Update app.py

app.py

@@ -2,7 +2,6 @@ import sentencepiece as spm
 import numpy as np
 import tensorflow as tf
 from tensorflow.keras.preprocessing.sequence import pad_sequences
-import re
 from valx import detect_profanity, detect_hate_speech
 import gradio as gr
 
@@ -13,31 +12,27 @@ model = tf.keras.models.load_model("dungen_dev_preview_model.keras")
 
 max_seq_len = 25
 
-def generate_text(seed_text, next_words=5, temperature=1.0):
-    seed_text = seed_text.lowercase() + ' | '
+def generate_text(seed_text, next_words=30, temperature=0.5):
+    seed_text = seed_text.lower() + ' | '
     hate_speech = detect_hate_speech(seed_text)
     profanity = detect_profanity([seed_text], language='All')
-    
-    if len(profanity) > 0:
+
+    if profanity:  # Simplified check
         gr.Warning("Profanity detected in the prompt, using the default prompt.")
         seed_text = 'game name | '
-    else:
-        if hate_speech == ['Hate Speech']:
-            gr.Warning('Hate speech detected in the seed text, using an empty seed text.')
-            seed_text = 'game name | '
-        elif hate_speech == ['Offensive Speech']:
-            gr.Warning('Offensive speech detected in the seed text, using an empty seed text.')
-            seed_text = 'game name | '
-            
+    elif hate_speech and hate_speech[0] in ['Hate Speech', 'Offensive Speech']: # Simplified check
+        gr.Warning('Harmful speech detected in the seed text, using default prompt.')
+        seed_text = 'game name | '
+
     generated_text = seed_text
     for _ in range(next_words):
         token_list = sp.encode_as_ids(generated_text)
         token_list = pad_sequences([token_list], maxlen=max_seq_len - 1, padding='pre')
         predicted = model.predict(token_list, verbose=0)[0]
 
-        # Apply temperature
+        # Apply temperature with numerical stability fix
         predicted = np.asarray(predicted).astype("float64")
-        predicted = np.log(predicted) / temperature
+        predicted = np.log(predicted + 1e-8) / temperature # Add small epsilon to prevent log(0)
         exp_preds = np.exp(predicted)
         predicted = exp_preds / np.sum(exp_preds)
 
@@ -49,30 +44,25 @@ def generate_text(seed_text, next_words=5, temperature=1.0):
             break
 
     decoded = sp.decode_pieces(sp.encode_as_pieces(generated_text))
-    decoded = decoded.replace("</s>", "")
-    decoded = decoded.replace("<unk>", "")
-    cleaned_text = decoded.strip()
-
-    hate_speech2 = detect_hate_speech(cleaned_text)
-    profanity2 = detect_profanity([cleaned_text], language='All')
+    decoded = decoded.replace("</s>", "").replace("<unk>", "").strip() # Combined replace and strip
     
-    if len(profanity2) > 0:
-        gr.Warning("Flagged potentially harmful output.")
-        cleaned_text = 'Flagged Output'
-    else:
-        if hate_speech2 == ['Hate Speech']:
-            gr.Warning('Flagged potentially harmful output.')
-            cleaned_text = 'Flagged Output'
-        elif hate_speech2 == ['Offensive Speech']:
-            gr.Warning('Flagged potentially harmful output.')
-            cleaned_text = 'Flagged Output'
+    hate_speech2 = detect_hate_speech(decoded)
+    profanity2 = detect_profanity([decoded], language='All')
     
-    return cleaned_text
+    if profanity2 or (hate_speech2 and hate_speech2[0] in ['Hate Speech', 'Offensive Speech']): #Combined check
+        gr.Warning("Flagged potentially harmful output.")
+        decoded = 'Flagged Output'
+
+    return decoded
 
 demo = gr.Interface(
     fn=generate_text,
-    inputs=[label="Prompt", value="a female character name", max_lines=1), gr.Slider(1,100, step=1, label='Next Words', value=30), gr.Slider(0.1, 1, value=0.5, label='Temperature', info='Controls randomness of generation, higher values = more creative, lower values = more probalistic')],
-    outputs=[gr.Dataframe(row_count = (2, "dynamic"), col_count=(1, "fixed"), label="Generated Names", headers=["Names"])],
+    inputs=[
+        gr.Textbox(label="Prompt", value="a female character name", max_lines=1),
+        gr.Slider(1, 100, step=1, label='Next Words', value=30),
+        gr.Slider(0.1, 1, value=0.5, label='Temperature', info='Controls randomness of generation, higher values = more creative, lower values = more probalistic')
+    ],
+    outputs=gr.Textbox(label="Generated Names"), #Simplified output
     title='Dungen Dev - Name Generator',
     description='A prompt-based name generator for game developers.'
 )