925cebd | moneta | 24 October 2018, 16:42:55 UTC | add new option for switching off computation of correlation matrix | 24 October 2018, 16:42:55 UTC |
67c06b2 | moneta | 22 August 2018, 15:23:03 UTC | use simply ScaleAdd when updating weightgradients | 22 August 2018, 15:23:03 UTC |
fdbcf7f | moneta | 22 August 2018, 13:19:06 UTC | fix a compilation warning | 22 August 2018, 13:19:06 UTC |
2b2dc7f | moneta | 22 August 2018, 13:10:57 UTC | optimize SGD and ADAM update calculations. For ADAM define specific update functions which are re-implemented for each architecture | 22 August 2018, 13:10:57 UTC |
9285d73 | moneta | 22 August 2018, 13:09:15 UTC | remove calculation of correlation matrix for large number of variables | 22 August 2018, 13:09:15 UTC |
a019513 | moneta | 21 August 2018, 08:10:43 UTC | Use fast zero GPU initialisation for Cuda matrices | 21 August 2018, 08:10:43 UTC |
47167e8 | moneta | 21 August 2018, 08:04:18 UTC | Fix a bug in GetMvaValues() when total events are not multiple of batch size Use for evaluation same batch size as for training and not a fixed one of 1000 | 21 August 2018, 08:04:18 UTC |
d09c108 | Giulio Eulisse | 20 August 2018, 19:52:58 UTC | Support unsigned integer types in RArrowDS | 20 August 2018, 21:58:34 UTC |
bcb4953 | Philippe Canal | 20 August 2018, 16:35:33 UTC | White space | 20 August 2018, 16:35:59 UTC |
ae9f39e | Brian Bockelman | 10 August 2018, 03:09:47 UTC | [TTree] Use TMath::Max instead of std::max. | 20 August 2018, 16:35:59 UTC |
ec2ce98 | Brian Bockelman | 09 August 2018, 20:55:14 UTC | [TTree] Touchups from review of PR 774. | 20 August 2018, 16:35:59 UTC |
2d8e9f6 | Brian Bockelman | 07 August 2018, 14:13:05 UTC | [TTree] Avoid narrowing conversion of constant. | 20 August 2018, 16:35:59 UTC |
63eb9e5 | Brian Bockelman | 06 August 2018, 14:27:36 UTC | [TTree] Replace use of ssize_t with Int_t. Matches the underlying type of the return value. | 20 August 2018, 16:35:59 UTC |
1cca868 | Brian Bockelman | 05 August 2018, 02:46:00 UTC | Update release notes for new kOnlyFlushAtCluster mode. | 20 August 2018, 16:35:59 UTC |
28e52d4 | Brian Bockelman | 05 August 2018, 02:34:50 UTC | [Tree] Test for one-basket-per-cluster mode. | 20 August 2018, 16:35:59 UTC |
ad8db39 | Brian Bockelman | 04 August 2018, 15:40:42 UTC | [Tree] Fixup build issue from bad rebase. | 20 August 2018, 16:35:59 UTC |
b89847f | Brian Bockelman | 07 September 2017, 13:44:51 UTC | [whitespace] clang-tidy fixups | 20 August 2018, 16:35:59 UTC |
95e116b | Brian Bockelman | 06 September 2017, 18:59:45 UTC | Use a different bit for new basket flush mode. This one appears to be unused by subclasses and passes the `ROOT::Detail::TStatusBitsChecker` test. | 20 August 2018, 16:35:59 UTC |
6bed7e3 | Brian Bockelman | 06 September 2017, 15:14:34 UTC | Allow the basket allocation tracking to be ifdef'd out. | 20 August 2018, 16:35:59 UTC |
1d87b72 | Brian Bockelman | 06 September 2017, 03:31:34 UTC | Fixup parenthesis in comment block. | 20 August 2018, 16:35:59 UTC |
f0f636d | Brian Bockelman | 06 September 2017, 13:53:50 UTC | Rename kFlushAtCluster to kOnlyFlushAtCluster. As pointed out by @pcanal, we always flush at cluster -- the new behavior with this bit is it allows us to _only_ flush at the cluster boundaries. | 20 August 2018, 16:35:59 UTC |
3f4816c | Brian Bockelman | 06 September 2017, 13:53:09 UTC | Whitespace-only changes from clang-format. | 20 August 2018, 16:35:59 UTC |
f3a062b | Brian Bockelman | 20 June 2017, 19:40:50 UTC | Add "one-basket-per-cluster" mode. By setting a new bit in the TTree object, kFlushAtCluster, one can enter the "one basket per event cluster" mode. This forces baskets to line up with event clusters, at the cost of extra memory use. This new mode simplifies the basket layout within the file, reducing the amount of time handling special cases for the bulk IO mode. Because the baskets *must* grow to the size of an event cluster, we do not invoke OptimizeBaskets when one-basket-per-cluster mode is enabled. As this mode is expected to cause increased memory usage (the memory utilized by the TTree is more strongly tied to the variations in event size), we combine this with a technique borrowed from a CMS patchset to more aggressively shrink basket sizes after very large objects. The new basket shrinking algorihtm will trigger whenever the basket is flushed. If the actual object size in the last clusters is significantly below the buffer size, then the basket will be shrunk. The ideal ratio of `(basket buffer size)/(occupied buffer size)` is controlled by a new tunable in the TTree, defaulting to 1.1. A lower setting reduces overall memory usage at the cost of extra allocations; a higher setting increases aggregate memory usage. In writing out a 10k event CMS file (total CPU time is 32 minutes): - Base case (without this patch): - 888MB RSS - 30 reallocations (shrinking of baskets due to low occupancy). - 0.173ms taken for reallocation. - New shrinking algorithm (this patch with defaults): - 866MB RSS - 4434 reallocations - 97.0ms - One basket per cluster mode with new shrinking algorithm: - 902MB RSS - 2882 reallocations - 93.6ms The CPU-time cost of the reallocation is 0.005% of total runtime (considering the file has to be read also, maybe 0.01% of output time?) This patch purposely leaves in the code that measures reallocation time in order to allow others to experiment; the intent is to disable / remove the code before a final release. | 20 August 2018, 16:35:59 UTC |
331982f | Danilo Piparo | 20 August 2018, 14:16:53 UTC | [VecOps] Add missing push_back signature | 20 August 2018, 15:59:05 UTC |
200a48d | Axel Naumann | 20 August 2018, 13:02:51 UTC | Revert "Revert "Fix typo in test."" This reverts commit a54b9e5ceb0574d5ebb49950f75114e4b5cd08a7. THis just shifts the error around; Danilo and Giulio are looking at it. | 20 August 2018, 13:02:51 UTC |
c6bf37b | Axel Naumann | 20 August 2018, 07:36:58 UTC | [RVec] Remove unused var. | 20 August 2018, 07:36:58 UTC |
da43fdd | Axel Naumann | 20 August 2018, 07:36:41 UTC | [TMVA] Remove unused var. | 20 August 2018, 07:36:41 UTC |
a54b9e5 | Axel Naumann | 20 August 2018, 07:28:45 UTC | Revert "Fix typo in test." This reverts commit e786005e24f5ecb265d68c3a19fe4039d6ce00a1. The type arrow::uint32 isn't supported and will cause an exception and test failure. | 20 August 2018, 07:28:45 UTC |
a1d5253 | Emmanouil Stergiadis | 15 August 2018, 13:40:23 UTC | added cuda stream support to parallelize over events in the batch | 20 August 2018, 07:09:42 UTC |
3ad391f | Danilo Piparo | 19 August 2018, 06:00:26 UTC | [RooFit] [ROOT-9623] Remove failing py tutorials | 19 August 2018, 06:00:26 UTC |
57a480f | Oksana Shadura | 16 August 2018, 16:23:40 UTC | Recovery of correct CMake invocation of rootcling We are fixing a correct invocation accidentally removed in ROOT_GENERATE_DICTIONARY() macro in '[cxxmodules] Move away from environment variable' commit. | 17 August 2018, 21:24:51 UTC |
8b7bc79 | Philippe Canal | 16 August 2018, 19:55:20 UTC | Additional fix for rule scheduling in TBranchElement. This is an additional fix for https://github.com/cms-sw/cmssw/pull/22594 If one of the class associated with a TBranchElement has a base class and/or sub-object class that has a rule that requires a cache (onfileObject) object to stage the original data *but* does not have branch in the TTree that also neeed the same cache object, we need to associate a cache object with the StreamerInfoAction sequence that needs it (associated with a higher level branch) and give it ownership of this cache object) | 17 August 2018, 16:07:52 UTC |
7afeae2 | Philippe Canal | 16 August 2018, 18:24:02 UTC | Avoid double search | 17 August 2018, 16:07:52 UTC |
645a914 | Philippe Canal | 16 August 2018, 18:18:42 UTC | Improve debug priting of StreamerElement info | 17 August 2018, 16:07:52 UTC |
830512d | Philippe Canal | 16 August 2018, 18:18:10 UTC | [NFC] White space | 17 August 2018, 16:07:52 UTC |
e786005 | Giulio Eulisse | 17 August 2018, 13:01:56 UTC | Fix typo in test. | 17 August 2018, 13:03:04 UTC |
18a5dcc | Danilo Piparo | 17 August 2018, 11:57:08 UTC | [PyROOT] [ROOT-9623] Remove RooFit tutorials which fail on osx python3 | 17 August 2018, 11:57:49 UTC |
8d17dd0 | Enrico Guiraud | 03 August 2018, 12:36:39 UTC | [VecOps] Add gtest for ROOT-9453 "Cannot instantiate RVec<bool>" | 17 August 2018, 11:28:06 UTC |
374be4c | Enrico Guiraud | 03 August 2018, 13:03:59 UTC | [VecOps] Make RAdoptAllocator<T> rebindable to RAdoptAllocator<bool> ...and viceversa. Only non-adopting RAdoptAllocators can be rebound. This makes it possible to copy-construct RVec<bool>'s from other RVec<bool>'s. | 17 August 2018, 11:28:06 UTC |
78e4bef | Enrico Guiraud | 03 August 2018, 12:30:39 UTC | [VecOps] Specialize RAdoptAllocator<bool> The different semantics of std::vector<bool> make memory adoption through a custom allocator more complex -- namely, RAdoptAllocator<bool> must be rebindable to RAdoptAllocator<unsigned long>, but if adopted memory is really a buffer of bools, reinterpretation of the adopted buffer as a different type is going to break things in horrible ways. As a workaround, RAdoptAllocator<bool> is specialized to be a simple allocator that forwards calls to std::allocator, never adopts memory, and can be rebound to any other allocator (as it never adopts memory it can rebind the same way that std::allocator can). This should resolve ROOT-9453 "Cannot instantiate RVec<bool>". | 17 August 2018, 11:28:06 UTC |
1e79c8d | Enrico Guiraud | 03 August 2018, 12:22:43 UTC | [VecOps][NFC] Add libROOTVecOps as dependency to RAdoptAllocator test | 17 August 2018, 11:28:06 UTC |
e3316bc | Enrico Guiraud | 03 August 2018, 12:09:28 UTC | [VecOps] Remove redundant RAdoptAllocator::construct overload | 17 August 2018, 11:28:06 UTC |
cc4915b | Enrico Guiraud | 03 August 2018, 11:07:05 UTC | [VecOps][NFC] Formatting of RAdoptAllocator | 17 August 2018, 11:28:06 UTC |
5927f14 | Oksana Shadura | 17 August 2018, 09:10:15 UTC | Updating clad build CMake arguments for the case when buiting_llvm and buitin_clang is OFF | 17 August 2018, 11:01:40 UTC |
d037c05 | Stefan Wunsch | 17 July 2018, 13:31:01 UTC | [VecOps] Add RVec helpers for sorting and selecting elements | 17 August 2018, 09:33:17 UTC |
ce0b29a | Giulio Eulisse | 16 August 2018, 14:27:39 UTC | Fix typo in RArrowDS The name of the type is Long_t, not Long32_t. | 17 August 2018, 09:16:15 UTC |
23aa75d | Lorenzo Moneta | 15 August 2018, 10:29:22 UTC | fix memory leaks | 17 August 2018, 07:29:32 UTC |
e7d134e | Oksana Shadura | 04 June 2018, 13:37:56 UTC | [layering] Cleaning TFile implementation file from TMath-specific functions and includes | 16 August 2018, 16:37:42 UTC |
d2f39d4 | Oksana Shadura | 04 June 2018, 13:35:02 UTC | [layering] Updating linkdef files for Core and MathCore, after moving sorting and binary search algorithms from MathCore to Core | 16 August 2018, 16:37:42 UTC |
6769ded | Oksana Shadura | 14 March 2018, 19:29:46 UTC | [layering] Removing TMath.h include, TFileCacheRead.cxx is depending only from Core library. | 16 August 2018, 16:37:42 UTC |
1b83473 | Oksana Shadura | 14 March 2018, 19:26:05 UTC | [layering] Moving BinarySearch and Sorting algorithms from math/mathcore/inc/TMath.h to TMathBase.h (evolution to ROOT "<algorithm>" header) | 16 August 2018, 16:37:42 UTC |
f8f4760 | Olivier Couet | 16 August 2018, 12:40:29 UTC | When drawing a `TGraph` if `TH1::SetDefaultSumw2()` was on, then the underlying histogram used to draw the `TGraph` axis was created with errors and therefore the histogram painter painted it with errors which was a non sense in that particular case. This is now fixed. It was discussed here: https://root-forum.cern.ch/t/horizontal-line-at-0-on-y-axis/30244/26. | 16 August 2018, 12:40:39 UTC |
beda404 | moneta | 14 August 2018, 16:10:05 UTC | Fix memory leak in UpdateWeights in Cuda | 14 August 2018, 16:10:43 UTC |
ab20fee | Danilo Piparo | 14 August 2018, 14:08:17 UTC | [PyROOT] Remove RooFit tests which fail (ROOT-9623) many of these fail because of double deletes. Something in the memory management needs attention. | 14 August 2018, 14:08:17 UTC |
d98b567 | Danilo Piparo | 14 August 2018, 10:22:30 UTC | [RooFit] Remove stray doc line | 14 August 2018, 10:22:30 UTC |
e52c7f8 | Danilo Piparo | 14 August 2018, 09:11:11 UTC | [RooFit] Add tutorial in python | 14 August 2018, 09:11:11 UTC |
8826824 | Danilo Piparo | 14 August 2018, 09:06:54 UTC | [RooFit] Add python tutorial after bug in factory was fixed. | 14 August 2018, 09:06:54 UTC |
5618a32 | Danilo Piparo | 14 August 2018, 09:01:41 UTC | [RooFit] Fix for ROOT-9581 when using a ws factory, the enumerators are found correctly but the enum constants may be inside a scope. These were found by CINT but not by cling (and rightfully so). This patch fixes this bug. | 14 August 2018, 09:01:41 UTC |
51555bd | Danilo Piparo | 14 August 2018, 08:28:44 UTC | [RooFit] Do not draw canvases in notebooks if nothing was drawn | 14 August 2018, 08:28:44 UTC |
f48ef1e | Danilo Piparo | 14 August 2018, 08:20:42 UTC | [RooFit] This is a DSL independent of PyROOT | 14 August 2018, 08:20:42 UTC |
e76667c | Danilo Piparo | 14 August 2018, 08:17:10 UTC | [Doc] The ROOTBOOK-izer tool is not anymore a beta. | 14 August 2018, 08:17:10 UTC |
269e690 | Sergey Linev | 13 August 2018, 09:24:17 UTC | xml: fix gcc8 warning with strncpy arguments At this place buffer was allocated using length of source string. gcc does not like when strncpy length argument directly derives from source string length. Actually, buffer should not depend from source length - here it was workaround for old problem, which is already fixed. Therefore just use constant buffer length | 13 August 2018, 16:00:05 UTC |
4e0fbf3 | Clemens Lange | 13 August 2018, 15:41:59 UTC | [PyROOT] Python implementation of RooFit tutorials | 13 August 2018, 15:41:59 UTC |
d73f479 | Sergey Linev | 13 August 2018, 08:27:00 UTC | json: fix compilation warning with gcc8 Solves https://sft.its.cern.ch/jira/browse/ROOT-9588 | 13 August 2018, 14:19:32 UTC |
0d87ba4 | moneta | 13 August 2018, 12:31:37 UTC | Improve test for CNN in MethodDL. Make image data with input pixels | 13 August 2018, 12:31:37 UTC |
d36b302 | Stefan Wunsch | 13 August 2018, 11:40:01 UTC | [Experimental PyROOT] Rename sources to follow convention of one file per class, introduce helpers for python-side pythonizations | 13 August 2018, 11:55:53 UTC |
09ee024 | Stefan Wunsch | 09 August 2018, 16:57:07 UTC | [Experimental PyROOT] Add array interface to STL vector and RVec | 13 August 2018, 11:30:19 UTC |
ae65108 | Stefan Wunsch | 09 August 2018, 12:01:17 UTC | [Experimental PyROOT] Restructure generic pythonizations | 13 August 2018, 11:30:19 UTC |
0b310c2 | Kim Albertsson | 08 August 2018, 14:30:21 UTC | [TMVA] Doc - Update pdf (cherry picked from commit 9bcf786a51c25d5fdfaef0d76b7f2f301c61cf14) | 13 August 2018, 07:55:30 UTC |
af619bb | Massimo Tumolo | 10 August 2018, 17:16:40 UTC | [DF] LoopManager tracks run actions The LoopManager now tracks the actions that have already run. This commit is needed by the Graph drawing PR. | 11 August 2018, 09:16:51 UTC |
847528b | Brian Bockelman | 09 August 2018, 19:53:26 UTC | [TTree] Change header types for consistency. | 10 August 2018, 17:51:13 UTC |
ec0ffbf | Brian Bockelman | 02 August 2018, 19:00:54 UTC | [Tree] Improve defaults for cluster iterator sizes. If the TTree has no cluster size recorded (no auto-flush and no event clusters are present) and no cache set, then the cluster iterator would default to an estimate of 1 event per cluster. With this change, the default becomes however many events fit into 30MB of compressed data; 30MB was chosen as it is the current default cache size in ROOT. Fixes: ROOT-9574 | 10 August 2018, 17:51:13 UTC |
c70aaf1 | Brian Bockelman | 02 August 2018, 01:51:11 UTC | [TTree] Create event clusters when TTree is flushed. Several of us were surprised to discover that calling `TTree::FlushBaskets()` does not actually create a new event cluster in terms of the TTree metadata -- even though it effectively creates an event cluster in terms of the physical disk layout! There may be cases where the user has special knowledge of the file where using the internally-calculated auto-flush intervals result in poor choices; CMS's NanoAOD is one such case. However, when CMS switched to calling FlushBaskets explicitly and disabling AutoFlush, this triggered unexpected behavior in RDataFrame as the file did not appear to have any proper event clusters. This change causes the `FlushBaskets` method to create an explicit event cluster. As we used the old behavior in some ROOT unit tests, the prior behavior is now accessible through a flag passed to the method. Fixes: ROOT-9442 | 10 August 2018, 17:51:13 UTC |
7a23161 | Yuka Takahashi | 08 August 2018, 15:39:47 UTC | Reland "[cxxmodules] Check correctly if the decl was cached." Improve the commit by tighten condition to take an expensive branch. In BaseSelectionRule::Match, for normal CXXRecordDecl, nothing need to be worried as you can just compare name or pointer and that's it. However when you have typedef, it means that you can't just compare name (it's different!) or pointer (it's also different). And here, ScopeSearch is not actually searching the argument. It's just renewing the fCXXRecordDecl from fName to keep it up with the current interpreter state. Also, rootcling needed to parse header to get the actual definition of typedef, because rootmap file given to the invocation was lacking the necessary information. For example it contained `typedef typedefB`, from this there is no way to know which RecordDecl is underneath. With modules, we don't use rootmap files. Thus we need to check if it's a typdef or not, and if it was a typedef, make sure that it didn't take the same branch with the same interp state (by checking (void*)-1) and it's not from modules. This improves cxxmodules res memory by 90MB. ``` before: res memory = 345.816 Mbytes after res memory = 267.582 Mbytes ``` Also, this fasten rootcling 2 times. ``` before: User time (seconds): 2.56 after: User time (seconds): 1.35 ``` | 10 August 2018, 15:57:04 UTC |
124e3fe | Enric Tejedor Saavedra | 07 August 2018, 11:58:28 UTC | [Exp PyROOT] Add unit test for TTree iterable pythonization | 10 August 2018, 13:45:19 UTC |
3e2163a | Enric Tejedor Saavedra | 07 August 2018, 11:56:41 UTC | [Exp PyROOT] Add unit test for TTree branch access pythonization | 10 August 2018, 13:45:19 UTC |
e519479 | Enric Tejedor Saavedra | 07 August 2018, 11:54:35 UTC | Add COPY_TO_BUILDDIR option to PYUNITTEST | 10 August 2018, 13:45:19 UTC |
55556e9 | Bertrand Bellenot | 10 August 2018, 07:46:04 UTC | Add https support in TGHtmlBrowser (requires ssl) | 10 August 2018, 13:02:24 UTC |
07a576a | moneta | 10 August 2018, 07:51:14 UTC | Fix computation of test error, which was incorrectly done after fixing conflicts | 10 August 2018, 09:09:40 UTC |
5d8c7cd | moneta | 09 August 2018, 21:49:20 UTC | Use the template type (Scalar_t) instead of double, otherwise it will not compile for networks based on float types | 10 August 2018, 09:09:40 UTC |
e2545b4 | moneta | 02 August 2018, 16:10:35 UTC | Additional fixes after rebasing to the master | 10 August 2018, 09:09:40 UTC |
7b81a3d | moneta | 02 August 2018, 15:26:39 UTC | Compilation fixes caused by rebasing to master. After applying this PR, network is now based on float types and one should use Scalar_t instead of hardcoded Double_t in the tests | 10 August 2018, 09:09:40 UTC |
bd9a3a6 | moneta | 01 August 2018, 16:29:51 UTC | Remove old code, not used anymore | 10 August 2018, 09:09:40 UTC |
e56d907 | moneta | 01 August 2018, 16:26:21 UTC | remove files added by mistake | 10 August 2018, 09:09:40 UTC |
d65982c | Lorenzo Moneta | 22 June 2018, 10:30:22 UTC | fix Dense layer part for TestMethodDL | 10 August 2018, 09:09:40 UTC |
9cee157 | Lorenzo Moneta | 21 June 2018, 14:12:33 UTC | Add possibility to extend XML buffer size. Set also output precision of matrix in xml dependent on the type | 10 August 2018, 09:09:40 UTC |
a5de3c9 | Lorenzo Moneta | 20 June 2018, 14:46:51 UTC | add missing include | 10 August 2018, 09:09:40 UTC |
8bea055 | Lorenzo Moneta | 20 June 2018, 14:38:00 UTC | compute initial test error at the beginning of training phase. Avoid an useless operation when adding the reg. term | 10 August 2018, 09:09:40 UTC |
2a9872f | Lorenzo Moneta | 20 June 2018, 08:55:32 UTC | Use TensorDataLoader when loading batches to use for evaluating the MVA values (in PredictDeepNet) | 10 August 2018, 09:09:40 UTC |
f7c7771 | moneta | 19 June 2018, 15:43:00 UTC | Implement GetMvaValues in MethodDL using a fixed batch size of 1000. The prediction is performed using the chosen architecture and is much faster than predicting a single event. | 10 August 2018, 09:09:40 UTC |
2f50b9a | Lorenzo Moneta | 13 June 2018, 09:50:36 UTC | Fixes compilation for Cuda to use now conversion to same scalar type between TCudaMatrix<T> and TMatrix<T> | 10 August 2018, 09:09:40 UTC |
41173e3 | moneta | 13 June 2018, 09:07:14 UTC | Changes in MethodDL to use as default Float for the scalar type for the Deep networks for both training and prediction. Change in MethodDNN to use same type (float) for the prediction. Changes in CudaMatrix and Cpu Matrix to have conversion to and for TMatrixT to the same template type. Anyway explicit conversion from TMatrixD and TMatrixF exist | 10 August 2018, 09:09:40 UTC |
03d1f8c | moneta | 12 June 2018, 16:01:16 UTC | Optimise the MT for CPU by setting a minimum number of matrix elements to operate in a task. If the matrix elements are smaller then do not run in parallel with tab the operation | 10 August 2018, 09:09:40 UTC |
58da281 | Lorenzo Moneta | 11 June 2018, 09:43:08 UTC | print used Blas library | 10 August 2018, 09:09:40 UTC |
f5ff6c8 | moneta | 11 June 2018, 08:12:03 UTC | Fix Error message when tmva-cpu is not found and architecture cpu is selected | 10 August 2018, 09:09:40 UTC |
739d71c | moneta | 06 June 2018, 13:14:18 UTC | Fix compilation for when tmva-cpu is off | 10 August 2018, 09:09:40 UTC |
450c87d | moneta | 06 June 2018, 12:25:59 UTC | Do not compute the regularisation for each back when computing the Loss function | 10 August 2018, 09:09:40 UTC |
812d677 | moneta | 06 June 2018, 09:48:15 UTC | Fix computation of regularisation for Loss | 10 August 2018, 09:09:40 UTC |
906a14e | moneta | 06 June 2018, 09:07:17 UTC | Fix a bug in setting the regularisation to None for subsequents training phase/ Add a new function to DeepNet, DeepNet::RegulariztionTerm to compute the regularisation term separately and it can be now conputed after looping on the batch to speed up time. It was seen as giving a significant contributions on the tech lab GPU (e.g. 1080) | 10 August 2018, 09:09:40 UTC |
55d69b2 | moneta | 01 June 2018, 09:24:09 UTC | Fix using MT in CpuMatrix. Improve MT for Regularization and Dropout , by using chunking. Use in Droput general DL random generator | 10 August 2018, 09:09:40 UTC |