Source code for tensorpack.tfutils.optimizer

# -*- coding: utf-8 -*-
# File: optimizer.py


from contextlib import contextmanager
import tensorflow as tf

from ..tfutils.common import get_tf_version_tuple
from ..compat import tfv1
from ..utils.develop import HIDE_DOC
from .gradproc import FilterNoneGrad, GradientProcessor

__all__ = ['apply_grad_processors', 'ProxyOptimizer',
           'PostProcessOptimizer', 'VariableAssignmentOptimizer',
           'AccumGradOptimizer']


[docs]class ProxyOptimizer(tfv1.train.Optimizer): """ A transparent proxy which delegates all methods of :class:`tf.train.Optimizer` """ def __init__(self, opt, name='ProxyOptimizer'): assert isinstance(opt, tfv1.train.Optimizer), opt super(ProxyOptimizer, self).__init__(False, name) self._opt = opt @HIDE_DOC def compute_gradients(self, *args, **kwargs): return self._opt.compute_gradients(*args, **kwargs) @HIDE_DOC def get_slot(self, *args, **kwargs): return self._opt.get_slot(*args, **kwargs) @HIDE_DOC def get_slot_names(self, *args, **kwargs): return self._opt.get_slot_names(*args, **kwargs) @HIDE_DOC def apply_gradients(self, *args, **kwargs): return self._opt.apply_gradients(*args, **kwargs)
[docs]def apply_grad_processors(opt, gradprocs): """ Wrapper around optimizers to apply gradient processors. Args: opt (tf.train.Optimizer): gradprocs (list[GradientProcessor]): gradient processors to add to the optimizer. Returns: a :class:`tf.train.Optimizer` instance which runs the gradient processors before updating the variables. """ assert isinstance(gradprocs, (list, tuple)), gradprocs for gp in gradprocs: assert isinstance(gp, GradientProcessor), gp class _ApplyGradientProcessor(ProxyOptimizer): def __init__(self, opt, gradprocs): self._gradprocs = gradprocs[:] super(_ApplyGradientProcessor, self).__init__(opt) def apply_gradients(self, grads_and_vars, global_step=None, name=None): g = self._apply(grads_and_vars) return self._opt.apply_gradients(g, global_step, name) def _apply(self, g): for proc in self._gradprocs: g = proc.process(g) return g return _ApplyGradientProcessor(opt, gradprocs)
[docs]class PostProcessOptimizer(ProxyOptimizer): """ An optimizer which applies some "post-processing operation" per variable (e.g. clipping, quantization) after the gradient update. """
[docs] def __init__(self, opt, func, colocate=True): """ Args: opt (tf.train.Optimizer): func (tf.Variable -> tf.Operation or None): the operation needed to perform for this variable after the gradient update. colocate (boolean): colocate the function with the variable. No effect since TF 1.13. """ super(PostProcessOptimizer, self).__init__(opt) self._func = func self._colocate = colocate
@HIDE_DOC def apply_gradients(self, grads_and_vars, global_step=None, name=None): update_op = super(PostProcessOptimizer, self).apply_gradients( grads_and_vars, global_step) ops = [] with tf.control_dependencies([update_op]): for _, var in grads_and_vars: with self._maybe_colocate(var): op = self._func(var) if op is not None: assert isinstance(op, tf.Operation), op ops.append(op) update_op = tf.group(update_op, *ops, name=name) return update_op @contextmanager def _maybe_colocate(self, var): G = tf.get_default_graph() if self._colocate and get_tf_version_tuple() <= (1, 12): with G.colocate_with(var): yield else: yield
[docs]class VariableAssignmentOptimizer(PostProcessOptimizer): """ An optimizer which assigns each variable a new value (e.g. clipping, quantization) after the gradient update. """
[docs] def __init__(self, opt, func): """ Args: opt (tf.train.Optimizer): func (tf.Variable -> tf.Tensor or None): the new value to be assigned to this variable after the gradient update. """ def f(v): t = func(v) if t is None: return t return tf.assign(v, t, use_locking=False).op super(VariableAssignmentOptimizer, self).__init__(opt, f)
[docs]class AccumGradOptimizer(ProxyOptimizer): """ An optimizer which accumulates gradients across :math:`k` :meth:`minimize` executions, and apply them together in every :math:`k` th :meth:`minimize` execution. This is roughly the same as using a :math:`k` times larger batch size plus a :math:`k` times larger learning rate, but uses much less memory. Note that this implementation may not support all models. E.g., it currently doesn't support sparse gradient update. This optimizer can be used in any TensorFlow code (with or without tensorpack). Example: .. code-block:: python from tensorpack.tfutils.optimizer import AccumGradOptimizer myopt = tf.train.GradientDescentOptimizer(0.01) myopt = AccumGradOptimizer(myopt, niter=5) train_op = myopt.minimize(loss) """
[docs] def __init__(self, opt, niter): """ Args: opt (tf.train.Optimizer): the underlying sub-optimizer. niter (int): number of iterations to accumulate gradients. """ super(AccumGradOptimizer, self).__init__(opt, 'AccumGrad') self._niter = int(niter)
def _create_accum_slots(self, var_list): slots = [] for v in var_list: # TODO an option to not colocate the accumulators with variables (to save more memory) s = self._zeros_slot(v, "accum", self._name) slots.append(s) return slots @HIDE_DOC def apply_gradients(self, grads_and_vars, global_step=None, name=None): grads_and_vars = FilterNoneGrad().process(grads_and_vars) vs = [] for g, v in grads_and_vars: assert isinstance(g, tf.Tensor) and isinstance(v, tf.Variable), \ "AccumGradOptimizer only works for dense update! " \ "Types of v and g are {} and {}".format(type(v), type(g)) vs.append(v) with tf.control_dependencies(None): slots = self._create_accum_slots(vs) slots_and_vars = [(s, gv[1]) for s, gv in zip(slots, grads_and_vars)] # Create the counter on the same device as the first variable. with tf.variable_scope(self._name), \ vs[0].graph.colocate_with(vs[0]): counter = tf.Variable( 0, name="counter", trainable=False, dtype=tf.int32) with tf.name_scope('AccumGradOptimizer'): ops = [] for s, gv in zip(slots, grads_and_vars): g, v = gv ops.append(s.assign_add(g)) update_counter = tf.assign_add(counter, 1, name='update_counter') update_slot_op = tf.group(update_counter, *ops, name='update_slot') def update_grad(): update_op = self._opt.apply_gradients(slots_and_vars) with tf.control_dependencies([update_op]): clear_ops = [tf.assign(s, tf.zeros_like(s)) for s in slots] return tf.group(*clear_ops, name='update_grad') pred = tf.equal(tf.mod(counter, self._niter), 0) with tf.control_dependencies([update_slot_op]): if name is None: name = 'cond_update_grad' op = tf.cond(pred, update_grad, tf.no_op) if global_step is not None: # Tensorpack maintains global_step by other means, # so this option is useless in tensorpack trainers. # But we include the implementation here for completeness global_step_increment = tf.assign_add(global_step, 1) op = tf.group(op, global_step_increment, name=name) else: op = tf.identity(op, name=name).op return op
if __name__ == '__main__': # run it with "python -m tensorpack.tfutils.optimizer" x = tf.get_variable('x', shape=[6]) cost = tf.reduce_sum(tf.abs(x), name='cost') opt = tf.train.GradientDescentOptimizer(0.01) opt = AccumGradOptimizer(opt, 5) min_op = opt.minimize(cost, global_step=tf.train.get_or_create_global_step()) sess = tf.Session() sess.run(tf.global_variables_initializer()) with sess.as_default(): for k in range(20): min_op.run() print(x.eval()) print(tf.train.get_or_create_global_step().eval())