def sg_optim(loss, **kwargs):
r"""Applies gradients to variables.
Args:
loss: A 0-D `Tensor` containing the value to minimize. list of 0-D tensor for Multiple GPU
kwargs:
optim: A name for optimizer. 'MaxProp' (default), 'AdaMax', 'Adam', 'RMSProp' or 'sgd'.
lr: A Python Scalar (optional). Learning rate. Default is .001.
beta1: A Python Scalar (optional). Default is .9.
beta2: A Python Scalar (optional). Default is .99.
momentum : A Python Scalar for RMSProp optimizer (optional). Default is 0.
category: A string or string list. Specifies the variables that should be trained (optional).
Only if the name of a trainable variable starts with `category`, it's value is updated.
Default is '', which means all trainable variables are updated.
"""
opt = tf.sg_opt(kwargs)
# default training options
opt += tf.sg_opt(optim='MaxProp', lr=0.001, beta1=0.9, beta2=0.99, momentum=0., category='')
# select optimizer
if opt.optim == 'MaxProp':
optim = tf.sg_optimize.MaxPropOptimizer(learning_rate=opt.lr, beta2=opt.beta2)
elif opt.optim == 'AdaMax':
optim = tf.sg_optimize.AdaMaxOptimizer(learning_rate=opt.lr, beta1=opt.beta1, beta2=opt.beta2)
elif opt.optim == 'Adam':
optim = tf.train.AdamOptimizer(learning_rate=opt.lr, beta1=opt.beta1, beta2=opt.beta2)
elif opt.optim == 'RMSProp':
optim = tf.train.RMSPropOptimizer(learning_rate=opt.lr, decay=opt.beta1, momentum=opt.momentum)
else:
optim = tf.train.GradientDescentOptimizer(learning_rate=opt.lr)
# get trainable variables
if isinstance(opt.category, (tuple, list)):
var_list = []
for cat in opt.category:
var_list.extend([t for t in tf.trainable_variables() if t.name.startswith(cat)])
else:
var_list = [t for t in tf.trainable_variables() if t.name.startswith(opt.category)]
#
# calc gradient
#
# multiple GPUs case
if isinstance(loss, (tuple, list)):
gradients = []
# loop for each GPU tower
for i, loss_ in enumerate(loss):
# specify device
with tf.device('/gpu:%d' % i):
# give new scope only to operation
with tf.name_scope('gpu_%d' % i):
# add gradient calculation operation for each GPU tower
gradients.append(tf.gradients(loss_, var_list))
# averaging gradient
gradient = []
for grad in zip(*gradients):
gradient.append(tf.add_n(grad) / len(loss))
# single GPU case
else:
gradient = tf.gradients(loss, var_list)
gradient, _ = tf.clip_by_global_norm(gradient, opt.clip_grad_norm)
# gradient update op
with tf.device('/gpu:0'):
grad_var = [(g, v) for g, v in zip(gradient, var_list)]
grad_op = optim.apply_gradients(grad_var, global_step=tf.sg_global_step())
# add summary using last tower value
for g, v in grad_var:
# exclude batch normal statics
if 'mean' not in v.name and 'variance' not in v.name \
and 'beta' not in v.name and 'gamma' not in v.name:
tf.sg_summary_gradient(v, g)
# extra update ops within category ( for example, batch normal running stat update )
if isinstance(opt.category, (tuple, list)):
update_op = []
for cat in opt.category:
update_op.extend([t for t in tf.get_collection(tf.GraphKeys.UPDATE_OPS) if t.name.startswith(cat)])
else:
update_op = [t for t in tf.get_collection(tf.GraphKeys.UPDATE_OPS) if t.name.startswith(opt.category)]
return tf.group(*([grad_op] + update_op))
def tower_loss2_old(xx, scope, reuse_vars=False):
# make embedding matrix for source and target
with tf.variable_scope('embs', reuse=reuse_vars):
emb_x = tf.sg_emb(name='emb_x',
voca_size=Hp.vs,
dim=Hp.hd,
dev=self._dev)
emb_y = tf.sg_emb(name='emb_y',
voca_size=Hp.vs,
dim=Hp.hd,
dev=self._dev)
x_sents = tf.unstack(xx, axis=1) #each element is (batch, sentlen)
# generate first an unconditioned sentence
n_input = Hp.hd
subrec1 = subrec_zero_state(Hp.bs, Hp.hd)
subrec2 = subrec_zero_state(Hp.bs, Hp.hd)
rnn_cell = LSTMCell(in_dim=n_input, dim=Hp.hd)
(rnn_state, rnn_h) = rnn_cell.zero_state(Hp.bs)
crnn_cell = ConvLSTMCell(in_dim=n_input, dim=Hp.hd)
(crnn_state, crnn_h) = crnn_cell.zero_state(n_input)
for sent in range(len(x_sents) - 1):
y = x_sents[i + 1]
x = x_sents[i] # (batch, sentlen) = (16, 200)
# shift target by one step for training source
y_src = tf.concat([tf.zeros((Hp.bs, 1), tf.sg_intx), y[:, :-1]], 1)
# embed table lookup
enc = x.sg_lookup(emb=emb_x) #(batch, sentlen, dim1)
# loop dilated conv block
for i in range(num_blocks):
enc = (enc.sg_res_block(
size=5, rate=1, name="enc1_%d" % (i),
reuse_vars=reuse_vars).sg_res_block(
size=5,
rate=2,
name="enc2_%d" % (i),
reuse_vars=reuse_vars).sg_res_block(
size=5,
rate=4,
name="enc4_%d" % (i),
reuse_vars=reuse_vars).sg_res_block(
size=5,
rate=8,
name="enc8_%d" % (i),
reuse_vars=reuse_vars).sg_res_block(
size=5,
rate=16,
name="enc16_%d" % (i),
reuse_vars=reuse_vars))
#quasi rnn layer [batch * 3, t, dim2 ]
conv = enc.sg_quasi_conv1d(is_enc=True,
size=2,
name="conv1",
reuse_vars=reuse_vars)
#attention layer
# recurrent layer # 1 + final encoder hidden state
concat = subrec1.sg_concat(target=conv, dim=0)
subrec1 = conv.sg_quasi_rnn(is_enc=True, att=True)
conv = pool.sg_quasi_conv1d(is_enc=True,
size=2,
name="conv2",
reuse_vars=reuse_vars)
concat = subrec2.sg_concat(target=conv, dim=0)
subrec2 = conv.sg_quasi_rnn(is_enc=True, att=True)
# conv LSTM
(crnn_state, crnn_h) = crnn_cell(subrec2, (crnn_state, crnn_h), 5)
# recurrent block
(rnn_state, rnn_h) = rnn_cell(crnn_h, (rnn_state, rnn_h))
# CNN decoder
dec = crnn_h.sg_concat(target=y_src.sg_lookup(emb=emb_y), name="dec")
for i in range(num_blocks):
dec = (dec.sg_res_block(
size=3,
rate=1,
causal=True,
name="dec1_%d" % (i),
reuse_vars=reuse_vars).sg_res_block(
size=3,
rate=2,
causal=True,
name="dec2_%d" % (i),
reuse_vars=reuse_vars).sg_res_block(
size=3,
rate=4,
causal=True,
name="dec4_%d" % (i),
reuse_vars=reuse_vars).sg_res_block(
size=3,
rate=8,
causal=True,
name="dec8_%d" % (i),
reuse_vars=reuse_vars).sg_res_block(
size=3,
rate=16,
causal=True,
name="dec16_%d" % (i),
reuse_vars=reuse_vars))
# final fully convolution layer for softmax
dec = dec.sg_conv1d_gpus(size=1, dim=Hp.vs,name="out",summary=False,\
dev = self._dev,reuse=reuse_vars)
ce_array = dec.sg_ce(target=y, mask=True, name="cross_ent_example")
cross_entropy_mean = tf.reduce_mean(ce_array, name='cross_entropy')
tf.add_to_collection('losses', cross_entropy_mean)
# Assemble all of the losses for the current tower only.
losses = tf.get_collection('losses', scope)
# Calculate the total loss for the current tower.
total_loss = tf.add_n(losses, name='total_loss')
return total_loss
size=1, act='tanh', bn=True, dim=20, dout=0.25).sg_conv1d(
size=1, act='tanh', bn=True, dim=2,
dout=0.25).sg_reshape(shape=(batch_size, 200)).sg_dense(
in_dim=200, dim=50,
act='relu', dout=0.20).sg_dense(in_dim=50,
dim=10,
act='relu',
dout=0.20).sg_dense(in_dim=10,
dim=2,
dout=0.20))
# CTC loss
#loss = logit.sg_ctc(target=y, seq_len=seq_len)
reg_lambda = 0.0002
trainable = tf.trainable_variables()
lossL2 = tf.add_n([tf.nn.l2_loss(v) for v in trainable]) * reg_lambda
loss = logit.sg_ce(target=y, one_hot=True) + lossL2
# train
config = tf.ConfigProto(allow_soft_placement=True,
inter_op_parallelism_threads=6,
intra_op_parallelism_threads=6)
sess = tf.Session(config=config)
tf.sg_init(sess)
learning_rate = tf.train.exponential_decay(0.00001,
tf.sg_global_step(),
100,
0.95,
staircase=False)
def rnn_body(time, subrec1, subrec2, rnn_state, rnn_h, crnn_state, crnn_h,
losses):
x = x_sent.read(time)
y = x_sent.read(time + 1) # (batch, sentlen) = (16, 200)
# shift target by one step for training source
y_src = tf.concat([tf.zeros((Hp.batch_size, 1), tf.int32), y[:, :-1]],
1)
reuse_vars = time == tf.constant(0) or reu_vars
# -------------------------- BYTENET ENCODER --------------------------
# embed table lookup
enc = x.sg_lookup(emb=emb_x) #(batch, sentlen, latentdim)
# loop dilated conv block
for i in range(num_blocks):
enc = (enc.sg_res_block(
size=5, rate=1, name="enc1_%d" % (i),
reuse_vars=reuse_vars).sg_res_block(
size=5,
rate=2,
name="enc2_%d" % (i),
reuse_vars=reuse_vars).sg_res_block(
size=5,
rate=4,
name="enc4_%d" % (i),
reuse_vars=reuse_vars).sg_res_block(
size=5,
rate=8,
name="enc8_%d" % (i),
reuse_vars=reuse_vars).sg_res_block(
size=5,
rate=16,
name="enc16_%d" % (i),
reuse_vars=reuse_vars))
# -------------------------- QCNN + QPOOL ENCODER with attention #1 --------------------------
#quasi cnn layer ZFO [batch * 3, t, dim2 ]
conv = enc.sg_quasi_conv1d(is_enc=True,
size=3,
name="qconv_1",
reuse_vars=reuse_vars)
#attention layer
# recurrent layer # 1 + final encoder hidden state
subrec1 = tf.tile((subrec1.sg_expand_dims(axis=1)), [1, Hp.maxlen, 1])
concat = conv.sg_concat(target=subrec1,
axis=0) # (batch*4, sentlen, latentdim)
pool = concat.sg_quasi_rnn(is_enc=True,
att=True,
name="qrnn_1",
reuse_vars=reuse_vars)
subrec1 = pool[:Hp.batch_size, -1, :] # last character in sequence
# -------------------------- QCNN + QPOOL ENCODER with attention #2 --------------------------
# quazi cnn ZFO (batch*3, sentlen, latentdim)
conv = pool.sg_quasi_conv1d(is_enc=True,
size=2,
name="qconv_2",
reuse_vars=reuse_vars)
# (batch, sentlen-duplicated, latentdim)
subrec2 = tf.tile((subrec2.sg_expand_dims(axis=1)), [1, Hp.maxlen, 1])
# (batch*4, sentlen, latentdim)
concat = conv.sg_concat(target=subrec2, axis=0)
pool = concat.sg_quasi_rnn(is_enc=True,
att=True,
name="qrnn_2",
reuse_vars=reuse_vars)
subrec2 = pool[:Hp.batch_size, -1, :] # last character in sequence
# -------------------------- ConvLSTM with RESIDUAL connection and MULTIPLICATIVE block --------------------------
#residual block
causal = False # for encoder
crnn_input = (pool[:Hp.batch_size, :, :].sg_bypass_gpus(
name='relu_0', act='relu', bn=(not causal),
ln=causal).sg_conv1d_gpus(name="dimred_0",
size=1,
dev="/cpu:0",
reuse=reuse_vars,
dim=Hp.hd / 2,
act='relu',
bn=(not causal),
ln=causal))
# conv LSTM
with tf.variable_scope("mem/clstm") as scp:
(crnn_state, crnn_h) = crnn_cell(crnn_input, (crnn_state, crnn_h),
size=5,
reuse_vars=reuse_vars)
# dimension recover and residual connection
rnn_input0 = pool[:Hp.batch_size,:,:] + crnn_h\
.sg_conv1d_gpus(name = "diminc_0",size=1,dev="/cpu:0", dim=Hp.hd,reuse=reuse_vars, act='relu', bn=(not causal), ln=causal)
# -------------------------- QCNN + QPOOL ENCODER with attention #3 --------------------------
# pooling for lstm input
# quazi cnn ZFO (batch*3, sentlen, latentdim)
conv = rnn_input0.sg_quasi_conv1d(is_enc=True,
size=2,
name="qconv_3",
reuse_vars=reuse_vars)
pool = conv.sg_quasi_rnn(is_enc=True,
att=False,
name="qrnn_3",
reuse_vars=reuse_vars)
rnn_input = pool[:Hp.batch_size, -1, :] # last character in sequence
# -------------------------- LSTM with RESIDUAL connection and MULTIPLICATIVE block --------------------------
# recurrent block
with tf.variable_scope("mem/lstm") as scp:
(rnn_state, rnn_h) = rnn_cell(rnn_input, (rnn_state, rnn_h))
rnn_h2 = tf.tile(((rnn_h + rnn_input).sg_expand_dims(axis=1)),
[1, Hp.maxlen, 1])
# -------------------------- BYTENET DECODER --------------------------
# CNN decoder
dec = y_src.sg_lookup(emb=emb_y).sg_concat(target=rnn_h2, name="dec")
for i in range(num_blocks):
dec = (dec.sg_res_block(
size=3,
rate=1,
causal=True,
name="dec1_%d" % (i),
reuse_vars=reuse_vars).sg_res_block(
size=3,
rate=2,
causal=True,
name="dec2_%d" % (i),
reuse_vars=reuse_vars).sg_res_block(
size=3,
rate=4,
causal=True,
name="dec4_%d" % (i),
reuse_vars=reuse_vars).sg_res_block(
size=3,
rate=8,
causal=True,
name="dec8_%d" % (i),
reuse_vars=reuse_vars).sg_res_block(
size=3,
rate=16,
causal=True,
name="dec16_%d" % (i),
reuse_vars=reuse_vars))
# final fully convolution layer for softmax
dec = dec.sg_conv1d_gpus(size=1,
dim=Hp.vs,
name="out",
summary=False,
dev=self._dev,
reuse=reuse_vars)
ce_array = dec.sg_ce(target=y, mask=True, name="cross_ent_example")
cross_entropy_mean = tf.reduce_mean(ce_array, name='cross_entropy')
losses = tf.add_n([losses, cross_entropy_mean], name='total_loss')
return (time + 1, subrec1, subrec2, rnn_state, rnn_h, crnn_state,
crnn_h, losses)