def _opt_fd():
_od = maybe_call(optimization_step_feed_dict, maybe_eval(self._global_step)) \
if optimization_step_feed_dict else {} # e.g. hyper-learning rate is a placeholder
_oo_fd = maybe_call(outer_objective_feed_dicts, maybe_eval(self._global_step)) \
if outer_objective_feed_dicts else {} # this is used in ForwardHG. In ReverseHG should't be needed
# but it doesn't matter
return merge_dicts(_od, _oo_fd)
def run(self, T_or_generator, inner_objective_feed_dicts=None, outer_objective_feed_dicts=None,
initializer_feed_dict=None, global_step=None, session=None, online=False):
ss = session or tf.get_default_session()
self._history.clear()
if not online:
self._history.append(ss.run(self.initialization, feed_dict=utils.maybe_call(
initializer_feed_dict, utils.maybe_eval(global_step, ss))))
for t in range(T_or_generator) if isinstance(T_or_generator, int) else T_or_generator:
self._history.append(ss.run(self.iteration,
feed_dict=utils.maybe_call(inner_objective_feed_dicts, t)))
# initialization of support variables (supports stochastic evaluation of outer objective via global_step ->
# variable)
ss.run(self._reverse_initializer, feed_dict=utils.maybe_call(outer_objective_feed_dicts,
utils.maybe_eval(global_step, ss)))
for pt, state_feed_dict in self._state_feed_dict_generator(reversed(self._history[:-1])):
# this should be fine also for truncated reverse... but check again the index t
t = len(self._history) - pt - 2 # if T is int then len(self.history) is T + 1 and this numerator
# shall start at T-1 (99.99 sure its correct)
ss.run(self._alpha_iter,
feed_dict=utils.merge_dicts(state_feed_dict,
utils.maybe_call(inner_objective_feed_dicts, t)
if inner_objective_feed_dicts else {}))
def _opt_fd():
_od = maybe_call(optimization_step_feed_dict, maybe_eval(self._global_step)) \
if optimization_step_feed_dict else {} # e.g. hyper-learning rate is a placeholder
_oo_fd = maybe_call(outer_objective_feed_dicts, maybe_eval(self._global_step)) \
if outer_objective_feed_dicts else {} # this is used in ForwardHG. In ReverseHG should't be needed
# but it doesn't matter
return merge_dicts(_od, _oo_fd)
def run(self,
T_or_generator,
inner_objective_feed_dicts=None,
outer_objective_feed_dicts=None,
initializer_feed_dict=None,
global_step=None,
session=None,
online=False,
callback=None):
ss = session or tf.get_default_session()
if online == False:
self._run_batch_initialization(
ss,
utils.maybe_call(initializer_feed_dict,
utils.maybe_eval(global_step, ss)))
elif online == 'wr':
self._run_wr_initialization(
ss,
utils.maybe_call(initializer_feed_dict,
utils.maybe_eval(global_step, ss)))
for t in utils.solve_int_or_generator(T_or_generator):
_fd = utils.maybe_call(inner_objective_feed_dicts, t)
self._forward_step(ss, _fd)
utils.maybe_call(callback, t, _fd, ss)
def run(self,
T_or_generator,
inner_objective_feed_dicts=None,
outer_objective_feed_dicts=None,
initializer_feed_dict=None,
global_step=None,
session=None,
online=False,
forward_callback=None):
ss = session or tf.get_default_session()
self._history.clear()
if not online:
_fd = utils.maybe_call(initializer_feed_dict,
utils.maybe_eval(global_step, ss))
self._history.append(ss.run(self.initialization, feed_dict=_fd))
else:
self._history.append(ss.run(self.state))
# optionally may track inner objective (to check for divergence)
for t in range(T_or_generator) if utils.isinteger(
T_or_generator) else T_or_generator:
_fd = utils.maybe_call(inner_objective_feed_dicts, t)
self._history.append(ss.run(self.iteration, feed_dict=_fd))
if forward_callback is not None:
forward_callback(t, _fd, ss)
# initialization of support variables (supports stochastic evaluation of outer objective via global_step ->
# variable)
# TODO (maybe tf bug or oddity) for some strange reason, if some variable's initializer depends on
# a placeholder, then the initializer of alpha SEEMS TO DEPEND ALSO ON THAT placeholder,
# as if the primary variable should be reinitialized as well, but, I've checked, the primary variable is NOT
# actually reinitialized. This doesn't make sense since the primary variable is already initialized
# and Tensorflow seems not to care... should maybe look better into this issue
reverse_init_fd = utils.maybe_call(outer_objective_feed_dicts,
utils.maybe_eval(global_step, ss))
# now adding also the initializer_feed_dict because of tf quirk...
maybe_init_fd = utils.maybe_call(initializer_feed_dict,
utils.maybe_eval(global_step, ss))
if maybe_init_fd is not None:
reverse_init_fd = utils.merge_dicts(reverse_init_fd, maybe_init_fd)
ss.run(self._reverse_initializer, feed_dict=reverse_init_fd)
for pt, state_feed_dict in self._state_feed_dict_generator(
reversed(self._history[:-1])):
# this should be fine also for truncated reverse... but check again the index t
t = len(
self._history
) - pt - 2 # if T is int then len(self.history) is T + 1 and this numerator
# shall start at T-1
ss.run(self._alpha_iter,
feed_dict=utils.merge_dicts(
state_feed_dict,
utils.maybe_call(inner_objective_feed_dicts, t)
if inner_objective_feed_dicts else {}))
def run(self,
T_or_generator,
inner_objective_feed_dicts=None,
outer_objective_feed_dicts=None,
initializer_feed_dict=None,
global_step=None,
session=None,
online=False,
callback=None):
# callback may be a pair, first for froward pass, second for reverse pass
callback = utils.as_tuple_or_list(callback)
# same thing for T
T_or_generator = utils.as_tuple_or_list(T_or_generator)
ss = session or tf.get_default_session()
del self._history[:]
if not online:
_fd = utils.maybe_call(initializer_feed_dict,
utils.maybe_eval(global_step, ss))
self._save_history(ss.run(self.initialization, feed_dict=_fd))
# else: # not totally clear if i should add this
# self._save_history(ss.run(list(self.state)))
T = 0 # this is useful if T_or_generator is indeed a generator...
for t in utils.solve_int_or_generator(T_or_generator[0]):
# nonlocal t # with nonlocal would not be necessary the variable T... not compatible with 2.7
_fd = utils.maybe_call(inner_objective_feed_dicts, t)
self._save_history(ss.run(self.iteration, feed_dict=_fd))
utils.maybe_call(callback[0], t, _fd, ss)
T = t
# initialization of support variables (supports stochastic evaluation of outer objective via global_step ->
# variable)
# TODO (maybe tf bug or oddity) for some strange reason, if some variable's initializer depends on
# a placeholder, then the initializer of alpha SEEMS TO DEPEND ALSO ON THAT placeholder,
# as if the primary variable should be reinitialized as well, but, I've checked, the primary variable is NOT
# actually reinitialized. This doesn't make sense since the primary variable is already initialized
# and Tensorflow seems not to care... should maybe look better into this issue
reverse_init_fd = utils.maybe_call(outer_objective_feed_dicts,
utils.maybe_eval(global_step, ss))
# now adding also the initializer_feed_dict because of tf quirk...
maybe_init_fd = utils.maybe_call(initializer_feed_dict,
utils.maybe_eval(global_step, ss))
reverse_init_fd = utils.merge_dicts(reverse_init_fd, maybe_init_fd)
ss.run(self._reverse_initializer, feed_dict=reverse_init_fd)
for pt, state_feed_dict in self._state_feed_dict_generator(
reversed(self._history[:-1]), T_or_generator[-1]):
# this should be fine also for truncated reverse... but check again the index t
t = T - pt - 1 # if T is int then len(self.history) is T + 1 and this numerator
# shall start at T-1
_fd = utils.merge_dicts(
state_feed_dict, utils.maybe_call(inner_objective_feed_dicts,
t))
ss.run(self._alpha_iter, _fd)
if len(callback) == 2: utils.maybe_call(callback[1], t, _fd, ss)
def run(self,
T_or_generator,
inner_objective_feed_dicts=None,
outer_objective_feed_dicts=None,
initializer_feed_dict=None,
global_step=None,
session=None,
online=False,
forward_callback=None):
ss = session or tf.get_default_session()
if not online:
ss.run(self.initialization,
feed_dict=utils.maybe_call(
initializer_feed_dict,
utils.maybe_eval(global_step, ss)))
ss.run(self._forward_initializer)
# optionally may track inner objective (to check for divergence)
for t in range(T_or_generator) if utils.isinteger(
T_or_generator) else T_or_generator:
_fd = utils.maybe_call(inner_objective_feed_dicts, t)
ss.run(self._z_iter, _fd)
ss.run(self.iteration, _fd)
if forward_callback is not None:
forward_callback(t, _fd, ss)
def run(self,
T_or_generator,
inner_objective_feed_dicts=None,
outer_objective_feed_dicts=None,
initializer_feed_dict=None,
global_step=None,
session=None,
online=False):
ss = session or tf.get_default_session()
if not online:
ss.run(self.initialization,
feed_dict=utils.maybe_call(
initializer_feed_dict,
utils.maybe_eval(global_step, ss)))
ss.run(self._forward_initializer)
for t in range(T_or_generator) if isinstance(T_or_generator,
int) else T_or_generator:
# ss.run()
ss.run(self._z_iter, utils.maybe_call(inner_objective_feed_dicts,
t))
ss.run(self.iteration,
utils.maybe_call(inner_objective_feed_dicts, t))
def _adjust_step(_t):
if online:
_T = utils.maybe_eval(global_step, ss)
if _T is None:
_T = 0
tot_t = T_or_generator[0]
if not isinstance(tot_t, int): return _t # when using a generator there is little to do...
return int(_t + tot_t*_T)
else: return _t
def run(self,
T_or_generator,
inner_objective_feed_dicts=None,
outer_objective_feed_dicts=None,
initializer_feed_dict=None,
global_step=None,
session=None,
online=False):
ss = session or tf.get_default_session()
self._history = []
if not online:
self._history.append(
ss.run(self.initialization,
feed_dict=utils.maybe_call(
initializer_feed_dict,
utils.maybe_eval(global_step, ss))))
for t in range(T_or_generator) if isinstance(T_or_generator,
int) else T_or_generator:
self._history.append(
ss.run(self.iteration,
feed_dict=utils.maybe_call(inner_objective_feed_dicts,
t)))
# initialization of support variables (supports stochastic evaluation of outer objective via global_step ->
# variable)
ss.run(self._reverse_initializer,
feed_dict=utils.maybe_call(outer_objective_feed_dicts,
utils.maybe_eval(global_step, ss)))
for pt, state_feed_dict in self._state_feed_dict_generator(
reversed(self._history[:-1])):
# this should be fine also for truncated reverse... but check again the index t
t = len(
self._history
) - pt - 2 # if T is int then len(self.history) is T + 1 and this numerator
# shall start at T-1 (99.99 sure its correct)
ss.run(self._alpha_iter,
feed_dict=utils.merge_dicts(
state_feed_dict,
utils.maybe_call(inner_objective_feed_dicts, t)
if inner_objective_feed_dicts else {}))
def run(self,
T_or_generator,
inner_objective_feed_dicts=None,
outer_objective_feed_dicts=None,
initializer_feed_dict=None,
global_step=None,
session=None,
online=False,
callback=None):
ss = session or tf.get_default_session()
inner_objective_feed_dicts = utils.as_tuple_or_list(
inner_objective_feed_dicts)
if not online:
self._run_batch_initialization(
ss,
utils.maybe_call(initializer_feed_dict,
utils.maybe_eval(global_step, ss)))
for t in utils.solve_int_or_generator(T_or_generator):
_fd = utils.maybe_call(inner_objective_feed_dicts[0], t)
self._forward_step(ss, _fd)
utils.maybe_call(callback, t, _fd, ss)
# end of optimization. Solve linear systems.
tol_val = utils.maybe_call(self.tolerance,
utils.maybe_eval(
global_step,
ss)) # decreasing tolerance (seq.)
# feed dictionaries (could...in theory, implement stochastic solution of this linear system...)
_fd = utils.maybe_call(inner_objective_feed_dicts[-1], -1)
_fd_outer = utils.maybe_call(outer_objective_feed_dicts,
utils.maybe_eval(global_step, ss))
_fd = utils.merge_dicts(_fd, _fd_outer)
for lin_sys in self._lin_sys:
lin_sys(tol_val).minimize(
ss, _fd) # implicitly warm restarts with previously found q
def run(self, T_or_generator, inner_objective_feed_dicts=None, outer_objective_feed_dicts=None,
initializer_feed_dict=None, global_step=None, session=None, online=False):
ss = session or tf.get_default_session()
if not online:
ss.run(self.initialization, feed_dict=utils.maybe_call(
initializer_feed_dict, utils.maybe_eval(global_step, ss)))
ss.run(self._forward_initializer)
for t in range(T_or_generator) if isinstance(T_or_generator, int) else T_or_generator:
# ss.run()
ss.run(self._z_iter, utils.maybe_call(inner_objective_feed_dicts, t))
ss.run(self.iteration, utils.maybe_call(inner_objective_feed_dicts, t))
