import gradio as gr import math import spacy from datasets import load_dataset from transformers import pipeline, T5Tokenizer from transformers import AutoTokenizer, AutoModel, AutoModelForSequenceClassification from transformers import TrainingArguments, Trainer, T5ForConditionalGeneration import torch import torch.nn.functional as F from torch.utils.data import DataLoader import numpy as np import evaluate import nltk from nltk.corpus import stopwords import subprocess import sys import random from textwrap import fill # !pip install https://huggingface.co/spacy/en_core_web_sm/resolve/main/en_core_web_sm-any-py3-none-any.whl subprocess.check_call([sys.executable, '-m', 'pip', 'install', 'https://huggingface.co/spacy/en_core_web_sm/resolve/main/en_core_web_sm-any-py3-none-any.whl']) # tokenizer = AutoTokenizer.from_pretrained('sentence-transformers/all-MiniLM-L6-v2') model_base = "results/checkpoint-17000" nltk.download('stopwords') nlp = spacy.load("en_core_web_sm") stops = stopwords.words("english") ROMAN_CONSTANTS = ( ( "", "I", "II", "III", "IV", "V", "VI", "VII", "VIII", "IX" ), ( "", "X", "XX", "XXX", "XL", "L", "LX", "LXX", "LXXX", "XC" ), ( "", "C", "CC", "CCC", "CD", "D", "DC", "DCC", "DCCC", "CM" ), ( "", "M", "MM", "MMM", "", "", "-", "", "", "" ), ( "", "i", "ii", "iii", "iv", "v", "vi", "vii", "viii", "ix" ), ( "", "x", "xx", "xxx", "xl", "l", "lx", "lxx", "lxxx", "xc" ), ( "", "c", "cc", "ccc", "cd", "d", "dc", "dcc", "dccc", "cm" ), ( "", "m", "mm", "mmm", "", "", "-", "", "", "" ), ) # answer = "Pizza" guesses = [] return_guesses = [] answer = "Moon" word1 = "Black" word2 = "White" word3 = "Sun" base_prompts = ["Sun is to Moon as ", "Black is to White as ", "Atom is to Element as", "Athens is to Greece as ", "Cat is to Dog as ", "Robin is to Bird as", "Hunger is to Ambition as "] #Mean Pooling - Take attention mask into account for correct averaging def mean_pooling(model_output, attention_mask): token_embeddings = model_output['token_embeddings'] #First element of model_output contains all token embeddings input_mask_expanded = attention_mask.unsqueeze(-1).expand(token_embeddings.size()).float() return torch.sum(token_embeddings * input_mask_expanded, 1) / torch.clamp(input_mask_expanded.sum(1), min=1e-9) def normalize(comment, lowercase, remove_stopwords): if lowercase: comment = comment.lower() comment = nlp(comment) lemmatized = list() for word in comment: lemma = word.lemma_.strip() if lemma: if not remove_stopwords or (remove_stopwords and lemma not in stops): lemmatized.append(lemma) return " ".join(lemmatized) # def tokenize_function(examples): # return tokenizer(examples["text"]) def compute_metrics(eval_pred): logits, labels = eval_pred predictions = np.argmax(logits, axis=-1) metric = evaluate.load("accuracy") return metric.compute(predictions=predictions, references=labels) def get_model(): global model_base # last_checkpoint = "./results/checkpoint-22500" finetuned_model = T5ForConditionalGeneration.from_pretrained(model_base) tokenizer = T5Tokenizer.from_pretrained(model_base) # model = SentenceTransformer(model_base) gpu_available = torch.cuda.is_available() device = torch.device("cuda" if gpu_available else "cpu") finetuned_model = finetuned_model.to(device) return finetuned_model, tokenizer def cosine_scores(model, sentence): global word1 global word2 global word3 # sentence1 = f"{word1} is to {word2} as" embeddings1 = model.encode(sentence, convert_to_tensor=True) def embeddings(model, sentences, tokenizer): global word1 global word2 global word3 global model_base gpu_available = torch.cuda.is_available() device = torch.device("cuda" if gpu_available else "cpu") # device = torch.device('cuda:0') # embeddings = model.encode(sentences) question = "Please answer to this question: " + sentences inputs = tokenizer(question, return_tensors="pt") print(inputs) # print(inputs.device) print(model.device) print(inputs['input_ids'].device) print(inputs['attention_mask'].device) inputs['attention_mask'] = inputs['attention_mask'].to(device) inputs['input_ids'] = inputs['input_ids'].to(device) outputs = model.generate(**inputs) answer = tokenizer.decode(outputs[0]) answer = answer[6:-4] # print(fill(answer, width=80)) print("ANSWER IS", answer) return answer def random_word(model, tokenizer): global model_base vocab = tokenizer.get_vocab() # with open(model_base + '/vocab.txt', 'r') as file: line = "" # content = file.readlines() length = tokenizer.vocab_size # print(vocab) while line == "": rand_line = random.randrange(0, length) # print("TRYING TO FIND", rand_line, "OUT OF", length, "WITH VOCAB OF TYPE", type(vocab)) for word, id in vocab.items(): if id == rand_line and word[0].isalpha() and word not in stops and word not in ROMAN_CONSTANTS: # if vocab[rand_line][0].isalpha() and vocab[rand_line][:-1] not in stops and vocab[rand_line][:-1] not in ROMAN_CONSTANTS: line = word elif id == rand_line: print(f"{word} is not alpha or is a stop word") # for num, aline in enumerate(file, 1997): # if random.randrange(num) and aline.isalpha(): # continue # # elif not aline.isalpha(): # line = aline print(line) return line def generate_prompt(model, tokenizer): global word1 global word2 global word3 global answer global base_prompts word1 = random_word(model, tokenizer) # word2 = random_word() word2 = embeddings(model, f"{base_prompts[random.randint(0, len(base_prompts) - 1)]}{word1} is to ___.", tokenizer) word3 = random_word(model, tokenizer) sentence = f"{word1} is to {word2} as {word3} is to ___." print(sentence) answer = embeddings(model, sentence, tokenizer) print("ANSWER IS", answer) return f"# {word1} is to {word2} as {word3} is to ___." # cosine_scores(model, sentence) def greet(name): return "Hello " + name + "!!" def check_answer(guess:str): global guesses global answer global return_guesses global word1 global word2 global word3 model, tokenizer = get_model() output = "" protected_guess = guess sentence = f"{word1} is to {word2} as [MASK] is to {guess}." other_word = embeddings(model, sentence, tokenizer) guesses.append(guess) for guess in return_guesses: output += ("- " + guess + "
") # output = output[:-1] prompt = f"{word1} is to {word2} as {word3} is to ___." # print("IS", protected_guess, "EQUAL TO", answer, ":", protected_guess.lower() == answer.lower()) if protected_guess.lower() == answer.lower(): return_guesses.append(f"{protected_guess}: {word1} is to {word2} as {word3} is to {protected_guess}.") output += f"- {return_guesses[-1]}
" new_prompt = generate_prompt(model, tokenizer) return new_prompt, "Correct!", output else: return_guess = f"{protected_guess}: {word1} is to {word2} as {other_word} is to {protected_guess}." return_guesses.append(return_guess) output += ("- " + return_guess + "
") return prompt, "Try again!", output def main(): global word1 global word2 global word3 global answer # answer = "Moon" global guesses # num_rows, data_type, value, example, embeddings = training() # sent_embeddings = embeddings() model, tokenizer = get_model() generate_prompt(model, tokenizer) prompt = f"{word1} is to {word2} as {word3} is to ____" print(prompt) print("TESTING EMBEDDINGS") with gr.Blocks() as iface: mark_question = gr.Markdown(prompt) with gr.Tab("Guess"): text_input = gr.Textbox() text_output = gr.Textbox() text_button = gr.Button("Submit") with gr.Accordion("Open for previous guesses"): text_guesses = gr.Markdown() # with gr.Tab("Testing"): # gr.Markdown(f"""The Embeddings are {sent_embeddings}.""") text_button.click(check_answer, inputs=[text_input], outputs=[mark_question, text_output, text_guesses]) # iface = gr.Interface(fn=greet, inputs="text", outputs="text") iface.launch() if __name__ == "__main__": main()