returnn 1.20250228.104237__tar.gz → 1.20250304.10039__tar.gz

{returnn-1.20250228.104237/returnn.egg-info → returnn-1.20250304.10039}/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-1.20250304.10039/_setup_info_generated.py

@@ -0,0 +1,2 @@
+version = '1.20250304.010039'
+long_version = '1.20250304.010039+git.3e53d74'

{returnn-1.20250228.104237 → returnn-1.20250304.10039}/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-1.20250228.104237 → returnn-1.20250304.10039}/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-1.20250228.104237 → returnn-1.20250304.10039}/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 → returnn-1.20250304.10039/returnn.egg-info}/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-1.20250228.104237 → returnn-1.20250304.10039}/tests/rf_utils.py

@@ -143,7 +143,10 @@ def run_model(
     for k, v_pt in out_pt_raw.items():
         v_tf = out_tf_raw[k]
         print(f"  comparing {k!r} {_array_repr(v_pt)} PT vs TF")
-        numpy.testing.assert_allclose(v_pt, v_tf, atol=1e-5, rtol=1e-5, err_msg=f"output {k!r} differs")
+        if not numpy.allclose(v_pt, v_tf, atol=1e-5, rtol=1e-5):
+            print(f"  PT:\n{v_pt}")
+            print(f"  TF:\n{v_tf}")
+            raise Exception(f"output {k!r} differs")
     return out_pt
 
 

returnn-1.20250228.104237/_setup_info_generated.py

@@ -1,2 +0,0 @@
-version = '1.20250228.104237'
-long_version = '1.20250228.104237+git.bd23951'