Update app.py

app.py

@@ -1,23 +1,12 @@
 import streamlit as st
-import numpy as np
-import pandas as pd
-import os
+from transformers import AutoTokenizer, AutoModelForCausalLM, AutoModel
 import torch
-import torch.nn as nn
-from transformers.activations import get_activation
-from transformers import AutoTokenizer, AutoModelForCausalLM
 
-
-st.title('GPT2: To see all prompt outlines: https://huggingface.co/BigSalmon/InformalToFormalLincoln64Paraphrase')
-
-device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+first = """informal english: corn fields are all across illinois, visible once you leave chicago.\nTranslated into the Style of Abraham Lincoln: corn fields ( permeate illinois / span the state of illinois / ( occupy / persist in ) all corners of illinois / line the horizon of illinois / envelop the landscape of illinois ), manifesting themselves visibly as one ventures beyond chicago.\n\ninformal english: """
 
 @st.cache(allow_output_mutation=True)
 def get_model():
 
-    tokenizer = AutoTokenizer.from_pretrained("BigSalmon/InformalToFormalLincoln89Paraphrase")
-    model = AutoModelForCausalLM.from_pretrained("BigSalmon/InformalToFormalLincoln89Paraphrase")
-
     #tokenizer = AutoTokenizer.from_pretrained("BigSalmon/InformalToFormalLincoln86Paraphrase")
     #model = AutoModelForCausalLM.from_pretrained("BigSalmon/InformalToFormalLincoln86Paraphrase")
     
@@ -83,109 +72,47 @@ def get_model():
     
     #tokenizer = AutoTokenizer.from_pretrained("BigSalmon/PointsToSentence")
     #model = AutoModelForCausalLM.from_pretrained("BigSalmon/PointsToSentence")
+    tokenizer = AutoTokenizer.from_pretrained("BigSalmon/InformalToFormalLincoln89Paraphrase")
+    model = AutoModelForCausalLM.from_pretrained("BigSalmon/InformalToFormalLincoln89Paraphrase")
+    tokenizer2 = AutoTokenizer.from_pretrained("BigSalmon/InformalToFormalLincoln90ParaphraseMedium")
+    model2 = AutoModelForCausalLM.from_pretrained("BigSalmon/InformalToFormalLincoln90ParaphraseMedium")
+    return model, model2, tokenizer, tokenizer2
     
-    return model, tokenizer
-    
-model, tokenizer = get_model()
-
-g = """informal english: garage band has made people who know nothing about music good at creating music.
-Translated into the Style of Abraham Lincoln: garage band ( offers the uninitiated in music the ability to produce professional-quality compositions / catapults those for whom music is an uncharted art the ability the realize masterpieces / stimulates music novice's competency to yield sublime arrangements / begets individuals of rudimentary musical talent the proficiency to fashion elaborate suites ).
-
-informal english: chrome extensions can make doing regular tasks much easier to get done.
-Translated into the Style of Abraham Lincoln: chrome extensions ( yield the boon of time-saving convenience / ( expedite the ability to / unlock the means to more readily ) accomplish everyday tasks / turbocharges the velocity with which one can conduct their obligations ).
+model, model2, tokenizer, tokenizer2 = get_model()
 
-informal english: broadband is finally expanding to rural areas, a great development that will thrust them into modern life.
-Translated into the Style of Abraham Lincoln: broadband is ( ( finally / at last / after years of delay ) arriving in remote locations / springing to life in far-flung outposts / inching into even the most backwater corners of the nation ) that will leap-frog them into the twenty-first century.
+st.text('''For Prompt Templates: https://huggingface.co/BigSalmon/InformalToFormalLincoln82Paraphrase''')
 
-informal english: google translate has made talking to people who do not share your language easier.
-Translated into the Style of Abraham Lincoln: google translate ( imparts communicability to individuals whose native tongue differs / mitigates the trials of communication across linguistic barriers / hastens the bridging of semantic boundaries / mollifies the complexity of multilingual communication / avails itself to the internationalization of discussion / flexes its muscles to abet intercultural conversation / calms the tides of linguistic divergence ).
-
-informal english: corn fields are all across illinois, visible once you leave chicago.
-Translated into the Style of Abraham Lincoln: corn fields ( permeate illinois / span the state of illinois / ( occupy / persist in ) all corners of illinois / line the horizon of illinois / envelop the landscape of illinois ), manifesting themselves visibly as one ventures beyond chicago.
-
-informal english: """
-
-number_of_outputs = st.sidebar.slider("Number of Outputs", 5, 100)
-log_nums = st.sidebar.slider("How Many Log Outputs?", 50, 600)
-
-def BestProbs(prompt):
-  prompt = prompt.strip()
-  text = tokenizer.encode(prompt)
-  myinput, past_key_values = torch.tensor([text]), None
-  myinput = myinput
-  logits, past_key_values = model(myinput, past_key_values = past_key_values, return_dict=False)
-  logits = logits[0,-1]
-  probabilities = torch.nn.functional.softmax(logits)
-  best_logits, best_indices = logits.topk(10)
-  best_words = [tokenizer.decode([idx.item()]) for idx in best_indices]
-  for i in best_words[0:10]:
-    print("_______")
-    st.write(f"${i} $\n")
-    f = (f"${i} $\n")
-    m = (prompt + f"{i}")
-    BestProbs2(m)
-  return f
+temp = st.sidebar.slider("Temperature", 0.7, 1.5)
+number_of_outputs = st.sidebar.slider("Number of Outputs", 5, 50)
+lengths = st.sidebar.slider("Length", 3, 500)
+bad_words = st.text_input("Words You Do Not Want Generated", " core lemon height time ")
+logs_outputs = st.sidebar.slider("Logit Outputs", 50, 300)
 
-def BestProbs2(prompt):
-  prompt = prompt.strip()
-  text = tokenizer.encode(prompt)
-  myinput, past_key_values = torch.tensor([text]), None
-  myinput = myinput
-  logits, past_key_values = model(myinput, past_key_values = past_key_values, return_dict=False)
-  logits = logits[0,-1]
-  probabilities = torch.nn.functional.softmax(logits)
-  best_logits, best_indices = logits.topk(20)
-  best_words = [tokenizer.decode([idx.item()]) for idx in best_indices]
-  for i in best_words[0:20]:
-    print(i)
-    st.write(i)
-    
-def LogProbs(prompt):
-  col1 = []
-  col2 = []
-  prompt = prompt.strip()
-  text = tokenizer.encode(prompt)
-  myinput, past_key_values = torch.tensor([text]), None
-  myinput = myinput
-  logits, past_key_values = model(myinput, past_key_values = past_key_values, return_dict=False)
-  logits = logits[0,-1]
-  probabilities = torch.nn.functional.softmax(logits)
-  best_logits, best_indices = logits.topk(10)
-  best_words = [tokenizer.decode([idx.item()]) for idx in best_indices]
-  for i in best_words[0:10]:
-    print("_______")
-    f = i
-    col1.append(f)
-    m = (prompt + f"{i}")
-    #print("^^" + f + " ^^")
-    prompt = m.strip()
-    text = tokenizer.encode(prompt)
-    myinput, past_key_values = torch.tensor([text]), None
-    myinput = myinput
-    logits, past_key_values = model(myinput, past_key_values = past_key_values, return_dict=False)
-    logits = logits[0,-1]
-    probabilities = torch.nn.functional.softmax(logits)
-    best_logits, best_indices = logits.topk(20)
-    best_words = [tokenizer.decode([idx.item()]) for idx in best_indices]
-    for i in best_words[0:20]:
-      #print(i)
-      col2.append(i)
-  #print(col1)
-  #print(col2)
-  d = {col1[0]: [col2[0], col2[1], col2[2], col2[3], col2[4], col2[5], col2[6], col2[7], col2[8], col2[9], col2[10], col2[11], col2[12], col2[13], col2[14], col2[15], col2[16], col2[17], col2[18], col2[19]],
-    col1[1]: [col2[20], col2[21], col2[22], col2[23], col2[24], col2[25], col2[26], col2[27], col2[28], col2[29], col2[30], col2[31], col2[32], col2[33], col2[34], col2[35], col2[36], col2[37], col2[38], col2[39]],
-    col1[2]: [col2[40], col2[41], col2[42], col2[43], col2[44], col2[45], col2[46], col2[47], col2[48], col2[49], col2[50], col2[51], col2[52], col2[53], col2[54], col2[55], col2[56], col2[57], col2[58], col2[59]],
-    col1[3]: [col2[60], col2[61], col2[62], col2[63], col2[64], col2[65], col2[66], col2[67], col2[68], col2[69], col2[70], col2[71], col2[72], col2[73], col2[74], col2[75], col2[76], col2[77], col2[78], col2[79]],
-    col1[4]: [col2[80], col2[81], col2[82], col2[83], col2[84], col2[85], col2[86], col2[87], col2[88], col2[89], col2[90], col2[91], col2[92], col2[93], col2[94], col2[95], col2[96], col2[97], col2[98], col2[99]],
-    col1[5]: [col2[100], col2[101], col2[102], col2[103], col2[104], col2[105], col2[106], col2[107], col2[108], col2[109], col2[110], col2[111], col2[112], col2[113], col2[114], col2[115], col2[116], col2[117], col2[118], col2[119]],
-    col1[6]: [col2[120], col2[121], col2[122], col2[123], col2[124], col2[125], col2[126], col2[127], col2[128], col2[129], col2[130], col2[131], col2[132], col2[133], col2[134], col2[135], col2[136], col2[137], col2[138], col2[139]],
-    col1[7]: [col2[140], col2[141], col2[142], col2[143], col2[144], col2[145], col2[146], col2[147], col2[148], col2[149], col2[150], col2[151], col2[152], col2[153], col2[154], col2[155], col2[156], col2[157], col2[158], col2[159]],
-    col1[8]: [col2[160], col2[161], col2[162], col2[163], col2[164], col2[165], col2[166], col2[167], col2[168], col2[169], col2[170], col2[171], col2[172], col2[173], col2[174], col2[175], col2[176], col2[177], col2[178], col2[179]],
-    col1[9]: [col2[180], col2[181], col2[182], col2[183], col2[184], col2[185], col2[186], col2[187], col2[188], col2[189], col2[190], col2[191], col2[192], col2[193], col2[194], col2[195], col2[196], col2[197], col2[198], col2[199]]}
-  df = pd.DataFrame(data=d)
-  print(df)
-  st.write(df)
-  return df
+def run_generate(text, bad_words):
+  yo = []
+  input_ids = tokenizer.encode(text, return_tensors='pt')
+  res = len(tokenizer.encode(text))
+  bad_words = bad_words.split()
+  bad_word_ids = []
+  for bad_word in bad_words: 
+    bad_word = " " + bad_word
+    ids = tokenizer(bad_word).input_ids
+    bad_word_ids.append(ids)
+  sample_outputs = model.generate(
+    input_ids,
+    do_sample=True, 
+    max_length= res + lengths, 
+    min_length = res + lengths, 
+    top_k=50,
+    temperature=temp,
+    num_return_sequences=number_of_outputs,
+    bad_words_ids=bad_word_ids
+  )
+  for i in range(number_of_outputs):
+    e = tokenizer.decode(sample_outputs[i])
+    e = e.replace(text, "")
+    yo.append(e)
+  return yo
   
 def BestProbs5(prompt):
   prompt = prompt.strip()
@@ -204,66 +131,88 @@ def BestProbs5(prompt):
     st.write(g)
     l = run_generate(g, "hey")
     st.write(l)
-    
-def run_generate(text, bad_words):
+  
+def run_generate2(text, bad_words):
   yo = []
-  input_ids = tokenizer.encode(text, return_tensors='pt')
-  res = len(tokenizer.encode(text))
+  input_ids = tokenizer2.encode(text, return_tensors='pt')
+  res = len(tokenizer2.encode(text))
   bad_words = bad_words.split()
-  bad_word_ids = [[7829], [40940]]
+  bad_word_ids = []
   for bad_word in bad_words: 
     bad_word = " " + bad_word
-    ids = tokenizer(bad_word).input_ids
+    ids = tokenizer2(bad_word).input_ids
     bad_word_ids.append(ids)
-  sample_outputs = model.generate(
+  sample_outputs = model2.generate(
     input_ids,
     do_sample=True, 
-    max_length= res + 5, 
-    min_length = res + 5, 
+    max_length= res + lengths, 
+    min_length = res + lengths, 
     top_k=50,
-    temperature=1.0,
-    num_return_sequences=3,
+    temperature=temp,
+    num_return_sequences=number_of_outputs,
     bad_words_ids=bad_word_ids
   )
-  for i in range(3):
-    e = tokenizer.decode(sample_outputs[i])
+  for i in range(number_of_outputs):
+    e = tokenizer2.decode(sample_outputs[i])
     e = e.replace(text, "")
     yo.append(e)
-  print(yo)
   return yo
-
+  
+def prefix_format(sentence):
+  words = sentence.split()
+  if "[MASK]" in sentence:
+    words2 = words.index("[MASK]")
+    #print(words2)
+    output = ("<Prefix> " + ' '.join(words[:words2]) + " <Prefix> " + "<Suffix> " + ' '.join(words[words2+1:]) + " <Suffix>" + " <Middle>")
+    st.write(output)
+  else:
+    st.write("Add [MASK] to sentence")
+ 
 with st.form(key='my_form'):
-    prompt = st.text_area(label='Enter sentence', value=g, height=500)
+    text = st.text_area(label='Enter sentence', value=first)
     submit_button = st.form_submit_button(label='Submit')
-    submit_button2 = st.form_submit_button(label='Fast Forward')
-    submit_button3 = st.form_submit_button(label='Fast Forward 2.0')
-    submit_button4 = st.form_submit_button(label='Get Top')
-
+    submit_button2 = st.form_submit_button(label='Submit Log Probs')
+    
+    submit_button3 = st.form_submit_button(label='Submit Other Model')
+    submit_button4 = st.form_submit_button(label='Submit Log Probs Other Model')
+    
+    submit_button5 = st.form_submit_button(label='Most Prob')
+    
+    submit_button6 = st.form_submit_button(label='Turn Sentence with [MASK] into <Prefix> Format')
+    
     if submit_button:
+      translated_text = run_generate(text, bad_words)
+      st.write(translated_text if translated_text else "No translation found")
+    if submit_button2:
       with torch.no_grad():
-        text = tokenizer.encode(prompt)
-        myinput, past_key_values = torch.tensor([text]), None
+        text2 = str(text)
+        print(text2)
+        text3 = tokenizer.encode(text2)
+        myinput, past_key_values = torch.tensor([text3]), None
         myinput = myinput
-        myinput= myinput.to(device)
         logits, past_key_values = model(myinput, past_key_values = past_key_values, return_dict=False)
         logits = logits[0,-1]
         probabilities = torch.nn.functional.softmax(logits)
-        best_logits, best_indices = logits.topk(log_nums)
-        best_words = [tokenizer.decode([idx.item()]) for idx in best_indices]
-        text.append(best_indices[0].item())
-        best_probabilities = probabilities[best_indices].tolist()
-        words = []              
+        best_logits, best_indices = logits.topk(logs_outputs)
+        best_words = [tokenizer.decode([idx.item()]) for idx in best_indices]      
         st.write(best_words)
-    if submit_button2:
-        print("----")
-        st.write("___")
-        m = LogProbs(prompt)
-        st.write("___")
-        st.write(m)
-        st.write("___")
     if submit_button3:
-        print("----")
-        st.write("___")
-        st.write(BestProbs)
+      translated_text = run_generate2(text, bad_words)
+      st.write(translated_text if translated_text else "No translation found")
     if submit_button4:
-      BestProbs5(prompt)
+      text2 = str(text)
+      print(text2)
+      text3 = tokenizer2.encode(text2)
+      myinput, past_key_values = torch.tensor([text3]), None
+      myinput = myinput
+      logits, past_key_values = model2(myinput, past_key_values = past_key_values, return_dict=False)
+      logits = logits[0,-1]
+      probabilities = torch.nn.functional.softmax(logits)
+      best_logits, best_indices = logits.topk(logs_outputs)
+      best_words = [tokenizer2.decode([idx.item()]) for idx in best_indices]      
+      st.write(best_words)
+    if submit_button5:
+      BestProbs5(text)
+    if submit_button6:
+      text2 = str(text)
+      prefix_format(text2)