Source code for tensorpack.predict.multigpu

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


import tensorflow as tf
from ..utils import logger
from ..graph_builder.model_desc import InputDesc
from ..input_source import PlaceholderInput
from ..tfutils.tower import PredictTowerContext
from .base import OnlinePredictor

__all__ = ['MultiTowerOfflinePredictor',
           'DataParallelOfflinePredictor']


[docs]class MultiTowerOfflinePredictor(OnlinePredictor): """ A multi-tower multi-GPU predictor. It builds one predictor for each tower. """
[docs] def __init__(self, config, towers): """ Args: config (PredictConfig): the config to use. towers: a list of relative GPU id. """ assert len(towers) > 0 self.graph = config._maybe_create_graph() self.predictors = [] self.return_input = config.return_input with self.graph.as_default(): handles = [] input = PlaceholderInput() input.setup(config.inputs_desc) for idx, t in enumerate(towers): tower_name = 'tower' + str(t) with tf.variable_scope(tf.get_variable_scope(), reuse=idx > 0), \ tf.device('/gpu:{}'.format(t)), \ PredictTowerContext(tower_name): config.tower_func(*input.get_input_tensors()) handles.append(config.tower_func.towers[-1]) self.sess = config.session_creator.create_session() config.session_init.init(self.sess) for h in handles: input_tensors = h.get_tensors(config.input_names) output_tensors = h.get_tensors(config.output_names) self.predictors.append(OnlinePredictor( input_tensors, output_tensors, config.return_input, self.sess))
def _do_call(self, dp): # use the first tower for compatible PredictorBase interface return self.predictors[0]._do_call(dp)
[docs] def get_predictor(self, n): """ Returns: OnlinePredictor: the nth predictor on the nth tower. """ l = len(self.predictors) if n >= l: logger.warn("n > #towers, will assign predictor to GPU by round-robin") return [self.predictors[k % l] for k in range(n)]
[docs] def get_predictors(self): """ Returns: list[OnlinePredictor]: a list of predictor """ return self.predictors
[docs]class DataParallelOfflinePredictor(OnlinePredictor): """ A data-parallel predictor. It builds one predictor that utilizes all GPUs. Note that it doesn't split/concat inputs/outputs automatically. Instead, its inputs are: ``[input[0] in tower[0], input[1] in tower[0], ..., input[0] in tower[1], input[1] in tower[1], ...]`` Similar for the outputs. """
[docs] def __init__(self, config, towers): """ Args: config (PredictConfig): the config to use. towers: a list of relative GPU id. """ self.graph = config._maybe_create_graph() with self.graph.as_default(): input_tensors = [] output_tensors = [] for idx, t in enumerate(towers): tower_name = 'tower' + str(t) inputs_desc = [InputDesc(desc.type, desc.shape, tower_name + '_' + desc.name) for desc in config.inputs_desc] input = PlaceholderInput() input.setup(inputs_desc) with tf.variable_scope(tf.get_variable_scope(), reuse=idx > 0), \ tf.device('/gpu:{}'.format(t)), \ PredictTowerContext(tower_name): config.tower_func(*input.get_input_tensors()) h = config.tower_func.towers[-1] input_tensors.extend(h.get_tensors(config.input_names)) output_tensors.extend(h.get_tensors(config.output_names)) sess = config.session_creator.create_session() config.session_init.init(sess) super(DataParallelOfflinePredictor, self).__init__( input_tensors, output_tensors, config.return_input, sess)