|
53 |
+
gr.Checkbox(label="Unbias", info="Correct Graham's bias?"),
|
54 |
+
gr.Slider(minimum=1, maximum=20, step=1, value=15, label="Interestingness threshold",
|
55 |
+
info="How many of the most interesting words to select in the probability calculation?"),
|
56 |
gr.TextArea(label="Email"),
|
57 |
],
|
58 |
outputs=[gr.Number(label="Spam probability")],
|
59 |
title="Bayes or Spam?",
|
60 |
+
description="Choose and configure your model, and predict if your email is a spam! 📨<br>COMING SOON: NN and LLM models.",
|
61 |
examples=[
|
62 |
+
[BAYES, "enron actuals for june 26, 2000"],
|
63 |
+
[BAYES, "stop the aging clock nerissa"],
|
64 |
],
|
65 |
article="This is a demo of the models in the [Bayes or Spam?](https://github.com/tbitai/bayes-or-spam) project.",
|
66 |
)
|