GatherLayer on batch axis

returnn/tf/layers/basic.py

@@ -1341,6 +1341,17 @@ def __init__(self, position, axis, **kwargs):
     # (BatchAxes.., InputAxesBeforeGatherAxis, PositionAxes.., InputAxesAfterGatherAxis..)
     self.output.placeholder = tf.gather(params=params, indices=indices, axis=gather_axis, batch_dims=batch_dims)
 
+    if input_data.dim_tags[old_gather_axis].is_batch_dim():
+      for axis in self.output.size_placeholder:
+        new_size = tf.gather(params=self.output.size_placeholder[axis], indices=position_data.placeholder)
+        from ..util.data import Dim
+        Dim(
+          kind=Dim.Types.Spatial, description="%s_gather_axis" % self.name,
+          dyn_size=new_size, batch=self.output.batch,
+          src_data=self.output, src_axis=axis, auto_generated=True)
+        self.output.size_placeholder[axis] = new_size
+
+
   @classmethod
   def _get_common_input_position_axes(cls, input_data, position_data, old_gather_axis):
     """

tests/test_TFNetworkLayer.py

@@ -5456,6 +5456,51 @@ def test_GatherLayer_broadcast_dim():
   })
 
 
+def test_GatherLayer_batch_dim():
+  with make_scope() as session:
+    import numpy as np
+    net = TFNetwork(extern_data=ExternData())
+    batch_dim, time_dim, feature_dim = 3, 4, 2
+    # [B, T, F]
+    random = np.random.RandomState(42)
+    values_seqs = random.rand(batch_dim, time_dim, feature_dim).astype('float32')
+    values_size = np.array([4, 2, 3])
+    values_placeholder = tf.constant(values_seqs, dtype=tf.float32)
+    values_size_placeholder = {0: tf.constant(values_size, dtype=tf.int32)}
+    values = InternalLayer(
+      name="values", network=net,
+      output=Data(
+        name="values",
+        batch_dim_axis=0, time_dim_axis=1, feature_dim_axis=2,
+        shape=[None, feature_dim],
+        placeholder=values_placeholder,
+        size_placeholder=values_size_placeholder,
+      ))
+    position_np = np.array([0, 2])
+    position = InternalLayer(
+      name="position", network=net,
+      output=Data(
+        name="position",
+        placeholder=tf.constant(position_np, dtype=tf.int64),
+        batch_dim_axis=0, shape=[], dtype="int64",
+      ))
+    values.output.sanity_check()
+    position.output.sanity_check()
+
+    # should become [B', T, F]
+    layer = GatherLayer(
+      name="gather", network=net,
+      sources=[values], position=position, axis="B",
+      output=GatherLayer.get_out_data_from_opts(
+        name="gather", sources=[values], position=position, axis="B"))
+    layer.output.sanity_check()
+    out_seqs, out_size = session.run([layer.output.placeholder, layer.output.size_placeholder.as_dict()])
+    assert isinstance(out_seqs, numpy.ndarray)
+
+    np.testing.assert_equal(values_seqs[position_np, :], out_seqs)
+    np.testing.assert_equal(values_size[position_np], out_size[0])
+
+
 def test_SliceNdLayer():
   n_batch = 5
   n_time = 7
@@ -9213,6 +9258,62 @@ def _get_mask_eval_layer(source, **_kwargs):
     assert numpy.isfinite(loss_v)
 
 
+def test_supervised_multilingual_training():
+  n_batch = 3
+  out_dim = 5
+
+  net_dict = {
+    "encoder": {"class": "copy", "from": "data"},
+
+    "loss_language_0": {
+      "class": "subnetwork",
+      "from": "data:classes",
+      "is_output_layer": True,
+      "subnetwork": {
+        # create language id indices based on where we see non-silence in the alignment.
+        # this should be done in the dataset in a neat way.
+        "aux_0": {"class": "gather", "position": 0, "axis": "T", "from": "data"},
+        "aux_1": {"class": "compare", "kind": "greater", "value": 2, "from": "aux_0"},
+        "idx": {  # shape (B',) with B' < B (B' is the number of utterances in the batch for this language)
+          "class": "eval",
+          "eval": "tf.squeeze(tf.where(source(0)), axis=-1)",
+          "out_type": {"dtype": "int64"},
+          "from": "aux_1",
+        },
+
+        # gather targets and encoder outputs
+        "tgt": {"class": "gather", "from": "data", "axis": "B", "position": "idx"},  # B', T (sparse)
+        "enc_raw": {"class": "gather", "from": "base:encoder", "axis": "B", "position": "idx"},  # B', T, F
+        "enc": {"class": "reinterpret_data", "size_base": "tgt", "from": "enc_raw"},  # B', T, F
+        "logits": {"class": "linear", "n_out": out_dim, "from": "enc"},  # B', T, n_out
+        "output": {
+          "class": "sparse_softmax_cross_entropy_with_logits",
+          "logits": "logits",
+          "targets": "tgt",
+          "loss": "as_is",
+        },
+      },
+    },
+  }
+
+  with make_scope() as session:
+    config = Config({
+      "extern_data": {
+        "data": {"dim": 7, "batch_dim_axis": 0, "time_dim_axis": 1, "feature_dim_axis": 2},
+        "classes": {"dim": out_dim, "sparse": True}},
+      "debug_print_layer_output_shape": True,  # "debug_print_layer_output": True,
+    })
+    net = TFNetwork(config=config, train_flag=True)
+    net.construct_from_dict(net_dict)
+    loss = net.get_total_loss()
+
+    tf_compat.v1.set_random_seed(1)
+    net.initialize_params(session)
+    loss_v = session.run(loss, feed_dict=make_feed_dict(net.extern_data, same_time=True, n_batch=n_batch))
+    print("loss:", loss_v)
+    assert numpy.isfinite(loss_v)
+
+
 if __name__ == "__main__":
   try:
     better_exchook.install()