while all_vars[decoder_var_start_id].name.find(
"transformer/decoding") == -1:
decoder_var_start_id += 1
encoder_variables = all_vars[:decoder_var_start_id]
decoder_variables = all_vars[decoder_var_start_id:]
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])
finalized_op_output_ids, finalized_op_sequence_lengths, op_output_ids, \
op_parent_ids, op_sequence_lengths = op_decoding(op_encoder_result,
memory_sequence_length,
embedding_table,
decoder_variables,
decoding_args)
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())
finalized_tf_output_ids_result, tf_output_ids_result, tf_parent_ids_result, \
tf_sequence_lengths_result = sess.run(
[finalized_tf_output_ids, tf_output_ids, tf_parent_ids, tf_sequence_lengths])
finalized_op_output_ids_result, op_output_ids_result, op_parent_ids_result, \
op_sequence_lengths_result = sess.run(
[finalized_op_output_ids, op_output_ids, op_parent_ids, op_sequence_lengths])
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:
config.graph_options.optimizer_options.global_jit_level = tf.OptimizerOptions.ON_1
with tf.Session(config=config) as sess:
sess.run(tf.global_variables_initializer())
sess.run(tf.tables_initializer())
if args.cross_check == 1:
finalized_tf_output_ids_result, tf_output_ids_result, tf_parent_ids_result, \
tf_sequence_lengths_result = sess.run(
[finalized_tf_output_ids, tf_output_ids, tf_parent_ids, tf_sequence_lengths])
## 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(
"transformer/decoding") == -1:
decoder_var_start_id += 1
encoder_variables = all_vars[:decoder_var_start_id]
decoder_variables = all_vars[decoder_var_start_id:]
finalized_op_output_ids, finalized_op_sequence_lengths, op_output_ids, \
op_parent_ids, op_sequence_lengths = op_decoding(tf_encoder_result,
memory_sequence_length,
target_inputter.embedding,
decoder_variables, # first one is embedding table
decoding_args)
op_target_ids = finalized_op_output_ids
op_target_length = finalized_op_sequence_lengths
op_target_tokens = target_vocab_rev.lookup(tf.cast(op_target_ids,
tf.int64))
## end of op decoding
opennmt_target_ids = tf.cast(opennmt_target_ids, tf.int32)
tf_target_ids = tf.cast(tf_target_ids, tf.int32)
op_target_ids = tf.cast(op_target_ids, tf.int32)
opennmt_target_length = tf.minimum(opennmt_target_length + 1,