Merge pull request #677 from lhotse-speech/feature/fix-assertion

Fix assertions, rename variables

Author: pzelasko

lhotse/bin/modes/features.py

@@ -277,9 +277,9 @@ def upload(
 
 
 def _upload_one(item: Features, url: str) -> Features:
-    feats_mtx = item.load()
+    feats_mat = item.load()
     feats_writer = LilcomURLWriter(url)
-    new_key = feats_writer.write(key=item.storage_key, value=feats_mtx)
+    new_key = feats_writer.write(key=item.storage_key, value=feats_mat)
     return fastcopy(
         item, storage_path=url, storage_key=new_key, storage_type=feats_writer.name
     )

lhotse/cut.py

@@ -4650,37 +4650,37 @@ def compute_and_store_features_batch(
                         waves, sampling_rate=cuts[0].sampling_rate
                     )
 
-                for cut, feat_mtx in zip(cuts, features):
+                for cut, feat_mat in zip(cuts, features):
                     if isinstance(cut, PaddingCut):
                         # For padding cuts, just fill out the fields in the manfiest
                         # and don't store anything.
                         cuts_writer.write(
                             fastcopy(
                                 cut,
-                                num_frames=feat_mtx.shape[0],
-                                num_features=feat_mtx.shape[1],
+                                num_frames=feat_mat.shape[0],
+                                num_features=feat_mat.shape[1],
                                 frame_shift=frame_shift,
                             )
                         )
                         continue
                     # Store the computed features and describe them in a manifest.
-                    if isinstance(feat_mtx, torch.Tensor):
-                        feat_mtx = feat_mtx.cpu().numpy()
-                    storage_key = feats_writer.write(cut.id, feat_mtx)
+                    if isinstance(feat_mat, torch.Tensor):
+                        feat_mat = feat_mat.cpu().numpy()
+                    storage_key = feats_writer.write(cut.id, feat_mat)
                     feat_manifest = Features(
                         start=cut.start,
                         duration=cut.duration,
                         type=extractor.name,
-                        num_frames=feat_mtx.shape[0],
-                        num_features=feat_mtx.shape[1],
+                        num_frames=feat_mat.shape[0],
+                        num_features=feat_mat.shape[1],
                         frame_shift=frame_shift,
                         sampling_rate=cut.sampling_rate,
                         channels=0,
                         storage_type=feats_writer.name,
                         storage_path=str(feats_writer.storage_path),
                         storage_key=storage_key,
                     )
-                    validate_features(feat_manifest, feats_data=feat_mtx)
+                    validate_features(feat_manifest, feats_data=feat_mat)
 
                     # Update the cut manifest.
                     if isinstance(cut, MonoCut):

lhotse/features/io.py

@@ -824,7 +824,7 @@ def __init__(
 
         if "b" not in mode:
             mode = mode + "b"
-        assert mode == "wb" or "ab"
+        assert mode in ("wb", "ab")
 
         # ".lca" -> "lilcom chunky archive"
         self.storage_path_ = Path(storage_path).with_suffix(".lca")

lhotse/qa.py

@@ -7,7 +7,7 @@
 
 from lhotse.array import Array, TemporalArray
 from lhotse.audio import Recording, RecordingSet
-from lhotse.cut import Cut, CutSet, MixedCut, PaddingCut
+from lhotse.cut import Cut, CutSet, MixedCut, MonoCut, PaddingCut
 from lhotse.features import FeatureSet, Features
 from lhotse.supervision import SupervisionSegment, SupervisionSet
 from lhotse.utils import compute_num_frames, overlaps
@@ -379,17 +379,20 @@ def validate_cut(c: Cut, read_data: bool = False) -> None:
                 c.num_samples == samples.shape[1]
             ), f"MonoCut {c.id}: expected {c.num_samples} samples, got {samples.shape[1]}"
 
-    # Conditions related to supervisions
-    for s in c.supervisions:
-        validate_supervision(s)
-        assert s.recording_id == c.recording_id, (
-            f"MonoCut {c.id}: supervision {s.id} has a mismatched recording_id "
-            f"(expected {c.recording_id}, supervision has {s.recording_id})"
-        )
-        assert s.channel == c.channel, (
-            f"MonoCut {c.id}: supervision {s.id} has a mismatched channel "
-            f"(expected {c.channel}, supervision has {s.channel})"
-        )
+    # Conditions related to supervisions.
+    # We only validate those for MonoCut; PaddingCut doesn't have supervisions,
+    # and MixedCut may consist of more than one recording/channel.
+    if isinstance(c, MonoCut):
+        for s in c.supervisions:
+            validate_supervision(s)
+            assert s.recording_id == c.recording_id, (
+                f"MonoCut {c.id}: supervision {s.id} has a mismatched recording_id "
+                f"(expected {c.recording_id}, supervision has {s.recording_id})"
+            )
+            assert s.channel == c.channel, (
+                f"MonoCut {c.id}: supervision {s.id} has a mismatched channel "
+                f"(expected {c.channel}, supervision has {s.channel})"
+            )
 
     # Conditions related to custom fields
     if c.custom is not None: