Update app.py

app.py

@@ -4,10 +4,9 @@ from transformers import T5Tokenizer, T5ForConditionalGeneration
 from transformers.modeling_outputs import BaseModelOutput
 
 # ==========================================
-# 1. SETUP: Load FLAN-T5 (Smarter than BART)
+# 1. SETUP
 # ==========================================
-# Options: "google/flan-t5-base" (fast) or "google/flan-t5-large" (smart)
-model_name = "google/flan-t5-large" # "google/flan-t5-large" 
+model_name = "google/flan-t5-large"
 
 print(f"Loading {model_name}...")
 tokenizer = T5Tokenizer.from_pretrained(model_name)
@@ -15,110 +14,92 @@ model = T5ForConditionalGeneration.from_pretrained(model_name)
 model.eval() 
 
 # ==========================================
-# 2. CORE LOGIC
+# 2. LOGIC
 # ==========================================
 
 def text_to_embedding(text):
-    # T5 requires a "prompt" structure usually, but for raw encoding, just text is fine
     inputs = tokenizer(text, return_tensors="pt")
-    
     with torch.no_grad():
-        # Get the output of the T5 Encoder
         encoder_outputs = model.encoder(**inputs)
-        
     return encoder_outputs.last_hidden_state
 
 def embedding_to_text(embedding_tensor):
-    # Wrap the vector so T5's decoder accepts it
     encoder_outputs_wrapped = BaseModelOutput(last_hidden_state=embedding_tensor)
-    
     with torch.no_grad():
-        # Generate text using the decoder
         generated_ids = model.generate(
             encoder_outputs=encoder_outputs_wrapped,
-            max_length=100, # T5 can handle slightly longer outputs
+            max_length=100,
             num_beams=5,
-            repetition_penalty=2.5, # T5 likes to repeat, this penalty stops "And cats. And cats."
+            repetition_penalty=2.5,
             early_stopping=True
         )
-    
-    decoded_text = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
-    return decoded_text
-
-# ==========================================
-# 3. GRADIO UI (Same as before)
-# ==========================================
-
+    return tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
 
 def run_mixing(text1, text2):
-    if not text1 or not text2:
-        return "Please enter two sentences."
-
-    print(f"Mixing: '{text1}' + '{text2}'")
+    if not text1 or not text2: return "Please enter two sentences."
+    
     v1 = text_to_embedding(text1)
     v2 = text_to_embedding(text2)
-
-    # Truncate to match lengths
+    
+    # Truncate to min length
     min_len = min(v1.shape[1], v2.shape[1])
     v1 = v1[:, :min_len, :]
     v2 = v2[:, :min_len, :]
-
-    # Average them
+    
+    # 50/50 Average
     v_mixed = (v1 + v2) / 2.0
-
     return embedding_to_text(v_mixed)
 
-# Update the mixing function to accept a 'ratio' (0.0 to 1.0)
 def run_weighted_mixing(text1, text2, mix_ratio):
-    if not text1 or not text2:
-        return "Please enter two sentences."
+    if not text1 or not text2: return "Please enter two sentences."
 
-    # 1. Get vectors
     v1 = text_to_embedding(text1)
     v2 = text_to_embedding(text2)
 
-    # 2. Align lengths
     min_len = min(v1.shape[1], v2.shape[1])
     v1 = v1[:, :min_len, :]
     v2 = v2[:, :min_len, :]
 
-    # 3. Weighted Average
-    # if ratio is 0.0 -> 100% Text1
-    # if ratio is 1.0 -> 100% Text2
-    # if ratio is 0.5 -> 50/50
+    # Weighted Average formula
     v_mixed = (v1 * (1 - mix_ratio)) + (v2 * mix_ratio)
-
     return embedding_to_text(v_mixed)
 
+# ==========================================
+# 3. GRADIO UI (FIXED STRUCTURE)
+# ==========================================
 
 with gr.Blocks(title="FLAN-T5 Latent Explorer", theme=gr.themes.Soft()) as demo:
     gr.Markdown("# 🧠 FLAN-T5 Latent Space Mixer")
-    gr.Markdown(f"Running `{model_name}`. This model is instruction-tuned and much smarter than BART.")
-    
-    with gr.Row():
-        with gr.Column():
-            t1 = gr.Textbox(label="Concept A", value="The King is powerful.")
-            t2 = gr.Textbox(label="Concept B", value="The woman is beautiful.")
-            btn = gr.Button("Mix Vectors", variant="primary")
-        with gr.Column():
-            out = gr.Textbox(label="Result", lines=2)
-
-    btn.click(run_mixing, inputs=[t1, t2], outputs=out)
-    # Update the mixing function to accept a 'ratio' (0.0 to 1.0)
-
-# In the Gradio UI section:
-with gr.TabItem("2. Vector Mixing"):
-    with gr.Row():
-        t1 = gr.Textbox(label="Start Sentence", value="The dog is happy.")
-        t2 = gr.Textbox(label="End Sentence", value="The cat is angry.")
-    
-    # Add a slider
-    slider = gr.Slider(minimum=0.0, maximum=1.0, value=0.5, step=0.1, label="Mixing Ratio (Left = Start, Right = End)")
-    
-    btn_mix = gr.Button("Morph")
-    out = gr.Textbox(label="Result")
+    gr.Markdown(f"Running `{model_name}`.")
 
-    btn_mix.click(run_weighted_mixing, inputs=[t1, t2, slider], outputs=out)
+    # We need a Tabs container to hold the TabItems
+    with gr.Tabs():
+        
+        # --- TAB 1: 50/50 MIX ---
+        with gr.TabItem("1. Simple Mix (50/50)"):
+            with gr.Row():
+                with gr.Column():
+                    t1_simple = gr.Textbox(label="Concept A", value="The King is powerful.")
+                    t2_simple = gr.Textbox(label="Concept B", value="The woman is beautiful.")
+                    btn_simple = gr.Button("Mix Vectors", variant="primary")
+                with gr.Column():
+                    out_simple = gr.Textbox(label="Result", lines=2)
+
+            btn_simple.click(run_mixing, inputs=[t1_simple, t2_simple], outputs=out_simple)
+
+        # --- TAB 2: WEIGHTED SLIDER ---
+        with gr.TabItem("2. Weighted Morph (Slider)"):
+            gr.Markdown("Slide between the two sentences to see how the meaning shifts.")
+            with gr.Row():
+                t1_morph = gr.Textbox(label="Start Sentence", value="The dog is happy.")
+                t2_morph = gr.Textbox(label="End Sentence", value="The cat is angry.")
+            
+            slider = gr.Slider(minimum=0.0, maximum=1.0, value=0.5, step=0.1, label="Mixing Ratio (Left = Start, Right = End)")
+            
+            btn_morph = gr.Button("Morph", variant="primary")
+            out_morph = gr.Textbox(label="Result")
+
+            btn_morph.click(run_weighted_mixing, inputs=[t1_morph, t2_morph, slider], outputs=out_morph)
 
 if __name__ == "__main__":
     demo.launch()