Update app.py

app.py

@@ -1,4 +1,5 @@
 import gradio as gr
+import torch
 from transformers import pipeline, set_seed
 
 # Setzen eines Seeds für Reproduzierbarkeit
@@ -8,46 +9,62 @@ set_seed(42)
 model = pipeline("text-generation", model="Loewolf/GPT_1")
 tokenizer = model.tokenizer
 
-def generate_text(input_text, temperature, top_k, top_p, length, system_prompt):
-    # Anpassen des Eingabetextes mit System-Prompt, falls vorhanden
-    adjusted_input_text = system_prompt + input_text if system_prompt else input_text
-
+def generate_text(input_text, temp, top_k, top_p, length):
     # Konvertieren des Eingabetextes in Token-IDs
-    input_ids = tokenizer.encode(adjusted_input_text, return_tensors="pt")
+    input_ids = tokenizer.encode(input_text, return_tensors="pt")
+
+    # Erstellung der Attention-Mask
+    attention_mask = torch.ones(input_ids.shape, dtype=torch.bool)
 
     # Einstellung der maximalen Länge
-    max_length = length if length else model.model.config.n_positions
+    max_length = model.model.config.n_positions if len(input_ids[0]) > model.model.config.n_positions else len(input_ids[0]) + length
 
     # Textgenerierung mit spezifischen Parametern
-    output = model.model.generate(
+    beam_output = model.model.generate(
         input_ids,
+        attention_mask=attention_mask,
         max_length=max_length,
-        temperature=temperature,
-        top_k=top_k,
-        top_p=top_p,
+        min_length=4,
+        num_beams=5,
         no_repeat_ngram_size=2,
+        early_stopping=True,
+        temperature=temp,
+        top_p=top_p,
+        top_k=top_k,
+        length_penalty=2.0,
+        do_sample=True,
+        eos_token_id=tokenizer.eos_token_id,
         pad_token_id=tokenizer.eos_token_id
     )
 
     # Konvertieren der generierten Token-IDs zurück in Text
-    return tokenizer.decode(output[0], skip_special_tokens=True)
+    return tokenizer.decode(beam_output[0], skip_special_tokens=True)
+
+def chat_with_model(user_input, history, temperature, top_k, top_p, length, system_prompt):
+    combined_input = f"{history}\nNutzer: {user_input}\n{system_prompt}:"
+    response = generate_text(combined_input, temperature, top_k, top_p, length)
+    new_history = f"{combined_input}\n{response}"
+    return "", new_history  # Leerer String für user_input, um das Eingabefeld zurückzusetzen
 
 # Erstellen der Gradio-Schnittstelle
-interface = gr.Interface(
-    fn=generate_text,
-    inputs=[
-        gr.inputs.Textbox(lines=2, placeholder="Geben Sie Ihren Text hier ein..."),
-        gr.inputs.Slider(minimum=0.1, maximum=1.0, step=0.1, default=0.9, label="Temperature"),
-        gr.inputs.Slider(minimum=1, maximum=100, step=1, default=50, label="Top K"),
-        gr.inputs.Slider(minimum=0.1, maximum=1.0, step=0.1, default=0.9, label="Top P"),
-        gr.inputs.Number(default=50, label="Länge"),
-        gr.inputs.Textbox(lines=2, placeholder="System-Prompt (optional)")
-    ],
-    outputs="text",
-    layout="vertical"
-)
+with gr.Blocks() as demo:
+    with gr.Row():
+        history = gr.Textbox(label="Chatverlauf", lines=10, interactive=False)
+        user_input = gr.Textbox(label="Deine Nachricht")
+        system_prompt = gr.Textbox(label="System Prompt", value="Löwolf GPT")
+        with gr.Column(scale=1):
+            temperature = gr.Slider(minimum=0, maximum=1, step=0.01, label="Temperature", value=0.9)
+            top_k = gr.Slider(minimum=0, maximum=100, step=1, label="Top K", value=50)
+            top_p = gr.Slider(minimum=0, maximum=1, step=0.01, label="Top P", value=0.9)
+            length = gr.Slider(minimum=1, maximum=100, step=1, label="Länge", value=20)
+        submit_btn = gr.Button("Senden")
+    submit_btn.click(
+        chat_with_model,
+        inputs=[user_input, history, temperature, top_k, top_p, length, system_prompt],
+        outputs=[user_input, history]
+    )
 
 # Starten der Gradio-App
-interface.launch()
+demo.launch()