Age | Commit message (Collapse) | Author | Files | Lines | |
---|---|---|---|---|---|
2015-01-25 | Merge pull request #1755 from jeffdonahue/softmax-optimization | Evan Shelhamer | 1 | -24/+24 | |
SoftmaxLayer GPU optimization | |||||
2015-01-25 | Merge pull request #1754 from jeffdonahue/softmax-loss-fix | Jeff Donahue | 1 | -0/+4 | |
SoftmaxWithLossLayer: use CreateLayer | |||||
2015-01-21 | hdf5_save_nd_dataset takes a const string& (instead of const string) | Jeff Donahue | 1 | -2/+2 | |
2015-01-21 | SoftmaxWithLossLayer: use CreateLayer so that a CuDNNSoftmaxLayer | Jeff Donahue | 1 | -0/+4 | |
is created if available | |||||
2015-01-21 | Back-merge fixes + docs | Evan Shelhamer | 1 | -7/+8 | |
and other fixes and documentation updates. | |||||
2015-01-19 | Unroll kernels in SoftmaxLayer...from terrible performance to mediocre | Jeff Donahue | 1 | -24/+24 | |
performance. | |||||
2015-01-19 | SetTotalBytesLimit to the max (2 GB minus 1 byte) | Jeff Donahue | 1 | -1/+3 | |
2015-01-17 | Make comments for sparse GaussianFiller match actual behavior | Jeff Donahue | 1 | -1/+1 | |
(Fixes #1497 reported by @denizyuret) | |||||
2015-01-16 | Merge pull request #1686 from longjon/net-const | Jon Long | 2 | -18/+21 | |
Improve const-ness of Net | |||||
2015-01-15 | Merge pull request #1236 from mlapin/legacy_nvcc_support | Evan Shelhamer | 4 | -4/+32 | |
Drop OpenCV includes from NVCC code for legacy reasons. | |||||
2015-01-15 | lint internal thread | Evan Shelhamer | 1 | -1/+2 | |
2015-01-15 | Merge pull request #1335 from ryotat/master | Evan Shelhamer | 1 | -7/+7 | |
Fix leaking thread and groom internal thread implementation. | |||||
2015-01-10 | fix typo in layer_factory.cpp | Jonathan L Long | 1 | -1/+1 | |
2015-01-09 | improve const-ness of Net | Jonathan L Long | 2 | -18/+21 | |
2015-01-02 | break out Step from Solver | Jonathan L Long | 1 | -35/+38 | |
2014-12-30 | [tests] don't use Gaussian labels in NetTest's TinyNet | Jonathan L Long | 1 | -0/+4 | |
Previously, labels were being nonsensically filled with a zero-mean Gaussian. Now they are just set to constant zero. Note that this is not necessary for TrickyNet, neither forward nor backward are ever called. | |||||
2014-12-29 | remove SoftmaxLossLayer CPU_ONLY stubs, since there is no GPU version | Jonathan L Long | 1 | -5/+1 | |
2014-12-29 | move softmax loss GPU todo comment to header file | Jonathan L Long | 1 | -27/+0 | |
There's no need to have GPU stubs just for a comment, and the absence of a GPU-specific implementation is made more obvious by the nonexistence of the .cu file. | |||||
2014-12-27 | update use of GetLayer -> LayerRegistry::CreateLayer | Jonathan L Long | 1 | -1/+2 | |
2014-12-22 | Warning of fallback only the first time for cudnn_pooling_layer | Sergio | 1 | -2/+4 | |
2014-12-21 | Fix init current_step | Sergio | 1 | -0/+1 | |
2014-12-21 | Added Multistep, Poly and Sigmoid learning rate decay policies | Sergio | 2 | -8/+37 | |
Conflicts: include/caffe/solver.hpp src/caffe/proto/caffe.proto src/caffe/solver.cpp | |||||
2014-12-21 | Display averaged loss over the last several iterations | qipeng | 2 | -2/+20 | |
2014-12-20 | Add CHECKs to prevent segfault for incorrect IMAGE_DATA layers. | Thomas Schenker | 1 | -0/+5 | |
Incomplete IMAGE_DATA layers cause a segmentation fault, e.g.: layers { name: "mnist" type: IMAGE_DATA top: "data" top: "label" image_data_param { batch_size: 100 } include: { phase: TRAIN } } or layers { name: "mnist" type: IMAGE_DATA top: "data" top: "label" data_param { source: "./data/train.txt" batch_size: 100 } include: { phase: TRAIN } } Also empty image list files cause a segfault. This is fixed by adding 3 CHECK instructions. | |||||
2014-12-10 | Check input line count in HDF5 data layer | Jonathan Ho | 1 | -0/+2 | |
2014-12-05 | compile for compute capability 5.0 | Evan Shelhamer | 1 | -1/+3 | |
2014-11-30 | remove redundant code in ConvolutionLayer::Reshape | Jonathan L Long | 1 | -3/+0 | |
2014-11-25 | Merge pull request #1455 from seanbell/tanh-fix | Jeff Donahue | 3 | -5/+103 | |
Fix tanh layer for badly scaled input values | |||||
2014-11-23 | use DCHECK in SoftmaxLossLayer so that bounds checking has no perf penalty ↵ | Jonathan L Long | 1 | -2/+2 | |
without DEBUG | |||||
2014-11-22 | in SoftmaxLossLayer, check label >= 0 in addition to upper bound | Jonathan L Long | 1 | -0/+1 | |
2014-11-19 | Fixed header order to satisfy linter | Sean Bell | 1 | -2/+1 | |
2014-11-19 | fixed tanh to not return NaN for input values outside the range [-40, 40] | Sean Bell | 3 | -5/+104 | |
2014-11-13 | fix relu cudnn test bug | Yangqing Jia | 1 | -2/+5 | |
2014-11-07 | relax benchmark milliseconds threshold | Yangqing Jia | 1 | -4/+6 | |
2014-11-07 | clean incorrect relu test code | Yangqing Jia | 1 | -6/+14 | |
2014-10-22 | Minor whitespace fix in logging message in HDF5 output layer | Martin Baeuml | 1 | -1/+1 | |
2014-10-21 | Forward declare boost::thread instead of including boost/thread.hpp in ↵ | Ryota Tomioka | 1 | -0/+6 | |
internal_thread.hpp. | |||||
2014-10-20 | Merge pull request #1326 from jackculpepper/absgradatzero | Evan Shelhamer | 2 | -3/+2 | |
[fix] define gradient at zero to be zero to fix ABSVAL NaN | |||||
2014-10-20 | Fixed a memory leak issue in InternalThread (and removed caffe::Thread calss). | Ryota Tomioka | 1 | -7/+1 | |
2014-10-19 | include opencv only in .cpp | Maksim Lapin | 4 | -4/+32 | |
2014-10-19 | define gradient at zero to be zero | Jack Culpepper | 2 | -3/+2 | |
2014-10-18 | LMDB doesn't support many concurrent read-only transactions, so this ↵ | Kevin James Matzen | 2 | -43/+72 | |
preallocates one read-only transaction and reuses it. It's very important that iterations are considered invalid after a commit has been performed. | |||||
2014-10-17 | back-merge | Sergey Karayev | 1 | -1/+1 | |
2014-10-16 | Fix init current_step | Sergio | 1 | -0/+1 | |
2014-10-15 | Merge pull request #190 from sguada/new_lr_policies | Sergio Guadarrama | 2 | -8/+37 | |
New lr policies, MultiStep and StepEarly | |||||
2014-10-15 | Added CPUTimer | Sergio | 4 | -28/+75 | |
Make timing more precise using double and microseconds | |||||
2014-10-15 | Upgrade compute_image_mean to use gflags, accept list_of_images, and print ↵ | Sergio | 1 | -2/+2 | |
mean_values | |||||
2014-10-15 | Change caffe time to do forward/backward and accumulate time per layer | Sergio | 4 | -60/+12 | |
2014-10-15 | Added cache_images to WindowDataLayer | Sergio | 2 | -6/+36 | |
Added root_folder to WindowDataLayer to locate images | |||||
2014-10-15 | Speed up WindowDataLayer and add mean_values | Sergio | 1 | -19/+46 | |