def __init__(self,
model_id_node,
model_cam_dir_node,
iterator_input_nodes,
blank_id):
super(IdAndCameraDirConvLoss, self).__init__(
_conv11_to_vector(model_id_node),
_conv11_to_vector(model_cam_dir_node),
iterator_input_nodes,
blank_id)
def make_validation_epoch_callbacks(epoch_logger):
loss_monitor = MeanOverEpoch(
loss_node,
callbacks=[print_validation_loss,
SavesH5AtMinimum(model,
training_state,
basepath + '_best.h5'),
SavesAtMinimum(things_to_save, basepath + '_best.pkl')])
epoch_logger.subscribe_to('validation mean loss', loss_monitor)
id_loss_monitor = MeanOverEpoch(loss_node.id_loss, callbacks=[])
epoch_logger.subscribe_to('validation cross-entropy',
id_loss_monitor)
pose_error_monitor = MeanOverEpoch(loss_node.cam_dir_loss,
callbacks=[])
epoch_logger.subscribe_to('validation orientation error',
pose_error_monitor)
# TODO: support NamedTuples for Node.inputs
target_id_node = input_nodes[1]
assert_equal(target_id_node.output_format.shape, (-1, ))
assert_equal(str(target_id_node.output_symbol.dtype), 'int32')
misclassification = Misclassification(
_conv11_to_vector(model.id_layers[-1]),
target_id_node)
misclassification_monitor = MeanOverEpoch(misclassification,
callbacks=[])
epoch_logger.subscribe_to('validation misclassification',
misclassification_monitor)
return [id_loss_monitor,
pose_error_monitor,
misclassification_monitor,
loss_monitor]
