学习笔记TF021:预测编码、字符级语言建模、ArXiv摘要

序列标注(sequence labelling)，输入序列每一帧预测一个类别。OCR(Optical Character Recognition 光学字符识别)。

MIT口语系统研究组Rob Kassel收集，斯坦福大学人工智能实验室Ben Taskar预处理OCR数据集(http://ai.stanford.edu/~btaskar/ocr/ )，包含大量单独手写小写字母，每个样本对应16X8像素二值图像。字线组合序列，序列对应单词。6800个，长度不超过14字母的单词。gzip压缩，内容用Tab分隔文本文件。Python csv模块直接读取。文件每行一个归一化字母属性，ID号、标签、像素值、下一字母ID号等。

下一字母ID值排序，按照正确顺序读取每个单词字母。收集字母，直到下一个ID对应字段未被设置为止。读取新序列。读取完目标字母及数据像素，用零图像填充序列对象，能纳入两个较大目标字母所有像素数据NumPy数组。

时间步之间共享softmax层。数据和目标数组包含序列，每个目标字母对应一个图像帧。RNN扩展，每个字母输出添加softmax分类器。分类器对每帧数据而非整个序列评估预测结果。计算序列长度。一个softmax层添加到所有帧：或者为所有帧添加几个不同分类器，或者令所有帧共享同一个分类器。共享分类器，权值在训练中被调整次数更多，训练单词每个字母。一个全连接层权值矩阵维数batch_size*in_size*out_size。现需要在两个输入维度batch_size、sequence_steps更新权值矩阵。令输入(RNN输出活性值)扁平为形状batch_size*sequence_steps*in_size。权值矩阵变成较大的批数据。结果反扁平化(unflatten)。

代价函数，序列每一帧有预测目标对，在相应维度平均。依据张量长度(序列最大长度)归一化的tf.reduce_mean无法使用。需要按照实际序列长度归一化，手工调用tf.reduce_sum和除法运算均值。

损失函数，tf.argmax针对轴2非轴1,各帧填充，依据序列实际长度计算均值。tf.reduce_mean对批数据所有单词取均值。

TensorFlow自动导数计算，可使用序列分类相同优化运算，只需要代入新代价函数。对所有RNN梯度裁剪，防止训练发散，避免负面影响。

训练模型，get_sataset下载手写体图像，预处理，小写字母独热编码向量。随机打乱数据顺序，分偏划分训练集、测试集。

单词相邻字母存在依赖关系(或互信息)，RNN保存同一单词全部输入信息到隐含活性值。前几个字母分类，网络无大量输入推断额外信息，双向RNN(bidirectional RNN)克服缺陷。
两个RNN观测输入序列，一个按照通常顺序从左端读取单词，另一个按照相反顺序从右端读取单词。每个时间步得到两个输出活性值。送入共享softmax层前，拼接。分类器从每个字母获取完整单词信息。tf.modle.rnn.bidirectional_rnn已实现。

实现双向RNN。划分预测属性到两个函数，只关注较少内容。_shared_softmax函数，传入函数张量data推断输入尺寸。复用其他架构函数，相同扁平化技巧在所有时间步共享同一个softmax层。rnn.dynamic_rnn创建两个RNN。
序列反转，比实现新反向传递RNN运算容易。tf.reverse_sequence函数反转帧数据中sequence_lengths帧。数据流图节点有名称。scope参数是rnn_dynamic_cell变量scope名称，默认值RNN。两个参数不同RNN，需要不同域。
反转序列送入后向RNN，网络输出反转，和前向输出对齐。沿RNN神经元输出维度拼接两个张量，返回。双向RNN模型性能更优。

    import requests
    import os
    from bs4 import BeautifulSoup

    from helpers import ensure_directory

    class ArxivAbstracts:

        ENDPOINT = 'http://export.arxiv.org/api/query'
        PAGE_SIZE = 100

        def __init__(self, cache_dir, categories, keywords, amount=None):
            self.categories = categories
            self.keywords = keywords
            cache_dir = os.path.expanduser(cache_dir)
            ensure_directory(cache_dir)
            filename = os.path.join(cache_dir, 'abstracts.txt')
            if not os.path.isfile(filename):
                with open(filename, 'w') as file_:
                    for abstract in self._fetch_all(amount):
                        file_.write(abstract + '\n')
            with open(filename) as file_:
                self.data = file_.readlines()

        def _fetch_all(self, amount):
            page_size = type(self).PAGE_SIZE
            count = self._fetch_count()
            if amount:
                count = min(count, amount)
            for offset in range(0, count, page_size):
                print('Fetch papers {}/{}'.format(offset + page_size, count))
                yield from self._fetch_page(page_size, count)

        def _fetch_page(self, amount, offset):
            url = self._build_url(amount, offset)
            response = requests.get(url)
            soup = BeautifulSoup(response.text)
            for entry in soup.findAll('entry'):
                text = entry.find('summary').text
                text = text.strip().replace('\n', ' ')
                yield text

        def _fetch_count(self):
            url = self._build_url(0, 0)
            response = requests.get(url)
            soup = BeautifulSoup(response.text, 'lxml')
            count = int(soup.find('opensearch:totalresults').string)
            print(count, 'papers found')
            return count

        def _build_url(self, amount, offset):
            categories = ' OR '.join('cat:' + x for x in self.categories)
            keywords = ' OR '.join('all:' + x for x in self.keywords)
            url = type(self).ENDPOINT
            url += '?search_query=(({}) AND ({}))'.format(categories, keywords)
            url += '&max_results={}&offset={}'.format(amount, offset)
            return url

    import random
    import numpy as np

    class Preprocessing:

        VOCABULARY = \
            " $%'()+,-./0123456789:;=?ABCDEFGHIJKLMNOPQRSTUVWXYZ" \
            "\\^_abcdefghijklmnopqrstuvwxyz{|}"

        def __init__(self, texts, length, batch_size):
            self.texts = texts
            self.length = length
            self.batch_size = batch_size
            self.lookup = {x: i for i, x in enumerate(self.VOCABULARY)}

        def __call__(self, texts):
            batch = np.zeros((len(texts), self.length, len(self.VOCABULARY)))
            for index, text in enumerate(texts):
                text = [x for x in text if x in self.lookup]
                assert 2 <= len(text) <= self.length
                for offset, character in enumerate(text):
                    code = self.lookup[character]
                    batch[index, offset, code] = 1
            return batch

        def __iter__(self):
            windows = []
            for text in self.texts:
                for i in range(0, len(text) - self.length + 1, self.length // 2):
                    windows.append(text[i: i + self.length])
            assert all(len(x) == len(windows[0]) for x in windows)
            while True:
                random.shuffle(windows)
                for i in range(0, len(windows), self.batch_size):
                    batch = windows[i: i + self.batch_size]
                    yield self(batch)

    import tensorflow as tf
    from helpers import lazy_property

    class PredictiveCodingModel:

        def __init__(self, params, sequence, initial=None):
            self.params = params
            self.sequence = sequence
            self.initial = initial
            self.prediction
            self.state
            self.cost
            self.error
            self.logprob
            self.optimize

        @lazy_property
        def data(self):
            max_length = int(self.sequence.get_shape()[1])
            return tf.slice(self.sequence, (0, 0, 0), (-1, max_length - 1, -1))

        @lazy_property
        def target(self):
            return tf.slice(self.sequence, (0, 1, 0), (-1, -1, -1))

        @lazy_property
        def mask(self):
            return tf.reduce_max(tf.abs(self.target), reduction_indices=2)

        @lazy_property
        def length(self):
            return tf.reduce_sum(self.mask, reduction_indices=1)

        @lazy_property
        def prediction(self):
            prediction, _ = self.forward
            return prediction

        @lazy_property
        def state(self):
            _, state = self.forward
            return state

        @lazy_property
        def forward(self):
            cell = self.params.rnn_cell(self.params.rnn_hidden)
            cell = tf.nn.rnn_cell.MultiRNNCell([cell] * self.params.rnn_layers)
            hidden, state = tf.nn.dynamic_rnn(
                inputs=self.data,
                cell=cell,
                dtype=tf.float32,
                initial_state=self.initial,
                sequence_length=self.length)
            vocabulary_size = int(self.target.get_shape()[2])
            prediction = self._shared_softmax(hidden, vocabulary_size)
            return prediction, state

        @lazy_property
        def cost(self):
            prediction = tf.clip_by_value(self.prediction, 1e-10, 1.0)
            cost = self.target * tf.log(prediction)
            cost = -tf.reduce_sum(cost, reduction_indices=2)
            return self._average(cost)

        @lazy_property
        def error(self):
            error = tf.not_equal(
                tf.argmax(self.prediction, 2), tf.argmax(self.target, 2))
            error = tf.cast(error, tf.float32)
            return self._average(error)

        @lazy_property
        def logprob(self):
            logprob = tf.mul(self.prediction, self.target)
            logprob = tf.reduce_max(logprob, reduction_indices=2)
            logprob = tf.log(tf.clip_by_value(logprob, 1e-10, 1.0)) / tf.log(2.0)
            return self._average(logprob)

        @lazy_property
        def optimize(self):
            gradient = self.params.optimizer.compute_gradients(self.cost)
            if self.params.gradient_clipping:
                limit = self.params.gradient_clipping
                gradient = [
                    (tf.clip_by_value(g, -limit, limit), v)
                    if g is not None else (None, v)
                    for g, v in gradient]
            optimize = self.params.optimizer.apply_gradients(gradient)
            return optimize

        def _average(self, data):
            data *= self.mask
            length = tf.reduce_sum(self.length, 0)
            data = tf.reduce_sum(data, reduction_indices=1) / length
            data = tf.reduce_mean(data)
            return data

        def _shared_softmax(self, data, out_size):
            max_length = int(data.get_shape()[1])
            in_size = int(data.get_shape()[2])
            weight = tf.Variable(tf.truncated_normal(
                [in_size, out_size], stddev=0.01))
            bias = tf.Variable(tf.constant(0.1, shape=[out_size]))
            # Flatten to apply same weights to all time steps.
            flat = tf.reshape(data, [-1, in_size])
            output = tf.nn.softmax(tf.matmul(flat, weight) + bias)
            output = tf.reshape(output, [-1, max_length, out_size])
            return output

    import os
    import re
    import tensorflow as tf
    import numpy as np

    from helpers import overwrite_graph
    from helpers import ensure_directory
    from ArxivAbstracts import ArxivAbstracts
    from Preprocessing import Preprocessing
    from PredictiveCodingModel import PredictiveCodingModel

    class Training:

        @overwrite_graph
        def __init__(self, params, cache_dir, categories, keywords, amount=None):
            self.params = params
            self.texts = ArxivAbstracts(cache_dir, categories, keywords, amount).data
            self.prep = Preprocessing(
                self.texts, self.params.max_length, self.params.batch_size)
            self.sequence = tf.placeholder(
                tf.float32,
                [None, self.params.max_length, len(self.prep.VOCABULARY)])
            self.model = PredictiveCodingModel(self.params, self.sequence)
            self._init_or_load_session()

        def __call__(self):
            print('Start training')
            self.logprobs = []
            batches = iter(self.prep)
            for epoch in range(self.epoch, self.params.epochs + 1):
                self.epoch = epoch
                for _ in range(self.params.epoch_size):
                    self._optimization(next(batches))
                self._evaluation()
            return np.array(self.logprobs)

        def _optimization(self, batch):
            logprob, _ = self.sess.run(
                (self.model.logprob, self.model.optimize),
                {self.sequence: batch})
            if np.isnan(logprob):
                raise Exception('training diverged')
            self.logprobs.append(logprob)

        def _evaluation(self):
            self.saver.save(self.sess, os.path.join(
                self.params.checkpoint_dir, 'model'), self.epoch)
            self.saver.save(self.sess, os.path.join(
                self.params.checkpoint_dir, 'model'), self.epoch)
            perplexity = 2 ** -(sum(self.logprobs[-self.params.epoch_size:]) /
                            self.params.epoch_size)
            print('Epoch {:2d} perplexity {:5.4f}'.format(self.epoch, perplexity))

        def _init_or_load_session(self):
            self.sess = tf.Session()
            self.saver = tf.train.Saver()
            checkpoint = tf.train.get_checkpoint_state(self.params.checkpoint_dir)
            if checkpoint and checkpoint.model_checkpoint_path:
                path = checkpoint.model_checkpoint_path
                print('Load checkpoint', path)
                self.saver.restore(self.sess, path)
                self.epoch = int(re.search(r'-(\d+)$', path).group(1)) + 1
            else:
                ensure_directory(self.params.checkpoint_dir)
                print('Randomly initialize variables')
                self.sess.run(tf.initialize_all_variables())
                self.epoch = 1

    from Training import Training
    from get_params import get_params

    Training(
        get_params(),
        cache_dir = './arxiv',
        categories = [
            'Machine Learning',
            'Neural and Evolutionary Computing',
            'Optimization'
        ],
        keywords = [
            'neural',
            'network',
            'deep'
        ]
        )()

    import tensorflow as tf
    import numpy as np

    from helpers import overwrite_graph
    from Preprocessing import Preprocessing
    from PredictiveCodingModel import PredictiveCodingModel

    class Sampling:

        @overwrite_graph
        def __init__(self, params):
            self.params = params
            self.prep = Preprocessing([], 2, self.params.batch_size)
            self.sequence = tf.placeholder(
                tf.float32, [1, 2, len(self.prep.VOCABULARY)])
            self.state = tf.placeholder(
                tf.float32, [1, self.params.rnn_hidden * self.params.rnn_layers])
            self.model = PredictiveCodingModel(
                self.params, self.sequence, self.state)
            self.sess = tf.Session()
            checkpoint = tf.train.get_checkpoint_state(self.params.checkpoint_dir)
            if checkpoint and checkpoint.model_checkpoint_path:
                tf.train.Saver().restore(
                    self.sess, checkpoint.model_checkpoint_path)
            else:
               print('Sampling from untrained model.')
            print('Sampling temperature', self.params.sampling_temperature)

        def __call__(self, seed, length=100):
            text = seed
            state = np.zeros((1, self.params.rnn_hidden * self.params.rnn_layers))
            for _ in range(length):
                feed = {self.state: state}
                feed[self.sequence] = self.prep([text[-1] + '?'])
                prediction, state = self.sess.run(
                    [self.model.prediction, self.model.state], feed)
                text += self._sample(prediction[0, 0])
            return text

        def _sample(self, dist):
            dist = np.log(dist) / self.params.sampling_temperature
            dist = np.exp(dist) / np.exp(dist).sum()
            choice = np.random.choice(len(dist), p=dist)
            choice = self.prep.VOCABULARY[choice]
            return choice

参考资料：
《面向机器智能的TensorFlow实践》

欢迎加我微信交流：qingxingfengzi
我的微信公众号:qingxingfengzigz
我老婆张幸清的微信公众号：qingqingfeifangz

---恢复内容结束---