sample_trainer_nest_rnn_gen.conf

#edit-mode: -*- python -*-
# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.


from paddle.trainer_config_helpers import *

settings(batch_size=15, learning_rate=0)

num_words = 5
beam_flag = get_config_arg('beam_search', bool, False)

sent_id = data_layer(name="sent_id", size=1)

# This layer has no actual use, but only to decide batch_size in generation.
# When generating, at least one Memory in RecurrentLayer MUST have a boot layer.
dummy_data = data_layer(name="dummy_data_input", size=2)

def outer_step(dummy_data):

    gen_inputs = [StaticInput(input=dummy_data, size=2, is_seq=True),
                  GeneratedInput(size=num_words,
                                 embedding_name="wordvec",
                                 embedding_size=num_words)]

    def inner_step(dummy_memory, predict_word):
        
        # simplified RNN for testing
        with mixed_layer(size=num_words) as layer:
            layer += full_matrix_projection(input=predict_word,
                                            param_attr=ParamAttr(name="transtable"))

        with mixed_layer(size=num_words, act=ExpActivation()) as out:
            out += trans_full_matrix_projection(input=layer,
                                                param_attr=ParamAttr(name="wordvec"))

        return out
    
    beam_gen = beam_search(name="rnn_gen",
                           step=inner_step,
                           input=gen_inputs,
                           bos_id=0,
                           eos_id=num_words-1,
                           beam_size=2 if beam_flag else 1,
                           num_results_per_sample=2 if beam_flag else 1,
                           max_length=10) 
    return beam_gen

beam_gen_concat = recurrent_group(name="rnn_gen_concat",
                                  step=outer_step,
                                  input=[SubsequenceInput(dummy_data)])

seqtext_printer_evaluator(input=beam_gen_concat,
                          id_input=sent_id,
                          dict_file="./trainer/tests/test_gen_dict.txt",
                          result_file="./trainer/tests/dump_text.test")
#outputs(beam_gen_concat)
# In this config, as dummy_data_input doesn't work on beam_gen (we can find dummy_memory
# is read-only memory, and isn't used by other layers of step), we show the Inputs and Outputs
# as follows. Note that "__beam_search_predict__" is the default output name of beam_search.
Inputs("sent_id","dummy_data_input")
Outputs("__beam_search_predict__")