returnn 1.20250228.104237__py3-none-any.whl → 1.20250304.10039__py3-none-any.whl

returnn/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: returnn
-Version: 1.20250228.104237
+Version: 1.20250304.10039
 Summary: The RWTH extensible training framework for universal recurrent neural networks
 Home-page: https://github.com/rwth-i6/returnn/
 Author: Albert Zeyer

returnn/_setup_info_generated.py

@@ -1,2 +1,2 @@
-version = '1.20250228.104237'
-long_version = '1.20250228.104237+git.bd23951'
+version = '1.20250304.010039'
+long_version = '1.20250304.010039+git.3e53d74'

returnn/tensor/_dim_extra.py

@@ -1271,9 +1271,9 @@ class _DimMixin:
             if self.batch:
                 x_dim = x_dim.get_for_batch_ctx(self.batch, self.control_flow_ctx)
             x_dim.complete_dyn_size(template_only=template_only, _backend=backend)
-            if x_dim.dyn_size_ext is None and not x_dim.dimension:
+            if x_dim.dyn_size_ext is None and x_dim.dimension is None:
                 return
-            y = _bin_op(y, x_dim.dimension or x_dim.dyn_size_ext)
+            y = _bin_op(y, x_dim.dimension if x_dim.dimension is not None else x_dim.dyn_size_ext)
             if not template_only and y.raw_tensor is not None:
                 y_max_value = _bin_op(y_max_value, x_dim.get_dim_value_tensor())
         assert y is not None, f"op {op}?"

returnn/tf/layers/basic.py

@@ -6587,6 +6587,7 @@ class ConvLayer(_ConcatInputLayer):
         input_split_feature_dim=None,
         input_add_feature_dim=False,
         use_time_mask=False,
+        mask_value: float = 0.0,
     ):
         """
         :param Data input_data:
@@ -6600,6 +6601,7 @@ class ConvLayer(_ConcatInputLayer):
         :param bool input_add_feature_dim: will add a dim at the end and use input-feature-dim == 1,
           and use the original input feature-dim as a spatial dim.
         :param bool use_time_mask:
+        :param mask_value: when ``use_time_mask`` is used, what value to use for the mask
         :return: (transformed input, num batch dims). all batch dims are at the front
         :rtype: (Data, int)
         """
@@ -6697,7 +6699,7 @@ class ConvLayer(_ConcatInputLayer):
                     continue
                 axis = input_data.get_axis_from_description(dim)
                 mask = input_data.get_sequence_mask_broadcast(axis=axis)
-                x = tf_util.where_bc(mask, x, 0.0)
+                x = tf_util.where_bc(mask, x, mask_value)
 
             input_data.placeholder = x
 
@@ -7061,6 +7063,7 @@ class PoolLayer(_ConcatInputLayer):
             in_dim=in_dim,
             in_spatial_dims=in_spatial_dims,
             use_time_mask=use_time_mask,
+            mask_value={"MAX": float("-inf"), "AVG": 0}[mode],
         )
         # We want to prepare the input data such that the batch-dim(s) is the very first,
         # the feature-dim is the very last ("NHWC" format) or right after batch-dim ("NCHW"),

returnn/torch/frontend/_backend.py

@@ -1925,15 +1925,19 @@ class TorchBackend(Backend[torch.Tensor]):
                     stride_ = strides[i] if isinstance(strides, (list, tuple)) else strides
                 else:
                     stride_ = 1
-                # What is the logic here? You might be aware, in case without striding,
+                # What is the logic here? Also see https://github.com/rwth-i6/returnn/issues/1693.
+                # You might be aware, in case without striding,
                 # we simply have pad_left = (s.dimension - 1) // 2,
-                # or the total amount of padding is s.dimension - 1.
+                # or the total amount of padding is s.dimension - 1
+                # (s.dimension is the filter size).
                 # So we add the same amount of padding on both sides (or one less on the left side if odd).
                 # The output seq length in case of "valid" padding is ⌈(in_len - s.dimension + 1) / stride⌉.
                 # The output seq length in case of "same" padding with no striding (stride = 1)
-                # is simply the same as the input length.
+                # is simply the same as the input length (that's why it's called "same").
                 # What is the output seq length in case of "same" padding with striding?
-                # It should be ⌈in_len / stride⌉.
+                # It should be out_len = ⌈in_len / stride⌉ (it should be independent of s.dimension).
+                # We can rewrite out_len as:
+                # out_len = ⌊(in_len + stride - 1) / stride⌋ = (in_len + stride - 1 - (in_len - 1) % stride) / stride
                 # So, then we need to add a total amount of padding of s.dimension - 1.
                 # However, doing it the same way as without striding is incorrect.
                 # Why? Because then the first window might have more padding than the final window.
@@ -1941,8 +1945,6 @@ class TorchBackend(Backend[torch.Tensor]):
                 # or maybe one less on the first window if odd.
                 # How many frames do the windows cover?
                 # in_len_covered = out_len * stride + (s.dimension - stride)
-                # We can rewrite out_len as:
-                # out_len = ⌊(in_len + stride - 1) / stride⌋ = (in_len + stride - 1 - (in_len - 1) % stride) / stride
                 # So we have:
                 # in_len_covered = (in_len + stride - 1 - (in_len - 1) % stride) + (s.dimension - stride)
                 #                = in_len + s.dimension - 1 - (in_len - 1) % stride

{returnn-1.20250228.104237.dist-info → returnn-1.20250304.10039.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: returnn
-Version: 1.20250228.104237
+Version: 1.20250304.10039
 Summary: The RWTH extensible training framework for universal recurrent neural networks
 Home-page: https://github.com/rwth-i6/returnn/
 Author: Albert Zeyer

{returnn-1.20250228.104237.dist-info → returnn-1.20250304.10039.dist-info}/RECORD

@@ -1,9 +1,9 @@
-returnn/PKG-INFO,sha256=WyCpdpBjUpCVa51XYBBPnFFDZgy5pufGhH0PkInknok,5215
+returnn/PKG-INFO,sha256=et7Z9NstTVvnWjiIMXhquw3eiMnMxYMfnEEbVc755xQ,5214
 returnn/__init__.py,sha256=biBtRsM0WZ406vShaeH-9WFoqJ8XwTbn6g0EeFJ7l8E,1012
 returnn/__main__.py,sha256=qBFbuB1yN3adgVM5pXt2-Yq9vorjRNchNPL8kDKx44M,31752
 returnn/__old_mod_loader__.py,sha256=nvsNY-xELdS_IPNkv66Q9Rmvg4dbGW0-EBRDcCmctos,7654
 returnn/__setup__.py,sha256=22kQn2fh11iPM0hLb2Fy5sLmoU1JGvmDxXRYuRgQkwU,4659
-returnn/_setup_info_generated.py,sha256=AsveviXvDuDQsTAWwDCtvPC3oMGwJWCILWlydtbAeak,77
+returnn/_setup_info_generated.py,sha256=d4hd9PkngTUKLJT4Q6GLMhVg4nXyV3Pym04_IKcblgc,77
 returnn/config.py,sha256=3tmKhB6FnQZaNdtcYsiB61JnEY--iZ2qmJ4yq0b6tE0,29140
 returnn/forward_iface.py,sha256=A_OJiaXsX4MlXQRzST86ylyxSUZbC402PQL1REcqHjM,911
 returnn/learning_rate_control.py,sha256=ZvWryAn_tv9DhV8sh1LV3eE34Yltl3On3mYZAG4hR9s,34684
@@ -154,7 +154,7 @@ returnn/sprint/extern_interface.py,sha256=l-v1X-Yg0UpTFe7Y3c4FwWOqpSNuv9Oy5EzqlK
 returnn/sprint/interface.py,sha256=_IGNQlOFcJcwsSeVkKcM-y8g2NDJv07jFhii47KfWtg,36490
 returnn/tensor/README.md,sha256=X6BqcRLrPLPnwF9yR69uqIFrMnNluj9pBkOPHwNgzuo,501
 returnn/tensor/__init__.py,sha256=on6j5PEOQpck50UcsR4nJzJSDmoVy34z1Oq4efv6Ax0,154
-returnn/tensor/_dim_extra.py,sha256=pMrzlD8LqlhV9VdBuYSzl38VC3O71HJ1rlJaH8JFxUo,122321
+returnn/tensor/_dim_extra.py,sha256=U7H690nc9WCL_fKX11AZLbTa1FhvJnDaA-vs3HM8_fI,122358
 returnn/tensor/_tensor_extra.py,sha256=DYJ6Dv3AdHcqFeiS_8fFsJG8ewu0dvFiXYT_TG1U1sI,164873
 returnn/tensor/_tensor_mixin_base.py,sha256=H5z86I0NejxrSgMH1c5oXQzBqS6L9HpvP4y7oegBaSc,643
 returnn/tensor/_tensor_op_overloads.py,sha256=kVVcnYtcZdW7Vjj78V1Im_yVX2M2r6dUTgeiAQZ37X0,5449
@@ -193,7 +193,7 @@ returnn/tf/frontend_low_level/__init__.py,sha256=34469k3KzMUIGowxReOZnbf6WdTjxY7
 returnn/tf/frontend_low_level/_backend.py,sha256=JwwRRIGnElqBC4bTImdB7w3U1u_SJESeZHYLmq86wog,24479
 returnn/tf/layers/__init__.py,sha256=Ngu-X84nWFgz7ndDu88DqoZ-5lUMMTQWH4g7N8pSoCg,72
 returnn/tf/layers/base.py,sha256=KcADpZUxqLkoFpQPMe_l9thRC7rpyBJIZCHITmnOd7M,153169
-returnn/tf/layers/basic.py,sha256=IVQ_6PkM-uuBN_vVg-VeGM74bb1pc6TjJhKf92pPS1I,610870
+returnn/tf/layers/basic.py,sha256=la0EwaHVzAbL6JOXs6QXnYQ74F3R16piYpT55VwVFT4,611063
 returnn/tf/layers/rec.py,sha256=K9vvyDJeDApYQDKabz7PaOTGHeSTloInkecxKTbqeTU,548357
 returnn/tf/layers/segmental_model.py,sha256=wUyDZGr-eTVIIQWcsHLML0wtOxuWn_NFKOIrUKQcvoI,21515
 returnn/tf/layers/signal_processing.py,sha256=vRlkN7k7otk9_Qdv0qr_l6V0VT5Q6dO2MxwZWb2HH2M,52693
@@ -216,7 +216,7 @@ returnn/torch/data/queued_data_iter.py,sha256=PoOsGHdHVZjTmcyfq_ZOw--P6hyfTdmAWI
 returnn/torch/data/returnn_dataset_wrapper.py,sha256=2CaDapzrlqahANuq-nyVAtv5ENHuM8A7okORwYJDisg,8006
 returnn/torch/data/tensor_utils.py,sha256=-Teqi--LLbt6q_5mDRdoHZHmPgSdC83W706ukif_YiU,1284
 returnn/torch/frontend/__init__.py,sha256=AA48HZnC17ASuKA0EWy8loZ-Bib_yUtqF4T1wYvjst4,62
-returnn/torch/frontend/_backend.py,sha256=TqyDWNP4XCvJNNGn8jyxaT8BOEjVE24QCUR3qsTIS3A,101242
+returnn/torch/frontend/_backend.py,sha256=rFCoCnzZoBtHPg7mWpO3yJOJMVesuWuA3_6GGSKMc5k,101452
 returnn/torch/frontend/_rand.py,sha256=1JgIkV2XmpgJD86zXZ-NCAe-QuoP2swr6NaS1oz3Qa8,1830
 returnn/torch/frontend/bridge.py,sha256=Z2_UW8AagezC7zsXDc5PKcd8G9WwisV7j9SWGHU0m4U,7840
 returnn/torch/frontend/raw_ops.py,sha256=lF0h-KtYYsdaaqQADylVZp9qzPskOOXA4MfmYDyx5IU,296
@@ -253,8 +253,8 @@ returnn/util/sig_proc.py,sha256=Tjz0VOAVyqu2qDCF5HZ1JjALjcFsHcNkcd96WgZeKfE,7265
 returnn/util/task_system.py,sha256=y4sMVXQ25Qd2z0rx03uOlXlkE-jbCYC1Sjfn-XlraVU,26003
 returnn/util/train_proc_manager.py,sha256=Pjht28k6uz6BNQ47uW6Gf880iyq5q4wx7P_K2tmoAM8,3266
 returnn/util/watch_memory.py,sha256=BR5P2kvBN6UI81cE0_1WAA6Hd1SByLbBaiDxvLhPOew,4213
-returnn-1.20250228.104237.dist-info/LICENSE,sha256=ywBD_U2aD4vpuoIgNAsjIGBYydl0tVKll3De0Z8s77c,11041
-returnn-1.20250228.104237.dist-info/METADATA,sha256=WyCpdpBjUpCVa51XYBBPnFFDZgy5pufGhH0PkInknok,5215
-returnn-1.20250228.104237.dist-info/WHEEL,sha256=P9jw-gEje8ByB7_hXoICnHtVCrEwMQh-630tKvQWehc,91
-returnn-1.20250228.104237.dist-info/top_level.txt,sha256=Lsn4WZc5Pbfk0-xDQOgnFCxOoqxL4CyeM3N1TFbJncw,8
-returnn-1.20250228.104237.dist-info/RECORD,,
+returnn-1.20250304.10039.dist-info/LICENSE,sha256=ywBD_U2aD4vpuoIgNAsjIGBYydl0tVKll3De0Z8s77c,11041
+returnn-1.20250304.10039.dist-info/METADATA,sha256=et7Z9NstTVvnWjiIMXhquw3eiMnMxYMfnEEbVc755xQ,5214
+returnn-1.20250304.10039.dist-info/WHEEL,sha256=P9jw-gEje8ByB7_hXoICnHtVCrEwMQh-630tKvQWehc,91
+returnn-1.20250304.10039.dist-info/top_level.txt,sha256=Lsn4WZc5Pbfk0-xDQOgnFCxOoqxL4CyeM3N1TFbJncw,8
+returnn-1.20250304.10039.dist-info/RECORD,,