SINGA-82 Refactor input layers using data store abstraction * Enable StoreInputLayer to partition on dim = 0. * Add random skip for StoreInputLayer, which randomly skip some records at the beginning to make sure workers within the same group and running in the same process train over different data records.
Project: http://git-wip-us.apache.org/repos/asf/incubator-singa/repo Commit: http://git-wip-us.apache.org/repos/asf/incubator-singa/commit/4afc7b89 Tree: http://git-wip-us.apache.org/repos/asf/incubator-singa/tree/4afc7b89 Diff: http://git-wip-us.apache.org/repos/asf/incubator-singa/diff/4afc7b89 Branch: refs/heads/master Commit: 4afc7b893816f0fbb3141e72e3921fcd46feb8e4 Parents: 5f010ca Author: Wei Wang <[email protected]> Authored: Tue Oct 6 22:43:50 2015 +0800 Committer: wang sheng <[email protected]> Committed: Wed Oct 7 15:19:59 2015 +0800 ---------------------------------------------------------------------- include/neuralnet/input_layer.h | 3 ++- include/utils/image_transform.h | 29 +++++++++++++++++++++++++++++ src/neuralnet/input_layer.cc | 12 ++++++++++++ 3 files changed, 43 insertions(+), 1 deletion(-) ---------------------------------------------------------------------- http://git-wip-us.apache.org/repos/asf/incubator-singa/blob/4afc7b89/include/neuralnet/input_layer.h ---------------------------------------------------------------------- diff --git a/include/neuralnet/input_layer.h b/include/neuralnet/input_layer.h index 4dfded0..65e42cb 100644 --- a/include/neuralnet/input_layer.h +++ b/include/neuralnet/input_layer.h @@ -69,7 +69,8 @@ class StoreInputLayer : virtual public InputLayer { virtual bool Parse(int k, int flag, const string& key, const string& val) = 0; protected: - int batchsize_; + int batchsize_ = 1; + int random_skip_ = 0; io::Store* store_ = nullptr; }; http://git-wip-us.apache.org/repos/asf/incubator-singa/blob/4afc7b89/include/utils/image_transform.h ---------------------------------------------------------------------- diff --git a/include/utils/image_transform.h b/include/utils/image_transform.h new file mode 100644 index 0000000..24fe75d --- /dev/null +++ b/include/utils/image_transform.h @@ -0,0 +1,29 @@ +/************************************************************ +* +* Licensed to the Apache Software Foundation (ASF) under one +* or more contributor license agreements. See the NOTICE file +* distributed with this work for additional information +* regarding copyright ownership. The ASF licenses this file +* to you under the Apache License, Version 2.0 (the +* "License"); you may not use this file except in compliance +* with the License. You may obtain a copy of the License at +* +* http://www.apache.org/licenses/LICENSE-2.0 +* +* Unless required by applicable law or agreed to in writing, +* software distributed under the License is distributed on an +* "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY +* KIND, either express or implied. See the License for the +* specific language governing permissions and limitations +* under the License. +* +*************************************************************/ +#include <glog/logging.h> +// TODO(wangwei) provide image transformation API, the implementation can be +// done by opencv, manual transform, or mshadow. +namespace singa { + +void ImageTransform(const float* in, const float* mean, bool mirror, int h_crop, + int w_crop, int h_offset, int w_offset, int channel, int height, int width, + float scale, float* out); +} /* singa */ http://git-wip-us.apache.org/repos/asf/incubator-singa/blob/4afc7b89/src/neuralnet/input_layer.cc ---------------------------------------------------------------------- diff --git a/src/neuralnet/input_layer.cc b/src/neuralnet/input_layer.cc index b4743f4..1599def 100644 --- a/src/neuralnet/input_layer.cc +++ b/src/neuralnet/input_layer.cc @@ -85,6 +85,11 @@ void StoreInputLayer::Setup(const LayerProto& conf, const vector<Layer*>& srclayers) { InputLayer::Setup(conf, srclayers); batchsize_ = conf.store_conf().batchsize(); + if (conf.partition_dim() == 0) { + batchsize_ /= conf.num_partitions(); + } + if (conf.store_conf().random_skip() > 0) + random_skip_ = rand() % conf.store_conf().random_skip(); } void StoreInputLayer::ComputeFeature(int flag, @@ -94,6 +99,13 @@ void StoreInputLayer::ComputeFeature(int flag, store_ = io::OpenStore(layer_conf_.store_conf().backend(), layer_conf_.store_conf().path(), io::kRead); + while (random_skip_ > 0) { + if (!store_->Read(&key, &val)) { + store_->SeekToFirst(); + CHECK(store_->Read(&key, &val)); + } + random_skip_ --; + } } for (int k = 0; k < batchsize_; k++){ if (!store_->Read(&key, &val)) {
