update classifier

app.py

@@ -57,8 +57,6 @@ def handle_request(detector_name):
         return return_data(0, '', info)
     except Exception as ex:
         print(datetime.datetime.now().isoformat(), ex, flush=True)
-        import os
-        os._exit(1)
         return return_data(400, 'Bad request', str(ex))
 
 

configs/glimpse.json

@@ -8,7 +8,11 @@
     "prompt": "prompt3",
     "rank_size": 1000,
     "top_k": 5,
-    "linear_k": 0.93,
-    "linear_b": 1.68,
+     "classifier": {
+        "mu0": -3.8289,
+        "sigma0": 2.0131,
+        "mu1": -0.0733,
+        "sigma1": 4.0262
+    },
     "cache_dir": "../cache"
 }

detector_base.py

@@ -6,9 +6,17 @@ import os
 from utils import load_json
 from types import SimpleNamespace
 import numpy as np
+from scipy.stats import norm
+
+
+# Considering balanced classification that p(D0) equals to p(D1), we have
+#   p(D1|x) = p(x|D1) / (p(x|D1) + p(x|D0))
+def compute_prob_norm(x, mu0, sigma0, mu1, sigma1):
+    pdf_value0 = norm.pdf(x, loc=mu0, scale=sigma0)
+    pdf_value1 = norm.pdf(x, loc=mu1, scale=sigma1)
+    prob = pdf_value1 / (pdf_value0 + pdf_value1)
+    return prob
 
-def sigmoid(x):
-    return 1 / (1 + np.exp(-x))
 
 class DetectorBase:
     def __init__(self, config_name):
@@ -30,7 +38,11 @@ class DetectorBase:
 
     def compute_prob(self, text):
         crit, ntoken = self.compute_crit(text)
-        prob = sigmoid(self.config.linear_k * crit + self.config.linear_b)
+        mu0 = self.config.classifier['mu0']
+        sigma0 = self.config.classifier['sigma0']
+        mu1 = self.config.classifier['mu1']
+        sigma1 = self.config.classifier['sigma1']
+        prob = compute_prob_norm(crit, mu0, sigma0, mu1, sigma1)
         return prob, crit, ntoken
 
     def __str__(self):