test_LinearLayer_in_dim_spatial, allow in_dim not feature dim (#783)

#597

Author: albertz

returnn/tf/layers/base.py

@@ -298,7 +298,7 @@ def _base_get_out_data_from_opts(cls, network, name,
                                    out_type=None, out_dim=None, n_out=NotSpecified,
                                    out_shape=None,
                                    target=None, _target_layers=None, size_target=None,
-                                   sources=(), loss=None,
+                                   sources=(), in_dim=None, loss=None,
                                    **kwargs):
     """
     Called via BaseLayer.get_out_data_from_opts().
@@ -313,6 +313,7 @@ def _base_get_out_data_from_opts(cls, network, name,
     :param dict[str,LayerBase]|None _target_layers: if target.startswith("layer:"), then this is target -> layer
     :param str|None size_target:
     :param list[LayerBase] sources:
+    :param DimensionTag|None in_dim:
     :param Loss|None loss:
     :param kwargs: remaining kwargs of self.__init__(), ignored here
     :return: Data template (placeholder not set)
@@ -338,6 +339,15 @@ def _base_get_out_data_from_opts(cls, network, name,
     if n_out is not NotSpecified:
       assert out_type["dim"] == n_out
     sources_data_list = [src.output for src in sources if src]
+    if in_dim:
+      assert len(sources_data_list) == 1
+      if sources_data_list[0].feature_dim_or_sparse_dim != in_dim:
+        # Allow to specify some in_dim which is not the feature dim.
+        # However, the follow-up code will expect it to be the feature dim, thus reassign it if possible.
+        assert in_dim in sources_data_list[0].dim_tags
+        axis = sources_data_list[0].get_axis_from_description(in_dim)
+        sources_data_list = [sources_data_list[0].copy()]
+        sources_data_list[0].feature_dim_axis = axis
     allow_broadcast_all_sources = NotSpecified
     if "shape" in out_type or "dim_tags" in out_type or out_shape is not None:
       allow_broadcast_all_sources = True

returnn/tf/layers/basic.py

@@ -94,7 +94,17 @@ def concat_sources(src_layers, out_dim=None, allow_broadcast_all_sources=NotSpec
   if len(src_layers) == 1:
     data = src_layers[0].output.copy()
     if out_dim:
-      assert out_dim == data.feature_dim_or_sparse_dim
+      if out_dim == data.feature_dim_or_sparse_dim:
+        pass  # good
+      elif out_dim in data.dim_tags:
+        # We found out_dim in the input but it is not marked as the feature dim.
+        # This is explicitly allowed. Follow-up code will expect this to be the feature-dim though,
+        # So we mark it accordingly.
+        assert not data.sparse
+        axis = data.get_axis_from_description(out_dim)
+        data.feature_dim_axis = axis
+      else:
+        raise Exception("%s not found in %s" % (out_dim, data))
     return data
   network = src_layers[0].network
   cache_key = (tuple(src_layers), out_dim, 0.0, None)

tests/test_TFNetworkLayer.py

@@ -136,6 +136,32 @@ def test_LinearLayer():
       session.run(net.get_default_output_layer().output.placeholder, feed_dict=make_feed_dict(net.extern_data))
 
 
+def test_LinearLayer_in_dim_spatial():
+  from returnn.tf.util.data import BatchDim
+  time_dim = DimensionTag(kind=DimensionTag.Types.Spatial, description="time")
+  static_spatial_dim = DimensionTag(kind=DimensionTag.Types.Feature, description="static-spatial", dimension=3)
+  feat_dim = DimensionTag(kind=DimensionTag.Types.Feature, description="in-feature", dimension=5)
+  out_dim = DimensionTag(kind=DimensionTag.Types.Feature, description="out-feature", dimension=7)
+  config = Config({
+    "extern_data": {
+      "data": {"dim_tags": [BatchDim, time_dim, static_spatial_dim, feat_dim]}  # [B,T,D1,D2]
+    }
+  })
+  for _ in range(2):
+    with make_scope() as session:
+      net = TFNetwork(config=config)
+      net.construct_from_dict({
+        "output": {"class": "linear", "from": "data", "in_dim": static_spatial_dim, "out_dim": out_dim}})
+      layer = net.get_default_output_layer()
+      print("Output:", layer.output)
+      assert layer.output.dim_tags_set_implicit == {BatchDim, time_dim, out_dim, feat_dim}
+      param = layer.params["W"]
+      assert isinstance(param, tf.Variable)
+      assert param.shape.as_list() == [static_spatial_dim.dimension, out_dim.dimension]
+      session.run(tf_compat.v1.global_variables_initializer())
+      session.run(layer.output.placeholder, feed_dict=make_feed_dict(net.extern_data))
+
+
 def test_LinearLayer_two_time_dims_allow_broadcast_all_sources():
   from returnn.tf.util.data import BatchDim
   with make_scope() as session: