前一章Doc2Vec里提到，其实Doc2Vec只是通过加入Doc_id捕捉了文本的主题信息，并没有真正考虑语序以及上下文语义，n-gram只能在局部解决这一问题，那么还有别的解决方案么？依旧是通用文本向量，skip-thought尝试应用encoder-decoder来学习包含上下文信息和语序的句子向量。魔改后的实现可以看这里( ´▽｀) github-DSXiangLi-Embedding-skip_thought
Skip-Thought模型分析
Skip-Thought顾名思义是沿用了skip-gram的路子,不熟悉的童鞋看这里 无所不能的Embedding1 - Word2vec模型详解&代码实现
skip-gram是用中间词来预测周围单词，skip-Thought是用中间句子来预测前一个句子和后一个句子，模型思路就是这么简单粗暴，具体实现就涉及到句子的信息要如何提取，以及loss function的选择。作者选择了encoder-decoder来提取句子信息，用翻译模型常用的log-perplrexity作为loss。
这里想提一句不同模型，在不同的样本上，训练出的文本向量所包含的信息是不同的。例如word2vec的假设就是context（windo_size内周围词）相似的单词更相似（向量空间距离更近）。skip-thought作者对于文本向量的假设是：能更好reconstruct前后句子的信息，就是当前句子的所含信息，换言之前后句子相似的句子，文本向量的空间距离更近。
Encoder部分负责提取中间句子的信息生成定长向量output_state，Decoder则基于ouput_state进行迭代生成前（后）句子。Encoder-Decoder支持任意记忆单元，这里作者选择了GRU-GRU。
简单回顾下GRU Cell，GRU有两个Gate，从两个角度衡量历史sequence信息和当前token的相关程度，\(\Gamma_r\)控制多少历史信息参与state的重新计算是reset gate，\(\Gamma_u\)控制多少历史信息直接进入当前state是update gate，这里安利一篇博客 Illustrated Guide to LSTM’s and GRU’s: A step by step explanation
Encoder部分经过GRU把长度为T的sequence信息压缩到hidden_size的\(h^{<T>}\)，这里\(h^{<T>}\)也是最终skip-thought为每一个句子生成的通用向量表达。
Decoder部分基于\(h^{<T>}\)向前预测下一个/上一个句子中的每一个单词。Decoder比Encoder略复杂，在于训练阶段和预测阶段对于input的处理存在差异。
训练阶段使用了100%的Teacher Forcing，每个cell的输入除了上一个cell的hidden state，还有预测句子中前一个真实token对应的embedding，如图
而在预测阶段真实序列未知，因此会转而使用前一个cell的output来预测前一个token，再用预测token的embedding作为输入，如图
对于翻译模型来说，在训练阶段使用TeacherForcing的好处是可以加速模型收敛，避免向前迭代预测的误差进一步放大。坏处自然是训练和预测时decoder的表现存在差异（Exposure Bias)，以及预测时decode的output会受到训练样本的约束。这里最常用的解决方案是Scheduled Sampling, 简单来说就是在训练阶段有P的概率输入用teacher forcing，1-P的概率用预测output。但是！skip-thought并没有使用这个解决方案，为啥嘞？反转来了V(^_^)V
skip-thought的Decoder还有两点特殊：
- 前/后句子用两个decoder来训练，两个decoder除了word embedding共享之外，参数独立
- encoder state不只作为decoder的initial_state，而是直接传入decoder的每一个cell，起到类似residual/connditional的作用，避免encoder的信息传着传着给传没了。这里感觉不用attention而是直接传入outpu_state也是为了保证这个output_state能最大程度的学到用来reconstruct前后句子的信息
loss部分作者用了语言模型的log-perplexity把前后句子的loss加总得到loss function
论文比较有意思的一个点还有vocabulary expansion，就是如何把word embedding扩展到训练集之外。作者尝试用linear-mapping的方式学习word2vec和skip-thought里面word-embedding的映射关系，就是找到word2vec和skip-thought交集的word, 对他们的embedding做regression $ X_{word2vec} \sim W \cdot X_{skipthought} $，这样对样本外但是word2vec内的单词直接用W映射就能得到skip-thougt的词向量
最终在生成文本向量的时候，作者给出了几种方案，遵循大力一定出奇迹的原则自然方案3效果更好
- 2400-dim的uni-skip, 就是以上encoder生成的output state
- 两个1200-dim的bi-skip向量拼接得到2400-dim的向量，是在处理训练样本时一个用正常语序训练，一个reverse语序训练
- 把上面两个2400-dim拼接得到4800-dim
skip-thought 模型实现
这里有点任性的对论文做了魔改。。。部分细节和论文已经天差地别，可以拿来了解encoder-decoder的实现但不保证完全reproduce skip-thought的结果。。。以下只保留代码核心部分，完整代码在 github-DSXiangLi-Embedding-skip_thought。 这里用了tensorflow seq2seq的框架，不熟悉的童鞋可以先看后面seq2seq的代码解析~
dataset
论文中是\((s_{i-1}, s_i, s_{i+1})\)作为一组样本，其中\(s_i\)是encoder source，\(s_{i-1}\)和\(s_{i+1}\)是decoder target，这里我直接处理成\((s_i,s_{i-1})\),\((s_i,s_{i+1})\)两组样本。
其中encoder source不需要多做处理，但是decoder source在Train和Eval时需要在sequence前后加入start和end_token标记序列的开始和结束，在Predict时需要加入start_token标记开始。最后通过word_table把token映射到token_id，再Padding到相同长度就齐活。
这里在Dataset的部分加入了获取word2vec embedding的部分, word2vec以外的单词默认random.uniform(-0.1,0.1)
class SkipThoughtDataset(BaseDataset):
def __init__(self, data_file, dict_file, epochs, batch_size, buffer_size, min_count, max_count,
special_token, max_len):
...
def parse_example(self, line, prepend, append):
features = {}
tokens = tf.string_split([tf.string_strip(line)]).values
if prepend:
tokens = tf.concat([[self.special_token.SEQ_START], tokens], 0)
if append:
tokens = tf.concat([tokens, [self.special_token.SEQ_END]], 0)
features['tokens'] = tokens
features['seq_len'] = tf.size(tokens)
return features
...
def make_source_dataset(self, file_path, data_type, is_predict, word_table_func):
prepend, append = self.prepend_append_logic(data_type, is_predict)
dataset = tf.data.TextLineDataset(file_path).\
map(lambda x: self.parse_example(x, prepend, append), num_parallel_calls=tf.data.experimental.AUTOTUNE).\
map(lambda x: word_table_func(x), num_parallel_calls=tf.data.experimental.AUTOTUNE)
return dataset
def build_dataset(self, is_predict=0):
def input_fn():
word_table_func = self.word_table_lookup(self.build_wordtable())
_ = self.build_tokentable() # initialize here to ensure lookup table is in the same graph
encoder_source = self.make_source_dataset(self.data_file['encoder'], 'encoder', is_predict, word_table_func)
decoder_source = self.make_source_dataset(self.data_file['decoder'], 'decoder', is_predict, word_table_func)
dataset = tf.data.Dataset.zip((encoder_source, decoder_source)).\
filter(self.sample_filter_logic)
if not is_predict:
dataset = dataset.\
repeat(self.epochs)
dataset = dataset. \
padded_batch( batch_size=self.batch_size,
padded_shapes=self.padded_shape,
padding_values=self.padding_values,
drop_remainder=True ). \
prefetch( tf.data.experimental.AUTOTUNE )
else:
dataset = dataset.batch(1)
return dataset
return input_fn
def load_pretrain_embedding(self):
if self.embedding is None:
word_vector = gensim.downloader.load(PretrainModel)
embedding = []
for i in self._dictionary.keys():
try:
embedding.append( word_vector.get_vector( i ) )
except KeyError:
embedding.append( np.random.uniform(low=-0.1, high=0.1, size=300))
self.embedding = np.array(embedding, dtype=np.float32)
return self.embedding
Encoder-Decoder
Encoder的部分很常规，确认cell类型，然后经过dynamic_rnn迭代，输出output和state
def gru_encoder(input_emb, input_len, params):
gru_cell = build_rnn_cell('gru', params)
# state: batch_size * hidden_size, output: batch_size * max_len * hidden_size
output, state = tf.nn.dynamic_rnn(
cell=gru_cell, # one rnn units
inputs=input_emb, # batch_size * max_len * feature_size
sequence_length=input_len, # batch_size * seq_len
initial_state=None,
dtype=params['dtype'],
time_major=False # whether reshape max_length to first dim
)
return ENCODER_OUTPUT(output=output, state=state)
Decoder的部分可以分成helper, decoder, 以及最终dynamic_decode的部分。比较容易踩坑的有几个点
- dynamic_decode部分的max_iteration在训练时可以不设定，train_helper会基于seq_len判断finish，在Eval时因为用GreedyEmbeddingHelper所以需要手动传入pad_len，在predict部分只需给定最大max_iter保证预测一定会停止就好
- Decoder的output layer必须要有, 因为需要做hidden_size -> vocab_size(softmax)的转换，用于预测每个cell的token输出
- 前面dataset的decoder_source我们在前后都加了开始/停止token，训练时需要移除最后一个token(对于trainHelper可以改inputs也可以改sequence_length), 这样在计算loss时可以和移除第一个token的target对齐。
这里针对上面提到的把encoder的output_state直接传入每个decoder cell做了实现，直接把encoder state和embedding input做了拼接作为输入。
def get_helper(encoder_output, input_emb, input_len, batch_size, embedding, mode, params):
if mode == tf.estimator.ModeKeys.TRAIN:
if params['conditional']:
# conditional train helper with encoder output state as direct input
# Reshape encoder state as auxiliary input: 1* batch_size * hidden -> batch_size * max_len * hidden
decoder_length = tf.shape(input_emb)[1]
state_shape = tf.shape(encoder_output.state)
encoder_state = tf.tile(tf.reshape(encoder_output.state, [state_shape[1],
state_shape[0],
state_shape[2]]),
[1, decoder_length, 1])
input_emb = tf.concat([encoder_state, input_emb], axis=-1)
helper = seq2seq.TrainingHelper( inputs=input_emb, # batch_size * max_len-1 * emb_size
sequence_length=input_len-1, # exclude last token
time_major=False,
name='training_helper' )
else:
helper = seq2seq.GreedyEmbeddingHelper( embedding=embedding_func( embedding ),
start_tokens=tf.fill([batch_size], params['start_token']),
end_token=params['end_token'] )
return helper
def get_decoder(decoder_cell, encoder_output, input_emb, input_len, embedding, output_layer, mode, params):
batch_size = tf.shape(encoder_output.output)[0]
if params['beam_width'] >1 :
# If beam search multiple prediction are uesd at each time step
decoder = seq2seq.BeamSearchDecoder( cell=decoder_cell,
embedding=embedding_func( embedding ),
initial_state=encoder_output,
beam_width=params['beam_width'],
start_tokens=tf.fill([batch_size], params['start_token']),
end_token=params['end_token'],
output_layer=output_layer )
else:
helper = get_helper(encoder_output, input_emb, input_len, batch_size, embedding, mode, params)
decoder = seq2seq.BasicDecoder( cell=decoder_cell,
helper=helper,
initial_state=encoder_output.state,
output_layer=output_layer )
return decoder
def gru_decoder(encoder_output, input_emb, input_len, embedding, params, mode):
gru_cell = build_rnn_cell( 'gru', params )
if mode == tf.estimator.ModeKeys.TRAIN:
max_iteration = None
elif mode == tf.estimator.ModeKeys.EVAL:
max_iteration = tf.reduce_max(input_len) # decode max sequence length(=padded_length)in EVAL
else:
max_iteration = params['max_decode_iter'] # decode pre-defined max_decode iter in predict
output_layer=tf.layers.Dense(units=params['vocab_size']) # used for infer helper sample or train loss calculation
decoder = get_decoder(gru_cell, encoder_output, input_emb, input_len, embedding, output_layer, mode, params)
output, state, seq_len = seq2seq.dynamic_decode(decoder=decoder,
output_time_major=False,
impute_finished=True,
maximum_iterations=max_iteration)
return DECODER_OUTPUT(output=output, state = state, seq_len=seq_len)
loss
loss这了自己实现的一版sequence_loss，把计算loss和按不同维度聚合拆成了两块。感觉tf.sequence_loss只针对train,对eval的部分并不友好，因为trainHelper可以保证source和target的长度一致，但是infer时调用GreedyEmbeddingHelper是无法保证输出长度的(不知道是不是我哪里理解错了，如果是请大神指正(o^^o)), 所以对eval部分也做了特殊处理。
def sequence_loss(logits, target, mask, mode):
with tf.variable_scope('Sequence_loss_matrix'):
n_class = tf.shape(logits)[2]
decode_len = tf.shape(logits)[1] # used for infer only, max_len is determined by decoder
logits = tf.reshape(logits, [-1, n_class])
if mode == tf.estimator.ModeKeys.TRAIN:
# In train, target
target = tf.reshape(target[:, 1:], [-1]) # (batch * (padded_len-1)) * 1
elif mode == tf.estimator.ModeKeys.EVAL:
# In eval, target has paded_len, logits have decode_len
target = tf.reshape(target[:, : decode_len], [-1]) # batch * (decode_len) *1
else:
raise Exception('sequence loss is only used in train or eval, not in pure prediction')
loss_mat = tf.nn.sparse_softmax_cross_entropy_with_logits(labels = target, logits = logits)
loss_mat = tf.multiply(loss_mat, tf.reshape(mask, [-1])) # apply padded mask on output loss
return loss_mat
def agg_sequence_loss(loss_mat, mask, axis):
with tf.variable_scope('Loss_{}'.format(axis)):
if axis == 'scaler':
loss = tf.reduce_sum(loss_mat)
n_sample = tf.reduce_sum(mask)
loss = loss/n_sample
else:
loss_mat = tf.reshape(loss_mat, tf.shape(mask)) # (batch_size * max_len) * 1-> batch_size * max_len
if axis == 'batch':
loss = tf.reduce_sum(loss_mat, axis=1) # batch
n_sample = tf.reduce_sum(mask, axis=1) # batch
loss = tf.math.divide_no_nan(loss, n_sample) # batch
elif axis == 'time':
loss = tf.reduce_sum(loss_mat, axis=0) # max_len
n_sample = tf.reduce_sum(mask, axis=0) # max_len
loss = tf.math.divide_no_nan(loss, n_sample) # max_len
else:
raise Exception('Only scaler/batch/time are supported in axis param')
return loss
model
encoder, decoder, loss都ready，拼一块就齐活了, 这里embedding我们用了前面加载的word2vec来进行初始化。
class QuickThought(object):
def __init__(self, params):
self.params = params
self.init()
def init(self):
with tf.variable_scope('embedding', reuse=tf.AUTO_REUSE):
self.embedding = tf.get_variable(dtype = self.params['dtype'],
initializer=tf.constant(self.params['pretrain_embedding']),
name='word_embedding' )
add_layer_summary(self.embedding.name, self.embedding)
def build_model(self, features, labels, mode):
encoder_output = self._encode(features)
decoder_output = self._decode(encoder_output, labels, mode )
loss_output = self.compute_loss( decoder_output, labels, mode )
...
def _encode(self, features):
with tf.variable_scope('encoding'):
encoder = ENCODER_FAMILY[self.params['encoder_type']]
seq_emb_input = tf.nn.embedding_lookup(self.embedding, features['tokens']) # batch_size * max_len * emb_size
encoder_output = encoder(seq_emb_input, features['seq_len'], self.params) # batch_size
return encoder_output
def _decode(self, encoder_output, labels, mode):
with tf.variable_scope('decoding'):
decoder = DECODER_FAMILY[self.params['decoder_type']]
if mode == tf.estimator.ModeKeys.TRAIN:
seq_emb_output = tf.nn.embedding_lookup(self.embedding, labels['tokens']) # batch_size * max_len * emb_size
input_len = labels['seq_len']
elif mode == tf.estimator.ModeKeys.EVAL:
seq_emb_output = None
input_len = labels['seq_len']
else:
seq_emb_output = None
input_len = None
decoder_output = decoder(encoder_output, seq_emb_output, input_len,\
self.embedding, self.params, mode)
return decoder_output
def compute_loss(self, decoder_output, labels, mode):
with tf.variable_scope('compute_loss'):
mask = sequence_mask(decoder_output, labels, self.params, mode)
loss_mat = sequence_loss(logits=decoder_output.output.rnn_output,
target=labels['tokens'],
mask=mask,
mode=mode)
loss = []
for axis in ['scaler', 'batch', 'time']:
loss.append(agg_sequence_loss(loss_mat, mask, axis))
return SEQ_LOSS_OUTPUT(loss_id=loss_mat,
loss_scaler=loss[0],
loss_per_batch=loss[1],
loss_per_time=loss[2])
tf.seq2seq 代码解析
稀里糊涂开始用seq2seq，结果盯着shape mismatch的报错险些看到地老天荒，索性咱老老实实看一遍tf的实现, 以下代码只保留了核心部分，完整的官方代码在这里哟 tf.seq2seq.contrib
Encoding
Encoding部分就是一个dynamic_rnn，先看下输入
- cell：任意类型记忆单元rnn,gru, lstm
- inputs：rnn输入，一般是[batch_size, max_len, emb_siz] 也就是padding的序列token,经过embedding映射之后作为输入
- sequence_length: 真实序列长度(不包含padding)，用于判断序列是遍历完
- initial_state: encoder最初state，None则默认是zero_state
- dtype: output数据类型，建议全局设置统一数据类型，不然会有各种mismatch，不要问我是怎么知道的>.<
- parallel_iteration：内存换速度，没有上下文依赖的op进行并行计算
- time_major：如果你的输入数据是[max_len, batch_size，emb_siz]则为True，一般为False在dynamic_rnn内部再做reshape。
dynamic_rnn主函数其实只做了输入/输出数据的处理部分，包括
- reshape_input：对应上面time_major=False, 把输入数据从[batch_size, max_len, emb_siz]转换为[max_len, batch_size，emb_siz]
- inital_state: 默认是batch size的zero_state
- reshape_output: output输出是[max_len, batch_size，hidden_siz]转换为[batch_size, max_len, hidden_size]
def dynamic_rnn(cell, inputs, sequence_length=None, initial_state=None,
dtype=None, parallel_iterations=None, swap_memory=False,
time_major=False, scope=None):
flat_input = nest.flatten(inputs)
if not time_major:
flat_input = [ops.convert_to_tensor(input_) for input_ in flat_input]
flat_input = tuple(_transpose_batch_time(input_) for input_ in flat_input)
batch_size = _best_effort_input_batch_size(flat_input)
state = cell.zero_state(batch_size, dtype)
inputs = nest.pack_sequence_as(structure=inputs, flat_sequence=flat_input)
(outputs, final_state) = _dynamic_rnn_loop(
cell,
inputs,
state,
parallel_iterations=parallel_iterations,
swap_memory=swap_memory,
sequence_length=sequence_length,
dtype=dtype)
if not time_major:
# (T,B,D) => (B,T,D)
outputs = nest.map_structure(_transpose_batch_time, outputs)
return (outputs, final_state)
核心计算部分都在_dynamic_rnn_loop，是一个while_loop, 所以需要定义三要素[loop_var, body, condition]
- loop_var：(time, output_ta, state)
- time:遍历到第几个token
- output_ta: 每个cell的输出，padding部分是zero-output
- state: 最后一个cell的输出，对于padding的序列，只计算到最后一个真实token，之后的state是直接copy through
这里output_ta的shape是(batch, max_len, hidden_units), 对于rnn和GRU，state就是最后一个output, 那shape自然是(1, batch, hidden_units), 但LSTM是有两个hidden state的，一个用于向前传递信息一个用于输出，所以这里state的shape会是（2, batch, hidden_units)
- body
loop的核心计算部分是lambda: cell(input_t, state)，也就是相应记忆单元的计算。当sequence_length给定时，_rnn_step的额外操作其实是对已经遍历完的序列直接copy through（zero_output, last_state)
def _time_step(time, output_ta_t, state):
input_t = tuple(ta.read(time) for ta in input_ta)
input_t = nest.pack_sequence_as(structure=inputs, flat_sequence=input_t)
call_cell = lambda: cell(input_t, state)
if sequence_length is not None:
(output, new_state) = _rnn_step(
time=time,
sequence_length=sequence_length,
min_sequence_length=min_sequence_length,
max_sequence_length=max_sequence_length,
zero_output=zero_output,
state=state,
call_cell=call_cell,
state_size=state_size,
skip_conditionals=True)
else:
(output, new_state) = call_cell()
# Pack state if using state tuples
output = nest.flatten(output)
output_ta_t = tuple(ta.write(time, out) for ta, out in zip(output_ta_t, output))
return (time + 1, output_ta_t, new_state)
- condition
停止loop的条件loop_bound=min(max_sequence_length, max(1,time_steps) , 其中time_step是输入的max_len维度，也就是padding length, max_sequence_length是输入batch的最大真实长度，如果是batch_padding这两个取值应该是一样的
time_steps = input_shape[0]
if sequence_length is not None:
min_sequence_length = math_ops.reduce_min(sequence_length)
max_sequence_length = math_ops.reduce_max(sequence_length)
else:
max_sequence_length = time_steps
loop_bound = math_ops.minimum(time_steps, math_ops.maximum(1, max_sequence_length))
_, output_final_ta, final_state = control_flow_ops.while_loop(
cond=lambda time, *_: time < loop_bound,
body=_time_step,
loop_vars=(time, output_ta, state),
parallel_iterations=parallel_iterations,
maximum_iterations=time_steps,
swap_memory=swap_memory)
Decoding
Decoding主要有三个组件，Decoder，Helper和dynamic_decode。还有比较特殊独立出来的BeamSearch和Attention，这两个后面用到再说
BasicDecoder
BasicDecoder主要接口有2个
- initialize生成decode阶段的最初input
- step实现每一步decode的计算,之后被dynamic_decode的while_loop调用
其中initialize拼接了helper的初始化返回再加上initial_state，也就是encoder最后一步的output_state，helper返回的部分我们放在后面说。
def initialize(self, name=None):
return self._helper.initialize() + (self._initial_state,)
step部分做了如下操作
- 输入上一步的output, state计算下一步的output，这是Decoder的核心计算
- 如果定义了output_layer，对output做transform，为啥需要output_layer嘞? 这个看到Helper你就明白了
- sample, next_inputs: 都是调用Helper的接口
- 输出: BasicDecoderOutput(rnn_output, sample_id), next_state, next_inputs, finished
class BasicDecoderOutput(
collections.namedtuple("BasicDecoderOutput", ("rnn_output", "sample_id"))):
pass
class BasicDecoder(decoder.Decoder):
"""Basic sampling decoder."""
def __init__(self, cell, helper, initial_state, output_layer=None):
def step(self, time, inputs, state, name=None):
with ops.name_scope(name, "BasicDecoderStep", (time, inputs, state)):
cell_outputs, cell_state = self._cell(inputs, state)
if self._output_layer is not None:
cell_outputs = self._output_layer(cell_outputs)
sample_ids = self._helper.sample(
time=time, outputs=cell_outputs, state=cell_state)
(finished, next_inputs, next_state) = self._helper.next_inputs(
time=time,
outputs=cell_outputs,
state=cell_state,
sample_ids=sample_ids)
outputs = BasicDecoderOutput(cell_outputs, sample_ids)
return (outputs, next_state, next_inputs, finished)
这里发现BasicDecoder的实现只包括了承上的部分，启下的部分都放在了Helper里面，下面我们具体看下Helper的next_input和Sample接口干了啥
Helper
我们主要看两个helper一个用于训练，一个用于预测，主要实现3个接口
- initialize：生成decode阶段的最初input
- sample：生成decode下一步的input id
- next_inputs：生成decode下一步的input
TrainHelper用于训练，sample接口实际并没有用，next_input把sample_id定义为unused_kwargs.
- initialize返回 (finished, next_inputs)
- finished: 判断当前batch每个sequence是否已经遍历完, sequence_length是不包含padded的实际sequencec长度
- 除非batch里所有seq_length的长度都是0，否则直接读取每个sequence的第一个token作为decoder的初始输入
- next_inputs输出(finished, next_inputs, state)
- finished: 判断当前batch每个sequence是否已经遍历完, sequence_length是不包含padded的实际sequencec长度
- next_inputs: 训练时使用Teaching Force，传入下一个decoder cell的就是前一个位置的实际token embedding，所以这里next_input直接读取input sequence的下一个值，如果finished都是True就返回0【其实返回啥都无所谓因为在loss那里padded的部分会被mask掉】
- state: 这里是打酱油的，直接pass-throuh
class TrainingHelper(Helper):
def __init__(self, inputs, sequence_length, time_major=False, name=None):
...
def initialize(self, name=None):
with ops.name_scope(name, "TrainingHelperInitialize"):
finished = math_ops.equal(0, self._sequence_length)
all_finished = math_ops.reduce_all(finished)
next_inputs = control_flow_ops.cond(
all_finished, lambda: self._zero_inputs,
lambda: nest.map_structure(lambda inp: inp.read(0), self._input_tas))
return (finished, next_inputs)
def next_inputs(self, time, outputs, state, name=None, **unused_kwargs):
"""next_inputs_fn for TrainingHelper."""
with ops.name_scope(name, "TrainingHelperNextInputs",
[time, outputs, state]):
next_time = time + 1
finished = (next_time >= self._sequence_length)
all_finished = math_ops.reduce_all(finished)
def read_from_ta(inp):
return inp.read(next_time)
next_inputs = control_flow_ops.cond(
all_finished, lambda: self._zero_inputs,
lambda: nest.map_structure(read_from_ta, self._input_tas))
return (finished, next_inputs, state)
GreedyHelper用于预测
initialize返回 (finished, next_inputs)
- finished: 都是False，因为infer阶段不用判断input_sequence长度
- next_inputs: 返回start_token对应的embedding，和训练保持一致
sample返回sample_id
负责根据每个decoder cell的output计算出现概率最大的token，作为下一个decoder cell的输入，这里也是上面提到需要output_layer的原因，因为需要hidden_size -> vocab_size的变换，才能进一步计算softmax
- next_input
- finished: 如果sequence预测为end_token则该sequence预测完成，判断batch里所有sequence是否预测完成
- next_inputs: 对sample_id做embedding_lookup作为下一步的输入，如果finished都是True就返回start_token
- state: 继续打酱油
class GreedyEmbeddingHelper(Helper):
def __init__(self, embedding, start_tokens, end_token):
self._start_tokens = ops.convert_to_tensor(
start_tokens, dtype=dtypes.int32, name="start_tokens")
self._end_token = ops.convert_to_tensor(
end_token, dtype=dtypes.int32, name="end_token")
self._start_inputs = self._embedding_fn(self._start_tokens)
。。。
def sample(self, time, outputs, state, name=None):
sample_ids = math_ops.cast(
math_ops.argmax(outputs, axis=-1), dtypes.int32)
return sample_ids
def initialize(self, name=None):
finished = array_ops.tile([False], [self._batch_size])
return (finished, self._start_inputs)
def next_inputs(self, time, outputs, state, sample_ids, name=None):
finished = math_ops.equal(sample_ids, self._end_token)
all_finished = math_ops.reduce_all(finished)
next_inputs = control_flow_ops.cond(
all_finished,
lambda: self._start_inputs,
lambda: self._embedding_fn(sample_ids))
return (finished, next_inputs, state)
dynamic_decode
承上启下的工具都齐活了，要实现对sequence的预测，只剩下一步就是loop，于是有了dynamic_decode，它其实就干了个while_loop的活，于是还是loop三兄弟[loop_vars, condition, body]
loop_vars=[initial_time, initial_outputs_ta, initial_state, initial_inputs, initial_finished, initial_sequence_lengths]
- initial_finished, initial_inputs, initial_state是上面decoder的initialize返回
- initial_time, initial_sequennce=0
- initial_output_ta是每个elemennt都是batch * decoder.output_size的不定长TensorArray, 这里output_size=(rnn_output_size,sample_id_shape),预测返回1个token的sample_id_shape都是scaler, 有output_layer时rnn_output_size=output_layer_size, default= hidden_size
condition: 判断是否所有finished都为True，都遍历完则停止loop
body: loop的核心计算逻辑
step：调用Decoder进行每一步的decode计算
finished: 这里finished主要由三个逻辑判断（tracks_own_finished我没用过先忽略了哈哈）其余两个是：
- helper的next_inputs回传的finished：trainHelper判断输入sequence是否遍历完，GreedyEmbeddingHelper判断预测token是否为end_token。
- max_iteration：只用于预测，为了避免预测token一直不是end_token导致预测无限循环下去，设置一个最大预测长度，训练时max_iteraion应该为空
sequence_length: 记录实际预测sequence长度，没有finished的sequence+1
impute_finished: 如果sequence已遍历完, 后面的output补0，后面的state不再计算直接pass through当前state
def body(time, outputs_ta, state, inputs, finished, sequence_lengths):
(next_outputs, decoder_state, next_inputs,
decoder_finished) = decoder.step(time, inputs, state)
if maximum_iterations is not None:
next_finished = math_ops.logical_or(
next_finished, time + 1 >= maximum_iterations)
next_sequence_lengths = array_ops.where(
math_ops.logical_and(math_ops.logical_not(finished), next_finished),
array_ops.fill(array_ops.shape(sequence_lengths), time + 1),
sequence_lengths)
# Zero out output values past finish
if impute_finished:
emit = nest.map_structure(
lambda out, zero: array_ops.where(finished, zero, out),
next_outputs,
zero_outputs)
else:
emit = next_outputs
# Copy through states past finish
def _maybe_copy_state(new, cur):
# TensorArrays and scalar states get passed through.
if isinstance(cur, tensor_array_ops.TensorArray):
pass_through = True
else:
new.set_shape(cur.shape)
pass_through = (new.shape.ndims == 0)
return new if pass_through else array_ops.where(finished, cur, new)
if impute_finished:
next_state = nest.map_structure(
_maybe_copy_state, decoder_state, state)
else:
next_state = decoder_state
outputs_ta = nest.map_structure(lambda ta, out: ta.write(time, out),
outputs_ta, emit)
return (time + 1, outputs_ta, next_state, next_inputs, next_finished,
next_sequence_lengths)
欢迎留言吐槽以及评论哟～
Reference
skip-thought
- Ryan Kiros, yukun Zhu. 2015. SKip-Thought Vectors
- Lajanugen logeswaran, Honglak Lee. 2018. An Efficient Framework for Learning Sentence Representations.
- https://zhuanlan.zhihu.com/p/64342563
- https://towardsdatascience.com/document-embedding-techniques-fed3e7a6a25d
- https://towardsdatascience.com/the-best-document-similarity-algorithm-in-2020-a-beginners-guide-a01b9ef8cf05
- https://blog.floydhub.com/when-the-best-nlp-model-is-not-the-best-choice/
- https://towardsdatascience.com/document-embedding-techniques-fed3e7a6a25d
- https://zhuanlan.zhihu.com/p/72575806
- Scheduled Sampling for Sequence Prediction with Recurrent Neural Networks