lazy loading to try and bypass the 30 min startup limit

api/predict.py

@@ -1,7 +1,6 @@
 import torch
-import re
 import numpy as np
-
+import re
 from transformers import (
     DistilBertTokenizer, DistilBertForSequenceClassification,
     RobertaTokenizer, RobertaForSequenceClassification,
@@ -21,6 +20,8 @@ uncertaintyThreshold = 0.67
 
 CLASS_NAMES = ['genuine', 'fake']
 
+models_loaded = False
+
 def validateText(text):
     if not isinstance(text, str):
         return False
@@ -37,9 +38,9 @@ def cleanReview(text):
     return text.strip()
 
 def loadResources():
-    global models, tokenizers, maxLengths
+    global models, tokenizers, maxLengths, models_loaded
     
-    if len(models) > 0:
+    if models_loaded:
         return
     
     print("loading ensemble models...", flush=True)
@@ -120,10 +121,12 @@ def loadResources():
             print(f"error loading model {i}: {str(e)}", flush=True)
             raise
     
+    models_loaded = True
     print("all ensemble models loaded", flush=True)
 
 def ensemblePredict(text):
-    loadResources()
+    if not models_loaded:
+        loadResources()
     
     if not isinstance(text, str):
         text = str(text)