Fix the OOM problem during evaluation (#1580)

paddlehub/finetune/trainer.py

@@ -301,11 +301,13 @@ class Trainer(object):
                 collate_fn=collate_fn)
 
             self.model.eval()
+            
             avg_loss = num_samples = 0
             sum_metrics = defaultdict(int)
             avg_metrics = defaultdict(int)
 
             with logger.processing('Evaluation on validation dataset'):
+                with paddle.no_grad():
                     for batch_idx, batch in enumerate(loader):
                         result = self.validation_step(batch, batch_idx)