Proto commits in stephenyan1231/caffe-public

These commits are when the Protocol Buffers files have changed: (only the last 100 relevant commits are shown)

2015-08-14

update lstm_2dlayer and renet_lstm_layer

2015-07-31

a complementary commit for implementation of LSTM_2DLayer and ReNetLSTMLayer

2015-07-03

fix bugs in the implementation of 2D LSTM layer

2015-06-30

a complementary commit for the previous commit: reimplement 2D LSTM in layer class LSTM_2DLayer

2015-05-17

add video-enhance/data/uniform_set/global_ftr_lmdb and feature mean files

2015-04-26

For class DataManager and DataVariableSizeManager, when selective list is provided, random shuffle is enabled by default.

2015-04-16

support matrix quantization for convolutional layer

2015-04-09

fix some bugs

2015-04-06

support matrix quantization for inner product layer

2015-04-02

add support for DataVariableSizeLayer to use encoded lmdb database

use encoded lmdb database. Supprt selective list in data layer

2015-03-31

add shift_stitch_layer shift_pooling_layer

2015-03-29

support overfeat-style dense testing. An example script is at examples/imagenet/test_VGG_ILSVRC_16_layers_dense_testing.sh

2015-03-27

add data_variable_size_layer.cpp data_variable_size_transformer.cpp

2015-03-26

support multi-scale training as in VGG paper.

HDCNN v1.1

2015-03-25

Add RecurrentLayer: an abstract superclass for other recurrent layer types

AccuracyLayer: add 'denominator' param

EltwiseLayer can take a blob of per-num coefficients

Add (very simple version of) ReshapeLayer

fix a bug that the phase(TRAIN or TEST) in data_transformer_ in DataManageris NOT initialized properly. Thus, in periodic testing stage during training, testing image is randomly cropped.

2015-03-15

(1)add spatial_accuracy_layer (2) prepare extract_VGG_ILSVRC_16_layers_parser_fc7.sh to demonstrate efficient pixelwise feature extraction using shift and stitch trick

2015-03-14

add shift_pooling_layer and shift_stitch_layer

2015-03-13

Add EmbedLayer for inner products with sparse input (one-hot vectors), with unit tests

zero-init param diffs and accumulate gradients (With layers whose backwards accumlate gradients), this effectively decouples the computational batch from the SGD minibatch. Each iteration accumulates gradients over iter_size batches, then parameters are updated.

HDF5DataLayer shuffle: minor cleanup; clarification in HDF5DataParameter

shuffle data

2015-03-12

PReLU Layer and its tests described in Kaiming He et al, "Delving Deep into Rectifiers: Surpassing Human-Level Performance on ImageNet Classification", arxiv 2015. Belows are commit message histories that I had while developing. PReLULayer takes FillerParameter for init PReLU testing consistency with ReLU Fix : PReLU test concistency check PReLU tests in-place computation, and it failed in GPU Fix: PReLU in-place backward in GPU PReLULayer called an incorrect API for copying data (caffe_gpu_memcpy). First argment of `caffe_gpu_memcpy` should be size of memory region in byte. I modified to use `caffe_copy` function. Fix: style errors Fix: number of axes of input blob must be >= 2 Use 1D blob, zero-D blob. Rename: hw -> dim

2015-03-09

AccuracyLayer: add ignore_label param

Fixup AccuracyLayer like SoftmaxLossLayer in #1970 -- fixes #2063

2015-03-08

first workable multigpu caffe

2015-03-03

DummyDataLayer outputs blobs of arbitrary shape

SoftmaxLayer: generalized Blob axes

ConcatLayer: generalized Blob axes

SliceLayer: generalized Blob axes

InnerProductLayer can multiply along any axis

Blobs are ND arrays (for N not necessarily equals 4). vector<int> shape_ instead of (num, channels, height, width).

Add BlobShape message; use for Net input shapes

2015-02-20

Merge pull request #1910 from philkr/encoded add force_encoded_color flag to the data layer and warn about mixed encoding

added a force_encoded_color flag to the data layer. Printing a warning if images of different channel dimensions are encoded together

2015-02-17

give phase to Net and Layer Give the responsibility for phase to Net and Layer, making phase an immutable choice at instantiation and dropping it from the Caffe singleton.

[pycaffe] allow Layer to be extended from Python This is done by adding PythonLayer as a boost::python HeldType.

2015-02-14

Add gradient clipping -- limit L2 norm of parameter gradients

2015-02-06

fix for layer-type-str: loss_param and DECONVOLUTION type should have been included in V1LayerParameter, get upgraded

2015-02-05

start layer parameter field IDs at 100 (always want them printed at the end, and want to allow more fields to be added in the future, so reserve fields 10-99 for that purpose)

get rid of NetParameterPrettyPrint as layer is now after inputs (whoohoo)

automagic upgrade for v1->v2

add message ParamSpec to replace param name, blobs_lr, weight_decay, ...

'layers' -> 'layer'

restore upgrade_proto

Layer type is a string

2015-02-01

Merge pull request #1615 from longjon/deconv-layer Add deconvolution layer with refactoring of convolution layer to share code

2015-01-30

Merge pull request #1654 from longjon/softmax-missing-values Add missing value support to SoftmaxLossLayer

debug_info in NetParameter so it can be enabled outside training

2015-01-17

Make comments for sparse GaussianFiller match actual behavior (Fixes #1497 reported by @denizyuret)

2015-01-11

add DeconvolutionLayer, using BaseConvolutionLayer

2014-12-31

add spatial normalization option to SoftmaxLossLayer With missing values (and batches of varying spatial dimension), normalizing each batch across instances can inappropriately give different instances different weights, so we give the option of simply normalizing by the batch size instead.

add missing value support to SoftmaxLossLayer

2014-12-22

Added Multistep, Poly and Sigmoid learning rate decay policies Conflicts: include/caffe/solver.hpp src/caffe/proto/caffe.proto src/caffe/solver.cpp

Display averaged loss over the last several iterations

2014-10-16

Merge pull request #190 from sguada/new_lr_policies New lr policies, MultiStep and StepEarly

Added cache_images to WindowDataLayer Added root_folder to WindowDataLayer to locate images

Add root_folder to ImageDataLayer

2014-10-15

Added encoded datum to io

2014-10-14

Refactored leveldb and lmdb code.

2014-10-10

Added Multistep, Poly and Sigmoid learning rate decay policies Conflicts: include/caffe/solver.hpp src/caffe/proto/caffe.proto src/caffe/solver.cpp

2014-10-06

correct naming in comment and message about average_loss

2014-10-04

Merge pull request #1070 from sguada/move_data_mean Refactor data_transform to allow datum, cv:Mat and Blob transformation

2014-10-03

Added mean_value to specify mean channel substraction Added example of use to models/bvlc_reference_caffenet/train_val_mean_value.prototxt

Added global_pooling to set the kernel size equal to the bottom size Added check for padding and stride with global_pooling

Initial cv::Mat transformation Added cv::Mat transformation to ImageDataLayer Conflicts: src/caffe/layers/image_data_layer.cpp Added transform Datum to Blob Conflicts: src/caffe/layers/base_data_layer.cpp src/caffe/layers/base_data_layer.cu Added transform cv::Mat to Blob Added transform Vector<Datum> to Blob Conflicts: src/caffe/data_transformer.cpp

2014-10-02

Add ExpLayer to calculate y = base ^ (scale * x + shift)

caffe.proto: do some minor cleanup (fix comments, alphabetization)

2014-10-01

Merge pull request #1096 from qipeng/smoothed-cost Display averaged loss over the last several iterations

2014-09-29

code cleaning

2014-09-22

adds a parameter to the LRN layer (denoted as "k" in [Krizhevsky et al., NIPS 2012])

added support for "k" LRN parameter to upgrade_proto

2014-09-19

Merge pull request #959 from nickcarlevaris/contrastive_loss Add contrastive loss layer, tests, and a siamese network example

The documentation is generated from this commit.

2014-09-17

Display averaged loss over the last several iterations

2014-09-14

fix spelling error in caffe.proto

fix out-of-date next ID comment for SolverParameter

2014-09-08

Added contrastive loss layer, associated tests, and a siamese network example using shared weights and the contrastive loss.

Implemented elementwise max layer

2014-09-07

add cuDNN to build

add engine parameter for multiple computational strategies add `engine` switch to layers for selecting a computational backend when there is a choice. Currently the standard Caffe implementation is the only backend.

groom proto: sort layer type parameters, put loss_weight after basics

2014-09-05

Merge pull request #1036 from longjon/test-initialization-param Add test_initialization option to allow skipping initial test

2014-09-04

add SILENCE layer -- takes one or more inputs and produces no output This is useful for suppressing undesired outputs.

add test_initialization option to allow skipping initial test

2014-09-03

Add transformer to the memory data layer

Move transform param one level up in the proto to reduce redundancy

2014-09-01

Added L1 regularization support for the weights

fixed caffe.proto after a mistaken rebase

fixes after rebase

proto conflit, lint, and math_functions (compiler complaint)

cleanup caffe.proto

2014-08-28

add two new layers MULTINOMIAL_LOGISTIC_SPARSITY_LOSS and PROBABLISTIC_AVERAGE_PROB

2014-08-25

Add "stable_prod_grad" option (on by default) to ELTWISE layer to compute the eltwise product gradient using a slower but stabler formula.

2014-08-24

add FLOAT_DATA layer to allow float-typed input data

2014-08-22

restore old data transformation parameters for compatibility