Major changes to the app to allow prompt engineering

app.py

@@ -1,14 +1,10 @@
 
 '''
-This script calls the ada model from openai api to predict the next few words.
+This script calls the model from openai api to predict the next few words.
 '''
 import os
-os.system("pip install --upgrade pip")
 from pprint import pprint
-os.system("pip install git+https://github.com/openai/whisper.git")
 import sys
-print("Sys: ", sys.executable)
-os.system("pip install openai")
 import openai
 import gradio as gr
 import whisper
@@ -17,68 +13,47 @@ import torch
 from transformers import AutoModelForCausalLM
 from transformers import AutoTokenizer
 import time
-# import streaming.py
-# from next_word_prediction import GPT2
 
-
-
-
-#gpt2 = AutoModelForCausalLM.from_pretrained("gpt2", return_dict_in_generate=True)
-#tokenizer = AutoTokenizer.from_pretrained("gpt2")
-
-### /code snippet
-
-
-# get gpt2 model
-#generator = pipeline('text-generation', model='gpt2')
-
-# whisper model specification 
-model = whisper.load_model("tiny")
-
-
-        
-def inference(audio, state=""):
-
-    #time.sleep(2)
-    #text = p(audio)["text"]
-    #state += text + " "
-    # load audio data
-    audio = whisper.load_audio(audio)
-    # ensure sample is in correct format for inference
-    audio = whisper.pad_or_trim(audio)
-
-    # generate a log-mel spetrogram of the audio data
-    mel = whisper.log_mel_spectrogram(audio).to(model.device)
+EXAMPLE_PROMPT = """This is a tool for helping someone with memory issues remember the next word. 
+The predictions follow a few rules:
+1) The predictions are suggestions of ways to continue the transcript as if someone forgot what the next word was.
+2) The predictions do not repeat themselves.
+3) The predictions focus on suggesting nouns, adjectives, and verbs.
+4) The predictions are related to the context in the transcript.
     
-    _, probs = model.detect_language(mel)
-
-    # decode audio data
-    options = whisper.DecodingOptions(fp16 = False)
-    # transcribe speech to text
-    result = whisper.decode(model, mel, options)
-    print("result pre gp model from whisper: ", result, ".text ", result.text, "and the data type: ", type(result.text))
-
-    PROMPT = """The following is an incomplete transcript of a brief conversation. Predict a list of the next most probable words to complete the sentence. 
-Some examples: 
-Transcript1: Tomorrow night we're going out to 
-Predictions1: the movies, a restaurant, a baseball game, the theater, a party for a friend
-Transcript2: I would like to order a cheeseburger with a side of
-Predictions2: french fries, milkshake, apple slices, salad, extra catsup 
-Transcript3: My friend Savanah is
-Predictions3: an electrical engineer, a marine biologist, a classical musician 
-Transcript4: I need to buy a birthday
-Predictions4: present, gift, cake, card
-
-Transcript5: """
-    text = PROMPT + result.text + "Prediction5: "
+EXAMPLES:
+Transcript: Tomorrow night we're going out to 
+Prediction: The Movies, A Restaurant, A Baseball Game, The Theater, A Party for a friend   
+Transcript: I would like to order a cheeseburger with a side of
+Prediction: Frnech fries, Milkshake, Apple slices, Side salad, Extra katsup 
+Transcript: My friend Savanah is
+Prediction: An elecrical engineer, A marine biologist, A classical musician 
+Transcript: I need to buy a birthday
+Prediction: Present, Gift, Cake, Card
+Transcript: """
+
+# whisper model specification
+asr_model = whisper.load_model("tiny")
+
+openai.api_key = os.environ["Openai_APIkey"]
+
+# Transcribe function
+def transcribe(audio_file):
+    print("Transcribing")
+    transcription = asr_model.transcribe(audio_file)["text"]
+    return transcription
+
+def debug_inference(audio, prompt, model, temperature, state=""):
+    # Transcribe with Whisper
+    print("The audio is:", audio)
+    transcript = transcribe(audio)
     
-    openai.api_key = os.environ["Openai_APIkey"]
+    text = prompt + transcript + "\nPrediction: "
     
     response = openai.Completion.create(
-                        model="text-ada-001",
-                        #model="text-curie-001",
+                        model=model,
                         prompt=text,
-                        temperature=1,
+                        temperature=temperature,
                         max_tokens=8,
                         n=5)
 
@@ -96,27 +71,17 @@ Transcript5: """
         infers = list(map(lambda x: x.replace("\n", ""), temp))
         #infered = list(map(lambda x: x.split(','), infers))
 
-        
-        
-
-    # result.text
-    #return getText, gr.update(visible=True), gr.update(visible=True), gr.update(visible=True)
-    return result.text, state, infers
-
-
+    return transcript, state, infers, text
 
 # get audio from microphone 
-
 gr.Interface(
-        fn=inference, 
-    inputs=[
-        gr.inputs.Audio(source="microphone", type="filepath"), 
-        "state"
-    ],
-    outputs=[
-        "textbox",
-        "state",
-        "textbox"
-    ],
-    live=True).launch()
-
+    fn=debug_inference, 
+    inputs=[gr.inputs.Audio(source="microphone", type="filepath"),
+            gr.inputs.Textbox(lines=15, placeholder="Enter a prompt here"),
+            gr.inputs.Dropdown(["text-ada-001", "text-davinci-002", "text-davinci-003", "gpt-3.5-turbo"], label="Model"),
+            gr.inputs.Slider(minimum=0.0, maximum=1.0, default=0.8, step=0.1, label="Temperature"),
+            "state"
+            ],
+    outputs=["textbox","state","textbox", "textbox"],
+    # examples=[["example_in-the-mood-to-eat.m4a", EXAMPLE_PROMPT, "text-ada-001", 0.8, ""],["","","",0.9,""]],
+    live=False).launch()