def __init__(self, returns_dict=False):
embeddings = [
("", [0, 0, 0, 0]), # OOV items are mapped to this embedding.
("hello world", [1, 2, 3, 4]),
("pair-programming", [5, 5, 5, 5]),
]
keys = tf.constant([item[0] for item in embeddings], dtype=tf.string)
indices = tf.constant(list(range(len(embeddings))), dtype=tf.int64)
tbl_init = KeyValueTensorInitializer(keys, indices)
self.table = HashTable(tbl_init, 0)
self.weights = tf.Variable(
list([item[1] for item in embeddings]), dtype=tf.float32)
self.variables = [self.weights]
self.trainable_variables = self.variables
self._returns_dict = returns_dict
class TextEmbedding(tf.train.Checkpoint):
def __init__(self, returns_dict=False):
embeddings = [
("", [0, 0, 0, 0]), # OOV items are mapped to this embedding.
("hello world", [1, 2, 3, 4]),
("pair-programming", [5, 5, 5, 5]),
]
keys = tf.constant([item[0] for item in embeddings], dtype=tf.string)
indices = tf.constant(list(range(len(embeddings))), dtype=tf.int64)
tbl_init = KeyValueTensorInitializer(keys, indices)
self.table = HashTable(tbl_init, 0)
self.weights = tf.Variable(
list([item[1] for item in embeddings]), dtype=tf.float32)
self.variables = [self.weights]
self.trainable_variables = self.variables
self._returns_dict = returns_dict
@tf.function(input_signature=[
tf.TensorSpec(dtype=tf.string, name="text", shape=[None])
])
def __call__(self, text_tensor):
indices_tensor = self.table.lookup(text_tensor)
embedding_tensor = tf.gather(self.weights, indices_tensor)
return dict(
outputs=embedding_tensor) if self._returns_dict else embedding_tensor
def __init__(self, mode, config):
self._mode = mode
self.config = config
if self._mode != Mode.Predict:
self.input = tf.placeholder(name='inputs', dtype=tf.int32, shape=[None, None])
self.pos_input = tf.placeholder(name='pos_input', dtype=tf.int32, shape=[None])
self.morph_input = tf.placeholder(name='morph_input', dtype=tf.int32, shape=[None, config.max_morph_tags])
else:
k, v = list(zip(*config.dec_char_numberer.value2num.items()))
i = KeyValueTensorInitializer(v, k,key_dtype=tf.int64,value_dtype=tf.string)
self.dec2char = HashTable(i, default_value="<UNK>")
self.enc_chars = self.lookup(config.enc_char_numberer)
self.morph_tags = self.lookup(config.morph_numberer)
self.pos_tags = self.lookup(config.pos_numberer)
self.input_plc =tf.placeholder(name='inputs', dtype=tf.string, shape=[None, None])
self.pos_plc = tf.placeholder(name='pos_input', dtype=tf.string, shape=[None])
self.morph_plc = tf.placeholder(name='morph_input', dtype=tf.string, shape=[None, config.max_morph_tags])
self.input = self.enc_chars.lookup(self.input_plc)
self.pos_input = self.pos_tags.lookup(self.pos_plc)
self.morph_input = self.morph_tags.lookup(self.morph_plc)
self.input_lengths = tf.placeholder(name='input_lengths', dtype=tf.int32, shape=[None])
self.enc_character_embeddings = tf.get_variable('enc_character_embeddings',
shape=[self.config.enc_vocab_size,
self.config.char_embedding_size])
self.dec_character_embeddings = tf.get_variable('dec_character_embeddings',
shape=[self.config.dec_vocab_size,
self.config.char_embedding_size])
self.pos_embeddings = tf.get_variable('pos_embeddings',
shape=[self.config.pos_vocab_size,
self.config.pos_embedding_size])
self.morph_embeddings = tf.get_variable('morph_embeddings',
shape=[self.config.morph_vocab_size,
self.config.morph_embedding_size])
if self._mode != Mode.Train:
self.start_tokens = tf.tile([config.start_idx], [tf.shape(self.input)[0]])
if self._mode != Mode.Predict:
# length +2 since lengths are <bow> + word + <eow>
self.dec_input = tf.placeholder(name='dec_in', dtype=tf.int32, shape=[None, None])
# exclude start token from targets for loss-computations since we feed the start token and don't want to
# predict it
self.decoder_targets = self.dec_input[:, 1:]
self.dec_lengths = tf.placeholder(name='dec_lengths', dtype=tf.int32, shape=[None])
def text_module_fn():
embeddings = [
("", [0, 0, 0, 0]), # OOV items are mapped to this embedding.
("hello world", [1, 2, 3, 4]),
("pair-programming", [5, 5, 5, 5]),
]
keys = tf.constant([item[0] for item in embeddings], dtype=tf.string)
indices = tf.constant(list(range(len(embeddings))), dtype=tf.int64)
tbl_init = KeyValueTensorInitializer(keys, indices)
table = HashTable(tbl_init, 0)
weights_initializer = tf.cast(
tf.constant(list([item[1] for item in embeddings])), tf.float32)
weights = tf_v1.get_variable(
"weights", dtype=tf.float32, initializer=weights_initializer)
text_tensor = tf_v1.placeholder(dtype=tf.string, name="text", shape=[None])
indices_tensor = table.lookup(text_tensor)
embedding_tensor = tf.gather(weights, indices_tensor)
hub.add_signature(inputs=text_tensor, outputs=embedding_tensor)
def input_fn():
with tf.variable_scope("input_fn"), tf.device("/cpu:0"):
caption_dataset = Dataset.from_tensor_slices(list(captions))
filename_dataset = Dataset.from_tensor_slices(list(filenames))
table_init = KeyValueTensorInitializer(
list(self.word_to_idx.keys()),
list(self.word_to_idx.values()),
key_dtype=tf.string,
value_dtype=tf.int32)
table = HashTable(table_init, default_value=0)
def split_sentence(sentence):
words = tf.string_split(tf.reshape(sentence, (1, ))).values
words = tf.concat([
tf.constant(["<START>"]), words,
tf.constant(["<END>"])
],
axis=0)
return table.lookup(words)
index_dataset = caption_dataset.map(split_sentence,
num_threads=8)
def decode_image(filename):
image = tf.image.decode_jpeg(tf.read_file(filename),
channels=3)
# image = tf.image.resize_images(image, [224, 224])
image = tf.to_float(image)
return image
image_dataset = filename_dataset.map(decode_image,
num_threads=8)
caption_structure = {
"raw": caption_dataset,
"index": index_dataset
}
return image_dataset, caption_structure
def __init__(self, args, session, updates=None):
self.args = args
self.sess = session
# updates
if not updates:
updates = 0
self.updates = updates
self.global_step = tf.get_variable(
'global_step',
shape=(),
dtype=tf.float32,
initializer=tf.constant_initializer(updates),
trainable=False)
self.step = tf.assign_add(self.global_step, 1)
# placeholders
table = HashTable(TextFileIdTableInitializer(
filename=os.path.join(args.output_dir, 'vocab.txt')),
default_value=Vocab.unk())
self.q1_string = tf.placeholder(tf.string, [None, None], name='q1_str')
self.q2_string = tf.placeholder(tf.string, [None, None], name='q2_str')
self.q1 = tf.placeholder_with_default(table.lookup(self.q1_string),
[None, None],
name='q1')
self.q2 = tf.placeholder_with_default(table.lookup(self.q2_string),
[None, None],
name='q2')
self.q1_len = tf.placeholder(tf.int32, [None], name='q1_len')
self.q2_len = tf.placeholder(tf.int32, [None], name='q2_len')
self.y = tf.placeholder(tf.int32, [None], name='y')
self.dropout_keep_prob = tf.placeholder(tf.float32, (),
name='dropout_keep_prob')
q1_mask = tf.expand_dims(tf.sequence_mask(self.q1_len,
dtype=tf.float32),
dim=-1)
q2_mask = tf.expand_dims(tf.sequence_mask(self.q2_len,
dtype=tf.float32),
dim=-1)
devices = self.get_available_gpus() or ['/device:CPU:0']
if not args.multi_gpu:
devices = devices[:1]
if len(devices) == 1:
splits = 1
else:
splits = [tf.shape(self.q1)[0] // len(devices)
] * (len(devices) - 1) + [-1] # handle uneven split
q1 = tf.split(self.q1, splits)
q2 = tf.split(self.q2, splits)
q1_mask = tf.split(q1_mask, splits)
q2_mask = tf.split(q2_mask, splits)
y = tf.split(self.y, splits)
# network
self.network = Network(args)
# optimizer
lr = tf.get_variable('lr',
shape=(),
dtype=tf.float32,
initializer=tf.constant_initializer(args.lr),
trainable=False)
lr_next = tf.cond(self.global_step < args.lr_warmup_steps,
true_fn=lambda: args.min_lr +
(args.lr - args.min_lr) / max(
1, args.lr_warmup_steps) * self.global_step,
false_fn=lambda: tf.maximum(
args.min_lr,
args.lr * args.lr_decay_rate**tf.floor(
(self.global_step - args.lr_warmup_steps
) / args.lr_decay_steps)))
tf.add_to_collection(tf.GraphKeys.UPDATE_OPS,
tf.assign(lr, lr_next, name='update_lr'))
self.lr = lr
self.opt = tf.train.AdamOptimizer(learning_rate=lr,
beta1=args.beta1,
beta2=args.beta2)
# training graph
tower_names = ['tower-{}'.format(i) for i in range(len(devices))
] if len(devices) > 1 else ['']
tower_logits = []
tower_grads = []
summaries = []
loss = 0
with tf.variable_scope(tf.get_variable_scope(), reuse=tf.AUTO_REUSE):
for i, device in enumerate(devices):
with tf.device(device):
with tf.name_scope(tower_names[i]) as scope:
logits = self.network(q1[i], q2[i], q1_mask[i],
q2_mask[i],
self.dropout_keep_prob)
tower_logits.append(logits)
loss = self.get_loss(logits, y[i])
tf.get_variable_scope().reuse_variables()
summaries = tf.get_collection(tf.GraphKeys.SUMMARIES,
scope)
grads = self.opt.compute_gradients(loss)
tower_grads.append(grads)
gradients = []
variables = []
for grad_and_vars in zip(*tower_grads):
if grad_and_vars[0][0] is None:
msg = 'WARNING: trainable variable {} receives no grad.\n'.format(
grad_and_vars[0][1].op.name)
sys.stderr.write(msg)
continue
grad = tf.stack([g for g, _ in grad_and_vars])
grad = tf.reduce_mean(grad, 0)
v = grad_and_vars[0][
1] # use the first tower's pointer to the (shared) variable
gradients.append(grad)
variables.append(v)
gradients, self.gnorm = tf.clip_by_global_norm(gradients,
self.args.grad_clipping)
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
with tf.control_dependencies(update_ops):
self.train_op = self.opt.apply_gradients(zip(gradients, variables))
logits = tf.concat(tower_logits, 0)
self.prob = tf.nn.softmax(logits, dim=1, name='prob')
self.pred = tf.argmax(input=logits, axis=1, name='pred')
self.loss = tf.identity(loss, name='loss')
summaries.append(tf.summary.scalar('training/lr', lr))
summaries.append(tf.summary.scalar('training/gnorm', self.gnorm))
summaries.append(tf.summary.scalar('training/loss', self.loss))
# add summary
self.summary = tf.summary.merge(summaries)
# saver
self.saver = tf.train.Saver(
[var for var in tf.global_variables() if 'Adam' not in var.name],
max_to_keep=args.max_checkpoints)
def __init__(self, hparam, export=False, with_char=False):
self.hparam = hparam
self.Embedding = None
self.embed_matrix = None
self.init_embedding = None
if export:
table = HashTable(TextFileIdTableInitializer(
filename='assets/vocab',
key_column_index=0,
value_column_index=1,
vocab_size=None,
delimiter='\t',
name='table'),
default_value=0)
self.q1_string = tf.placeholder(tf.string,
[None, self.hparam.seq_length],
name='string_input1')
self.q2_string = tf.placeholder(tf.string,
[None, self.hparam.seq_length],
name='string_input2')
self.premise = table.lookup(self.q1_string)
self.hypothesis = table.lookup(self.q2_string)
if with_char:
char_table = HashTable(TextFileIdTableInitializer(
filename='assets/char_vocab',
key_column_index=0,
value_column_index=1,
vocab_size=None,
delimiter='\t',
name='char_table'),
default_value=0)
self.ch1_string = tf.placeholder(
tf.string, [None, self.hparam.char_seq_length],
name='char_string_input1')
self.ch2_string = tf.placeholder(
tf.string, [None, self.hparam.char_seq_length],
name='char_string_input1')
self.ch1 = char_table.lookup(self.ch1_string)
self.ch2 = char_table.lookup(self.ch2_string)
else:
self.ch1, self.ch2 = None, None
else:
self.premise = tf.placeholder(tf.int32,
[None, self.hparam.seq_length],
'premise')
self.hypothesis = tf.placeholder(tf.int32,
[None, self.hparam.seq_length],
'hypothesis')
self.ch1, self.ch2 = None, None
self.premise_mask, self.hypothesis_mask = None, None
self.y = None
self.pred = None
self.logits = None
self.dropout_keep_prob = None
self.loss = None
self.train_op = None
self.is_training = None
self.output_w = None
self.output_b = None
self.lsf_q1, self.lsf_q2 = None, None
self.random_size = None
def lookup(self, numberer):
k, v = list(zip(*numberer.value2num.items()))
i = KeyValueTensorInitializer(k, v)
return HashTable(i, default_value=numberer.unknown_idx)
def __init__(self, args, session, updates=None):
self.args = args
self.sess = session
# updates
if not updates:
updates = 0
# tf.get_variable(): Make the variable be shared. If tf.Variable() is used, the system treat the variables with the same name as
# two vairables
self.global_step = tf.get_variable(
'global_step',
shape=(),
dtype=tf.float32,
initializer=tf.constant_initializer(updates),
trainable=False)
self.step = tf.assign_add(self.global_step,
1) # update self.global by add 1 to it
# placeholders
table = HashTable(
TextFileIdTableInitializer(
filename=os.path.join(args.output_dir, 'vocab.txt')),
default_value=Vocab.unk(
)) # string to id table, generates one key-value pair per line
self.q1_string = tf.placeholder(tf.string, [None, None], name='q1_str')
self.q2_string = tf.placeholder(tf.string, [None, None], name='q2_str')
self.q1 = tf.placeholder_with_default(table.lookup(self.q1_string),
[None, None],
name='q1')
self.q2 = tf.placeholder_with_default(table.lookup(self.q2_string),
[None, None],
name='q2')
self.q1_len = tf.placeholder(tf.int32, [None], name='q1_len')
self.q2_len = tf.placeholder(tf.int32, [None], name='q2_len')
self.y = tf.placeholder(tf.int32, [None], name='y')
self.dropout_keep_prob = tf.placeholder(tf.float32, (),
name='dropout_keep_prob')
q1_mask = tf.expand_dims(tf.sequence_mask(self.q1_len,
dtype=tf.float32),
dim=-1) # 返回一个表示每个单元的前N个位置的mask张量
q2_mask = tf.expand_dims(tf.sequence_mask(self.q2_len,
dtype=tf.float32),
dim=-1)
devices = self.get_available_gpus() or ['/device:CPU:0']
if not args.multi_gpu:
devices = devices[:1]
if len(devices) == 1:
splits = 1
else:
splits = [tf.shape(self.q1)[0] // len(devices)
] * (len(devices) - 1) + [-1] # handle uneven split
q1 = tf.split(self.q1, splits)
q2 = tf.split(self.q2, splits)
q1_mask = tf.split(q1_mask, splits)
q2_mask = tf.split(q2_mask, splits)
y = tf.split(self.y, splits)
# network
self.network = Network(args)
# optimizer
lr = tf.get_variable('lr',
shape=(),
dtype=tf.float32,
initializer=tf.constant_initializer(args.lr),
trainable=False)
lr_next = tf.cond(self.global_step < args.lr_warmup_steps,
true_fn=lambda: args.min_lr +
(args.lr - args.min_lr) / max(
1, args.lr_warmup_steps) * self.global_step,
false_fn=lambda: tf.maximum(
args.min_lr,
args.lr * args.lr_decay_rate**tf.floor(
(self.global_step - args.lr_warmup_steps
) / args.lr_decay_steps)))
# 在一个计算图中，可以通过集合（collection）来管理不同类别的资源
# tf.GraphKeys.UPDATE_OPS: ops的集合(图表运行时执行的操作,如乘法,ReLU等),而不是变量
tf.add_to_collection(tf.GraphKeys.UPDATE_OPS,
tf.assign(lr, lr_next, name='update_lr'))
self.lr = lr
self.opt = tf.train.AdamOptimizer(learning_rate=lr,
beta1=args.beta1,
beta2=args.beta2)
# training graph
tower_names = ['tower-{}'.format(i) for i in range(len(devices))
] if len(devices) > 1 else ['']
tower_logits = []
tower_grads = []
summaries = []
loss = 0
# 当系统检测到我们用了一个之前已经定义的变量时，就开启共享，否则就重新创建变量
with tf.variable_scope(tf.get_variable_scope(), reuse=tf.AUTO_REUSE):
for i, device in enumerate(devices):
with tf.device(device):
with tf.name_scope(tower_names[i]) as scope:
logits = self.network(q1[i], q2[i], q1_mask[i],
q2_mask[i],
self.dropout_keep_prob)
tower_logits.append(logits)
loss = self.get_loss(logits, y[i])
tf.get_variable_scope().reuse_variables()
summaries = tf.get_collection(tf.GraphKeys.SUMMARIES,
scope)
grads = self.opt.compute_gradients(loss)
tower_grads.append(grads)
gradients = []
variables = []
for grad_and_vars in zip(*tower_grads): # *list: 将一组行转换为一组列
if grad_and_vars[0][0] is None:
msg = 'WARNING: trainable variable {} receives no grad.\n'.format(
grad_and_vars[0][1].op.name)
sys.stderr.write(msg) # output error msg to window
continue
grad = tf.stack([g for g, _ in grad_and_vars])
grad = tf.reduce_mean(grad, 0)
v = grad_and_vars[0][
1] # use the 1st tower's pointer to the (shared) variable
gradients.append(grad)
variables.append(v)
gradients, self.gnorm = tf.clip_by_global_norm(gradients,
self.args.grad_clipping)
"""
tf.clip_by_global_norm(t_list, clip_norm, use_norm=None, name=None)
t_list 是梯度张量， clip_norm 是截取的比率，和上面的 clip_gradient是同一个东西。 这个函数返回截取过的梯度张量和一个所有张量的全局范数。
这样就保证了在一次迭代更新中，所有权重的梯度的平方和在一个设定范围以内
"""
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
"""
该函数接受的参数control_inputs，是Operation或者Tensor构成的list。返回的是一个上下文管理器，该上下文管理器用来控制在该上下文中的操作的依赖。
也就是说，上下文管理器下定义的操作是依赖control_inputs中的操作的，control_dependencies用来控制control_inputs中操作执行后，
才执行上下文管理器中定义的操作
"""
with tf.control_dependencies(update_ops): # 先执行update_ops
self.train_op = self.opt.apply_gradients(zip(gradients, variables))
logits = tf.concat(tower_logits, 0)
self.prob = tf.nn.softmax(logits, dim=1, name='prob')
self.pred = tf.argmax(input=logits, axis=1, name='pred')
self.loss = tf.identity(loss, name='loss')
summaries.append(tf.summary.scalar('training/lr', lr))
summaries.append(tf.summary.scalar('training/gnomr', self.gnorm))
summaries.append(tf.summary.scalar('training/loss', self.loss))
# add summary
self.summary = tf.summary.merge(summaries)
# saver
self.saver = tf.train.Saver(
[var for var in tf.global_variables() if 'Adam' not in var.name],
max_to_keep=args.max_checkpoints)
