Dim auto_generated flag

Allows for better Dim is_equal
which does not rely on the description.

#634

Author: albertz

returnn/tf/layers/basic.py

@@ -1039,7 +1039,7 @@ def get_out_data_from_opts(
     if out_dim:
       assert out_dim.dimension == new_dim
     else:
-      out_dim = Dim(kind=dim_tag.kind, description="%s:slice" % name, dimension=new_dim)
+      out_dim = Dim(kind=dim_tag.kind, description="%s:slice" % name, dimension=new_dim, auto_generated=True)
     return input_data.copy_template_replace_dim_tag(axis=axis, new_dim_tag=out_dim, name="%s_output" % name)
 
 
@@ -1219,7 +1219,7 @@ def get_out_data_from_opts(cls, name, sources=(), start=None, size=None, axis="T
         out_spatial_dim = Dim(
           kind=Dim.Types.Spatial,
           description="sliced-time:%s" % name,
-          dimension=size)
+          dimension=size, auto_generated=True)
     gather_positions_data = gather_positions_data.copy_add_dim_by_tag(
       out_spatial_dim, unbroadcast=True, axis=start_data.batch_ndim)
     position = InternalLayer(
@@ -2613,7 +2613,7 @@ def get_out_data_from_opts(cls, name, network, shape, maxval, minval=0, dtype="i
       elif isinstance(d, int):
         d = Dim(
           kind=Dim.Types.Spatial if i < len(shape) - 1 else Dim.Types.Feature,
-          description="%s:static:%i" % (name, i),
+          description="%s:static:%i" % (name, i), auto_generated=True,
           dimension=d)
       else:
         raise TypeError("Layer %r: invalid type %s in shape %r" % (name, type(d), shape))
@@ -2679,7 +2679,7 @@ def get_out_data_from_opts(cls, name, limit, start=0, delta=1, dtype=None, spars
       else:
         dtype = "int32"
     dim = len(range(start, limit, delta))
-    tag = Dim(kind=Dim.Types.Spatial, dimension=dim, description="%s:range" % name)
+    tag = Dim(kind=Dim.Types.Spatial, dimension=dim, description="%s:range" % name, auto_generated=True)
     if out_spatial_dim:
       tag.declare_same_as(out_spatial_dim)
     sparse_dim = None
@@ -2799,7 +2799,7 @@ def get_out_data_from_opts(cls, name, sources, dtype="int32", sparse=False, out_
       dim_tag = Dim.get_tag_from_size_tensor(source.placeholder)
     if not dim_tag:
       dim_tag = Dim(
-        kind=Dim.Types.Spatial, description="%s_input_len" % name,
+        kind=Dim.Types.Spatial, description="%s_input_len" % name, auto_generated=True,
         batch=source.batch, control_flow_ctx=source.control_flow_ctx,
         dyn_size_ext=source)
       if source.placeholder is not None:
@@ -3079,15 +3079,15 @@ def get_out_data_from_opts(cls, name, network, sources, window_size=None, window
               filter_size=window_size, stride=stride, dilation_rate=1, padding=padding)
           out_spatial_dim = Dim(
             kind=Dim.Types.Spatial, description="%s:spatial" % name,
-            dimension=dim, derived_from_tag=in_spatial_dim,
+            dimension=dim, derived_from_tag=in_spatial_dim, auto_generated=True,
             batch=data.batch, control_flow_ctx=data.control_flow_ctx)
       data = data.copy_template_replace_dim_tag(axis=axis, new_dim_tag=out_spatial_dim)
       new_dim_axis = axis + 1  # add new axis right after
     if window_dim:
       assert window_dim.dimension == window_size
     else:
       window_dim = Dim(
-        kind=Dim.Types.Spatial, description="%s:window" % name, dimension=window_size)
+        kind=Dim.Types.Spatial, description="%s:window" % name, dimension=window_size, auto_generated=True)
     return data.copy_add_dim_by_tag(axis=new_dim_axis, dim_tag=window_dim, unbroadcast=True)
 
   # noinspection PyMethodOverriding
@@ -3585,9 +3585,11 @@ def _get_axis_size_splits_num_splits(cls, name, input_data, axis=None,
         err_prefix, out_dims, dim, input_data)
     if not out_dims:
       assert size_splits
-      out_dims = [Dim(
-        kind=input_data.dim_tags[axis].kind, description="%s_split%i" % (name, idx),
-        dimension=size_splits[idx]) for idx in range(len(size_splits))]
+      out_dims = [
+        Dim(
+          kind=input_data.dim_tags[axis].kind, description="%s_split%i" % (name, idx),
+          dimension=size_splits[idx], auto_generated=True)
+        for idx in range(len(size_splits))]
     return axis, out_dims
 
   def _make_split_layer(self, idx):
@@ -3867,6 +3869,7 @@ def get_out_data_from_opts(cls, name, axis, dims, pad_to_multiples=None, sources
             kind=axis_dim_tag.kind,
             description="%s_split_dims%i_rem" % (name, rem_dim_idx),
             dimension=resolved_shape_dims[rem_dim_idx],
+            auto_generated=True,
             derived_from_tag=axis_dim_tag,
             batch=axis_dim_tag.batch, control_flow_ctx=axis_dim_tag.control_flow_ctx)
           if rem_dim.dimension is None and axis_dim_tag.dyn_size_ext is not None:
@@ -3883,7 +3886,7 @@ def get_out_data_from_opts(cls, name, axis, dims, pad_to_multiples=None, sources
           Dim(
             kind=axis_dim_tag.kind if not axis_dim_tag.is_batch_dim() else Dim.Types.Spatial,
             description="%s_split_dims%i" % (name, i),
-            dimension=shape_dim)
+            dimension=shape_dim, auto_generated=True)
           if rem_dim is None or i != rem_dim_idx else rem_dim
           for i, shape_dim in enumerate(resolved_shape_dims))
     out_batch = data.batch
@@ -4238,7 +4241,7 @@ def get_out_data_from_opts(cls, name, axis, dim=1, sources=(), **kwargs):
     else:
       new_dim = Dim(
         kind=Dim.Types.Feature if init_axis.lower() == "f" else Dim.Types.Spatial,
-        description="%s_expand_dims" % name,
+        description="%s_expand_dims" % name, auto_generated=True,
         dimension=dim)
     data = data.copy_template(name="%s_output" % name)
     data = data.copy_add_dim_by_tag(new_dim, unbroadcast=True, axis=axis)
@@ -4394,7 +4397,7 @@ def get_out_data_from_opts(cls, name, sources, axis, repetitions, out_dim=None,
       if isinstance(repetitions, int):
         out_dim = tag * repetitions
       else:
-        out_dim = Dim(description="repeated:%s" % name, kind=tag.kind, derived_from_tag=tag)
+        out_dim = Dim(description="repeated:%s" % name, kind=tag.kind, derived_from_tag=tag, auto_generated=True)
     return data.copy_template_replace_dim_tag(axis=data.get_batch_axis(0), new_dim_tag=out_dim)
 
 
@@ -4804,12 +4807,12 @@ def map_axis_name(s):
         pass
       else:
         out.sparse_dim = Dim(
-          kind=Dim.Types.Feature, dimension=set_sparse_dim, description="%s:set-sparse-dim" % name)
+          kind=Dim.Types.Feature, dimension=set_sparse_dim, description="%s:set-sparse-dim" % name, auto_generated=True)
     if increase_sparse_dim:
       assert out.sparse
       out.sparse_dim = Dim(
         kind=Dim.Types.Feature, dimension=out.sparse_dim.dimension + 1,
-        description="%s:inc-sparse-dim" % name)
+        description="%s:inc-sparse-dim" % name, auto_generated=True)
     if batch_dim_base:
       out.batch = batch_dim_base.output.batch
     return out
@@ -5263,7 +5266,7 @@ def get_out_data_from_opts(
             filter_size=filter_size[i], stride=strides[i], dilation_rate=dilation_rate[i], padding=padding)
         dim_tags.append(Dim(
           kind=Dim.Types.Spatial, description="%s:conv:s%i" % (name, i), dimension=new_dim,
-          derived_from_tag=old_tag, undefined=not old_tag))
+          derived_from_tag=old_tag, undefined=not old_tag, auto_generated=True))
     if not out_dim:
       assert n_out
       out_dim = FeatureDim("%s:channel" % name, dimension=n_out)
@@ -5767,7 +5770,7 @@ def get_out_data_from_opts(cls, name, sources, network,
             padding=padding, output_padding=output_padding[i]) - remove_padding[i] * 2
         dim_tags.append(Dim(
           kind=Dim.Types.Spatial, description="%s:conv:s%i" % (name, i), dimension=new_dim,
-          derived_from_tag=old_tag, undefined=not old_tag))
+          derived_from_tag=old_tag, undefined=not old_tag, auto_generated=True))
     if not out_dim:
       assert n_out
       out_dim = FeatureDim("%s:channel" % name, dimension=n_out)
@@ -5983,7 +5986,8 @@ def get_out_data_from_opts(cls, name, sources, mode="", axes=None, axis=None, ke
     out_time_dim_axis = x.time_dim_axis
     if keep_dims:
       for i in axes:
-        y_dim_tags[i] = Dim(kind=y_dim_tags[i].kind, dimension=1, description="%s:keep-dim-%i" % (name, i))
+        y_dim_tags[i] = Dim(
+          kind=y_dim_tags[i].kind, dimension=1, description="%s:keep-dim-%i" % (name, i), auto_generated=True)
     else:
       if out_batch_dim_axis in axes:
         out_batch_dim_axis = None
@@ -6184,7 +6188,7 @@ def get_out_data_from_opts(cls, name, sources, axis=None, out_spatial_dim=None,
     out = common_source.copy_template(name="%s_output" % name)
     if not out_spatial_dim:
       out_spatial_dim = Dim(
-        kind=Dim.Types.Spatial, description="%s:stack" % name, dimension=len(sources))
+        kind=Dim.Types.Spatial, description="%s:stack" % name, dimension=len(sources), auto_generated=True)
     assert out_spatial_dim.dimension == len(sources)
     out = out.copy_add_dim_by_tag(axis=axis, dim_tag=out_spatial_dim, unbroadcast=True)
     return out
@@ -6316,7 +6320,8 @@ def get_out_data_from_opts(cls, name, sources, axes, padding=None, size=None, ke
       dim_tags = list(data.dim_tags)
       for i, a in enumerate(axes):
         dim_tags[a] = Dim(
-          kind=dim_tags[a].kind, description="%s:weighted-sum:%i" % (name, i), dimension=res_dims[i])
+          kind=dim_tags[a].kind, description="%s:weighted-sum:%i" % (name, i), dimension=res_dims[i],
+          auto_generated=True)
       data = data.copy_template_new_dim_tags(dim_tags, keep_special_axes=True)
     else:
       assert all([d == 1 for d in res_dims])
@@ -6630,7 +6635,7 @@ def get_out_data_from_opts(cls, name, sources, axis="T", out_dim=None, **kwargs)
     data = data.copy_move_axis(old_axis=axis, new_axis=0)
     data = data.copy_with_batch_dim_axis(1)
     if not out_dim:
-      out_dim = Dim(kind=in_dim.kind, description="%s:chunking" % name)
+      out_dim = Dim(kind=in_dim.kind, description="%s:chunking" % name, auto_generated=True)
     data = data.copy_template_replace_dim_tag(axis=0, new_dim_tag=out_dim)
     data.time_dim_axis = 0
     return data
@@ -7189,9 +7194,9 @@ def __init__(self, axis, amount, pad=True, adjust_size_info=True, **kwargs):
         self.output.size_placeholder[axis_wob] + size_delta, 0, tf.shape(shifted)[axis])
       from ..util.data import Dim
       Dim(
-        kind=Dim.Types.Spatial, description="shift_axis",
+        kind=Dim.Types.Spatial, description="%s_shift_axis" % self.name,
         dyn_size=new_size, batch=self.output.batch,
-        src_data=self.output, src_axis=axis)
+        src_data=self.output, src_axis=axis, auto_generated=True)
       self.output.size_placeholder[axis_wob] = new_size
 
   @classmethod
@@ -7210,7 +7215,7 @@ def get_out_data_from_opts(cls, name, amount, axis, pad, sources=(), **kwargs):
     axis = out.get_axis_from_description(axis)
     tag = out.dim_tags[axis]
     dim = None if tag.dimension is None else max(0, tag.dimension - abs(amount))
-    tag = Dim(kind=tag.kind, description="%s_shift_axis" % name, dimension=dim)
+    tag = Dim(kind=tag.kind, description="%s_shift_axis" % name, dimension=dim, auto_generated=True)
     return out.copy_template_replace_dim_tag(axis=axis, new_dim_tag=tag)
 
 
@@ -7319,7 +7324,7 @@ def get_out_data_from_opts(cls, factor, axis, sources, name, out_dim=None, **kwa
     if out_dim:
       assert out_dim.dimension == dim
     else:
-      out_dim = Dim(kind=tag.kind, description="%s_resize" % name, dimension=dim)
+      out_dim = Dim(kind=tag.kind, description="%s_resize" % name, dimension=dim, auto_generated=True)
     return out.copy_template_replace_dim_tag(axis=axis, new_dim_tag=out_dim)
 
 
@@ -7410,7 +7415,8 @@ def get_out_data_from_opts(cls, name, sources, axis="T", out_dim=None, **kwargs)
     axis = out.get_axis_from_description(axis, allow_int=False)
     in_dim = out.dim_tags[axis]
     if not out_dim:
-      out_dim = Dim(kind=in_dim.kind, description="%s_removed_items", dimension=None, derived_from_tag=in_dim)
+      out_dim = Dim(
+        kind=in_dim.kind, description="%s_removed_items", dimension=None, derived_from_tag=in_dim, auto_generated=True)
     return out.copy_template_replace_dim_tag(axis=axis, new_dim_tag=out_dim)
 
 
@@ -8582,18 +8588,17 @@ def get_out_data_from_opts(cls, name, network,
       elif isinstance(d, int):
         d = Dim(
           kind=Dim.Types.Spatial if i < len(shape) - 1 else Dim.Types.Feature,
-          description="%s:static:%i" % (name, i),
+          description="%s:static:%i" % (name, i), auto_generated=True,
           dimension=d)
       else:
         raise TypeError("Layer %r: invalid type %s in shape %r" % (name, type(d), shape))
       dim_tags.append(d)
     if add_time_axis:
       dim_tags.insert(
-        0, Dim(kind=Dim.Types.Time, description="%s:dummy-time" % name, dimension=1))
+        0, Dim(kind=Dim.Types.Time, description="%s:dummy-time" % name, dimension=1, auto_generated=True))
     if add_batch_axis:
       dim_tags.insert(
-        0, Dim(
-          kind=Dim.Types.Batch, description="batch", batch=network.get_global_batch_info()))
+        0, Dim(kind=Dim.Types.Batch, description="batch", batch=network.get_global_batch_info()))
     return Data(
       name="%s_output" % name, dim_tags=dim_tags, dtype=dtype,
       batch=network.get_global_batch_info() if add_batch_axis else None)

returnn/tf/layers/rec.py

@@ -3638,7 +3638,8 @@ def get_loop_acc_layer(name):
           else:
             time_dim_tag = Dim(
               kind=Dim.Types.Spatial,
-              description="dyn-time:%s/%s" % (self.parent_rec_layer.get_full_ctx_name(), search_choices))
+              description="dyn-time:%s/%s" % (self.parent_rec_layer.get_full_ctx_name(), search_choices),
+              auto_generated=True)
         elif is_out_time_dim:
           self.time_dim_tag.declare_same_as(time_dim_tag)
         output = (
@@ -6864,7 +6865,7 @@ def get_out_data_from_opts(cls, n_out, name, sources, **kwargs):
       else:
         time_dim = None
       time_tag = Dim(
-        kind=Dim.Types.Spatial, description="%s_self_att_time" % name, dimension=time_dim)
+        kind=Dim.Types.Spatial, description="%s_self_att_time" % name, dimension=time_dim, auto_generated=True)
       dim_tags = (batch_dim_tag, time_tag, feat_tag)
     else:
       dim_tags = (batch_dim_tag, feat_tag)
@@ -7242,7 +7243,7 @@ def get_out_data_from_opts(cls, name, sources,
       vocab = Vocabulary(vocab_file=vocab_file, unknown_label=vocab_unknown_label)
       tag = Dim(
         kind=Dim.Types.Feature, description="%s_ken_lm_vocab" % name,
-        dimension=vocab.num_labels, vocab=vocab)
+        dimension=vocab.num_labels, vocab=vocab, auto_generated=True)
       data = data.copy_add_dim_by_tag(tag, axis=-1, unbroadcast=True)
     return data
 
@@ -7808,7 +7809,7 @@ def _create_template(cls, name, network, sources, masked_from, unit,
         if not out_spatial_dim:
           out_spatial_dim = Dim(
             kind=Dim.Types.Spatial, description="%s:masked:time" % name,
-            derived_from_tag=source_data.get_time_dim_tag())
+            derived_from_tag=source_data.get_time_dim_tag(), auto_generated=True)
         source_data = source_data.copy_template_replace_dim_tag(
           axis=0,
           new_dim_tag=out_spatial_dim)
@@ -9225,17 +9226,17 @@ def get_out_data_from_opts(cls, name, sources, n_out, **kwargs):
     data = get_concat_sources_data_template(sources, name="%s_output" % name)
     # The result will be without batch dim.
     feature_dim_tag = Dim(
-      kind=Dim.Types.Feature, description="%s_rel_pos_enc_feat" % name, dimension=n_out)
+      kind=Dim.Types.Feature, description="%s_rel_pos_enc_feat" % name, dimension=n_out, auto_generated=True)
     if data.have_time_axis():
       time_dim_tag = data.get_time_dim_tag()
       # TODO using same dim tag twice will not be supported at some future point...
       data = data.copy_template_new_dim_tags((time_dim_tag, time_dim_tag, feature_dim_tag))
     else:
       # length will be ``network.get_rec_step_index() + 1``.
       dummy_dim_tag = Dim(
-        kind=Dim.Types.Spatial, description="%s_rel_pos_enc_dummy" % name, dimension=1)
+        kind=Dim.Types.Spatial, description="%s_rel_pos_enc_dummy" % name, dimension=1, auto_generated=True)
       time_dim_tag = Dim(
-        kind=Dim.Types.Spatial, description="%s_rel_pos_enc_time" % name, dimension=None)
+        kind=Dim.Types.Spatial, description="%s_rel_pos_enc_time" % name, dimension=None, auto_generated=True)
       data = data.copy_template_new_dim_tags((dummy_dim_tag, time_dim_tag, feature_dim_tag))
     return data
 

returnn/tf/layers/signal_processing.py

@@ -457,9 +457,9 @@ def _compute_size_placeholder():
       new_size = nr_of_full_frames + nf_of_paded_frames
       from ..util.data import Dim
       Dim(
-        kind=Dim.Types.Spatial, description="MultiChannelMultiResolutionStft",
+        kind=Dim.Types.Spatial, description="%s:MultiChannelMultiResolutionStft" % self.name,
         dyn_size=new_size, batch=self.output.batch,
-        src_data=self.output, src_axis=self.output.get_batch_axis(0))
+        src_data=self.output, src_axis=self.output.get_batch_axis(0), auto_generated=True)
       size_placeholder_dict[0] = new_size
       return size_placeholder_dict
 

returnn/tf/util/data.py

@@ -58,6 +58,7 @@ def __init__(self, kind=Types.Unspecified, description=None,
                vocab=None,
                dyn_size=None, dyn_size_ext=None,
                undefined=False, generic=False, special=False,
+               auto_generated=False,
                match_priority=0,
                derived_from_tag=None, derived_from_op=None,
                batch=None, control_flow_ctx=None,
@@ -75,6 +76,9 @@ def __init__(self, kind=Types.Unspecified, description=None,
     :param bool special: Like `generic`, this can not be a dim tag of :class:`Data`.
       But this dim tag also does not match anything except itself.
       So it can be used to represent special placeholders with special meanings like ``single_step``.
+    :param bool auto_generated: This is auto-generated by RETURNN because it was not explicitly specified by the user.
+      E.g. for ConvLayer and others. This implies certain behavior on equality, such as comparing the description,
+      to allow for several independent creations of the dim tag during template construction.
     :param Dim|None derived_from_tag:
       Whether this new tag is reduced, down/up sampled, padded etc from this given other tag.
       In situations where dim tags are being matched (Data.get_common_data),
@@ -121,6 +125,7 @@ def __init__(self, kind=Types.Unspecified, description=None,
     self._undefined = undefined
     self.generic = generic
     self.special = special
+    self.auto_generated = auto_generated
     # We can have different tag variants per batch info (e.g. with beam), or per control flow ctx.
     # They each have same_as = self. The same_base should have the base (global) batch info.
     self._same_for_batch_ctx = {}  # type: typing.Dict[typing.Tuple[BatchInfo,typing.Optional[ControlFlowContext]],Dim]  # nopep8
@@ -739,7 +744,7 @@ def is_equal(self, other, ignore_feature_dim=False, allow_same_feature_dim=False
     # We currently use the description because the identity would not be the same
     # in case of template construction where a dim tag is once created for a template layer,
     # and then later again for the real layer.
-    if self.description == other.description:
+    if self.auto_generated and other.auto_generated and self.description == other.description:
       return True
     return False
 
@@ -781,7 +786,9 @@ def __hash__(self):
       return hash(base)
     if self.derived_from_op:
       return hash(self.derived_from_op)
-    return hash((base.kind, base.dimension, base.description))
+    if self.auto_generated:
+      return hash((base.kind, base.dimension, base.description))
+    return hash(id(base))
 
   def get_same_base(self):
     """
@@ -1168,7 +1175,8 @@ def _make_constant_static_dim(cls, value, kind=None):
       dimension=value,
       kind=kind or Dim.Types.Unspecified,
       description="unnamed_%sdim_%i" % (kind.name + "_" if kind else "", value),
-      derived_from_op=Dim.Op(kind="constant", inputs=[], attribs={"value": value}))
+      derived_from_op=Dim.Op(kind="constant", inputs=[], attribs={"value": value}),
+      auto_generated=True)
 
   def _is_constant_static_dim(self):
     return self.derived_from_op and self.derived_from_op.kind == "constant"