update rounding, train logging (facebookresearch#431)

examples/train_model.py

@@ -225,9 +225,9 @@ def train(self):
                 if opt['num_epochs'] > 0 and self.max_parleys is not None and (
                     (self.max_parleys > 0 and self.parleys >= self.max_parleys)
                     or self.total_epochs >= opt['num_epochs']):
+                    self.log()
                     print('[ num_epochs completed:{} time elapsed:{}s ]'.format(
                         opt['num_epochs'], self.train_time.time()))
-                    self.log()
                     break
                 if opt['max_train_time'] > 0 and self.train_time.time() > opt['max_train_time']:
                     print('[ max_train_time elapsed:{}s ]'.format(self.train_time.time()))

parlai/agents/seq2seq/seq2seq.py

@@ -357,7 +357,7 @@ def predict(self, xs, ys=None, cands=None, valid_cands=None, lm=False):
             loss.backward()
             self.update_params()
             losskey = 'loss' if not lm else 'lmloss'
-            loss_dict = {losskey: loss.mul_(len(xs)).data[0]}
+            loss_dict = {losskey: loss.mul_(len(xs)).data}
         else:
             self.model.eval()
             predictions, scores, text_cand_inds = self.model(xs, ys, cands,

parlai/core/utils.py

@@ -88,12 +88,13 @@ def round_sigfigs(x, sigfigs=4):
     try:
         if x == 0:
             return 0
-    except RuntimeError:
+        if x in [float('inf'), float('-inf'), float('NaN')]:
+            return x
+        return round(x, -math.floor(math.log10(abs(x)) - sigfigs + 1))
+    except TypeError:
         # handle 1D torch tensors
-        x = x[0]
-    if x in [float('inf'), float('-inf'), float('NaN')]:
-        return x
-    return round(x, -math.floor(math.log10(abs(x)) - sigfigs + 1))
+        # if anything else breaks here please file an issue on Github
+        return round_sigfigs(x[0], sigfigs)
 
 
 def flatten(teacher, context_length=-1, include_labels=True):