start_id=start_of_sentence_id,
end_id=end_of_sentence_id,
encoder_hidden_dim=memory_hidden_dim,
dtype=tf_datatype)
embedding_table = np.random.rand(vocab_size, hidden_dim).astype(
np_datatype) # a [vocab_size, hidden_dim] table
embedding_table = tf.convert_to_tensor(embedding_table)
memory, memory_sequence_length = generate_encoder_result(
batch_size, max_seq_len, memory_hidden_dim, tf_datatype)
finalized_tf_output_ids, finalized_tf_sequence_lengths, tf_output_ids, \
tf_parent_ids, tf_sequence_lengths = tf_decoding(memory,
memory_sequence_length,
embedding_table,
decoding_args,
0,
kernel_initializer_range,
bias_initializer_range)
all_vars = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES)
finalized_op_output_ids, finalized_op_sequence_lengths, op_output_ids, \
op_parent_ids, op_sequence_lengths = op_decoding(memory,
memory_sequence_length,
embedding_table,
all_vars,
decoding_args)
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
if use_XLA == 1:
start_id=start_of_sentence_id,
end_id=end_of_sentence_id,
encoder_hidden_dim=memory_hidden_dim,
dtype=tf_datatype)
embedding_table = np.random.randn(vocab_size, hidden_dim).astype(
np_datatype) * 0.01 # a [vocab_size, hidden_dim] table
embedding_table = tf.convert_to_tensor(embedding_table)
memory, memory_sequence_length = generate_encoder_result(
batch_size, max_seq_len, memory_hidden_dim, tf_datatype)
finalized_tf_output_ids, finalized_tf_sequence_lengths, tf_output_ids, \
tf_parent_ids, tf_sequence_lengths = tf_decoding(memory,
memory_sequence_length,
embedding_table,
decoding_args,
args.decoder_type,
kernel_initializer_range,
bias_initializer_range,
atol_threshold)
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
with tf.Session(config=config) as sess:
sess.run(tf.global_variables_initializer())
sess.run(tf.tables_initializer())
sess.run(finalized_tf_output_ids)
if args.test_time == 1:
time_cost = time_test(sess, finalized_tf_output_ids, iterations=50)
types = ["TF", "OP", "TF+OP"]
print("[INFO] time costs of {} decoder: {} ms.".format(
memory=memory,
memory_sequence_length=memory_sequence_length)
target_vocab_rev = target_inputter.vocabulary_lookup_reverse()
target_tokens = target_vocab_rev.lookup(tf.cast(target_ids, tf.int64))
opennmt_target_length = target_length
opennmt_target_tokens = target_tokens
opennmt_target_ids = target_ids
opennmt_variables = tf.global_variables()
## TF Decoding ###
finalized_tf_output_ids, finalized_tf_sequence_lengths, tf_output_ids, \
tf_parent_ids, tf_sequence_lengths = tf_decoding(tf_encoder_result,
memory_sequence_length,
target_inputter.embedding,
decoding_args,
decoder_type=1,
kernel_initializer_range=kernel_initializer_range,
bias_initializer_range=bias_initializer_range)
tf_target_ids = finalized_tf_output_ids
tf_target_length = finalized_tf_sequence_lengths
tf_target_tokens = target_vocab_rev.lookup(tf.cast(tf_target_ids,
tf.int64))
## end of tf decoding ##
## op decoding ##
all_vars = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES)
decoder_var_start_id = 0
while all_vars[decoder_var_start_id].name.find(
max_seq_len=max_seq_len,
vocab_size=vocab_size,
start_id=start_of_sentence_id,
end_id=end_of_sentence_id,
encoder_hidden_dim=encoder_head_num *
encoder_size_per_head,
dtype=tf_datatype)
tf_encoder_result = tf_encoder(input_tensor=from_tensor,
encoder_args=encoder_args,
attention_mask=attention_mask)
tf_encoder_result = tf.reshape(
tf_encoder_result, [batch_size, max_seq_len, encoder_hidden_dim])
tf_decoding_result, _, _, _, _ = tf_decoding(
tf_encoder_result, memory_sequence_length, embedding_table,
decoding_args, args.decoder_type, kernel_initializer_range,
bias_initializer_range, atol_threshold)
encoder_variables = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES)
op_encoder_result = op_encoder(inputs=from_tensor,
encoder_args=encoder_args,
encoder_vars=encoder_variables,
attention_mask=attention_mask)
op_encoder_result = tf.reshape(
op_encoder_result, [batch_size, max_seq_len, encoder_hidden_dim])
op_decoding_result, _, _, _, _ = tf_decoding(
op_encoder_result, memory_sequence_length, embedding_table,
decoding_args, args.decoder_type, kernel_initializer_range,
bias_initializer_range, atol_threshold)