add timer info

deepspeech/training/gradclip.py

@@ -48,7 +48,6 @@ class ClipGradByGlobalNormWithLog(paddle.nn.ClipGradByGlobalNorm):
             sum_square_list.append(sum_square)
 
             # debug log
-            if i < 10:
             logger.debug(
                 f"Grad Before Clip: {p.name}: {float(sum_square.sqrt()) }")
 
@@ -77,7 +76,6 @@ class ClipGradByGlobalNormWithLog(paddle.nn.ClipGradByGlobalNorm):
             params_and_grads.append((p, new_grad))
 
             # debug log
-            if i < 10:
             logger.debug(
                 f"Grad After Clip: {p.name}: {float(new_grad.square().sum().sqrt())}"
             )

deepspeech/training/timer.py

@@ -27,7 +27,7 @@ class Timer():
             do some thing
     """
 
-    def __init__(self, message):
+    def __init__(self, message=None):
         self.message = message
 
     def duration(self) -> str:
@@ -40,6 +40,7 @@ class Timer():
         return self
 
     def __exit__(self, type, value, traceback):
+        if self.message:
             logger.info(self.message.format(self.duration()))
 
     def __call__(self) -> float:

deepspeech/training/trainer.py

@@ -185,7 +185,6 @@ class Trainer():
 
     def train(self):
         """The training process control by epoch."""
-        with Timer("Load/Init Model: {}"):
         from_scratch = self.resume_or_scratch()
         if from_scratch:
             # save init model, i.e. 0 epoch
@@ -196,6 +195,7 @@ class Trainer():
 
         logger.info(f"Train Total Examples: {len(self.train_loader.dataset)}")
         while self.epoch < self.config.training.n_epoch:
+            with Timer("Epoch-Train Time Cost: {}"):
                 self.model.train()
                 try:
                     data_start_time = time.time()
@@ -214,6 +214,7 @@ class Trainer():
                     logger.error(e)
                     raise e
 
+            with Timer("Eval Time Cost: {}"):
                 total_loss, num_seen_utts = self.valid()
                 if dist.get_world_size() > 1:
                     num_seen_utts = paddle.to_tensor(num_seen_utts)