returnn 1.20231130.220256__tar.gz → 1.20231204.103837__tar.gz

{returnn-1.20231130.220256/returnn.egg-info → returnn-1.20231204.103837}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: returnn
-Version: 1.20231130.220256
+Version: 1.20231204.103837
 Summary: The RWTH extensible training framework for universal recurrent neural networks
 Home-page: https://github.com/rwth-i6/returnn/
 Author: Albert Zeyer

returnn-1.20231204.103837/_setup_info_generated.py

@@ -0,0 +1,2 @@
+version = '1.20231204.103837'
+long_version = '1.20231204.103837+git.eeea993'

{returnn-1.20231130.220256 → returnn-1.20231204.103837}/returnn/tf/layers/basic.py

@@ -6120,6 +6120,7 @@ class ConvLayer(_ConcatInputLayer):
         filter_perm=None,
         bias=None,
         use_time_mask=False,
+        pad_seq_len_to_power=None,
         **kwargs,
     ):
         """
@@ -6158,6 +6159,10 @@ class ConvLayer(_ConcatInputLayer):
         :param dict[str,str]|None filter_perm: transposes the filter (input filter as layer)
         :param LayerBase|None bias: if given, will not create an own parameter, but use this as the bias
         :param bool use_time_mask:
+        :param Optional[float] pad_seq_len_to_power: pad sequence length to power of given number
+            to reduce number of different sequence lengths.
+            See https://github.com/rwth-i6/returnn/issues/1450 and
+            https://github.com/tensorflow/tensorflow/issues/62441.
         """
         from returnn.util import BehaviorVersion
 
@@ -6301,6 +6306,22 @@ class ConvLayer(_ConcatInputLayer):
         else:
             x = input_data.placeholder
 
+        if pad_seq_len_to_power is None:
+            pad_seq_len_to_power = self.network.get_config().float("conv_pad_seq_len_to_power", None)
+        if pad_seq_len_to_power is not None:
+            pad_seq_len_to_power = float(pad_seq_len_to_power)
+            padding_for_power = []
+            for ax in range(input_data.batch_ndim):
+                if input_data.is_axis_dynamic(ax):
+                    seq_len = tf.cast(tf.shape(x)[ax], tf.float32)
+                    padded_len = tf.math.ceil(
+                        pad_seq_len_to_power ** (tf.math.ceil(tf.math.log(seq_len) / tf.math.log(pad_seq_len_to_power)))
+                    )
+                    padding_for_power.append((0, padded_len - seq_len))
+                else:
+                    padding_for_power.append((0, 0))
+            x = tf.pad(x, padding_for_power)
+
         extended_batch_shape = None
         if num_batch_dims > 1:
             x_shape = tf.shape(x)
@@ -6361,6 +6382,16 @@ class ConvLayer(_ConcatInputLayer):
             )
         if num_batch_dims > 1:
             y = tf.reshape(y, tf.concat([extended_batch_shape, tf.shape(y)[1:]], axis=0))
+
+        if pad_seq_len_to_power is not None:
+            slice_size = []
+            for ax in range(self.output.batch_ndim):
+                if self.output.is_axis_dynamic(ax):
+                    slice_size.append(tf.reduce_max(self.output.get_dynamic_size(ax)))
+                else:
+                    slice_size.append(tf.shape(y)[ax])
+            y = tf.slice(y, begin=[0] * len(slice_size), size=slice_size)
+
         # y shape is [batch] + dynamic_dims + [n_out].
         if with_bias is NotSpecified:
             if bias or BehaviorVersion.get() >= 10:

{returnn-1.20231130.220256 → returnn-1.20231204.103837/returnn.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: returnn
-Version: 1.20231130.220256
+Version: 1.20231204.103837
 Summary: The RWTH extensible training framework for universal recurrent neural networks
 Home-page: https://github.com/rwth-i6/returnn/
 Author: Albert Zeyer

returnn-1.20231130.220256/_setup_info_generated.py

@@ -1,2 +0,0 @@
-version = '1.20231130.220256'
-long_version = '1.20231130.220256+git.8a7d673'