Unable to freeze bidirectional LSTM

[stefan-falk]

I am using the following to define some input/output nodes for a t2t model:

def load_translation_model(ckpt_dir, config):
    d = get_model_hparams(ckpt_dir)
    hparams = d['hparams']
    problem = d['problem']
    model = d['model']

    # Modifying the graph
    decode_length = 100
    input_node_name = 'encoder_inputs'
    output_node_name = 'decoder_outputs'

    input_node = tf.placeholder(dtype=tf.int32, shape=[None, None, 1, 1], name=input_node_name)

    encoded_dict = problem.preprocess_example(
        {'inputs': input_node, 'targets': [0]}, tf.estimator.ModeKeys.PREDICT, hparams
    )

    outputs = model.infer(
        encoded_dict, beam_size=config['beam_size'], alpha=config['alpha'], decode_length=decode_length
    )
    output_node = tf.identity(outputs['outputs'], name=output_node_name)
    inputs = {input_node_name: input_node}
    outputs = {output_node_name: output_node}
    return model, inputs, outputs

Taking this and freeze the graph after saving the model work (at least for the Transformer model). However, it fails as I try to freeze a lstm_seq2seq_attention_bidirectional_encoder model.

    freeze.freeze_graph(
        input_graph=input_graph,
        input_checkpoint=ckpt_path,
        output_graph=output_graph,
        output_node_names=','.join(output_node_names),
        input_binary=True,
        input_saver='',
        restore_op_name='save/restore_all',
        filename_tensor_name='save/Const:0',
        clear_devices=True,
        initializer_nodes=''
    )

The issue seems to go deeper. I have found this (https://github.com/tensorflow/tensorflow/issues/24591) which seems to be related to this issue.

Does anybody else have this problem or knows how to solve this?

Full trace of the error:

Traceback (most recent call last):
  File "/home/sfalk/tmp/pycharm_project_265/asr/model/persistence.py", line 392, in <module>
    main()
  File "/home/sfalk/tmp/pycharm_project_265/asr/model/persistence.py", line 377, in main
    dump_latest_checkpoint(output_dir, ckpt_path, config, graph_builder=graph_builder)
  File "/home/sfalk/tmp/pycharm_project_265/asr/model/persistence.py", line 290, in dump_latest_checkpoint
    freeze_graph(graph_file, frozen_graph_file, ckpt_path, output_nodes.keys())
  File "/home/sfalk/tmp/pycharm_project_265/asr/model/persistence.py", line 193, in freeze_graph
    initializer_nodes=''
  File "/home/sfalk/miniconda3/envs/t2t/lib/python3.5/site-packages/tensorflow/python/tools/freeze_graph.py", line 363, in freeze_graph
    checkpoint_version=checkpoint_version)
  File "/home/sfalk/miniconda3/envs/t2t/lib/python3.5/site-packages/tensorflow/python/tools/freeze_graph.py", line 190, in freeze_graph_with_def_protos
    var_list=var_list, write_version=checkpoint_version)
  File "/home/sfalk/miniconda3/envs/t2t/lib/python3.5/site-packages/tensorflow/python/training/saver.py", line 1102, in __init__
    self.build()
  File "/home/sfalk/miniconda3/envs/t2t/lib/python3.5/site-packages/tensorflow/python/training/saver.py", line 1114, in build
    self._build(self._filename, build_save=True, build_restore=True)
  File "/home/sfalk/miniconda3/envs/t2t/lib/python3.5/site-packages/tensorflow/python/training/saver.py", line 1151, in _build
    build_save=build_save, build_restore=build_restore)
  File "/home/sfalk/miniconda3/envs/t2t/lib/python3.5/site-packages/tensorflow/python/training/saver.py", line 773, in _build_internal
    saveables = self._ValidateAndSliceInputs(names_to_saveables)
  File "/home/sfalk/miniconda3/envs/t2t/lib/python3.5/site-packages/tensorflow/python/training/saver.py", line 680, in _ValidateAndSliceInputs
    for converted_saveable_object in self.SaveableObjectsForOp(op, name):
  File "/home/sfalk/miniconda3/envs/t2t/lib/python3.5/site-packages/tensorflow/python/training/saver.py", line 654, in SaveableObjectsForOp
    variable, "", name)
  File "/home/sfalk/miniconda3/envs/t2t/lib/python3.5/site-packages/tensorflow/python/training/saver.py", line 128, in __init__
    self.handle_op = var.op.inputs[0]
  File "/home/sfalk/miniconda3/envs/t2t/lib/python3.5/site-packages/tensorflow/python/framework/ops.py", line 2128, in __getitem__
    return self._inputs[i]
IndexError: list index out of range

[jiarongqiu]

I am facing the same issue when trying to freeze a simple RNN

[jiarongqiu]

My problem is solved. It is caused by tf.nn.rnn_cell.DropoutWrapper. Just delete it in the model definition part.

[stefan-falk]

@jiarongqiu Thanks for reporting back with a solution. However, another model I tried to freeze was this:

def get_model(input_dict_size, output_dict_size, max_input_length, max_output_length):
    # Defining the encoder
    encoder = Embedding(input_dict_size, 128, input_length=max_input_length, mask_zero=True)(encoder_input)
    encoder = Bidirectional(LSTM(128, return_sequences=True, unroll=True), merge_mode='concat')(encoder)
    encoder_last = encoder[:, -1, :]
    # Defining the decoder
    decoder_input = Input(shape=(max_output_length,))
    decoder = Embedding(output_dict_size, 256, input_length=max_output_length, mask_zero=True)(decoder_input)
    decoder = LSTM(256, return_sequences=True, unroll=True)(decoder, initial_state=[encoder_last, encoder_last])
    return decoder

I am not calling tf.nn.rnn_cell.DropoutWrapper - unless it's being used internally?