[MXNET-171] Fix a bug that was causing training accuracy to be printed as nan sometimes (apache#10437)

* Fix a bug that was causing training accuracy to be printed as nan sometimes.

* Avoid the additional 'arg_eval_metric' variable. There should be no
overhead except for the batch in an epoch.

* Fix lint.

* For the last batch, Capture metrics before callback and use it to print epoch metrics

* Remove unused import

Author: indhub

python/mxnet/module/base_module.py

@@ -522,6 +522,9 @@ def fit(self, train_data, eval_data=None, eval_metric='acc',
                 if monitor is not None:
                     monitor.toc_print()
 
+                if end_of_batch:
+                    eval_name_vals = eval_metric.get_name_value()
+
                 if batch_end_callback is not None:
                     batch_end_params = BatchEndParam(epoch=epoch, nbatch=nbatch,
                                                      eval_metric=eval_metric,
@@ -531,7 +534,7 @@ def fit(self, train_data, eval_data=None, eval_metric='acc',
                 nbatch += 1
 
             # one epoch of training is finished
-            for name, val in eval_metric.get_name_value():
+            for name, val in eval_name_vals:
                 self.logger.info('Epoch[%d] Train-%s=%f', epoch, name, val)
             toc = time.time()
             self.logger.info('Epoch[%d] Time cost=%.3f', epoch, (toc-tic))