From a2a79c7992f8ef45e1843d419236c5b0b3eb5ac8 Mon Sep 17 00:00:00 2001
From: suxia <yangsuxia@163.com>
Date: Wed, 10 Aug 2022 17:59:31 +0800
Subject: [PATCH] Add function display_and_save_batch in
 egs/wenetspeech/ASR/pruned_transducer_stateless2/train.py

---
 .../ASR/pruned_transducer_stateless2/train.py | 71 ++++++++++++++-----
 1 file changed, 53 insertions(+), 18 deletions(-)

diff --git a/egs/wenetspeech/ASR/pruned_transducer_stateless2/train.py b/egs/wenetspeech/ASR/pruned_transducer_stateless2/train.py
index faf25eda1..d6c4de58f 100644
--- a/egs/wenetspeech/ASR/pruned_transducer_stateless2/train.py
+++ b/egs/wenetspeech/ASR/pruned_transducer_stateless2/train.py
@@ -701,25 +701,29 @@ def train_one_epoch(
         params.batch_idx_train += 1
         batch_size = len(batch["supervisions"]["text"])
 
-        with torch.cuda.amp.autocast(enabled=params.use_fp16):
-            loss, loss_info = compute_loss(
-                params=params,
-                model=model,
-                graph_compiler=graph_compiler,
-                batch=batch,
-                is_training=True,
-                warmup=(params.batch_idx_train / params.model_warm_step),
-            )
-        # summary stats
-        tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, loss_info = compute_loss(
+                    params=params,
+                    model=model,
+                    graph_compiler=graph_compiler,
+                    batch=batch,
+                    is_training=True,
+                    warmup=(params.batch_idx_train / params.model_warm_step),
+                )
+            # summary stats
+            tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
 
-        # NOTE: We use reduction==sum and loss is computed over utterances
-        # in the batch and there is no normalization to it so far.
-        scaler.scale(loss).backward()
-        scheduler.step_batch(params.batch_idx_train)
-        scaler.step(optimizer)
-        scaler.update()
-        optimizer.zero_grad()
+            # NOTE: We use reduction==sum and loss is computed over utterances
+            # in the batch and there is no normalization to it so far.
+            scaler.scale(loss).backward()
+            scheduler.step_batch(params.batch_idx_train)
+            scaler.step(optimizer)
+            scaler.update()
+            optimizer.zero_grad()
+        except:    # noqa
+            display_and_save_batch(batch, params=params)
+            raise
 
         if params.print_diagnostics and batch_idx == 5:
             return
@@ -957,6 +961,34 @@ def run(rank, world_size, args):
         torch.distributed.barrier()
         cleanup_dist()
 
+def display_and_save_batch(
+    batch: dict,
+    params: AttributeDict,
+) -> None:
+    """Display the batch statistics and save the batch into disk.
+
+    Args:
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      params:
+        Parameters for training. See :func:`get_params`.
+      sp:
+        The BPE model.
+    """
+    from lhotse.utils import uuid4
+
+    filename = f"{params.exp_dir}/batch-{uuid4()}.pt"
+    logging.info(f"Saving batch to {filename}")
+    torch.save(batch, filename)
+
+    supervisions = batch["supervisions"]
+    features = batch["inputs"]
+
+    logging.info(f"features shape: {features.shape}")
+
+    num_tokens = params.vocab_size
+    logging.info(f"num tokens: {num_tokens}")
 
 def scan_pessimistic_batches_for_oom(
     model: nn.Module,
@@ -998,6 +1030,7 @@ def scan_pessimistic_batches_for_oom(
                     f"Failing criterion: {criterion} "
                     f"(={crit_values[criterion]}) ..."
                 )
+            display_and_save_batch(batch, params=params)
             raise
 
 
@@ -1021,3 +1054,5 @@ torch.set_num_interop_threads(1)
 
 if __name__ == "__main__":
     main()
+
+