|
-
- """
- A MultistepAdamWeightDecayOptimizer can use larger batch_size in BERT
- which updates var after n steps
-
- """
-
- from __future__ import absolute_import
- from __future__ import division
- from __future__ import print_function
-
- import re
- import tensorflow as tf
- from tensorflow.python.training import optimizer
- from tensorflow.python.framework import ops
-
- def create_optimizer(loss, init_lr, num_train_steps, num_warmup_steps, use_tpu):
- """Creates an optimizer training op."""
- global_step = tf.train.get_or_create_global_step()
-
- learning_rate = tf.constant(value=init_lr, shape=[], dtype=tf.float32)
-
- # Implements linear decay of the learning rate.
- learning_rate = tf.train.polynomial_decay(
- learning_rate,
- global_step,
- num_train_steps,
- end_learning_rate=0.0,
- power=1.0,
- cycle=False)
-
- # Implements linear warmup. I.e., if global_step < num_warmup_steps, the
- # learning rate will be `global_step/num_warmup_steps * init_lr`.
- if num_warmup_steps:
- global_steps_int = tf.cast(global_step, tf.int32)
- warmup_steps_int = tf.constant(num_warmup_steps, dtype=tf.int32)
-
- global_steps_float = tf.cast(global_steps_int, tf.float32)
- warmup_steps_float = tf.cast(warmup_steps_int, tf.float32)
-
- warmup_percent_done = global_steps_float / warmup_steps_float
- warmup_learning_rate = init_lr * warmup_percent_done
-
- is_warmup = tf.cast(global_steps_int < warmup_steps_int, tf.float32)
- learning_rate = (
- (1.0 - is_warmup) * learning_rate + is_warmup * warmup_learning_rate)
-
- # It is recommended that you use this optimizer for fine tuning, since this
- # is how the model was trained (note that the Adam m/v variables are NOT
- # loaded from init_checkpoint.)
- accoptimizer = MultistepAdamWeightDecayOptimizer(
- learning_rate=learning_rate,
- weight_decay_rate=0.01,
- beta_1=0.9,
- beta_2=0.999,
- epsilon=1e-6,
- n=8,
- exclude_from_weight_decay=["LayerNorm", "layer_norm", "bias"])
-
- if use_tpu:
- accoptimizer = tf.contrib.tpu.CrossShardOptimizer(accoptimizer)
-
- tvars = tf.trainable_variables()
- grads = tf.gradients(loss, tvars)
-
- # This is how the model was pre-trained.
- (grads, _) = tf.clip_by_global_norm(grads, clip_norm=1.0)
- # print(grads)
-
- train_op = accoptimizer.apply_gradients(
- zip(grads, tvars), global_step=global_step)
-
- # Normally the global step update is done inside of `apply_gradients`.
- # However, `AdamWeightDecayOptimizer` doesn't do this. But if you use
- # a different optimizer, you should probably take this line out.
- new_global_step = global_step + 1
- print_op = tf.print(new_global_step,'----',learning_rate,'----',grads[-2][:5])
- with tf.control_dependencies([print_op]):
- train_op = tf.group(train_op, [global_step.assign(new_global_step)])
-
- return train_op
-
-
- class MultistepAdamWeightDecayOptimizer(optimizer.Optimizer):
- """A Multistep Adam optimizer that includes "correct" L2 weight decay."""
-
- def __init__(self,
- learning_rate=2e-5,
- weight_decay_rate=0.01,
- beta_1=0.9,
- beta_2=0.999,
- epsilon=1e-6,
- n=1, ##n steps per update
- exclude_from_weight_decay=None,
- name="MultistepAdamWeightDecayOptimizer"):
- """Constructs a MultistepAdamWeightDecayOptimizer."""
- super(MultistepAdamWeightDecayOptimizer, self).__init__(False, name)
-
- self.learning_rate = learning_rate
- self.weight_decay_rate = weight_decay_rate
- self.beta_1 = beta_1
- self.beta_2 = beta_2
- self.epsilon = epsilon
- self._n = n # Call Adam optimizer every n batches with accumulated grads
- self.exclude_from_weight_decay = exclude_from_weight_decay
-
- self._n_t = None # n as tensor
-
- def _prepare(self):
- super(MultistepAdamWeightDecayOptimizer, self)._prepare()
- self._n_t = tf.convert_to_tensor(self._n, name="n")
-
- def _create_slots(self, var_list):
- """Create slot variables for MultistepAdamWeightDecayOptimizer with accumulated gradients.
-
- Like super class method, but additionally creates slots for the gradient
- accumulator `grad_acc` and the counter variable.
- """
- super(MultistepAdamWeightDecayOptimizer, self)._create_slots(var_list)
- first_var = min(var_list, key=lambda x: x.name)
- self._create_non_slot_variable(initial_value=0 if self._n == 1 else 1,
- name="iter",
- colocate_with=first_var)
- for v in var_list:
- self._zeros_slot(v, "grad_acc", self._name)
-
- def _get_iter_variable(self):
- if tf.contrib.eager.in_eager_mode():
- graph = None
- else:
- graph = tf.get_default_graph()
- return self._get_non_slot_variable("iter", graph=graph)
-
- def apply_gradients(self, grads_and_vars, global_step=None, name=None):
- """See base class."""
- update_ops = []
-
- var_list = [v for g, v in grads_and_vars if g is not None]
-
- with ops.init_scope():
- self._create_slots(var_list)
- self._prepare()
-
- for (grad, param) in grads_and_vars:
- if grad is None or param is None:
- continue
- grad_acc = self.get_slot(param, "grad_acc")
- param_name = self._get_variable_name(param.name)
- m = tf.get_variable(name=param_name + "/adam_m",shape=param.shape.as_list(),
- dtype=tf.float32,trainable=False,initializer=tf.zeros_initializer())
- v = tf.get_variable(name=param_name + "/adam_v",shape=param.shape.as_list(),
- dtype=tf.float32,trainable=False,initializer=tf.zeros_initializer())
-
- ##apply adam for v
- def _apply_adam(grad_acc, grad, param, m, v):
- total_grad = (grad_acc + grad) / tf.cast(self._n_t, grad.dtype)
- # Standard Adam update.
- next_m = (
- tf.multiply(self.beta_1, m) + tf.multiply(1.0 - self.beta_1, total_grad))
- next_v = (
- tf.multiply(self.beta_2, v) + tf.multiply(1.0 - self.beta_2,
- tf.square(total_grad)))
- update = next_m / (tf.sqrt(next_v) + self.epsilon)
- if self._do_use_weight_decay(param_name):
- update += self.weight_decay_rate * param
- update_with_lr = self.learning_rate * update
- next_param = param - update_with_lr
- adam_op = tf.group(param.assign(next_param), m.assign(next_m),
- v.assign(next_v))
- with tf.control_dependencies([adam_op]):
- grad_acc_to_zero_op = grad_acc.assign(tf.zeros_like(grad_acc),
- use_locking=self._use_locking)
- return tf.group(adam_op, grad_acc_to_zero_op)
-
- ## accumulate gradients for var
- def _accumulate_gradient(grad_acc, grad):
- assign_op = tf.assign_add(grad_acc, grad, use_locking=self._use_locking)
- return tf.group(assign_op)
- ##apply adam or accumulate gradients for 'Var'
- update_op = tf.cond(tf.equal(self._get_iter_variable(), 0),
- lambda: _apply_adam(grad_acc, grad, param, m, v),
- lambda: _accumulate_gradient(grad_acc, grad))
- update_ops.append(update_op)
-
- ##do extra update ops for some var
- apply_updates = self._finish(update_ops, name_scope=name)
- return apply_updates
-
- def _finish(self, update_ops, name_scope):
- """
- iter <- iter + 1 mod n
- """
- iter_ = self._get_iter_variable()
- with tf.control_dependencies(update_ops):
- with tf.colocate_with(iter_):
- update_iter = iter_.assign(tf.mod(iter_ + 1, self._n_t),
- use_locking=self._use_locking)
- return tf.group(
- *update_ops+[update_iter], name=name_scope)
-
- def _do_use_weight_decay(self, param_name):
- """Whether to use L2 weight decay for `param_name`."""
- if not self.weight_decay_rate:
- return False
- if self.exclude_from_weight_decay:
- for r in self.exclude_from_weight_decay:
- if re.search(r, param_name) is not None:
- return False
- return True
-
- def _get_variable_name(self, param_name):
- """Get the variable name from the tensor name."""
- m = re.match("^(.*):\\d+$", param_name)
- if m is not None:
- param_name = m.group(1)
- return param_name
|