returnn 1.20240726.140811__tar.gz → 1.20240730.135048__tar.gz

{returnn-1.20240726.140811 → returnn-1.20240730.135048}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: returnn
-Version: 1.20240726.140811
+Version: 1.20240730.135048
 Summary: The RWTH extensible training framework for universal recurrent neural networks
 Home-page: https://github.com/rwth-i6/returnn/
 Author: Albert Zeyer

returnn-1.20240730.135048/_setup_info_generated.py

@@ -0,0 +1,2 @@
+version = '1.20240730.135048'
+long_version = '1.20240730.135048+git.61ad52a'

{returnn-1.20240726.140811 → returnn-1.20240730.135048}/returnn/datasets/audio.py

@@ -151,8 +151,6 @@ class OggZipDataset(CachedDataset2):
             self.num_outputs["classes"] = [self.targets.num_labels, 1]
         if self.feature_extractor:
             self.num_outputs["data"] = [self.num_inputs, 2]
-        else:
-            self.num_outputs["data"] = [0, 2]
         self._data: Optional[List[Dict[str, Any]]] = None  # lazily loaded
         self._fixed_random_subset = fixed_random_subset
         self._fixed_random_subset_seed = fixed_random_subset_seed
@@ -402,15 +400,46 @@ class OggZipDataset(CachedDataset2):
         self._lazy_init()
         return len(self._data)
 
-    def get_data_shape(self, key):
+    def get_data_dtype(self, key: str) -> str:
+        """:return: dtype of data entry with `key`"""
+        if key == "data":
+            return "float32"
+        elif key == "classes":
+            return "int32"
+        elif key == "raw":
+            return "string"
+        elif key == "orth":
+            return "uint8"
+        else:
+            raise ValueError(f"{self}: unknown data key: {key}")
+
+    def get_data_keys(self) -> List[str]:
+        """:return: available data keys"""
+        keys = []
+        if self.feature_extractor is not None:
+            keys.append("data")
+        if self.targets is not None:
+            keys.append("classes")
+        return [*keys, "orth", "raw"]
+
+    def get_data_shape(self, key: str):
         """
         :returns get_data(*, key).shape[1:], i.e. num-frames excluded
         :rtype: list[int]
         """
-        if key == "data" and self.feature_extractor is not None:
+        if key == "data":
+            assert self.feature_extractor is not None
             if self.feature_extractor.num_channels is not None:
                 return [self.feature_extractor.num_channels, self.feature_extractor.get_feature_dimension()]
-        return super(OggZipDataset, self).get_data_shape(key)
+            return [self.feature_extractor.get_feature_dimension()]
+        elif key in ["classes", "orth", "raw"]:
+            return []
+        else:
+            raise ValueError(f"{self}: unknown data key {key}")
+
+    def is_data_sparse(self, key: str) -> bool:
+        """:return: whether data entry with `key` is sparse"""
+        return key == "classes"
 
     def _get_transcription(self, corpus_seq_idx: int):
         """
@@ -467,13 +496,14 @@ class OggZipDataset(CachedDataset2):
         """
         self._lazy_init()
         seq_tag = self._get_tag_from_info_dict(self._data[corpus_seq_idx])
+        features = {}
         if self.feature_extractor:
             with self._open_audio_file(corpus_seq_idx) as audio_file:
-                features = self.feature_extractor.get_audio_features_from_raw_bytes(audio_file, seq_name=seq_tag)
-        else:
-            features = numpy.zeros((), dtype=numpy.float32)  # currently the API requires some dummy values...
+                data = self.feature_extractor.get_audio_features_from_raw_bytes(audio_file, seq_name=seq_tag)
+            features["data"] = data
         targets, txt = self._get_transcription(corpus_seq_idx)
-        targets = numpy.array(targets, dtype="int32")
+        if self.targets is not None:
+            features["classes"] = numpy.array(targets, dtype="int32")
         raw_txt = str_to_numpy_array(txt)
         orth = txt.encode("utf8")
         if PY3:
@@ -483,8 +513,7 @@ class OggZipDataset(CachedDataset2):
             orth = list(map(ord, orth))
         orth = numpy.array(orth, dtype="uint8")
         return DatasetSeq(
-            features=features,
-            targets={"classes": targets, "raw": raw_txt, "orth": orth},
+            features={**features, "raw": raw_txt, "orth": orth},
             seq_idx=corpus_seq_idx,
             seq_tag=seq_tag,
         )

{returnn-1.20240726.140811 → returnn-1.20240730.135048}/returnn/frontend/rand.py

@@ -205,7 +205,10 @@ def random(
             if seed is None:
                 seed = get_static_step_based_seed()
     if dtype is None:
-        dtype = rf.get_default_float_dtype()
+        if sparse_dim:
+            dtype = rf.get_default_array_index_dtype()
+        else:
+            dtype = rf.get_default_float_dtype()
     return _global_backend.random(
         dims=dims,
         dtype=dtype,

{returnn-1.20240726.140811 → returnn-1.20240730.135048}/returnn/tensor/dim.py

@@ -25,7 +25,6 @@ between a spatial dim and a feature dim (the batch dim is often handled differen
 
 """
 
-
 from __future__ import annotations
 from typing import Optional, Union
 
@@ -77,6 +76,8 @@ class Dim(_DimMixin):
             self.size = dimension
             self.dyn_size_ext = None
         elif isinstance(dimension, _t.Tensor):
+            if not dimension.dtype.startswith("int") and not dimension.dtype.startswith("uint"):
+                raise TypeError(f"unexpected dtype for dimension: {dimension.dtype}")
             self.capacity = capacity
             self.size = None
             self.dyn_size_ext = dimension.copy()

{returnn-1.20240726.140811 → returnn-1.20240730.135048}/returnn.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: returnn
-Version: 1.20240726.140811
+Version: 1.20240730.135048
 Summary: The RWTH extensible training framework for universal recurrent neural networks
 Home-page: https://github.com/rwth-i6/returnn/
 Author: Albert Zeyer

returnn-1.20240726.140811/_setup_info_generated.py

@@ -1,2 +0,0 @@
-version = '1.20240726.140811'
-long_version = '1.20240726.140811+git.396757d'