Upload 13 files

aglib/meliad/transformer/configs/recurrent/bias_dual.gin

@@ -0,0 +1,15 @@
+
+decoder_stack.DecoderStack:
+  dstack_window_length = 0
+  recurrent_layer_indices = (-3,)
+  feedback_recurrence = False
+
+transformer_layer.TransformerLayer:
+  window_length = 512
+  recurrent_num_states = 512
+  recurrent_gate_type = "bias"
+  recurrent_single_gate = False
+  recurrent_skip_ffn = False
+
+training_loop.Trainer:
+  generate_every_steps = 0   # Not yet supported with recurrence.

aglib/meliad/transformer/configs/recurrent/bias_single.gin

@@ -0,0 +1,15 @@
+
+decoder_stack.DecoderStack:
+  dstack_window_length = 0
+  recurrent_layer_indices = (-3,)
+  feedback_recurrence = False
+
+transformer_layer.TransformerLayer:
+  window_length = 512
+  recurrent_num_states = 512
+  recurrent_gate_type = "bias"
+  recurrent_single_gate = True
+  recurrent_skip_ffn = False
+
+training_loop.Trainer:
+  generate_every_steps = 0   # Not yet supported with recurrence.

aglib/meliad/transformer/configs/recurrent/bias_skip.gin

@@ -0,0 +1,15 @@
+
+decoder_stack.DecoderStack:
+  dstack_window_length = 0
+  recurrent_layer_indices = (-3,)
+  feedback_recurrence = False
+
+transformer_layer.TransformerLayer:
+  window_length = 512
+  recurrent_num_states = 512
+  recurrent_gate_type = "bias"
+  recurrent_single_gate = False
+  recurrent_skip_ffn = True
+
+training_loop.Trainer:
+  generate_every_steps = 0   # Not yet supported with recurrence.

aglib/meliad/transformer/configs/recurrent/block_feedback_512.gin

@@ -0,0 +1,9 @@
+
+# Apply after the gate config.
+
+transformer_layer.TransformerLayer:
+  window_length = 512
+
+decoder_stack.DecoderStack:
+  dstack_window_length = 512
+  feedback_recurrence = True

aglib/meliad/transformer/configs/recurrent/lstm_dual.gin

@@ -0,0 +1,15 @@
+
+decoder_stack.DecoderStack:
+  dstack_window_length = 0
+  recurrent_layer_indices = (-3,)
+  feedback_recurrence = False
+
+transformer_layer.TransformerLayer:
+  window_length = 512
+  recurrent_num_states = 512
+  recurrent_gate_type = "lstm"
+  recurrent_single_gate = False
+  recurrent_skip_ffn = False
+
+training_loop.Trainer:
+  generate_every_steps = 0   # Not yet supported with recurrence.

aglib/meliad/transformer/configs/recurrent/lstm_single.gin

@@ -0,0 +1,15 @@
+
+decoder_stack.DecoderStack:
+  dstack_window_length = 0
+  recurrent_layer_indices = (-3,)
+  feedback_recurrence = False
+
+transformer_layer.TransformerLayer:
+  window_length = 512
+  recurrent_num_states = 512
+  recurrent_gate_type = "lstm"
+  recurrent_single_gate = True
+  recurrent_skip_ffn = False
+
+training_loop.Trainer:
+  generate_every_steps = 0   # Not yet supported with recurrence.

aglib/meliad/transformer/configs/recurrent/lstm_skip.gin

@@ -0,0 +1,15 @@
+
+decoder_stack.DecoderStack:
+  dstack_window_length = 0
+  recurrent_layer_indices = (-3,)
+  feedback_recurrence = False
+
+transformer_layer.TransformerLayer:
+  window_length = 512
+  recurrent_num_states = 512
+  recurrent_gate_type = "lstm"
+  recurrent_single_gate = False
+  recurrent_skip_ffn = True
+
+training_loop.Trainer:
+  generate_every_steps = 0   # Not yet supported with recurrence.

aglib/meliad/transformer/configs/recurrent/states_1024.gin

@@ -0,0 +1,5 @@
+
+# Apply after the gate config.
+
+transformer_layer.TransformerLayer:
+  recurrent_num_states = 1024

aglib/meliad/transformer/configs/recurrent/states_128.gin

@@ -0,0 +1,5 @@
+
+# Apply after the gate config.
+
+transformer_layer.TransformerLayer:
+  recurrent_num_states = 128

aglib/meliad/transformer/configs/recurrent/states_2048.gin

@@ -0,0 +1,5 @@
+
+# Apply after the gate config.
+
+transformer_layer.TransformerLayer:
+  recurrent_num_states = 2048

aglib/meliad/transformer/configs/recurrent/states_256.gin

@@ -0,0 +1,5 @@
+
+# Apply after the gate config.
+
+transformer_layer.TransformerLayer:
+  recurrent_num_states = 256

aglib/meliad/transformer/configs/recurrent/states_512.gin

@@ -0,0 +1,5 @@
+
+# Apply after the gate config.
+
+transformer_layer.TransformerLayer:
+  recurrent_num_states = 512

aglib/meliad/transformer/configs/recurrent/tall_24L.gin

@@ -0,0 +1,8 @@
+
+# Use a "tall" 24L configuration, with two recurrent layers.
+# Apply after the gate config.
+
+NUM_LAYERS = 24
+
+decoder_stack.DecoderStack:
+  recurrent_layer_indices = (9,19,)