From d30a28c76e0c5679cabc6c30ea7fa20065cd16cb Mon Sep 17 00:00:00 2001
From: caoying03 <caoying03@baidu.com>
Date: Tue, 25 Jul 2017 09:05:03 +0800
Subject: [PATCH] proj init.

---
 globally_normalized_reader/.gitignore       |   3 +
 globally_normalized_reader/basic_modules.py | 135 +++++++++++++
 globally_normalized_reader/config.py        |  25 +++
 globally_normalized_reader/model.py         | 199 ++++++++++++++++++++
 globally_normalized_reader/reader.py        |  43 +++++
 globally_normalized_reader/train.py         | 139 ++++++++++++++
 6 files changed, 544 insertions(+)
 create mode 100644 globally_normalized_reader/.gitignore
 create mode 100755 globally_normalized_reader/basic_modules.py
 create mode 100644 globally_normalized_reader/config.py
 create mode 100755 globally_normalized_reader/model.py
 create mode 100755 globally_normalized_reader/reader.py
 create mode 100755 globally_normalized_reader/train.py

diff --git a/globally_normalized_reader/.gitignore b/globally_normalized_reader/.gitignore
new file mode 100644
index 00000000..c345f460
--- /dev/null
+++ b/globally_normalized_reader/.gitignore
@@ -0,0 +1,3 @@
+data
+*.txt
+*.pyc
diff --git a/globally_normalized_reader/basic_modules.py b/globally_normalized_reader/basic_modules.py
new file mode 100755
index 00000000..99e1d905
--- /dev/null
+++ b/globally_normalized_reader/basic_modules.py
@@ -0,0 +1,135 @@
+#!/usr/bin/env python
+#coding=utf-8
+import pdb
+import collections
+
+import paddle.v2 as paddle
+from paddle.v2.layer import parse_network
+
+__all__ = [
+    "stacked_bidirectional_lstm",
+    "lstm_by_nested_sequence",
+]
+
+
+def stacked_bidirectional_lstm(inputs, size, depth, drop_rate=0., prefix=""):
+    if not isinstance(inputs, collections.Sequence):
+        inputs = [inputs]
+
+    lstm_last = []
+    for dirt in ["fwd", "bwd"]:
+        for i in range(depth):
+            input_proj = paddle.layer.mixed(
+                name="%s_in_proj_%0d_%s__" % (prefix, i, dirt),
+                size=size * 4,
+                bias_attr=paddle.attr.Param(initial_std=0.),
+                input=[paddle.layer.full_matrix_projection(lstm)] if i else [
+                    paddle.layer.full_matrix_projection(in_layer)
+                    for in_layer in inputs
+                ])
+            lstm = paddle.layer.lstmemory(
+                input=input_proj,
+                bias_attr=paddle.attr.Param(initial_std=0.),
+                param_attr=paddle.attr.Param(initial_std=5e-4),
+                reverse=(dirt == "bwd"))
+        lstm_last.append(lstm)
+
+    final_states = paddle.layer.concat(input=[
+        paddle.layer.last_seq(input=lstm_last[0]),
+        paddle.layer.first_seq(input=lstm_last[1]),
+    ])
+    return final_states, paddle.layer.concat(
+        input=lstm_last,
+        layer_attr=paddle.attr.ExtraLayerAttribute(drop_rate=drop_rate), )
+
+
+def lstm_by_nested_sequence(input_layer, hidden_dim, name="", reverse=False):
+    '''
+    This is a LSTM implemended by nested recurrent_group.
+    Paragraph is a nature nested sequence:
+    1. each paragraph is a sequence of sentence.
+    2. each sentence is a sequence of words.
+
+    This function ueses the nested recurrent_group to implement LSTM.
+    1. The outer group iterates over sentence in a paragraph.
+    2. The inner group iterates over words in a sentence.
+    3. A LSTM is used to encode sentence, its final outputs is used to
+       initialize memory of the LSTM that is used to encode the next sentence.
+    4. Parameters are shared among these sentence-encoding LSTMs.
+    5. Consequently, this function is just equivalent to concatenate all
+       sentences in a paragraph into one (long) sentence, and use one LSTM to
+       encode this new long sentence.
+    '''
+
+    def lstm_outer_step(lstm_group_input, hidden_dim, reverse, name=''):
+        outer_memory = paddle.layer.memory(
+            name="__inner_%s_last__" % name, size=hidden_dim)
+
+        def lstm_inner_step(input_layer, hidden_dim, reverse, name):
+            inner_memory = paddle.layer.memory(
+                name="__inner_state_%s__" % name,
+                size=hidden_dim,
+                boot_layer=outer_memory)
+            input_proj = paddle.layer.fc(
+                size=hidden_dim * 4, bias_attr=False, input=input_layer)
+            return paddle.networks.lstmemory_unit(
+                input=input_proj,
+                name="__inner_state_%s__" % name,
+                out_memory=inner_memory,
+                size=hidden_dim,
+                act=paddle.activation.Tanh(),
+                gate_act=paddle.activation.Sigmoid(),
+                state_act=paddle.activation.Tanh())
+
+        inner_out = paddle.layer.recurrent_group(
+            name="__inner_%s__" % name,
+            step=lstm_inner_step,
+            reverse=reverse,
+            input=[lstm_group_input, hidden_dim, reverse, name])
+
+        if reverse:
+            inner_last_output = paddle.layer.first_seq(
+                input=inner_out,
+                name="__inner_%s_last__" % name,
+                agg_level=paddle.layer.AggregateLevel.TO_SEQUENCE)
+        else:
+            inner_last_output = paddle.layer.last_seq(
+                input=inner_out,
+                name="__inner_%s_last__" % name,
+                agg_level=paddle.layer.AggregateLevel.TO_SEQUENCE)
+        return inner_out
+
+    return paddle.layer.recurrent_group(
+        input=[
+            paddle.layer.SubsequenceInput(input_layer), hidden_dim, reverse,
+            name
+        ],
+        step=lstm_outer_step,
+        name="__outter_%s__" % name,
+        reverse=reverse)
+
+
+def stacked_bi_lstm_by_nested_seq(input_layer, depth, hidden_dim, prefix=""):
+    lstm_final_outs = []
+    for dirt in ["fwd", "bwd"]:
+        for i in range(depth):
+            lstm_out = lstm_by_nested_sequence(
+                input_layer=(lstm_out if i else input_layer),
+                hidden_dim=hidden_dim,
+                name="__%s_%s_%02d__" % (prefix, dirt, i),
+                reverse=(dirt == "bwd"))
+        lstm_final_outs.append(lstm_out)
+    return paddle.layer.concat(input=lstm_final_outs)
+
+
+if __name__ == "__main__":
+    vocab_size = 1024
+    emb_dim = 128
+    embedding = paddle.layer.embedding(
+        input=paddle.layer.data(
+            name="word",
+            type=paddle.data_type.integer_value_sub_sequence(vocab_size)),
+        size=emb_dim)
+    print(parse_network(
+        stacked_bi_lstm_by_nested_seq(
+            input_layer=embedding, depth=3, hidden_dim=128, prefix="__lstm")))
diff --git a/globally_normalized_reader/config.py b/globally_normalized_reader/config.py
new file mode 100644
index 00000000..9b0ec4f0
--- /dev/null
+++ b/globally_normalized_reader/config.py
@@ -0,0 +1,25 @@
+#!/usr/bin/env python
+#coding=utf-8
+
+__all__ = ["ModelConfig"]
+
+
+class ModelConfig(object):
+    beam_size = 3
+    vocab_size = 102400
+    embedding_dim = 256
+    embedding_droprate = 0.3
+
+    lstm_depth = 3
+    lstm_hidden_dim = 300
+    lstm_hidden_droprate = 0.3
+
+    passage_indep_embedding_dim = 300
+    passage_aligned_embedding_dim = 128
+
+    beam_size = 5
+
+
+class TrainerConfig(object):
+    learning_rate = 1e-3
+    data_dir = "data/featurized"
diff --git a/globally_normalized_reader/model.py b/globally_normalized_reader/model.py
new file mode 100755
index 00000000..21e86b65
--- /dev/null
+++ b/globally_normalized_reader/model.py
@@ -0,0 +1,199 @@
+#!/usr/bin/env python
+#coding=utf-8
+import pdb
+
+import paddle.v2 as paddle
+from paddle.v2.layer import parse_network
+import basic_modules
+from config import ModelConfig
+
+__all__ = ["GNR"]
+
+
+def build_pretrained_embedding(name,
+                               data_type,
+                               vocab_size,
+                               emb_dim,
+                               emb_drop=0.):
+    one_hot_input = paddle.layer.data(
+        name=name, type=paddle.data_type.integer_value_sequence(vocab_size))
+    return paddle.layer.embedding(
+        input=one_hot_input,
+        size=emb_dim,
+        param_attr=paddle.attr.Param(
+            name="GloveVectors", is_static=True),
+        layer_attr=paddle.attr.ExtraLayerAttribute(drop_rate=emb_drop), )
+
+
+def encode_question(input_embedding, config, prefix):
+    lstm_final, lstm_outs = basic_modules.stacked_bidirectional_lstm(
+        inputs=input_embedding,
+        size=config.lstm_hidden_dim,
+        depth=config.lstm_depth,
+        drop_rate=config.lstm_hidden_droprate,
+        prefix=prefix)
+
+    # passage-independent embeddings
+    candidates = paddle.layer.fc(input=lstm_outs,
+                                 bias_attr=False,
+                                 size=config.passage_indep_embedding_dim,
+                                 act=paddle.activation.Linear())
+    weights = paddle.layer.fc(input=lstm_outs,
+                              size=1,
+                              act=paddle.activation.SequenceSoftmax())
+    weighted_candidates = paddle.layer.scaling(input=candidates, weight=weights)
+    passage_indep_embedding = paddle.layer.pooling(
+        input=weighted_candidates, pooling_type=paddle.pooling.Sum())
+    return paddle.layer.concat(
+        input=[lstm_final, passage_indep_embedding]), lstm_outs
+
+
+def question_aligned_passage_embedding(question_lstm_outs, document_embeddings,
+                                       config):
+    def outer_sentence_step(document_embeddings, question_lstm_outs, config):
+        '''
+        in this recurrent_group, document_embeddings has scattered into sequence,
+        '''
+
+        def inner_word_step(word_embedding, question_lstm_outs,
+                            question_outs_proj, config):
+            '''
+            in this recurrent_group, sentence embedding has scattered into word
+            embeddings.
+            '''
+            doc_word_expand = paddle.layer.expand(
+                input=word_embedding,
+                expand_as=question_lstm_outs,
+                expand_level=paddle.layer.ExpandLevel.FROM_NO_SEQUENCE)
+
+            weights = paddle.layer.fc(
+                input=[question_lstm_outs, doc_word_expand],
+                size=1,
+                act=paddle.activation.SequenceSoftmax())
+            weighted_candidates = paddle.layer.scaling(
+                input=question_outs_proj, weight=weights)
+            return paddle.layer.pooling(
+                input=weighted_candidates, pooling_type=paddle.pooling.Sum())
+
+        question_outs_proj = paddle.layer.fc(
+            input=question_lstm_outs,
+            bias_attr=False,
+            size=config.passage_aligned_embedding_dim)
+        return paddle.layer.recurrent_group(
+            input=[
+                paddle.layer.SubsequenceInput(document_embeddings),
+                paddle.layer.StaticInput(question_lstm_outs),
+                paddle.layer.StaticInput(question_outs_proj),
+                config,
+            ],
+            step=inner_word_step,
+            name="iter_over_word")
+
+    return paddle.layer.recurrent_group(
+        input=[
+            paddle.layer.SubsequenceInput(document_embeddings),
+            paddle.layer.StaticInput(question_lstm_outs), config
+        ],
+        step=outer_sentence_step,
+        name="iter_over_sen")
+
+
+def encode_documents(input_embedding, same_as_question, question_vector,
+                     question_lstm_outs, config, prefix):
+    question_expanded = paddle.layer.expand(
+        input=question_vector,
+        expand_as=input_embedding,
+        expand_level=paddle.layer.ExpandLevel.FROM_NO_SEQUENCE)
+    question_aligned_embedding = question_aligned_passage_embedding(
+        question_lstm_outs, input_embedding, config)
+    return paddle.layer.concat(input=[
+        input_embedding, question_expanded, same_as_question,
+        question_aligned_embedding
+    ])
+
+
+def search_answer(doc_lstm_outs, sentence_idx, start_idx, end_idx, config):
+    last_state_of_sentence = paddle.layer.last_seq(
+        input=doc_lstm_outs, agg_level=paddle.layer.AggregateLevel.TO_SEQUENCE)
+
+    # HERE do not use sequence softmax activition.
+    sentence_scores = paddle.layer.fc(input=last_state_of_sentence,
+                                      size=1,
+                                      act=paddle.activation.Exp())
+    topk_sentence_ids = paddle.layer.kmax_sequence_score(
+        input=sentence_scores, beam_size=config.beam_size)
+    topk_sen = paddle.layer.sub_nested_seq(
+        input=last_state_of_sentence, selected_indices=topk_sentence_ids)
+
+    # expand beam to search start positions on selected sentences
+    start_pos_scores = paddle.layer.fc(input=topk_sen,
+                                       size=1,
+                                       act=paddle.activation.Exp())
+    topk_start_pos_ids = paddle.layer.kmax_sequence_score(
+        input=sentence_scores, beam_size=config.beam_size)
+    topk_start_spans = paddle.layer.seq_slice(
+        input=topk_sen, starts=topk_start_pos_ids, ends=None)
+
+    # expand beam to search end positions on selected start spans
+    _, end_span_embedding = basic_modules.stacked_bidirectional_lstm(
+        inputs=topk_start_spans,
+        size=config.lstm_hidden_dim,
+        depth=config.lstm_depth,
+        drop_rate=config.lstm_hidden_droprate,
+        prefix="__end_span_embeddings__")
+    end_pos_scores = paddle.layer.fc(input=end_span_embedding,
+                                     size=1,
+                                     act=paddle.activation.Exp())
+    topk_end_pos_ids = paddle.layer.kmax_sequence_score(
+        input=end_pos_scores, beam_size=config.beam_size)
+    cost = paddle.layer.cross_entropy_over_beam(
+        input=[
+            sentence_scores, topk_sentence_ids, start_pos_scores,
+            topk_start_pos_ids, end_pos_scores, topk_end_pos_ids
+        ],
+        label=[sentence_idx, start_idx, end_idx])
+    return cost
+
+
+def GNR(config):
+    # encoding question words
+    question_embeddings = build_pretrained_embedding(
+        "question", paddle.data_type.integer_value_sequence, config.vocab_size,
+        config.embedding_dim, config.embedding_droprate)
+    question_vector, question_lstm_outs = encode_question(
+        input_embedding=question_embeddings, config=config, prefix="__ques")
+
+    # encoding document words
+    document_embeddings = build_pretrained_embedding(
+        "documents", paddle.data_type.integer_value_sub_sequence,
+        config.vocab_size, config.embedding_dim, config.embedding_droprate)
+    same_as_question = paddle.layer.data(
+        name="same_as_question",
+        type=paddle.data_type.integer_value_sub_sequence(2))
+    document_words_ecoding = encode_documents(
+        input_embedding=document_embeddings,
+        question_vector=question_vector,
+        question_lstm_outs=question_lstm_outs,
+        same_as_question=same_as_question,
+        config=config,
+        prefix="__doc")
+
+    doc_lstm_outs = basic_modules.stacked_bi_lstm_by_nested_seq(
+        input_layer=document_words_ecoding,
+        hidden_dim=config.lstm_hidden_dim,
+        depth=config.lstm_depth,
+        prefix="__doc_lstm")
+
+    # define labels
+    sentence_idx = paddle.layer.data(
+        name="sen_idx", type=paddle.data_type.integer_value(1))
+    start_idx = paddle.layer.data(
+        name="start_idx", type=paddle.data_type.integer_value(1))
+    end_idx = paddle.layer.data(
+        name="end_idx", type=paddle.data_type.integer_value(1))
+    return search_answer(doc_lstm_outs, sentence_idx, start_idx, end_idx,
+                         config)
+
+
+if __name__ == "__main__":
+    print(parse_network(GNR(ModelConfig)))
diff --git a/globally_normalized_reader/reader.py b/globally_normalized_reader/reader.py
new file mode 100755
index 00000000..20e7128b
--- /dev/null
+++ b/globally_normalized_reader/reader.py
@@ -0,0 +1,43 @@
+#!/usr/bin/env python
+#coding=utf-8
+import pdb
+import os
+import random
+import json
+
+
+def train_reader(data_list, is_train=True):
+    def reader():
+        # every pass shuffle the data list again
+        if is_train:
+            random.shuffle(data_list)
+
+        for train_sample in data_list:
+            data = json.load(open(train_sample, "r"))
+            sent_len = data['sent_lengths']
+
+            doc_len = len(data['context'])
+            same_as_question_word = [[[x]]
+                                     for x in data['same_as_question_word']]
+
+            ans_sentence = [0] * doc_len
+            ans_sentence[data['ans_sentence']] = 1
+
+            ans_start = [0] * doc_len
+            ans_start[data['ans_start']] = 1
+
+            ans_end = [0] * doc_len
+            ans_end[data['ans_end']] = 1
+            yield (data['question'], data['context'], same_as_question_word,
+                   ans_sentence, ans_start, ans_end)
+
+    return reader
+
+
+if __name__ == "__main__":
+    from train import choose_samples
+
+    train_list, dev_list = choose_samples("data/featurized")
+    for i, item in enumerate(train_reader(train_list)()):
+        print(item)
+        if i > 5: break
diff --git a/globally_normalized_reader/train.py b/globally_normalized_reader/train.py
new file mode 100755
index 00000000..8ef4bab1
--- /dev/null
+++ b/globally_normalized_reader/train.py
@@ -0,0 +1,139 @@
+#!/usr/bin/env python
+#coding=utf-8
+from __future__ import print_function
+
+import pdb
+import os
+import sys
+import logging
+import random
+import glob
+import gzip
+
+import reader
+import paddle.v2 as paddle
+from paddle.v2.layer import parse_network
+from model import GNR
+from config import ModelConfig, TrainerConfig
+
+logger = logging.getLogger("paddle")
+logger.setLevel(logging.INFO)
+
+
+def load_pretrained_parameters(path, height, width):
+    return
+
+
+def save_model(save_path, parameters):
+    with gzip.open(save_path, "w") as f:
+        parameters.to_tar(f)
+
+
+def load_initial_model(model_path, parameters):
+    with gzip.open(model_path, "rb") as f:
+        parameters.init_from_tar(f)
+
+
+def choose_samples(path):
+    """
+    Load filenames for train, dev, and augmented samples.
+    """
+    if not os.path.exists(os.path.join(path, "train")):
+        print(
+            "Non-existent directory as input path: {}".format(path),
+            file=sys.stderr)
+        sys.exit(1)
+
+    # Get paths to all samples that we want to load.
+    train_samples = glob.glob(os.path.join(path, "train", "*"))
+    valid_samples = glob.glob(os.path.join(path, "dev", "*"))
+
+    train_samples.sort()
+    valid_samples.sort()
+
+    random.shuffle(train_samples)
+
+    return train_samples, valid_samples
+
+
+def build_reader(data_dir):
+    """
+    Build the data reader for this model.
+    """
+    train_samples, valid_samples = choose_samples(data_dir)
+    pdb.set_trace()
+
+    train_reader = paddle.batch(
+        paddle.reader.shuffle(
+            reader.train_reader(train_samples), buf_size=102400),
+        batch_size=config.batch_size)
+
+    # testing data is not shuffled
+    test_reader = paddle.batch(
+        reader.train_reader(valid_samples, is_train=False),
+        batch_size=config.batch_size)
+    return train_reader, test_reader
+
+
+def build_event_handler(config, parameters, trainer, test_reader):
+    """
+    Build the event handler for this model.
+    """
+
+    # End batch and end pass event handler
+    def event_handler(event):
+        """The event handler."""
+        if isinstance(event, paddle.event.EndIteration):
+            if (not event.batch_id % 100) and event.batch_id:
+                save_model("checkpoint_param.latest.tar.gz", parameters)
+
+            if not event.batch_id % 5:
+                logger.info(
+                    "Pass %d, Batch %d, Cost %f, %s" %
+                    (event.pass_id, event.batch_id, event.cost, event.metrics))
+
+        if isinstance(event, paddle.event.EndPass):
+            save_model(config.param_save_filename_format % event.pass_id,
+                       parameters)
+            with gzip.open(param_path, 'w') as handle:
+                parameters.to_tar(handle)
+
+            result = trainer.test(reader=test_reader)
+            logger.info("Test with Pass %d, %s" %
+                        (event.pass_id, result.metrics))
+
+    return event_handler
+
+
+def train(model_config, trainer_config):
+    paddle.init(use_gpu=True, trainer_count=1)
+
+    # define the optimizer
+    optimizer = paddle.optimizer.Adam(
+        learning_rate=trainer_config.learning_rate,
+        regularization=paddle.optimizer.L2Regularization(rate=1e-3),
+        model_average=paddle.optimizer.ModelAverage(average_window=0.5))
+
+    # define network topology
+    losses = GNR(model_config)
+    parameters = paddle.parameters.create(losses)
+    # print(parse_network(losses))
+    trainer = paddle.trainer.SGD(
+        cost=losses, parameters=parameters, update_equation=optimizer)
+    """
+    parameters.set('GloveVectors',
+                   load_pretrained_parameters(parameter_path, height, width))
+    """
+
+    # define data reader
+    train_reader, test_reader = build_reader(trainer_config.data_dir)
+
+    event_handler = build_event_handler(conf, parameters, trainer, test_reader)
+    trainer.train(
+        reader=train_reader,
+        num_passes=conf.epochs,
+        event_handler=event_handler)
+
+
+if __name__ == "__main__":
+    train(ModelConfig, TrainerConfig)
-- 
GitLab