Sentence Classification Model.py

# Importing the correct libraries.

!pip install datasets huggingface_hub
import torch
from datasets import load_dataset
from transformers import AutoTokenizer, AutoModelForSequenceClassification, AdamW
from torch.utils.data import DataLoader
from sklearn.metrics import accuracy_score, f1_score

# Importing the datasets and defining a function for preprocessing.

sst2 = load_dataset("stanfordnlp/sst2")
imdb = load_dataset("stanfordnlp/imdb")

def data_preprocessor_sst2(data, token):
  return data.map(lambda x: token(x['sentence'], truncation = True, padding = 'max_length', max_length = 128), batched = True)
def data_preprocessor_imdb(data, token):
  return data.map(lambda x: token(x['text'], truncation = True, padding = 'max_length', max_length = 128), batched = True)
  
# Loading Bert and Distilbert tokenizers to tokenize the datasets.

token_B = AutoTokenizer.from_pretrained("bert-base-uncased")
token_DB = AutoTokenizer.from_pretrained("distilbert-base-uncased")

train_sst2 = data_preprocessor_sst2(sst2['train'], token_B)
test_sst2 = data_preprocessor_sst2(sst2['validation'], token_B)
test_imdb = data_preprocessor_imdb(imdb['test'], token_B)

# Formatting the datasets to fit the model and transformers.

train_sst2.set_format(type = 'torch', columns = ['input_ids', 'attention_mask', 'label'])
test_sst2.set_format(type = 'torch', columns = ['input_ids', 'attention_mask', 'label'])
test_imdb.set_format(type = 'torch', columns = ['input_ids', 'attention_mask', 'label'])

# Creating a data loader for each of the splits.

def data_loader(data, batch = 32):
  return DataLoader(data, batch_size = batch, shuffle = True)

train_DL = data_loader(train_sst2)
test_DL = data_loader(test_sst2)
test_DL_imdb = data_loader(test_imdb)

# Loading the Bert and Distilbert models.

model_B = AutoModelForSequenceClassification.from_pretrained("bert-base-uncased", num_labels = 2)
model_DB = AutoModelForSequenceClassification.from_pretrained("distilbert-base-uncased", num_labels = 2)

# Building a function to train and optimize the model.

def model_trainer(model, data, epochs = 1):
  gpu = torch.device("cuda" if torch.cuda.is_available() else "cpu")
  model.to(gpu)
  optimizer = AdamW(model.parameters(), lr = 0.00005)
  model.train()
  for x in range(epochs):
    loss_total = 0
    for y in data:
      optimizer.zero_grad()
      input_ids, attention_mask, labels = (y['input_ids'].to(gpu), y['attention_mask'].to(gpu), y['label'].to(gpu))
      forward_info = model(input_ids, attention_mask = attention_mask, labels = labels)
      forward_loss = forward_info.loss
      loss_total += forward_loss.item()
      forward_loss.backward()
      optimizer.step()
    print(f"Epoch - Loss = {loss_total / len(data)}")
  return model
  
# Building a function to evaluate the model.

def model_evaluator(model, data):
  model.eval()
  gpu = torch.device("cuda" if torch.cuda.is_available() else "cpu")
  model.to(gpu)
  preds, trues = [], []
  with torch.no_grad():
    for y in data:
      input_ids, attention_mask, labels = (y['input_ids'].to(gpu), y['attention_mask'].to(gpu), y['label'].to(gpu))
      forward_info = model(input_ids, attention_mask = attention_mask)
      forward_loss = forward_info.logits
      preds.extend(torch.argmax(forward_loss, dim = 1).cpu().numpy())
      trues.extend(labels.cpu().numpy())
  score_acc = accuracy_score(trues, preds)
  score_f1 = f1_score(trues, preds)
  return score_acc, score_f1
  
# Training the Bert and Distilbert models.

model_B = model_trainer(model_B, train_DL)
model_DB = model_trainer(model_DB, train_DL)

# Evaluating the scores of the models on the SST2 dataset.

bert_acc_sst2, bert_f1_sst2 = model_evaluator(model_B, test_DL)
dist_acc_sst2, dist_f1_sst2 = model_evaluator(model_DB, test_DL)

print(f"Bert Accuracy: {bert_acc_sst2}, Bert F1: {bert_f1_sst2}")
print(f"Distilbert Accuracy: {dist_acc_sst2}, Distilbert F1: {dist_f1_sst2}")

# Evaluating the scores of the models on the IMDB dataset.

bert_acc_imdb, bert_f1_imdb = model_evaluator(model_B, test_DL_imdb)
dist_acc_imdb, dist_f1_imdb = model_evaluator(model_DB, test_DL_imdb)

print(f"Bert Accuracy: {bert_acc_imdb}, Bert F1: {bert_f1_imdb}")
print(f"Distilbert Accuracy: {dist_acc_imdb}, Distilbert F1: {dist_f1_imdb}")

# Defining a function to classify the sentences

def classify(sentence: str, model, token):
    gpu = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    model.to(gpu)
    model.eval()
    inputs = token(sentence, return_tensors = "pt", truncation = True, padding = True, max_length = 128).to(gpu)
    with torch.no_grad():
        forward_info = model(**inputs)
    pred = forward_info.logits.argmax(dim = 1).item()
    return pred
    
# Testing sentence 1.

sentence1 = "for all its highfalutin title and corkscrew narrative , the movie turns out to be not much more than a shaggy human tale ."
print(f"BERT Prediction: {classify(sentence1, model_B, token_B)}")
print(f"DistilBERT Prediction: {classify(sentence1, model_DB, token_DB)}")

# Testing sentence 2.

sentence2 = "its underlying mythology is a hodgepodge of inconsistencies that pose the question : since when did dumb entertainment have to be this dumb ?"
print(f"BERT Prediction: {classify(sentence2, model_B, token_B)}")
print(f"DistilBERT Prediction: {classify(sentence2, model_DB, token_DB)}")

# Testing sentence 3.

sentence3 = '''
the actors do n’t inhabit their roles– they ’re trapped by them ,
forced to change behavior in bizarre unjustified fashion and spout dialog that consists mostly of platitudes .
'''
print(f"BERT Prediction: {classify(sentence3, model_B, token_B)}")
print(f"DistilBERT Prediction: {classify(sentence3, model_DB, token_DB)}")

# Testing sentence 4.

sentence4 = "an absorbing trip into the minds and motivations of people under stress as well as a keen , unsentimental look at variations on the theme of motherhood ."
print(f"BERT Prediction: {classify(sentence4, model_B, token_B)}")
print(f"DistilBERT Prediction: {classify(sentence4, model_DB, token_DB)}")

# Testing sentence 5.

sentence5 = "one of those rare , exhilarating cinematic delights that gets even better in hindsight , as you mull over its every nuance in your mind ."
print(f"BERT Prediction: {classify(sentence5, model_B, token_B)}")
print(f"DistilBERT Prediction: {classify(sentence5, model_DB, token_DB)}")