Working EDGAR-P model

continuations.py

@@ -0,0 +1,124 @@
+from utility import *
+
+def continue_story_svo_customprompt_past_tense(model, story, question, verbose = False, width = 10):
+  instructions = "The following are results from a children's reasoning test.  \nThe child will be read a short story and then answer a question about what happened before the story.  \nThe answer cannot contradict the story.  \nWe have collected the results below:\n"
+  story1="Story 1: Jim sat by the swings as Tom slowly approached. Tom gave the book to Jim during recess.\n"
+  question1="Question: Why did Jim receive the book?\n"
+  answer1="Answers:\n1. Tom desired to return the book to his friend.\n2. Jim bought the book from Tom.\n3. Tom noticed Jim dropped his book.\n4. Jim needed a book to study for his exam.\n5. Tom wanted to give Jim his favorite book.\n\n"
+  story2 = "Story 2: Mary was happy to finally cross the street.\n"
+  question2 = "Question: Why was Mary running?\n"
+  answer2="Answers:\n1. Mary was running from a monster.\n2. Mary was running a marathon.\n3. Mary wanted to get away from her parents.\n\n" 
+  story3 = "Story 3: " + story + "\n"
+  question3 = "Question: " + question +"\n"
+  inp = instructions + story1 + question1 +  answer1 + story2 + question2 + answer2 + story3 + question3
+  #inp = "Q. Why was 6 afraid of 7?"
+  if verbose:
+    print("constructed inp ", inp, "\n"*2)
+  con = construct(inp, "Answers:\n1.")
+
+  #con = construct(inp, " A.")
+  #print("construct is ", con, "\n"*2)
+  out = generate(model, con, max_length=850, horizon=story, horizon_penalty=1.8, beams=5, repetition_penalty=2.8, do_beams=True)
+
+  #return out
+  out = out.split("Story 3")[1]
+  #Remove the next story
+  out = out.split("Story 4")[0]
+  #Filter to correct answers
+  #print("prefiltered out ", out)
+  out = out.split("Answers:")[1]
+  out = out.split("Wrong")[0]
+  #print("filtered out is ", out)
+  #If the user does not specify a width
+  if width == -1:
+    #Capture all of them
+    width = 100
+  responses = list()
+  #Reads through the outputted list and returns every item
+  for i in range(1, width):
+    try:
+      start = "\n"+str(i)
+      end = "\n"+str(i+1)
+      responses.append(out.split(start)[1].split(end)[0])
+    except:
+      break
+  #Take responses that are long enough
+  responses = list(filter(lambda x: len(x) > 5, responses))
+  #print("og responses are ", responses)
+  #Remove first space
+  for i in range(len(responses)):
+    try:
+      if responses[i][:2] == '. ':
+        responses[i] = responses[i][2:] 
+      elif responses[i][:2] == ') ':
+        responses[i] = responses[i][2:]
+      elif responses[i][0] == ' ':
+        responses[i] = responses[i][1:]
+
+      responses[i] = " ".join(responses[i].split())
+      responses[i] = clean_story(responses[i])
+    except:
+      continue
+  return responses
+
+def continue_story_svo_customprompt_future_tense(model, story, question, verbose = False, width = 10):
+  instructions = "The following are results from a children's reasoning test.  \nThe child will be read a short story and then answer a question about what happened before the story.  \nThe answer cannot contradict the story.  \nWe have collected the results below:\n"
+  story1="Story 1: Jim sat by the swings as Tom slowly approached. Tom gave the book to Jim during recess.\n"
+  question1="Question: What did Jim do after recieving the book?\n"
+  answer1="Answers:\n1. Jim jumped for joy!\n2. Jim threw the book away.\n3. Tom noticed that Jim looked unhappy.\n\n"
+  story2 = "Story 2: Mary was happy to finally cross the street.\n"
+  question2 = "Question: What happened after Mary crossed the street?\n"
+  answer2="Answers:\n1. She sighed in relief.\n2. Her smile quickly faded when she remembered she left her wallet at the restaurant.\n3. She looked over her shoulder to make sure she was not being followed.\n\n" 
+  story3 = "Story 3: " + story + "\n"
+  question3 = "Question: " + question +"\n"
+  inp = instructions + story1 + question1 +  answer1 + story2 + question2 + answer2 + story3 + question3
+  #inp = "Q. Why was 6 afraid of 7?"
+  if verbose:
+    print("constructed inp ", inp, "\n"*2)
+  con = construct(inp, "Answers:\n1.")
+
+  #con = construct(inp, " A.")
+  #print("construct is ", con, "\n"*2)
+  out = generate(model, con, max_length=850, horizon=story, horizon_penalty=1.8, beams=5, repetition_penalty=2.8, do_beams=True)
+
+  #return out
+  out = out.split("Story 3")[1]
+  #Remove the next story
+  out = out.split("Story 4")[0]
+  #Filter to correct answers
+  #print("prefiltered out ", out)
+  out = out.split("Answers:")[1]
+  out = out.split("Wrong")[0]
+  #print("filtered out is ", out)
+  #If the user does not specify a width
+  if width == -1:
+    #Capture all of them
+    width = 100
+  responses = list()
+  #Reads through the outputted list and returns every item
+  for i in range(1, width):
+    try:
+      start = "\n"+str(i)
+      end = "\n"+str(i+1)
+      responses.append(out.split(start)[1].split(end)[0])
+    except:
+      break
+  #Take responses that are long enough
+  responses = list(filter(lambda x: len(x) > 5, responses))
+  #print("og responses are ", responses)
+  #Remove first space
+  for i in range(len(responses)):
+    try:
+      if responses[i][:2] == '. ':
+        responses[i] = responses[i][2:] 
+      elif responses[i][:2] == ') ':
+        responses[i] = responses[i][2:]
+      elif responses[i][0] == ' ':
+        responses[i] = responses[i][1:]
+
+      responses[i] = " ".join(responses[i].split())
+      responses[i] = clean_story(responses[i])
+    except:
+      continue
+  return responses
+

main.py

@@ -0,0 +1,258 @@
+from utility import *
+from questions import *
+from continuations import *
+
+
+# load the language model
+model = load_gptj()
+
+
+def hasSVOExpand(sent):
+  text = nlp(sent) 
+  svos = findSVOs(text)
+  if svos: 
+    return True 
+  else:
+    return False
+
+
+backwards = False
+
+#check on random sample of stories
+import random 
+from tqdm import tqdm
+r = list(range(len(stories)))
+random.shuffle(r)
+
+#at least one output  0.98
+#total svoable percentages  0.8954248366013072
+
+at_least_one_svoable_output = 0 
+total_svoable_percents = 0 
+total_sentences_checked = 0 
+num_stories = 5
+
+#for inp_story_full in tqdm(stories[:num_stories], position = 0, leave = True):
+for i in tqdm(r[:num_stories]):
+  inp_story_full = stories[i]
+
+  inp_story = inp_story_full[-1]
+  if backwards:
+    q = get_questions_past_tense(model, inp_story)[0]
+    starts = continue_story_svo_customprompt_past_tense(model, inp_story, q)
+
+  else:
+    q = get_questions_future_tense(model, inp_story)[0]
+    starts = continue_story_svo_customprompt_future_tense(model, inp_story, q)
+
+
+
+  svoability = ['x'for start in starts if hasSVOExpand(start)]
+
+  if svoability:
+    at_least_one_svoable_output += 1  
+  total_svoable_percents += len(svoability) 
+  total_sentences_checked += len(starts) 
+
+print("at least one output ", at_least_one_svoable_output/num_stories)
+print("total svoable percentages ", total_svoable_percents/total_sentences_checked) 
+
+
+import json 
+expansionfilenames = [
+    "Archiveprompts/Prose_Enhancer_-_Past_Tense_1st_2021-07-13T19_20_37.875Z.story",
+    "Archiveprompts/Prose_Enhancer_-_Past_Tense_2nd_2021-07-13T19_21_00.433Z.story",
+    "Archiveprompts/Prose_Enhancer_-_Past_Tense_3rd_2021-07-13T19_21_12.748Z.story",
+    "Archiveprompts/Prose_Enhancer_-_Past_Tense_3rd_2021-07-13T19_21_51.028Z.story",
+        ]
+
+expansion_prompt_jsons = []
+for filename in expansionfilenames:
+    with open(filename, 'r') as jsonfile:
+        expansion_prompt_jsons.append(json.load(jsonfile))
+
+expansion_prompts = [x['content']['story']['datablocks'][1]['dataFragment']['data'] for x in expansion_prompt_jsons]
+
+def expand_text_simple(model, simple_sentence, expansion_prompt, verbose = False, width = 10):
+
+  inp = expansion_prompt
+  #inp = "Q. Why was 6 afraid of 7?"
+  inp = inp.replace("[Insert Text Here]", simple_sentence)
+  #inp = expansion_prompt_text.replace("[Insert Text Here]", "I shot the dog.")
+
+  #print("input is ", inp, "\n\n")
+  input_ids = tokenizer.encode(inp, return_tensors='pt').cuda()
+  model_output = model.generate(
+    input_ids,
+    do_sample=True, 
+    max_length=input_ids.shape[1] + 150, 
+    top_k=50, 
+    top_p=0.95, 
+    early_stopping = True,
+    num_return_sequences=1
+  )
+
+  #25 tokens for expansion
+  text_output = tokenizer.decode(model_output[0], skip_special_tokens=True)
+  #print("fulltext ", text_output)
+  prose = text_output[len(inp):]
+  #print(prose)
+  return prose
+
+def expand_story_simple(story, expansion_prompt, verbose = False):
+    story_text = []
+    for sentence in story:
+        print("sentence is ", sentence)
+        prose = expand_text_simple(sentence, expansion_prompt, verbose = verbose)
+        prose = prose.split('\n')[0].strip()
+        story_text.append(prose)
+        print("prose is ", type(prose), prose)
+    return story_text
+
+#This stays on CPU
+tokenizer_deberta = AutoTokenizer.from_pretrained("microsoft/deberta-v2-xxlarge-mnli")
+model_deberta = AutoModelForSequenceClassification.from_pretrained("microsoft/deberta-v2-xxlarge-mnli")
+
+def not_contradict(sentA, sentB):
+  if backwards:
+    to_run = "[CLS]" + sentA  + "[SEP]" + sentB + "[SEP]"
+  else:
+    to_run = "[CLS]" + sentB  + "[SEP]" + sentA + "[SEP]"
+  inputs = tokenizer_deberta(to_run, return_tensors="pt")
+  with torch.no_grad():
+    outputs = model_deberta(**inputs)
+
+  outputs = softmax(np.array(outputs.logits.squeeze().cpu().tolist()))
+  choice = np.argmax(outputs)
+  if choice == 0:
+    return -100
+  return outputs[2]
+
+print(not_contradict("The sky is blue.", "The sky is not blue.")) #scores w/ BERT
+print(not_contradict("The sky is blue.", "The sky is not red."))
+print(not_contradict("The sky is blue.", "The ground is green."))
+
+#Beams should be [[inp_story, [""]]] when we start
+def beam_search(beams, width=20, diversity_width=2, graph=None, reverse_rank=True):
+  print("beginning beam search fn")
+  candidates = list()
+  #Story is the story for this beam, q is questions already answered
+  for story, q_prev, score in tqdm(beams):
+    story_sents = sent_tokenize(story)
+    #print(story)
+
+    #Accumulate questions. q needs to start as [""]
+    if backwards:
+      questions = get_questions_past_tense(model, " ".join(story_sents[:min(len(story_sents),3)]), q_prev)
+    else:
+      questions = get_questions_future_tense(model, " ".join(story_sents[min(len(story_sents),3):]), q_prev)
+
+    print("accumulated questions are ", questions)
+    extensions = list()
+    for q in questions:
+      if backwards:
+        continuation = continue_story_svo_customprompt_past_tense(model, story, q, width=-1)
+      else:
+        continuation = continue_story_svo_customprompt_future_tense(model, story, q, width=-1)
+
+      if q_prev != ['']:
+        q_cur = [[q] + q_prev]*len(continuation)
+      else:
+        q_cur = [[q]] * len(continuation)
+      extensions += zip(continuation, q_cur)
+    #Sort by most likely to imply and take the top k
+    if backwards:
+      implication_story = " ".join(story_sents[0:min(len(story_sents), 1)]) #The first k sentences sliding window
+    else:
+      implication_story = " ".join(story_sents[min(len(story_sents), 1):]) #The first k sentences sliding window
+    extensions = list(filter(lambda x: hasSVOExpand(x[0]), extensions)) #Filter on if there is an SVO tuple
+    extensions_ranks = list(map(lambda x: not_contradict(x[0], implication_story) + score, extensions)) #Rank. Include prior score via sum. This keeps track of beam scores over time.
+    extensions_zip = list(filter(lambda x: x[1] > -100, zip(extensions, extensions_ranks))) #Zip
+    extensions_zip = sorted(extensions_zip, key=lambda x: x[1], reverse=reverse_rank)
+
+    extensions_zip = extensions_zip[:min(diversity_width, len(extensions_zip))] #Take top k
+    if backwards:
+      extensions = list(map(lambda x: (x[0][0]+" "+story, x[0][1], x[1]), extensions_zip)) #Sort
+    else:
+      extensions = list(map(lambda x: (story+" "+x[0][0], x[0][1], x[1]), extensions_zip)) #Sort
+    #new_stories = list(map(lambda x: (x[0]+" "+story, x[1]), extensions)) #Concat
+    #print(extensions)
+    #Debug mode
+    if graph is not None:
+      for i in range(len(extensions)):
+        graph.node(extensions[i][0])
+        graph.edge(story_sents[0], extensions[i][0], label=extensions[i][1][0])
+
+    #print("\n".join(new_stories))
+
+    #Internally rank the new stories to preserve diversity
+    candidates += extensions
+
+  sorted_l = list(map(lambda x: (x[0], x[1], x[2]), sorted(candidates, key=lambda x: x[2], reverse=reverse_rank)))
+  return sorted_l[:min(len(sorted_l), width)], graph
+
+#Order will be given by beam rank, so this method eliminates the correct beam if duplicates are found
+def remove_duplicates(beams):
+  d = {}
+  beams.reverse()
+  for story, questions, scores in beams:
+    d[story] = (questions, scores)
+  to_ret = list()
+  for k in d.keys():
+    to_ret.append((k, d[k][0], d[k][1]))
+  return to_ret
+
+reverse = True
+inp_story="The battle had been raging on for hours. William set his phasers to kill, he knew he had to make amends."
+f=None
+
+beams, f = beam_search([(inp_story, [""], 0.0)], width=10, diversity_width=2, graph=f, reverse_rank=reverse)
+print("beams are ", beams)
+print("graph is ", f)
+
+for i in range(3):  
+  beams = remove_duplicates(beams)
+  print("beams are ", beams)
+  print("\nSTEP: " + str(i) + "\n\n")
+  print("\n".join(list(map(lambda x: x[0], beams))))
+  beams, f = beam_search(beams, width=5, diversity_width=10,graph=f, reverse_rank=reverse)
+
+
+def gen_story_from_last_sentence(last_sentence, story_length=10):
+    reverse = True
+    inp_story=last_sentence
+    f=None
+
+    beams, f = beam_search([(inp_story, [""], 0.0)], width=10, diversity_width=2, graph=f, reverse_rank=reverse)
+    print("beams are ", beams)
+    print("graph is ", f)
+
+    for i in range(story_length - 1):
+      beams = remove_duplicates(beams)
+      print("beams are ", beams)
+      print("\nSTEP: " + str(i) + "\n\n")
+      print("\n".join(list(map(lambda x: x[0], beams))))
+      beams, f = beam_search(beams, width=5, diversity_width=10,graph=f, reverse_rank=reverse)
+
+    return beams
+
+
+story1 = "Hansel's hand still trembles as he pushes open the twice-cooked door. \n The last time he saw the house he was glancing back over his shoulder as he and his sister fled into the trees. \n"
+story2 = "Stella loves to eat peanuts! She goes to the market every day, and is best friends with the owner."
+story3 = "There was a princess with long hair locked in a tower far away. \n A prince came to save her by climbing her hair and taking her out of the tower."
+story4 = "I woke up realizing that I became a cat. \n I caught a mouse and at the next moment I realized that I'm a human again!"
+story5 = "One night, I decided to go to McDonalds to get some ice cream. \n But when I got to the store, they said the machine was down; I was sad."
+story6 = "He turned out the light and went into Jem's room. \n He would be there all night, and he would be there when Jem waked up in the morning. \n But little did Travis know, Jem knew about this all along."
+story7 = "The hero charged at the dragon, both disappearing into the void. The world is saved and people will not forget the sacrifice of the nameless hero."
+
+last_sentences = [story1, story2, story3, story4, story5, story6, story7]
+
+import pickle 
+for idx, sentence in enumerate(last_sentences):
+    story_beams = gen_story_from_last_sentence(sentence, story_length = 10)
+
+    with open('storypickle' + str(idx) + '.pickle', 'wb') as handle:
+        pickle.dump(story_beams, handle, protocol=pickle.HIGHEST_PROTOCOL)
+        print("Dumping \n")
+        print(story_beams)
+