fix: comply with function signature for tfa.seq2seq.dynamic_decode

ludwig/models/modules/recurrent_modules.py

@@ -397,7 +397,7 @@ def recurrent_decoder(encoder_outputs, targets, max_sequence_length, vocab_size,
 
         # ================ Decoding ================
         def decode(initial_state, cell, sampler, beam_width=1,
-                   projection_layer=None):
+                   projection_layer=None, inputs=None):
             # The decoder itself
             if beam_width > 1:
                 # Tile inputs for beam search decoder
@@ -414,22 +414,20 @@ def decode(initial_state, cell, sampler, beam_width=1,
             else:
                 decoder = BasicDecoder(
                     cell=cell, sampler=sampler,
-                    # todo tf2: remove obsolete code #initial_state=initial_state,
                     output_layer=projection_layer)
-
-                initial_state = cell.get_initial_state(
-                    batch_size=batch_size, dtype=tf.float32
-                )
+                    # todo tf2: remove obsolete code #initial_state=initial_state,
 
                 # todo tf2: need to figure out 'inputs' to next function
-                decoder.initialize(inputs=None, initial_state=initial_state)
+                decoder.initialize(inputs, initial_state=initial_state)
 
             # The decoding operation
             outputs = tfa.seq2seq.dynamic_decode(
                 decoder=decoder,
                 output_time_major=False,
                 impute_finished=False if beam_width > 1 else True,
-                maximum_iterations=max_sequence_length
+                maximum_iterations=max_sequence_length,
+                decoder_init_input=inputs,
+                decoder_init_kwargs={'initial_state': initial_state}
             )
 
             return outputs
@@ -451,15 +449,25 @@ def decode(initial_state, cell, sampler, beam_width=1,
 
         else:
             train_sampler = tfa.seq2seq.sampler.TrainingSampler()
+            train_sampler.initialize(targets_embedded,
+                                     sequence_length=targets_sequence_length_with_eos)
             # todo tf2: cleanout obsolete code
             # train_helper = tfa.seq2seq.sampler.TrainingSampler(
             #     inputs=targets_embedded,
             #     sequence_length=targets_sequence_length_with_eos)
+
+            # # todo tf2: test code
+            # initial_state = cell.get_initial_state(
+            #     batch_size=batch_size, dtype=tf.float32
+            # )
+
             final_outputs_train, final_state_train, final_sequence_lengths_train = decode(
                 initial_state,
                 cell,
                 train_sampler,  # todo: tf2 to be removed #train_helper,
-                projection_layer=projection_layer)
+                projection_layer=projection_layer,
+                inputs=encoder_outputs
+            )
             eval_logits = final_outputs_train.rnn_output
             train_logits = final_outputs_train.projection_input
             # train_predictions = final_outputs_train.sample_id