Merge pull request #24 from danishpruthi/master

Fixed the gradient updates on dev-set bug

Author: neubig

02-lm/loglin-lm.py

@@ -94,7 +94,6 @@ def generate_sent():
     my_loss = calc_sent_loss(sent)
     dev_loss += my_loss.value()
     dev_words += len(sent)
-    trainer.update()
   print("iter %r: dev loss/word=%.4f, ppl=%.4f, time=%.2fs" % (ITER, dev_loss/dev_words, math.exp(dev_loss/dev_words), time.time()-start))
   # Generate a few sentences
   for _ in range(5):

02-lm/nn-lm-batch.py

@@ -113,7 +113,6 @@ def generate_sent():
     my_loss = calc_sent_loss(sent)
     dev_loss += my_loss.value()
     dev_words += len(sent)
-    trainer.update()
   # Keep track of the development accuracy and reduce the learning rate if it got worse
   if last_dev < dev_loss:
     trainer.learning_rate /= 2

02-lm/nn-lm-optim.py

@@ -111,7 +111,6 @@ def generate_sent():
     my_loss = calc_sent_loss(sent)
     dev_loss += my_loss.value()
     dev_words += len(sent)
-    trainer.update()
   # CHANGE 3: Keep track of the development accuracy and reduce the learning rate if it got worse
   if last_dev < dev_loss:
     trainer.learning_rate /= 2

02-lm/nn-lm.py

@@ -103,7 +103,6 @@ def generate_sent():
     my_loss = calc_sent_loss(sent)
     dev_loss += my_loss.value()
     dev_words += len(sent)
-    trainer.update()
   print("iter %r: dev loss/word=%.4f, ppl=%.4f, time=%.2fs" % (ITER, dev_loss/dev_words, math.exp(dev_loss/dev_words), time.time()-start))
   # Generate a few sentences
   for _ in range(5):