Sentiment analysis with transformers

sentiment_with_bert.py

from transformers import BertForSequenceClassification, BertTokenizer

## Step 1 - Pre-processing
MODEL = 'YOUR MODEL'
TEXT = ("YOUR INPUTS")
tokenizer = BertTokenizer.from_pretrained(MODEL)
tokens = tokenizer.encode_plus(TEXT, max_length=512, # max number of tokens in each sample
                               truncation=True, 	 # what to do with extra token over max_length
                               padding='max_length', # for shorter sequences, pad with 0's
                               add_special_tokens=True, # add special tokens by default
                               return_tensors='pt')  # return TensorFlow tensors (tf)/PyTorch (pt)/Numpy (np)
# from the tokens we need:
#   input_ids - token ID representations and 
#   attention_mask - tells which words to calcuate attention for

## Step 2 - Feed into the model
model = BertForSequenceClassification.from_pretrained(MODEL)
activations = model(**tokens) # spread the keyword arguments

## Step 3 - Get sentiment
import tensorflow as tf
probabilities = tf.nn.softmax(activations[0].detach().numpy()) # convert activations first
predictions = tf.math.argmax(probabilities, axis=1)            # pick the max
predictions.numpy()

# alternatively if using PyTorch
import torch
probs = torch.nn.functional.softmax(activations[0], dim=-1)
pred = torch.argmax(probs)
pred.item()