def summary(self, feeds=None):
from ..scalar import current_step
if self.as_tensor() is None:
return
value = get_default_session().run(self.as_tensor(),
self.get_feed_dict(feeds))
self.nodes['summary_writer'].add_summary(value, current_step())
def summary(self, feeds=None):
from ..scalar import current_step
mrf = self._get_multiple_run_feeds(feeds)
feed_dict = self.get_feed_dict(feeds)
feed_dict.update(mrf)
value = get_default_session().run(self.nodes[SummaryKeys.MERGED],
feed_dict=feed_dict)
self.nodes['summary_writer'].add_summary(value, current_step())
def save(self, feeds):
from ..scalar import global_step
if self._saver is None:
self._saver = tf.train.Saver()
from dxpy.learn.session import get_default_session
sess = get_default_session()
step = sess.run(global_step())
print("[SAVE] model to: {}.".format(self._model_path()))
self._saver.save(sess, self._model_path(), global_step=step)
def _get_multiple_run_feeds(self, feeds=None):
result = dict()
for i in range(self.nb_max_runs):
current_result = get_default_session().run(
self.inputs, self.get_feed_dict(feeds))
for k in self.inputs:
if k in self.nb_runs and self.nb_runs[k] > i:
result[self.multi_runs[k][i]] = current_result[k]
return result
def load(self, feeds):
from ..scalar import global_global_step
import sys
if self._saver is None:
self._saver = tf.train.Saver()
from dxpy.learn.session import get_default_session
sess = get_default_session()
path_load, flag = self.__resolve_path_load(feeds)
if flag is False:
if isinstance(path_load, int):
msg = "[ERROR][LOAD] Save for given step {} not found. Skip restore."
print(msg.format(path_load), file=sys.stderr)
return
else:
msg = "[ERROR][LOAD] Checkpoint file {} not found. Skip restore."
print(msg.format(path_load), file=sys.stderr)
return
print("[LOAD] model from: {}.".format(path_load))
self._saver.restore(sess, path_load)
def __create_writer(self, feeds):
self.register_node(
'summary_writer',
tf.summary.FileWriter(self.param('path', feeds),
get_default_session().graph))
def set_value(self, feeds):
from dxpy.learn.session import get_default_session
get_default_session().run(self.assign_op,
feed_dict={self.nodes['new_value']: feeds})
def get_value():
from dxpy.learn.session import get_default_session
return get_default_session().run(global_step())
def _train(self, feeds):
sess = get_default_session()
sess.run(self.as_tensor(), feed_dict=feeds)
def get_value(self):
from dxpy.learn.session import get_default_session
return get_default_session().run(self.as_tensor())
def session(self):
from dxpy.learn.session import get_default_session
return get_default_session()
def get_value():
from dxpy.learn.session import get_default_session
return get_default_session().run(keep_prob())
