generated layers, remove conv and transposed conv

Those were the remaining modules with params. #82
rwth-i6 · Jan 6, 2022 · 28450df · 28450df
1 parent 49ae94d
commit 28450df
Show file tree

Hide file tree

Showing 2 changed files with 3 additions and 218 deletions.
diff --git a/nn/_generate_layers.py b/nn/_generate_layers.py
@@ -45,6 +45,8 @@
   "dropout",
   "batch_norm",
   "linear",
+  "conv",
+  "transposed_conv",
   "rec",
   "self_attention",
   "concat_attention",

diff --git a/nn/_generated_layers.py b/nn/_generated_layers.py
@@ -1,6 +1,6 @@
 """
 This file is auto-generated by _generate_layers.py.
-RETURNN: 1.20220105.113203+git.8470324
+RETURNN: 1.20220106.130144+git.248f8eb
 
 These are the RETURNN layers directly wrapped.
 Note that we intentionally exclude some layers or options for more consistency.
@@ -1147,123 +1147,6 @@ def cast(
     **args}, name=name or 'cast')
 
 
-class Conv(_Base):
-  """
-  A generic convolution layer which supports 1D, 2D and 3D convolution.
-  Pooling can be done in the separate "pool" layer.
-  """
-  returnn_layer_class = 'conv'
-  has_recurrent_state = False
-  has_variables = True
-
-  # noinspection PyShadowingBuiltins,PyShadowingNames
-  def __init__(self,
-               *,
-               out_dim: Dim,
-               filter_size: Sequence[int],
-               padding: str,
-               strides: Union[int, Sequence[int]] = NotSpecified,
-               dilation_rate: Union[int, Sequence[int]] = NotSpecified,
-               groups: int = NotSpecified,
-               input_expand_dims: int = NotSpecified,
-               input_add_feature_dim: bool = NotSpecified,
-               input_split_feature_dim: Optional[int] = NotSpecified,
-               in_dim: Optional[Dim] = NotSpecified,
-               out_spatial_dims: Optional[Sequence[Dim]] = NotSpecified,
-               with_bias: Union[bool, NotSpecified] = NotSpecified,
-               forward_weights_init: Any = NotSpecified,
-               bias_init: Any = NotSpecified,
-               filter_perm: Optional[Dict[str, str]] = NotSpecified,
-               **kwargs):
-    """
-    :param Dim out_dim:
-    :param tuple[int] filter_size: (width,), (height,width) or (depth,height,width) for 1D/2D/3D conv.
-      the input data ndim must match, or you can add dimensions via input_expand_dims or input_add_feature_dim.
-      it will automatically swap the batch-dim to the first axis of the input data.
-    :param str padding: "same" or "valid"
-    :param int|tuple[int] strides: strides for the spatial dims,
-      i.e. length of this tuple should be the same as filter_size, or a single int.
-    :param int|tuple[int] dilation_rate: dilation for the spatial dims
-    :param int groups: grouped convolution
-    :param int input_expand_dims: number of spatial dims to add to the input
-    :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 None|int input_split_feature_dim: if set, like input_add_feature_dim it will add a new feature dim
-      which is of value input_split_feature_dim, and the original input feature dim
-      will be divided by input_split_feature_dim, thus it must be a multiple of that value.
-    :param Dim|None in_dim:
-    :param Sequence[Dim]|None out_spatial_dims:
-    :param bool|NotSpecified with_bias: if True, will add a bias to the output features. False by default
-    :param forward_weights_init:
-    :param bias_init:
-    :param dict[str,str]|None filter_perm: transposes the filter (input filter as layer)
-    """
-    super().__init__(**kwargs)
-    self.out_dim = out_dim
-    self.filter_size = filter_size
-    self.padding = padding
-    self.strides = strides
-    self.dilation_rate = dilation_rate
-    self.groups = groups
-    self.input_expand_dims = input_expand_dims
-    self.input_add_feature_dim = input_add_feature_dim
-    self.input_split_feature_dim = input_split_feature_dim
-    self.in_dim = in_dim
-    self.out_spatial_dims = out_spatial_dims
-    self.with_bias = with_bias
-    self.forward_weights_init = forward_weights_init
-    self.bias_init = bias_init
-    self.filter_perm = filter_perm
-
-  def get_opts(self):
-    """
-    Return all options
-    """
-    opts = {
-      'out_dim': self.out_dim,
-      'filter_size': self.filter_size,
-      'padding': self.padding,
-      'strides': self.strides,
-      'dilation_rate': self.dilation_rate,
-      'groups': self.groups,
-      'input_expand_dims': self.input_expand_dims,
-      'input_add_feature_dim': self.input_add_feature_dim,
-      'input_split_feature_dim': self.input_split_feature_dim,
-      'in_dim': self.in_dim,
-      'out_spatial_dims': self.out_spatial_dims,
-      'with_bias': self.with_bias,
-      'forward_weights_init': self.forward_weights_init,
-      'bias_init': self.bias_init,
-      'filter_perm': self.filter_perm,
-    }
-    opts = {key: value for (key, value) in opts.items() if value is not NotSpecified}
-    return {**opts, **super().get_opts()}
-
-  # noinspection PyShadowingBuiltins,PyShadowingNames
-  def __call__(self,
-               source: LayerRef,
-               *,
-               in_spatial_dims: Sequence[Dim],
-               filter: Optional[LayerRef] = NotSpecified,
-               bias: Optional[LayerRef] = NotSpecified,
-               ) -> Layer:
-    """
-    Make layer dict
-    """
-    assert isinstance(source, LayerRef)
-    args = {
-      'in_spatial_dims': in_spatial_dims,
-      'filter': filter,
-      'bias': bias,
-    }
-    args = {key: value for (key, value) in args.items() if value is not NotSpecified}
-    return make_layer({
-      'class': 'conv',
-      'from': source,
-      **args,
-      **self.get_opts()}, module=self)
-
-
 # noinspection PyShadowingBuiltins,PyShadowingNames
 def pool(
          source: LayerRef,
@@ -1343,106 +1226,6 @@ def dct(
     **args}, name=name or 'dct')
 
 
-class TransposedConv(_Base):
-  """
-  Transposed convolution, sometimes also called deconvolution.
-  See :func:`tf.nn.conv2d_transpose` (currently we support 1D/2D).
-  """
-  returnn_layer_class = 'transposed_conv'
-  has_recurrent_state = False
-  has_variables = True
-
-  # noinspection PyShadowingBuiltins,PyShadowingNames
-  def __init__(self,
-               *,
-               out_dim: Dim,
-               filter_size: Sequence[int],
-               strides: Optional[Sequence[int]] = NotSpecified,
-               padding: str = NotSpecified,
-               remove_padding: Union[Sequence[int], int] = NotSpecified,
-               output_padding: Optional[Union[Sequence[Optional[int]], int]] = NotSpecified,
-               in_dim: Optional[Dim] = NotSpecified,
-               out_spatial_dims: Optional[Sequence[Dim]] = NotSpecified,
-               with_bias: bool = NotSpecified,
-               forward_weights_init: Any = NotSpecified,
-               bias_init: Any = NotSpecified,
-               filter_perm: Optional[Dict[str, str]] = NotSpecified,
-               **kwargs):
-    """
-    :param Dim out_dim:
-    :param list[int] filter_size:
-    :param list[int]|None strides: specifies the upscaling. by default, same as filter_size
-    :param str padding: "same" or "valid"
-    :param list[int]|int remove_padding:
-    :param list[int|None]|int|None output_padding:
-    :param Dim|None in_dim:
-    :param Sequence[Dim]|None out_spatial_dims:
-    :param bool with_bias: whether to add a bias. enabled by default.
-      Note that the default is different from ConvLayer!
-    :param forward_weights_init:
-    :param bias_init:
-    :param dict[str,str]|None filter_perm: transposes the filter (input filter as layer)
-    """
-    super().__init__(**kwargs)
-    self.out_dim = out_dim
-    self.filter_size = filter_size
-    self.strides = strides
-    self.padding = padding
-    self.remove_padding = remove_padding
-    self.output_padding = output_padding
-    self.in_dim = in_dim
-    self.out_spatial_dims = out_spatial_dims
-    self.with_bias = with_bias
-    self.forward_weights_init = forward_weights_init
-    self.bias_init = bias_init
-    self.filter_perm = filter_perm
-
-  def get_opts(self):
-    """
-    Return all options
-    """
-    opts = {
-      'out_dim': self.out_dim,
-      'filter_size': self.filter_size,
-      'strides': self.strides,
-      'padding': self.padding,
-      'remove_padding': self.remove_padding,
-      'output_padding': self.output_padding,
-      'in_dim': self.in_dim,
-      'out_spatial_dims': self.out_spatial_dims,
-      'with_bias': self.with_bias,
-      'forward_weights_init': self.forward_weights_init,
-      'bias_init': self.bias_init,
-      'filter_perm': self.filter_perm,
-    }
-    opts = {key: value for (key, value) in opts.items() if value is not NotSpecified}
-    return {**opts, **super().get_opts()}
-
-  # noinspection PyShadowingBuiltins,PyShadowingNames
-  def __call__(self,
-               source: LayerRef,
-               *,
-               in_spatial_dims: Sequence[Dim],
-               filter: Optional[LayerRef] = NotSpecified,
-               bias: Optional[LayerRef] = NotSpecified,
-               ) -> Layer:
-    """
-    Make layer dict
-    """
-    assert isinstance(source, LayerRef)
-    args = {
-      'in_spatial_dims': in_spatial_dims,
-      'filter': filter,
-      'bias': bias,
-    }
-    args = {key: value for (key, value) in args.items() if value is not NotSpecified}
-    return make_layer({
-      'class': 'transposed_conv',
-      'from': source,
-      **args,
-      **self.get_opts()}, module=self)
-
-
 # noinspection PyShadowingBuiltins,PyShadowingNames
 def reduce(
            source: LayerRef,