fix about tensorboard (#516)

* fix metricstracker

* fix style

Author: yaozengwei

egs/librispeech/ASR/conformer_ctc/train.py

@@ -447,6 +447,17 @@ def compute_loss(
 
     info["loss"] = loss.detach().cpu().item()
 
+    # `utt_duration` and `utt_pad_proportion` would be normalized by `utterances`  # noqa
+    info["utterances"] = feature.size(0)
+    # averaged input duration in frames over utterances
+    info["utt_duration"] = supervisions["num_frames"].sum().item()
+    # averaged padding proportion over utterances
+    info["utt_pad_proportion"] = (
+        ((feature.size(1) - supervisions["num_frames"]) / feature.size(1))
+        .sum()
+        .item()
+    )
+
     return loss, info
 
 

egs/librispeech/ASR/conformer_ctc2/train.py

@@ -605,6 +605,15 @@ def compute_loss(
     # Note: We use reduction=sum while computing the loss.
     info["loss"] = loss.detach().cpu().item()
 
+    # `utt_duration` and `utt_pad_proportion` would be normalized by `utterances`  # noqa
+    info["utterances"] = feature.size(0)
+    # averaged input duration in frames over utterances
+    info["utt_duration"] = feature_lens.sum().item()
+    # averaged padding proportion over utterances
+    info["utt_pad_proportion"] = (
+        ((feature.size(1) - feature_lens) / feature.size(1)).sum().item()
+    )
+
     return loss, info
 
 

egs/librispeech/ASR/streaming_conformer_ctc/train.py

@@ -430,6 +430,17 @@ def compute_loss(
 
     info["loss"] = loss.detach().cpu().item()
 
+    # `utt_duration` and `utt_pad_proportion` would be normalized by `utterances`  # noqa
+    info["utterances"] = feature.size(0)
+    # averaged input duration in frames over utterances
+    info["utt_duration"] = supervisions["num_frames"].sum().item()
+    # averaged padding proportion over utterances
+    info["utt_pad_proportion"] = (
+        ((feature.size(1) - supervisions["num_frames"]) / feature.size(1))
+        .sum()
+        .item()
+    )
+
     return loss, info
 
 

egs/librispeech/ASR/tdnn_lstm_ctc/train.py

@@ -349,6 +349,17 @@ def compute_loss(
     info["frames"] = supervision_segments[:, 2].sum().item()
     info["loss"] = loss.detach().cpu().item()
 
+    # `utt_duration` and `utt_pad_proportion` would be normalized by `utterances`  # noqa
+    info["utterances"] = feature.size(0)
+    # averaged input duration in frames over utterances
+    info["utt_duration"] = supervisions["num_frames"].sum().item()
+    # averaged padding proportion over utterances
+    info["utt_pad_proportion"] = (
+        ((feature.size(2) - supervisions["num_frames"]) / feature.size(2))
+        .sum()
+        .item()
+    )
+
     return loss, info
 
 

egs/librispeech/ASR/transducer/train.py

@@ -403,6 +403,15 @@ def compute_loss(
     # Note: We use reduction=sum while computing the loss.
     info["loss"] = loss.detach().cpu().item()
 
+    # `utt_duration` and `utt_pad_proportion` would be normalized by `utterances`  # noqa
+    info["utterances"] = feature.size(0)
+    # averaged input duration in frames over utterances
+    info["utt_duration"] = feature_lens.sum().item()
+    # averaged padding proportion over utterances
+    info["utt_pad_proportion"] = (
+        ((feature.size(1) - feature_lens) / feature.size(1)).sum().item()
+    )
+
     return loss, info
 
 

egs/librispeech/ASR/transducer_lstm/train.py

@@ -407,6 +407,15 @@ def compute_loss(
     # Note: We use reduction=sum while computing the loss.
     info["loss"] = loss.detach().cpu().item()
 
+    # `utt_duration` and `utt_pad_proportion` would be normalized by `utterances`  # noqa
+    info["utterances"] = feature.size(0)
+    # averaged input duration in frames over utterances
+    info["utt_duration"] = feature_lens.sum().item()
+    # averaged padding proportion over utterances
+    info["utt_pad_proportion"] = (
+        ((feature.size(1) - feature_lens) / feature.size(1)).sum().item()
+    )
+
     return loss, info
 
 

egs/librispeech/ASR/transducer_stateless/train.py

@@ -429,6 +429,15 @@ def compute_loss(
     # Note: We use reduction=sum while computing the loss.
     info["loss"] = loss.detach().cpu().item()
 
+    # `utt_duration` and `utt_pad_proportion` would be normalized by `utterances`  # noqa
+    info["utterances"] = feature.size(0)
+    # averaged input duration in frames over utterances
+    info["utt_duration"] = feature_lens.sum().item()
+    # averaged padding proportion over utterances
+    info["utt_pad_proportion"] = (
+        ((feature.size(1) - feature_lens) / feature.size(1)).sum().item()
+    )
+
     return loss, info
 
 

egs/librispeech/ASR/transducer_stateless2/train.py

@@ -417,6 +417,15 @@ def compute_loss(
     # Note: We use reduction=sum while computing the loss.
     info["loss"] = loss.detach().cpu().item()
 
+    # `utt_duration` and `utt_pad_proportion` would be normalized by `utterances`  # noqa
+    info["utterances"] = feature.size(0)
+    # averaged input duration in frames over utterances
+    info["utt_duration"] = feature_lens.sum().item()
+    # averaged padding proportion over utterances
+    info["utt_pad_proportion"] = (
+        ((feature.size(1) - feature_lens) / feature.size(1)).sum().item()
+    )
+
     return loss, info
 
 

egs/librispeech/ASR/transducer_stateless_multi_datasets/train.py

@@ -476,6 +476,15 @@ def compute_loss(
     # Note: We use reduction=sum while computing the loss.
     info["loss"] = loss.detach().cpu().item()
 
+    # `utt_duration` and `utt_pad_proportion` would be normalized by `utterances`  # noqa
+    info["utterances"] = feature.size(0)
+    # averaged input duration in frames over utterances
+    info["utt_duration"] = feature_lens.sum().item()
+    # averaged padding proportion over utterances
+    info["utt_pad_proportion"] = (
+        ((feature.size(1) - feature_lens) / feature.size(1)).sum().item()
+    )
+
     return loss, info
 
 

icefall/utils.py

@@ -544,9 +544,10 @@ def __str__(self) -> str:
                 else:
                     raise ValueError(f"Unexpected key: {k}")
         frames = "%.2f" % self["frames"]
-        ans_frames += "over " + str(frames) + " frames; "
-        utterances = "%.2f" % self["utterances"]
-        ans_utterances += "over " + str(utterances) + " utterances."
+        ans_frames += "over " + str(frames) + " frames. "
+        if ans_utterances != "":
+            utterances = "%.2f" % self["utterances"]
+            ans_utterances += "over " + str(utterances) + " utterances."
 
         return ans_frames + ans_utterances