tensorflow - tf.contrib.seq2seq.Bahdanau の使用方法注意

Question

tf 1.1 で注意を払って seq2seq モデルの簡単なコードを作成しようとしています。パラメータ「クエリメカニズムの深さ」が何であるかわかりません。Attention Mechanisms の作成時に、次のようなエラーが表示されます。

TypeError: int() argument must be a string, a bytes-like object or a number, not 'TensorShape'

これが私のコードです。私は正しい軌道に乗っていますか？詳細なドキュメントは見つかりませんでした。

import tensorflow as tf
from tensorflow.contrib.rnn import LSTMCell, LSTMStateTuple, BasicLSTMCell, DropoutWrapper, MultiRNNCell, EmbeddingWrapper, static_rnn 
import tensorflow.contrib.seq2seq as seq2seq
import attention_wrapper as wrapper


tf.reset_default_graph()
try:
    sess.close()
except:

    pass
sess = tf.InteractiveSession()


## Place holders

encode_input = [tf.placeholder(tf.int32, 
                                shape=(None,),
                                name = "ei_%i" %i)
                                for i in range(input_seq_length)]

labels = [tf.placeholder(tf.int32,
                                shape=(None,),
                                name = "l_%i" %i)
                                for i in range(output_seq_length)]

decode_input = [tf.zeros_like(encode_input[0], dtype=np.int32, name="GO")] + labels[:-1]



############ Encoder
lstm_cell = BasicLSTMCell(embedding_dim)
encoder_cell = EmbeddingWrapper(lstm_cell, embedding_classes=input_vocab_size, embedding_size=embedding_dim)
encoder_outputs, encoder_state = static_rnn(encoder_cell, encode_input, dtype=tf.float32) 

############ Decoder
# Attention Mechanisms. Bahdanau is additive style attention
attn_mech = tf.contrib.seq2seq.BahdanauAttention(
    num_units = input_seq_length, # depth of query mechanism
    memory = encoder_outputs, # hidden states to attend (output of RNN)
    normalize=False, # normalize energy term
    name='BahdanauAttention')

lstm_cell_decoder = BasicLSTMCell(embedding_dim)

# Attention Wrapper: adds the attention mechanism to the cell
attn_cell = wrapper.AttentionWrapper(
    cell = lstm_cell_decoder,# Instance of RNNCell
    attention_mechanism = attn_mech, # Instance of AttentionMechanism
    attention_size = embedding_dim, # Int, depth of attention (output) tensor
    attention_history=False, # whether to store history in final output
    name="attention_wrapper")


# Decoder setup
decoder = tf.contrib.seq2seq.BasicDecoder(
          cell = lstm_cell_decoder,
          helper = helper, # A Helper instance
          initial_state = encoder_state, # initial state of decoder
          output_layer = None) # instance of tf.layers.Layer, like Dense

# Perform dynamic decoding with decoder object
outputs, final_state = tf.contrib.seq2seq.dynamic_decode(decoder)