[BUG] - Bleu of machine translation

Add Link

https://pytorch.org/tutorials/beginner/translation_transformer.html

Describe the bug

• Expected Blue score of 35 -30

• got near 0.6 - tested on test set got from Multi30k -Attaching here. (format: de|en)

• Same code as the tutorial

• code used to calculate Bleu score:

  from nltk.translate.bleu_score import sentence_bleu import numpy as np def bleu4(candidate, reference): score = sentence_bleu([reference], candidate, weights=(0.25, 0.25, 0.25, 0.25)) # score = sentence_bleu([reference], candidate) return score

• Modified BlocK:

  `from timeit import default_timer as timer
  NUM_EPOCHS = 100
  
  for epoch in range(1, NUM_EPOCHS+1):
      start_time = timer()
      train_loss = train_epoch(transformer, optimizer)
      end_time = timer()
      val_loss = evaluate(transformer)
      print((f"Epoch: {epoch}, Train loss: {train_loss:.3f}, Val loss: {val_loss:.3f}, "f"Epoch time = {(end_time - start_time):.3f}s"))
  
      if epoch % 10 == 0:
          transformer.eval()
          i=0
          total_bleu = 0
          test = open("test.txt", "r").readlines()
          for r in test:
              data = r.strip().split('|')
              reference = data[1].split()
              candidate = translate(transformer, data[0]).split()
              total_bleu += bleu4(candidate, reference)
              print ("----------------- START -----------------")
              print ("GT: ", reference)
              print ("OUT: ", candidate)
              print ("BLEU for this example: ", bleu4(candidate, reference))
              print ("Average BLEU: ", total_bleu/(i+1))
  
              print ("----------------- END -----------------")
              i+=1
          print ("------FINAL BLEU: ", total_bleu/(i))
          with open("bleu_ours.txt", "a") as f:
              f.write("Epoch: " + str(epoch) + " BLEU: " + str(total_bleu/(i)) + "\n")
          if epoch % 50 == 0:
              torch.save(transformer.state_dict(), "transformer_ours"+str(epoch)+".pt")
  

  `

Describe your environment

Linux
Name: torch
Version: 2.0.1

cc @pytorch/team-text-core @Nayef211

Start with the linked translation transformer tutorial and compare its evaluation flow with the modified training block. Reproduce the reported result using test.txt and the shown NLTK sentence_bleu call, then determine whether the tutorial or its evaluation guidance needs correction; done means the expected BLEU calculation and any required tutorial change are clearly established.