Update train.py

1.benchmark needn't eval 
2.let benchmark test can modify the training epoch

paddlespeech/kws/exps/mdtc/train.py

@@ -91,7 +91,7 @@ if __name__ == '__main__':
     steps_per_epoch = len(train_sampler)
     timer = Timer(steps_per_epoch * config['epochs'])
     timer.start()
-
+    config['epochs'] = args.benchmark_max_step
     for epoch in range(1, config['epochs'] + 1):
         model.train()
 
@@ -137,33 +137,9 @@ if __name__ == '__main__':
 
         if epoch % config[
                 'save_freq'] == 0 and batch_idx + 1 == steps_per_epoch and local_rank == 0:
-            dev_sampler = paddle.io.BatchSampler(
+            
-                dev_ds,
-                batch_size=config['batch_size'],
-                shuffle=False,
-                drop_last=False)
-            dev_loader = paddle.io.DataLoader(
-                dev_ds,
-                batch_sampler=dev_sampler,
-                num_workers=config['num_workers'],
-                return_list=True,
-                use_buffer_reader=True,
-                collate_fn=collate_features, )
-
-            model.eval()
             num_corrects = 0
             num_samples = 0
-            with logger.processing('Evaluation on validation dataset'):
-                for batch_idx, batch in enumerate(dev_loader):
-                    keys, feats, labels, lengths = batch
-                    logits = model(feats)
-                    loss, corrects, acc = criterion(logits, labels, lengths)
-                    num_corrects += corrects
-                    num_samples += feats.shape[0]
-
-            eval_acc = num_corrects / num_samples
-            print_msg = '[Evaluation result]'
-            print_msg += ' dev_acc={:.4f}'.format(eval_acc)
 
             logger.eval(print_msg)