def _update_with_movers(self, feed_dict, feed_map):
# If a tensor handle that is fed to a device incompatible placeholder,
# we move the tensor to the right device, generate a new tensor handle,
# and update `feed_dict` to use the new handle.
handle_movers = []
for feed_name, val in feed_map.items():
mover = session_ops._get_handle_mover(self.graph, *val)
if mover:
handle_movers.append((feed_name, val[1], mover))
# Transfer a tensor to the right device if needed.
if not handle_movers:
return []
else:
feeds = {}
fetches = []
for _, handle, mover in handle_movers:
feeds[mover[0]] = handle
fetches.append(mover[1])
handles = self.run(fetches, feed_dict=feeds)
for handle_mover, handle in zip(handle_movers, handles):
np_val = np.array(handle.handle, dtype=np.object)
feed_dict[handle_mover[0]] = np_val
return handles
def _update_with_movers(self, feed_dict, feed_map):
# If a tensor handle that is fed to a device incompatible placeholder,
# we move the tensor to the right device, generate a new tensor handle,
# and update `feed_dict` to use the new handle.
handle_movers = []
for feed_name, val in feed_map.items():
mover = session_ops._get_handle_mover(self.graph, *val)
if mover:
handle_movers.append((feed_name, val[1], mover))
# Transfer a tensor to the right device if needed.
if not handle_movers:
return []
else:
feeds = {}
fetches = []
for _, handle, mover in handle_movers:
feeds[mover[0]] = handle
fetches.append(mover[1])
handles = self.run(fetches, feed_dict=feeds)
for handle_mover, handle in zip(handle_movers, handles):
np_val = np.array(handle.handle, dtype=np.object)
feed_dict[handle_mover[0]] = np_val
return handles
