def call(self, x, y, mask, training=False):
self.step += 1
x_ = x
x = dropout(x, keep_prob=self.keep_prob, training=training)
y = dropout(y, keep_prob=self.keep_prob, training=training)
if self.step == 0:
if not self.identity:
self.linear = layers.Dense(melt.get_shape(x, -1),
activation=tf.nn.relu)
else:
self.linear = None
# NOTICE shared linear!
if self.linear is not None:
x = self.linear(x)
y = self.linear(y)
scores = tf.matmul(x, tf.transpose(y, [0, 2, 1]))
if mask is not None:
JX = melt.get_shape(x, 1)
mask = tf.tile(tf.expand_dims(mask, axis=1), [1, JX, 1])
scores = softmax_mask(scores, mask)
alpha = tf.nn.softmax(scores)
self.alpha = alpha
y = tf.matmul(alpha, y)
if self.combine is None:
return y
else:
return self.combine(x_, y, training=training)
def call(self, x, training=False):
x = x['comment']
batch_size = melt.get_shape(x, 0)
length = melt.length(x)
#with tf.device('/cpu:0'):
x = self.embedding(x)
num_units = [
melt.get_shape(x, -1) if layer == 0 else 2 * self.num_units
for layer in range(self.num_layers)
]
#print('----------------length', tf.reduce_max(length), inputs.comment.shape)
mask_fws = [
melt.dropout(tf.ones([batch_size, 1, num_units[layer]],
dtype=tf.float32),
keep_prob=self.keep_prob,
training=training,
mode=None) for layer in range(self.num_layers)
]
mask_bws = [
melt.dropout(tf.ones([batch_size, 1, num_units[layer]],
dtype=tf.float32),
keep_prob=self.keep_prob,
training=training,
mode=None) for layer in range(self.num_layers)
]
#x = self.encode(x, length, mask_fws=mask_fws, mask_bws=mask_bws)
x = self.encode(x)
x = self.pooling(x, length)
#x = self.pooling(x)
x = self.logits(x)
return x
def call(self, x, y, training=False):
self.step += 1
if melt.get_shape(x, -1) != melt.get_shape(y, -1):
if self.step == 0:
self.dense = layers.Dense(melt.get_shape(x, -1),
activation=None,
name='sfu_dense')
y = self.dense(y)
return self.sfu(x, [y, x * y, x - y], training=training)
def call(self, input, training=False):
x1 = input['query']
x2 = input['passage']
length1 = melt.length(x1)
length2 = melt.length(x2)
#with tf.device('/cpu:0'):
x1 = self.embedding(x1)
x2 = self.embedding(x2)
x = x1
batch_size = melt.get_shape(x1, 0)
num_units = [
melt.get_shape(x, -1) if layer == 0 else 2 * self.num_units
for layer in range(self.num_layers)
]
#print('----------------length', tf.reduce_max(length), inputs.comment.shape)
mask_fws = [
melt.dropout(tf.ones([batch_size, 1, num_units[layer]],
dtype=tf.float32),
keep_prob=self.keep_prob,
training=training,
mode=None) for layer in range(self.num_layers)
]
mask_bws = [
melt.dropout(tf.ones([batch_size, 1, num_units[layer]],
dtype=tf.float32),
keep_prob=self.keep_prob,
training=training,
mode=None) for layer in range(self.num_layers)
]
x = self.encode(x1,
length1,
x2,
length2,
mask_fws=mask_fws,
mask_bws=mask_bws)
x = self.pooling(x, length1, length2)
#x = self.pooling(x)
if FLAGS.use_type:
x = tf.concat([x, tf.expand_dims(tf.to_float(input['type']), 1)],
1)
if not FLAGS.split_type:
x = self.logits(x)
else:
x1 = self.logits(x)
x2 = self.logits2(x)
x = tf.cond(tf.cast(input['type'] == 0, tf.bool), lambda:
(x1 + x2) / 2., lambda: x2)
return x
def call(self, x):
self.step += 1
if self.step == 0:
self.dense = layers.Dense(
melt.get_shape(x, -1) * self.ratio, self.activation,
self.use_bais)
return self.dense(x)
def encode(self, seq, seq_len=None, output_method='all'):
with tf.variable_scope(self.scope):
num_filters = self.num_units
seqs = [seq]
batch_size = melt.get_batch_size(seq)
kernel_sizes = [3, 5, 7, 9, 11, 13]
#kernel_sizes = [3] * 7
assert self.num_layers <= len(kernel_sizes)
for layer in range(self.num_layers):
input_size_ = melt.get_shape(seq, -1) if layer == 0 else num_filters
seq = melt.dropout(seq, self.keep_prob, self.is_train)
seq = tf.layers.conv1d(seqs[-1], num_filters, kernel_size=kernel_sizes[layer], padding='same', activation=tf.nn.relu)
# mask = melt.dropout(tf.ones([batch_size, 1, input_size_], dtype=tf.float32),
# keep_prob=self.keep_prob, is_train=self.is_train, mode=None)
#seq = tf.layers.conv1d(seqs[-1] * mask, num_filters, kernel_size=3, padding='same', activation=tf.nn.relu)
#seq = tf.layers.conv1d(seqs[-1] * mask, num_filters, kernel_size=kernel_sizes[layer], padding='same', activation=tf.nn.relu)
# if self.is_train and self.keep_prob < 1:
# seq = tf.nn.dropout(seq, self.keep_prob)
#seq = melt.layers.batch_norm(seq, self.is_train, name='layer_%d' % layer)
seqs.append(seq)
outputs = tf.concat(seqs[1:], 2)
# not do any dropout in convet just dropout outside
# if self.is_train and self.keep_prob < 1:
# outputs = tf.nn.dropout(outputs, self.keep_prob)
# compact for rnn with sate return
return melt.rnn.encode_outputs(outputs, seq_len, output_method)
def call(self, input, training=False):
x = input['rcontent'] if FLAGS.rcontent else input['content']
#print(x.shape)
batch_size = melt.get_shape(x, 0)
length = melt.length(x)
#with tf.device('/cpu:0'):
x = self.embedding(x)
x = self.encode(x, length, training=training)
# must mask pooling when eval ? but seems much worse result
#if not FLAGS.mask_pooling and training:
if not FLAGS.mask_pooling:
length = None
x = self.pooling(x, length)
if FLAGS.use_type:
x = tf.concat([x, tf.expand_dims(tf.to_float(input['type']), 1)],
1)
if not FLAGS.split_type:
x = self.logits(x)
else:
x1 = self.logits(x)
x2 = self.logits2(x)
x = tf.cond(tf.cast(input['type'] == 0, tf.bool), lambda:
(x1 + x2) / 2., lambda: x2)
return x
def aug(x, x_mask):
# print('---------------do unk aug')
# print('ori_x....', x)
if x_mask is None:
x_mask = x > 0
x_mask = tf.to_int64(x_mask)
ratio = tf.random_uniform([
1,
], 0, FLAGS.unk_aug_max_ratio)
mask = tf.random_uniform(
[melt.get_shape(x, 0),
melt.get_shape(x, 1)]) > ratio
mask = tf.to_int64(mask)
rmask = FLAGS.unk_id * (1 - mask)
x = (x * mask + rmask) * x_mask
#print('aug_x....', x)
return x
def roc_auc_scores(y_pred, y_true):
num_classes = melt.get_shape(y_pred, -1)
y_preds = tf.split(y_pred, num_classes, axis=1)
y_trues = tf.split(y_true, num_classes, axis=1)
losses = []
for y_pred, y_true in zip(y_preds, y_trues):
losses.append(roc_auc_score(y_pred, y_true))
#losses.append(art_loss(y_pred, y_true))
return tf.stack(losses)
def call(self, input, training=False):
x = input['content']
x = self.unk_aug(x, training=training)
c_mask = tf.cast(x, tf.bool)
batch_size = melt.get_shape(x, 0)
c_len, max_c_len = melt.length2(x)
if FLAGS.rnn_no_padding:
logging.info('------------------no padding! train or eval')
#c_len = tf.ones([batch_size], dtype=x.dtype) * tf.cast(melt.get_shape(x, -1), x.dtype)
c_len = max_c_len
x = self.encode(input, c_len, max_c_len, training=training)
# not help
if self.hier_encode is not None:
x = self.hier_encode(x, c_len)
# yes just using label emb..
label_emb = self.label_embedding(None)
label_seq = tf.tile(tf.expand_dims(label_emb, 0), [batch_size, 1, 1])
label_seq = self.label_encode(label_seq,
tf.ones([batch_size], dtype=tf.int64) *
self.label_emb_height,
training=training)
for i in range(FLAGS.hop):
x = self.att_dot_attentions[i](
x,
label_seq,
mask=tf.ones([batch_size, self.label_emb_height], tf.bool),
training=training)
x = self.att_encodes[i](x, c_len, training=training)
x = self.match_dot_attentions[i](x, mask=c_mask, training=training)
x = self.match_encodes[i](x, c_len, training=training)
x = self.pooling(x, c_len, calc_word_scores=self.debug)
#x = self.pooling(x)
# not help much
if self.dense is not None:
x = self.dense(x)
x = self.dropout(x, training=training)
if not FLAGS.use_label_emb:
x = self.logits(x)
else:
x = self.label_dense(x)
# TODO..
x = melt.dot(x, self.label_embedding(None))
x = tf.reshape(x, [batch_size, NUM_ATTRIBUTES, self.num_classes])
return x
def call(self, x, y, training=False):
self.step += 1
#with tf.variable_scope(self.scope):
res = tf.concat([x, y], axis=2)
dim = melt.get_shape(res, -1)
d_res = dropout(res, keep_prob=self.keep_prob, training=training)
if self.step == 0:
self.dense = layers.Dense(dim,
use_bias=False,
activation=tf.nn.sigmoid)
gate = self.dense(d_res)
return res * gate
def call(self, x, mask, training=False):
self.step += 1
x_ = x
x = dropout(x, keep_prob=self.keep_prob, training=training)
if self.step == 0:
if not self.identity:
self.linear = layers.Dense(melt.get_shape(x, -1),
activation=tf.nn.relu)
else:
self.linear = None
# NOTICE shared linear!
if self.linear is not None:
x = self.linear(x)
scores = tf.matmul(x, tf.transpose(x, [0, 2, 1]))
# x = tf.constant([[[1,2,3], [4,5,6],[7,8,9]],[[1,2,3],[4,5,6],[7,8,9]]], dtype=tf.float32) # shape=(2, 3, 3)
# z = tf.matrix_set_diag(x, tf.zeros([2, 3]))
if not self.diag:
# TODO better dim
dim0 = melt.get_shape(scores, 0)
dim1 = melt.get_shape(scores, 1)
scores = tf.matrix_set_diag(scores, tf.zeros([dim0, dim1]))
if mask is not None:
JX = melt.get_shape(x, 1)
mask = tf.tile(tf.expand_dims(mask, axis=1), [1, JX, 1])
scores = softmax_mask(scores, mask)
alpha = tf.nn.softmax(scores)
self.alpha = alpha
x = tf.matmul(alpha, x)
if self.combine is None:
return y
else:
return self.combine(x_, x, training=training)
def call(self, input, training=False):
q = input['query']
c = input['passage']
q_len = melt.length(q)
c_len = melt.length(c)
q_mask = tf.cast(q, tf.bool)
q_emb = self.embedding(q)
c_emb = self.embedding(c)
x = c_emb
batch_size = melt.get_shape(x, 0)
num_units = [melt.get_shape(x, -1) if layer == 0 else 2 * self.num_units for layer in range(self.num_layers)]
mask_fws = [melt.dropout(tf.ones([batch_size, 1, num_units[layer]], dtype=tf.float32), keep_prob=self.keep_prob, training=training, mode=None) for layer in range(self.num_layers)]
mask_bws = [melt.dropout(tf.ones([batch_size, 1, num_units[layer]], dtype=tf.float32), keep_prob=self.keep_prob, training=training, mode=None) for layer in range(self.num_layers)]
c = self.encode(c_emb, c_len, mask_fws=mask_fws, mask_bws=mask_bws)
q = self.encode(q_emb, q_len, mask_fws=mask_fws, mask_bws=mask_bws)
qc_att = self.att_dot_attention(c, q, mask=q_mask, training=training)
num_units = [melt.get_shape(qc_att, -1) if layer == 0 else 2 * self.num_units for layer in range(self.num_layers)]
mask_fws = [melt.dropout(tf.ones([batch_size, 1, num_units[layer]], dtype=tf.float32), keep_prob=self.keep_prob, training=training, mode=None) for layer in range(1)]
mask_bws = [melt.dropout(tf.ones([batch_size, 1, num_units[layer]], dtype=tf.float32), keep_prob=self.keep_prob, training=training, mode=None) for layer in range(1)]
x = self.att_encode(qc_att, c_len, mask_fws=mask_fws, mask_bws=mask_bws)
x = self.pooling(x, c_len)
if FLAGS.use_type:
x = tf.concat([x, tf.expand_dims(tf.to_float(input['type']), 1)], 1)
if not FLAGS.split_type:
x = self.logits(x)
else:
x1 = self.logits(x)
x2 = self.logits2(x)
x = tf.cond(tf.cast(input['type'] == 0, tf.bool), lambda: (x1 + x2) / 2., lambda: x2)
return x
def call(self, input, c_len=None, max_c_len=None, training=False):
self.step += 1
x = input['content'] if isinstance(input, dict) else input
batch_size = melt.get_shape(x, 0)
model = modeling.BertModel(
config=self.bert_config,
is_training=training,
input_ids=x,
input_mask=(x > 0) if c_len is not None else None)
if self.step == 0 and self.init_checkpoint:
self.restore()
x = model.get_sequence_output()
return x
def call(self, x, training=False):
self.step += 1
if self.step == 0:
n_state = melt.get_shape(x, -1)
self.g = self.add_variable("g",
shape=[n_state],
initializer=tf.constant_initializer(1))
self.b = self.add_variable("b",
shape=[n_state],
initializer=tf.constant_initializer(1))
e, axis = self.e, self.axis
u = tf.reduce_mean(x, axis=axis, keepdims=True)
s = tf.reduce_mean(tf.square(x - u), axis=axis, keepdims=True)
x = (x - u) * tf.rsqrt(s + e)
x = x * self.g + self.b
return x
def call(self, outputs, sequence_length=None, axis=1):
self.step += 1
if self.step == 0 and self.dense is None:
self.dense = layers.Dense(melt.get_shape(outputs, -1),
activation=self.activation)
x = self.dense(outputs)
logits = self.logits(x)
alphas = tf.nn.softmax(
logits) if sequence_length is None else melt.masked_softmax(
logits, sequence_length)
encoding = tf.reduce_sum(outputs * alphas, 1)
# [batch_size, sequence_length, 1] -> [batch_size, sequence_length]
self.alphas = tf.squeeze(alphas, -1)
#self.alphas = alphas
tf.add_to_collection('self_attention', self.alphas)
return encoding
def encode(self, seq, seq_len=None, output_method='all'):
with tf.variable_scope(self.scope, reuse=tf.AUTO_REUSE):
if self.use_position_encoding:
hidden_size = melt.get_shape(seq, -1)
# Scale embedding by the sqrt of the hidden size
seq *= hidden_size ** 0.5
# Create binary array of size [batch_size, length]
# where 1 = padding, 0 = not padding
padding = tf.to_float(tf.sequence_mask(seq_len))
# Set all padding embedding values to 0
seq *= tf.expand_dims(padding, -1)
pos_encoding = model_utils.get_position_encoding(
tf.shape(seq)[1], tf.shape(seq)[-1])
seq = seq + pos_encoding
num_filters = self.num_filters
seqs = [seq]
#batch_size = melt.get_batch_size(seq)
#kernel_sizes = [3, 5, 7, 9, 11, 13]
kernel_sizes = [3] * 7
assert self.num_layers <= len(kernel_sizes)
for layer in range(self.num_layers):
#input_size_ = melt.get_shape(seq, -1) if layer == 0 else num_filters
seq = melt.dropout(seq, self.keep_prob, self.is_train)
seq = tf.layers.conv1d(seqs[-1], num_filters, kernel_size=kernel_sizes[layer], padding='same', activation=tf.nn.relu)
# mask = melt.dropout(tf.ones([batch_size, 1, input_size_], dtype=tf.float32),
# keep_prob=self.keep_prob, is_train=self.is_train, mode=None)
#seq = tf.layers.conv1d(seqs[-1] * mask, num_filters, kernel_size=3, padding='same', activation=tf.nn.relu)
#seq = tf.layers.conv1d(seqs[-1] * mask, num_filters, kernel_size=kernel_sizes[layer], padding='same', activation=tf.nn.relu)
# if self.is_train and self.keep_prob < 1:
# seq = tf.nn.dropout(seq, self.keep_prob)
#seq = melt.layers.batch_norm(seq, self.is_train, name='layer_%d' % layer)
seqs.append(seq)
outputs = tf.concat(seqs[1:], 2)
# not do any dropout in convet just dropout outside
# if self.is_train and self.keep_prob < 1:
# outputs = tf.nn.dropout(outputs, self.keep_prob)
# compact for rnn with sate return
return melt.rnn.encode_outputs(outputs, seq_len, output_method)
def call(self, input, c_len=None, max_c_len=None, training=False):
assert isinstance(input, dict)
x = input['content']
batch_size = melt.get_shape(x, 0)
if c_len is None or max_c_len is None:
c_len, max_c_len = melt.length2(x)
if self.rnn_no_padding:
logging.info('------------------no padding! train or eval')
c_len = max_c_len
x = self.embedding(x)
if FLAGS.use_char:
cx = input['char']
cx = tf.reshape(cx, [batch_size * max_c_len, FLAGS.char_limit])
chars_len = melt.length(cx)
cx = self.char_embedding(cx)
cx = self.char_encode(cx, chars_len, training=training)
cx = self.char_pooling(cx, chars_len)
cx = tf.reshape(cx, [batch_size, max_c_len, 2 * self.num_units])
if self.char_combiner == 'concat':
x = tf.concat([x, cx], axis=2)
elif self.char_combiner == 'sfu':
x = self.char_sfu_combine(x, cx, training=training)
if FLAGS.use_pos:
px = input['pos']
px = self.pos_embedding(px)
x = tf.concat([x, px], axis=2)
if FLAGS.use_ner:
nx = input['ner']
nx = self.ner_embedding(nx)
x = tf.concat([x, nx], axis=2)
x = self.encode(x, c_len, training=training)
return x
def call(self, x, fusions, training=False):
self.step += 1
assert len(fusions) > 0
vectors = tf.concat(
[x] + fusions, axis=-1
) # size = [batch_size, ..., input_dim * (len(fusion_vectors) + 1)]
dim = melt.get_shape(x, -1)
dv = dropout(vectors, keep_prob=self.keep_prob, training=training)
if self.step == 0:
self.composition_dense = layers.Dense(dim,
use_bias=True,
activation=tf.nn.tanh,
name='compostion_dense')
self.gate_dense = layers.Dense(dim,
use_bias=True,
activation=tf.nn.sigmoid,
name='gate_dense')
r = self.composition_dense(dv)
g = self.gate_dense(dv)
return g * r + (1 - g) * x
def call(self, input, training=False):
self.step += 1
x = input['content']
x = self.unk_aug(x, training=training)
batch_size = melt.get_shape(x, 0)
c_mask = tf.cast(x, tf.bool)
# TODO move to __init__
model = modeling.BertModel(config=self.bert_config,
is_training=training,
input_ids=x,
input_mask=c_mask,
use_one_hot_embeddings=FLAGS.use_tpu)
if self.step == 0 and self.init_checkpoint:
self.restore()
c_len = melt.length(x)
if FLAGS.encoder_output_method == 'last':
x = model.get_pooled_output()
else:
x = model.get_sequence_output()
if training:
x = tf.nn.dropout(x, keep_prob=0.9)
logging.info('---------------bert_lr_ratio', FLAGS.bert_lr_ratio)
x = x * FLAGS.bert_lr_ratio + tf.stop_gradient(x) * (
1 - FLAGS.bert_lr_ratio)
if FLAGS.transformer_add_rnn:
assert FLAGS.encoder_output_method != 'last'
x = self.rnn_encode(x, c_len)
if FLAGS.encoder_output_method != 'last':
x = self.pooling(x, c_len)
x2 = model.get_pooled_output()
x = tf.concat([x, x2], -1)
x = self.logits(x)
x = tf.reshape(x, [batch_size, NUM_ATTRIBUTES, NUM_CLASSES])
return x
def focal_loss(target_tensor,
prediction_tensor,
weights=None,
alpha=0.25,
gamma=2):
r"""Compute focal loss for predictions.
Multi-labels Focal loss formula:
FL = -alpha * (z-p)^gamma * log(p) -(1-alpha) * p^gamma * log(1-p)
,which alpha = 0.25, gamma = 2, p = sigmoid(x), z = target_tensor.
Args:
prediction_tensor: A float tensor of shape [batch_size, num_anchors,
num_classes] representing the predicted logits for each class
target_tensor: A float tensor of shape [batch_size, num_anchors,
num_classes] representing one-hot encoded classification targets
weights: A float tensor of shape [batch_size, num_anchors]
alpha: A scalar tensor for focal loss alpha hyper-parameter
gamma: A scalar tensor for focal loss gamma hyper-parameter
Returns:
loss: A (scalar) tensor representing the value of the loss function
"""
sigmoid_p = tf.nn.sigmoid(prediction_tensor)
zeros = array_ops.zeros_like(sigmoid_p, dtype=sigmoid_p.dtype)
#target_tensor = tf.to_float(target_tensor)
target_tensor = tf.one_hot(target_tensor,
melt.get_shape(prediction_tensor, -1))
# For poitive prediction, only need consider front part loss, back part is 0;
# target_tensor > zeros <=> z=1, so poitive coefficient = z - p.
pos_p_sub = array_ops.where(target_tensor > zeros,
target_tensor - sigmoid_p, zeros)
# For negative prediction, only need consider back part loss, front part is 0;
# target_tensor > zeros <=> z=1, so negative coefficient = 0.
neg_p_sub = array_ops.where(target_tensor > zeros, zeros, sigmoid_p)
per_entry_cross_ent = - alpha * (pos_p_sub ** gamma) * tf.log(tf.clip_by_value(sigmoid_p, 1e-8, 1.0)) \
- (1 - alpha) * (neg_p_sub ** gamma) * tf.log(tf.clip_by_value(1.0 - sigmoid_p, 1e-8, 1.0))
#return tf.reduce_sum(per_entry_cross_ent)
return tf.reduce_mean(per_entry_cross_ent)
def call(self, seq, seq_len=None, masks=None,
output_method=OutputMethod.all,
training=False):
if self.use_position_encoding:
hidden_size = melt.get_shape(seq, -1)
# Scale embedding by the sqrt of the hidden size
seq *= hidden_size ** 0.5
# Create binary array of size [batch_size, length]
# where 1 = padding, 0 = not padding
padding = tf.to_float(tf.sequence_mask(seq_len))
# Set all padding embedding values to 0
seq *= tf.expand_dims(padding, -1)
pos_encoding = model_utils.get_position_encoding(
tf.shape(seq)[1], tf.shape(seq)[-1])
seq = seq + pos_encoding
num_filters = self.num_filters
seqs = [seq]
#batch_size = melt.get_batch_size(seq)
for layer in range(self.num_layers):
if masks is None:
seq_ = melt.dropout(seq, self.keep_prob, training)
else:
seq_ = seq * masks[layer]
seq = self.conv1ds[layer](seq_)
seqs.append(seq)
outputs = tf.concat(seqs[1:], 2)
# not do any dropout in convet just dropout outside
# if self.is_train and self.keep_prob < 1:
# outputs = tf.nn.dropout(outputs, self.keep_prob)
# compact for rnn with sate return
return melt.rnn.encode_outputs(outputs, seq_len, output_method)
def encode(self,
inputs,
seq_len,
emb=None,
concat_layers=True,
output_method=OutputMethod.all):
if emb is not None:
inputs = tf.nn.embedding_lookup(emb, inputs)
outputs = [inputs]
keep_prob = self.keep_prob
num_units = self.num_units
is_train = self.is_train
with tf.variable_scope(self.scope, reuse=tf.AUTO_REUSE):
for layer in range(self.num_layers):
input_size_ = melt.get_shape(
inputs, -1) if layer == 0 else 2 * num_units
batch_size = melt.get_batch_size(inputs)
with tf.variable_scope("fw_{}".format(layer)):
gru_fw = tf.contrib.rnn.GRUCell(num_units)
if not self.share_dropout:
mask_fw = dropout(tf.ones([batch_size, 1, input_size_],
dtype=tf.float32),
keep_prob=keep_prob,
is_train=is_train,
mode=self.dropout_mode)
else:
if self.dropout_mask_fw[layer] is None:
mask_fw = dropout(
tf.ones([batch_size, 1, input_size_],
dtype=tf.float32),
keep_prob=keep_prob,
is_train=is_train,
mode=self.dropout_mode)
self.dropout_mask_fw[layer] = mask_fw
else:
mask_fw = self.dropout_mask_fw[layer]
if self.train_init_state:
if self.init_fw[layer] is None:
self.init_fw[layer] = tf.tile(
tf.get_variable("init_state", [1, num_units],
tf.float32,
tf.zeros_initializer()),
[batch_size, 1])
out_fw, state = tf.nn.dynamic_rnn(
gru_fw,
outputs[-1] * mask_fw,
seq_len,
initial_state=self.init_fw[layer],
dtype=tf.float32)
with tf.variable_scope("bw_{}".format(layer)):
gru_bw = tf.contrib.rnn.GRUCell(num_units)
if not self.share_dropout:
mask_bw = dropout(tf.ones([batch_size, 1, input_size_],
dtype=tf.float32),
keep_prob=keep_prob,
is_train=is_train,
mode=self.dropout_mode)
else:
if self.dropout_mask_bw[layer] is None:
mask_bw = dropout(
tf.ones([batch_size, 1, input_size_],
dtype=tf.float32),
keep_prob=keep_prob,
is_train=is_train,
mode=self.dropout_mode)
self.dropout_mask_bw[layer] = mask_bw
else:
mask_bw = self.dropout_mask_bw[layer]
if self.train_init_state:
if self.init_bw[layer] is None:
self.init_bw[layer] = tf.tile(
tf.get_variable("init_state", [1, num_units],
tf.float32,
tf.zeros_initializer()),
[batch_size, 1])
inputs_bw = tf.reverse_sequence(outputs[-1] * mask_bw,
seq_lengths=seq_len,
seq_dim=1,
batch_dim=0)
out_bw, _ = tf.nn.dynamic_rnn(
gru_bw,
inputs_bw,
seq_len,
initial_state=self.init_bw[layer],
dtype=tf.float32)
out_bw = tf.reverse_sequence(out_bw,
seq_lengths=seq_len,
seq_dim=1,
batch_dim=0)
outputs.append(tf.concat([out_fw, out_bw], axis=2))
if concat_layers:
res = tf.concat(outputs[1:], axis=2)
else:
res = outputs[-1]
res = encode_outputs(res, seq_len, output_method=output_method)
return res
def call(self,
inputs,
sequence_length,
inputs2,
sequence_length2,
mask_fws,
mask_bws,
concat_layers=True,
output_method=OutputMethod.all,
training=False):
outputs = [inputs]
outputs2 = [inputs2]
keep_prob = self.keep_prob
num_units = self.num_units
batch_size = melt.get_batch_size(inputs)
for layer in range(self.num_layers):
input_size_ = melt.get_shape(inputs,
-1) if layer == 0 else 2 * num_units
gru_fw, gru_bw = self.gru_fws[layer], self.gru_bws[layer]
if self.train_init_state:
init_fw = self.init_fw_layer(layer, batch_size)
else:
init_fw = None
mask_fw = mask_fws[layer]
out_fw, state_fw = gru_fw(outputs[-1] * mask_fw, init_fw)
out_fw2, state_fw2 = gru_fw(outputs2[-1] * mask_fw, state_fw)
mask_bw = mask_bws[layer]
inputs_bw = tf.reverse_sequence(
outputs[-1] * mask_bw,
sequence_lengthgths=sequence_length,
seq_axis=1,
batch_axis=0)
inputs_bw2 = tf.reverse_sequence(
outputs2[-1] * mask_bw,
sequence_lengthgths=sequence_length2,
seq_axis=1,
batch_axis=0)
if self.train_init_state:
init_bw = self.init_bw_layer(layer, batch_size)
else:
init_bw = None
out_bw, state_bw = gru_bw(inputs_bw, init_bw)
out_bw2, state_bw2 = gru_bw(inputs_bw2, state_bw)
outputs.append(tf.concat([out_fw, out_bw], axis=2))
outputs2.append(tf.concat([out_fw2, out_bw2], axis=2))
if concat_layers:
res = tf.concat(outputs[1:], axis=2)
res2 = tf.concat(outputs2[1:], axis=2)
else:
res = outputs[-1]
res2 = outpus2[-1]
res = tf.concat([res, res2], axis=1)
res = encode_outputs(res,
output_method=output_method,
sequence_length=sequence_length)
self.state = (state_fw2, state_bw2)
return res
def call(self, input):
# TODO tf2 keras seem to auto append last dim so need this
melt.try_squeeze_dim(input)
if not FLAGS.batch_parse:
util.adjust(input, self.mode)
# print(input)
embs = []
if 'history' in input:
hlen = melt.length(input['history'])
hlen = tf.math.maximum(hlen, 1)
bs = melt.get_shape(input['did'], 0)
# user
if FLAGS.use_uid:
uemb = self.uemb(input['uid'])
embs += [uemb]
if FLAGS.use_did:
demb = self.demb(input['did'])
embs += [demb]
if FLAGS.use_time_emb:
embs += [
self.hour_emb(input['hour']),
self.weekday_emb(input['weekday']),
]
if FLAGS.use_fresh_emb:
fresh = input['fresh']
fresh_day = tf.cast(fresh / (3600 * 12), fresh.dtype)
fresh_hour = tf.cast(fresh / 3600, fresh.dtype)
embs += [
self.fresh_day_emb(fresh_day),
self.fresh_hour_emb(fresh_hour)
]
if FLAGS.use_position_emb:
embs += [self.position_emb(input['position'])]
if FLAGS.use_news_info and 'cat' in input:
# print('------entity_emb', self.entity_emb.emb.weights) # check if trainable is fixed in eager mode
embs += [
self.cat_emb(input['cat']),
self.scat_emb(input['sub_cat']),
self.pooling(self.entity_type_emb(input['title_entity_types']),
melt.length(input['title_entity_types'])),
self.pooling(
self.entity_type_emb(input['abstract_entity_types']),
melt.length(input['abstract_entity_types'])),
]
if FLAGS.use_entities and 'title_entities' in input:
embs += [
self.pooling(self.entity_emb(input['title_entities']),
melt.length(input['title_entities'])),
self.pooling(self.entity_emb(input['abstract_entities']),
melt.length(input['abstract_entities'])),
]
if FLAGS.use_history_info and 'history_cats' in input:
embs += [
self.his_simple_pooling(self.cat_emb(input['history_cats']),
melt.length(input['history_cats'])),
self.his_simple_pooling(
self.scat_emb(input['history_sub_cats']),
melt.length(input['history_sub_cats'])),
]
if FLAGS.use_history_entities:
try:
embs += [
self.his_simple_pooling(
self.entity_type_emb(
input['history_title_entity_types']),
melt.length(input['history_title_entity_types'])),
self.his_simple_pooling(
self.entity_type_emb(
input['history_abstract_entity_types']),
melt.length(
input['history_abstract_entity_types'])),
]
if FLAGS.use_entities and 'title_entities' in inpout:
embs += [
self.his_simple_pooling(
self.entity_emb(
input['history_title_entities']),
melt.length(input['history_title_entities'])),
self.his_simple_pooling(
self.entity_emb(
input['history_abstract_entities']),
melt.length(
input['history_abstract_entities'])),
]
except Exception:
pass
if FLAGS.use_history and FLAGS.use_did:
dids = input['history']
if FLAGS.his_strategy == 'bst' or FLAGS.his_pooling == 'mhead':
mask = tf.cast(tf.equal(dids, 0), dids.dtype)
dids += mask
hlen = tf.ones_like(hlen) * 50
hembs = self.demb(dids)
his_embs = hembs
his_embs = self.his_encoder(his_embs, hlen)
self.his_embs = his_embs
his_emb = self.his_pooling(demb, his_embs, hlen)
embs += [his_emb]
if FLAGS.use_title:
cur_title = self.title_encoder(self.title_lookup(input['ori_did']))
dids = input['ori_history']
if FLAGS.max_titles:
dids = dids[:, :FLAGS.max_titles]
his_title = self.titles_encoder(self.title_lookup(dids), hlen,
cur_title)
embs += [cur_title, his_title]
# 用impression id 会dev test不一致 不直接用id
if FLAGS.use_impressions:
embs += [self.mean_pooling(self.demb(input['impressions']))]
if FLAGS.use_dense:
dense_emb = self.deal_dense(input)
embs += [dense_emb]
# logging.debug('-----------embs:', len(embs))
embs = tf.stack(embs, axis=1)
if FLAGS.batch_norm:
embs = self.batch_norm(embs)
if FLAGS.l2_normalize_before_pooling:
x = tf.math.l2_normalize(embs)
x = self.feat_pooling(embs)
if FLAGS.dropout:
x = self.dropout(x)
if FLAGS.use_dense:
x = tf.concat([x, dense_emb], axis=1)
if FLAGS.use_his_concat:
x = tf.concat([x, his_concat], axis=1)
x = self.mlp(x)
self.logit = self.dense(x)
self.prob = tf.math.sigmoid(self.logit)
self.impression_id = input['impression_id']
self.position = input['position']
self.history_len = input['hist_len']
self.impression_len = input['impression_len']
self.input_ = input
return self.logit
def call(self,
x,
sequence_length=None,
mask_fws=None,
mask_bws=None,
concat_layers=None,
output_method=None,
training=False):
concat_layers = concat_layers or self.concat_layers
output_mehtod = output_method or self.output_method
if self.residual_connect:
x = self.residual_linear(x)
outputs = [x]
#states = []
keep_prob = self.keep_prob
num_units = self.num_units
batch_size = melt.get_batch_size(x)
if sequence_length is None:
len_ = melt.get_shape(x, 1)
sequence_length = tf.ones([
batch_size,
], dtype=tf.int64) * len_
for layer in range(self.num_layers):
input_size_ = melt.get_shape(x,
-1) if layer == 0 else 2 * num_units
gru_fw, gru_bw = self.gru_fws[layer], self.gru_bws[layer]
if self.train_init_state:
#init_fw = tf.tile(self.init_fw[layer], [batch_size, 1])
#init_fw = tf.tile(self.init_fw_layer(layer), [batch_size, 1])
init_fw = self.init_fw_layer(layer, batch_size)
if self.cell == 'lstm':
init_fw = (init_fw, self.init_fw2_layer(layer, batch_size))
else:
init_fw = None
if self.recurrent_dropout:
if mask_fws is not None:
mask_fw = mask_fws[layer]
else:
if not self.share_dropout:
mask_fw = dropout(tf.ones([batch_size, 1, input_size_],
dtype=tf.float32),
keep_prob=keep_prob,
training=training,
mode=None)
else:
if self.dropout_mask_fw[layer] is None or (
tf.executing_eagerly() and batch_size !=
self.dropout_mask_fw[layer].shape[0]):
mask_fw = dropout(
tf.ones([batch_size, 1, input_size_],
dtype=tf.float32),
keep_prob=keep_prob,
training=training,
mode=None)
self.dropout_mask_fw[layer] = mask_fw
else:
mask_fw = self.dropout_mask_fw[layer]
inputs_fw = outputs[-1] * mask_fw
else:
inputs_fw = dropout(outputs[-1],
keep_prob=keep_prob,
training=training,
mode=None)
# https://stackoverflow.com/questions/48233400/lstm-initial-state-from-dense-layer
# gru and lstm different ... state lstm need tuple (,) states as input state\
if self.cell == 'gru':
out_fw, state_fw = gru_fw(inputs_fw, init_fw)
else:
out_fw, state_fw1, state_fw2 = gru_fw(inputs_fw, init_fw)
state_fw = (state_fw1, state_fw2)
if self.train_init_state:
#init_bw = tf.tile(self.init_bw[layer], [batch_size, 1])
#init_bw = tf.tile(self.init_bw_layer(layer), [batch_size, 1])
init_bw = self.init_bw_layer(layer, batch_size)
if self.cell == 'lstm':
init_bw = (init_bw, self.init_bw2_layer(layer, batch_size))
else:
init_bw = None
if mask_bws is not None:
mask_bw = mask_bws[layer]
else:
if not self.share_dropout:
mask_bw = dropout(tf.ones([batch_size, 1, input_size_],
dtype=tf.float32),
keep_prob=keep_prob,
training=training,
mode=None)
else:
if self.dropout_mask_bw[layer] is None or (
tf.executing_eagerly() and batch_size !=
self.dropout_mask_bw[layer].shape[0]):
mask_bw = dropout(tf.ones([batch_size, 1, input_size_],
dtype=tf.float32),
keep_prob=keep_prob,
training=training,
mode=None)
self.dropout_mask_bw[layer] = mask_bw
else:
mask_bw = self.dropout_mask_bw[layer]
if self.recurrent_dropout:
inputs_bw = outputs[-1] * mask_bw
else:
if self.bw_dropout:
inputs_bw = dropout(outputs[-1],
keep_prob=keep_prob,
training=training,
mode=None)
else:
inputs_bw = inputs_fw
inputs_bw = tf.reverse_sequence(inputs_bw,
seq_lengths=sequence_length,
seq_axis=1,
batch_axis=0)
if self.cell == 'gru':
out_bw, state_bw = gru_bw(inputs_bw, init_bw)
else:
out_bw, state_bw1, state_bw2 = gru_bw(inputs_bw, init_bw)
state_bw = (state_bw1, state_bw2)
out_bw = tf.reverse_sequence(out_bw,
seq_lengths=sequence_length,
seq_axis=1,
batch_axis=0)
outputs.append(tf.concat([out_fw, out_bw], axis=2))
if self.residual_connect:
outputs[-1] = self.batch_norm(outputs[-2] + outputs[-1])
if concat_layers:
res = tf.concat(outputs[1:], axis=2)
else:
res = outputs[-1]
res = encode_outputs(res,
output_method=output_method,
sequence_length=sequence_length)
self.state = (state_fw, state_bw)
if not self.return_state:
return res
else:
return res, self.state
def call(self, outputs, sequence_length=None, axis=1):
x = melt.top_k_pooling(outputs, self.top_k, sequence_length,
axis).values
return tf.reshape(x, [-1, melt.get_shape(outputs, -1) * self.top_k])
def take_step(self, i, prev, state):
print('-------------i', i)
if self.output_fn is not None:
#[batch_size * beam_size, num_units] -> [batch_size * beam_size, num_classes]
try:
output = self.output_fn(prev)
except Exception:
output = self.output_fn(prev, state)
else:
output = prev
self.output = output
#[batch_size * beam_size, num_classes], here use log sofmax
if self.need_softmax:
logprobs = tf.nn.log_softmax(output)
else:
logprobs = tf.log(tf.maximum(output, 1e-12))
if self.num_classes is None:
self.num_classes = tf.shape(logprobs)[1]
#->[batch_size, beam_size, num_classes]
logprobs_batched = tf.reshape(logprobs,
[-1, self.beam_size, self.num_classes])
logprobs_batched.set_shape((None, self.beam_size, None))
# Note: masking out entries to -inf plays poorly with top_k, so just subtract out a large number.
nondone_mask = tf.reshape(
tf.cast(
tf.equal(tf.range(self.num_classes), self.done_token),
tf.float32) * -1e18,
[1, 1, self.num_classes])
if self.past_logprobs is None:
#[batch_size, beam_size, num_classes] -> [batch_size, num_classes]
#-> past_logprobs[batch_size, beam_size], indices[batch_size, beam_size]
self.past_logprobs, indices = tf.nn.top_k(
(logprobs_batched + nondone_mask)[:, 0, :],
self.beam_size)
step_logprobs = self.past_logprobs
else:
#logprobs_batched [batch_size, beam_size, num_classes] -> [batch_size, beam_size, num_classes]
#past_logprobs [batch_size, beam_size] -> [batch_size, beam_size, 1]
step_logprobs_batched = logprobs_batched
logprobs_batched = logprobs_batched + tf.expand_dims(self.past_logprobs, 2)
#get [batch_size, beam_size] each
self.past_logprobs, indices = tf.nn.top_k(
#[batch_size, beam_size * num_classes]
tf.reshape(logprobs_batched + nondone_mask,
[-1, self.beam_size * self.num_classes]),
self.beam_size)
#get current step logprobs [batch_size, beam_size]
step_logprobs = tf.gather_nd(tf.reshape(step_logprobs_batched,
[-1, self.beam_size * self.num_classes]),
melt.batch_values_to_indices(indices))
# For continuing to the next symbols [batch_size, beam_size]
symbols = indices % self.num_classes
#from wich beam it comes [batch_size, beam_size]
parent_refs = indices // self.num_classes
if self.past_symbols is None:
#here when i == 1, when i==0 will not do take step it just do one rnn() get output and use it for i==1 here
#here will not need to gather state for inital state of each beam is the same
#[batch_size, beam_size] -> [batch_size, beam_size, 1]
self.past_symbols = tf.expand_dims(symbols, 2)
self.past_step_logprobs = tf.expand_dims(step_logprobs, 2)
else:
# NOTE: outputing a zero-length sequence is not supported for simplicity reasons
#hasky/jupter/tensorflow/beam-search2.ipynb below for mergeing path
#here when i >= 2
# tf.reshape(
# (tf.range(3 * 5) // 5) * 5,
# [3, 5]
# ).eval()
# array([[ 0, 0, 0, 0, 0],
# [ 5, 5, 5, 5, 5],
# [10, 10, 10, 10, 10]], dtype=int32)
parent_refs_offsets = tf.reshape(
(tf.range(self.batch_size * self.beam_size)
// self.beam_size) * self.beam_size,
[self.batch_size, self.beam_size])
#self.past_symbols [batch_size, beam_size, i - 1] -> past_symbols_batch_major [batch_size * beam_size, i - 1]
past_symbols_batch_major = tf.reshape(self.past_symbols, [-1, i-1])
past_step_logprobs_batch_major = tf.reshape(self.past_step_logprobs, [-1, i - 1])
#[batch_size, beam_size]
past_indices = parent_refs + parent_refs_offsets
#-> [batch_size, beam_size, i - 1]
beam_past_symbols = tf.gather(past_symbols_batch_major, #[batch_size * beam_size, i - 1]
past_indices #[batch_size, beam_size]
)
beam_past_step_logprobs = tf.gather(past_step_logprobs_batch_major, past_indices)
#we must also choose corresponding past state as new start
past_indices = tf.reshape(past_indices, [-1])
#TODO not support tf.TensorArray right now, can not use aligment_history in attention_wrapper
def try_gather(x, indices):
#if isinstance(x, tf.Tensor) and x.shape.ndims >= 2:
assert isinstance(x, tf.Tensor)
if x.shape.ndims >= 2:
return tf.gather(x, indices)
else:
return x
state = nest.map_structure(lambda x: try_gather(x, past_indices), state)
if hasattr(state, 'alignments'):
attention_size = melt.get_shape(state.alignments, -1)
alignments = tf.reshape(state.alignments, [-1, self.beam_size, attention_size])
print('alignments', alignments)
if not self.fast_greedy:
#[batch_size, beam_size, max_len]
path = tf.concat([self.past_symbols,
tf.ones_like(tf.expand_dims(symbols, 2)) * self.done_token,
tf.tile(tf.ones_like(tf.expand_dims(symbols, 2)) * self.pad_token,
[1, 1, self.max_len - i])], 2)
step_logprobs_path = tf.concat([self.past_step_logprobs,
tf.expand_dims(step_logprobs_batched[:, :, self.done_token], 2),
tf.tile(tf.ones_like(tf.expand_dims(step_logprobs, 2)) * -float('inf'),
[1, 1, self.max_len - i])], 2)
#[batch_size, 1, beam_size, max_len]
path = tf.expand_dims(path, 1)
step_logprobs_path = tf.expand_dims(step_logprobs_path, 1)
self.paths_list.append(path)
self.step_logprobs_list.append(step_logprobs_path)
#[batch_size * beam_size, i - 1] -> [batch_size, bam_size, i] the best beam_size paths until step i
self.past_symbols = tf.concat([beam_past_symbols, tf.expand_dims(symbols, 2)], 2)
self.past_step_logprobs = tf.concat([beam_past_step_logprobs, tf.expand_dims(step_logprobs, 2)], 2)
# For finishing the beam
#[batch_size, beam_size]
logprobs_done = logprobs_batched[:, :, self.done_token]
if not self.fast_greedy:
self.logprobs_list.append(logprobs_done / i ** self.length_normalization_factor)
else:
done_parent_refs = tf.cast(tf.argmax(logprobs_done, 1), tf.int32)
done_parent_refs_offsets = tf.range(self.batch_size) * self.beam_size
done_past_symbols = tf.gather(past_symbols_batch_major,
done_parent_refs + done_parent_refs_offsets)
#[batch_size, max_len]
symbols_done = tf.concat([done_past_symbols,
tf.ones_like(done_past_symbols[:,0:1]) * self.done_token,
tf.tile(tf.zeros_like(done_past_symbols[:,0:1]),
[1, self.max_len - i])
], 1)
#[batch_size, beam_size] -> [batch_size,]
logprobs_done_max = tf.reduce_max(logprobs_done, 1)
if self.length_normalization_factor > 0:
logprobs_done_max /= i ** self.length_normalization_factor
#[batch_size, max_len]
self.finished_beams = tf.where(logprobs_done_max > self.logprobs_finished_beams,
symbols_done,
self.finished_beams)
self.logprobs_finished_beams = tf.maximum(logprobs_done_max, self.logprobs_finished_beams)
#->[batch_size * beam_size,]
symbols_flat = tf.reshape(symbols, [-1])
self.final_state = state
return symbols_flat, state
def _get_aligments(self, state): attention_size = melt.get_shape(state.alignments, -1) alignments = tf.reshape(state.alignments, [-1, self.beam_size, attention_size]) return alignments
def argtopk_pooling(outputs, top_k, sequence_length=None, axis=1):
x = top_k_pooling(outputs, top_k, sequence_length, axis).indices
#return tf.reshape(x, [melt.get_shape(outputs, 0), -1])
return tf.reshape(x, [-1, melt.get_shape(outputs, -1) * top_k])
