aboutsummaryrefslogtreecommitdiffhomepage
Commit message (Collapse)AuthorAge
* Upgraded to the latest version of Eigen that supports convolutions on fp16Gravatar Benoit Steiner2016-05-25
| | | | Change: 123238579
* StreamExecutor add CUDA support for cudnnConvolutionBackwardBiasGravatar A. Unique TensorFlower2016-05-25
| | | | Change: 123233121
* Use tf.GraphKeys.UPDATE_OPS as the default updates_collection for batch_norm.Gravatar A. Unique TensorFlower2016-05-25
| | | | Change: 123227324
* Fix RandomShuffle for huge tensors.Gravatar Josh Levenberg2016-05-25
| | | | Change: 123226210
* Make it possible to override the node color in dot graphs via DotOptions.Gravatar A. Unique TensorFlower2016-05-25
| | | | Change: 123212554
* Speed up install of pip packages; solve test failuresGravatar Shanqing Cai2016-05-25
| | | | | | | | | | | | | Previously, pip install of scipy and sklearn builds from source and takes a long time. This CL makes the packages install from binary whl files downloaded from PyPI. Docker build (Dockerfile.cpu) performance improvement (no caching): Without this CL: 24'23'' With this CL: 9'48'' The CL also solves two build failures that surfaced lately, including: http://ci.tensorflow.org/view/Nightly/job/nigntly-matrix-linux-gpu/TF_BUILD_CONTAINER_TYPE=GPU,TF_BUILD_IS_OPT=OPT,TF_BUILD_IS_PIP=PIP,TF_BUILD_PYTHON_VERSION=PYTHON3,label=gpu-linux/105/console http://ci.tensorflow.org/view/Nightly/job/nightly-matrix-cpu/TF_BUILD_CONTAINER_TYPE=CPU,TF_BUILD_IS_OPT=OPT,TF_BUILD_IS_PIP=PIP,TF_BUILD_PYTHON_VERSION=PYTHON3,label=cpu-slave/104/console Change: 123201123
* Update generated Python Op docs.Gravatar A. Unique TensorFlower2016-05-25
| | | | Change: 123197297
* Update ops-related pbtxt files.Gravatar A. Unique TensorFlower2016-05-25
| | | | Change: 123197154
* fp16-enable all image resize ops. (Most of them already had code builtGravatar A. Unique TensorFlower2016-05-25
| | | | | with fp16, and just needed declaration in the op.) Change: 123196867
* Fix an bug when either linear or dnn columns were empty.Gravatar Martin Wicke2016-05-24
| | | | Change: 123180783
* tf.learn: Remove unused trainer.py.Gravatar Illia Polosukhin2016-05-24
| | | | Change: 123180545
* Fixed predictions in Linear/Dnn/Combined estimators.Gravatar Mustafa Ispir2016-05-24
| | | | | | Aligned usage of metrics with the metrics API. Added custom metrics and prediction tests. Change: 123178457
* Turn on CUDNN autotune by default.Gravatar Xiaoqiang Zheng2016-05-24
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | * For Soumith's Convnet benchmarks, * GoogleNet V1 forward only becomes 2.82% faster. * VGG forward+backward becomes 3.48% faster. * GoogleNet V1 forwrad+backwar becomes 1.66% faster. * For Inception model at batch size 32, on Titan-X, * The step size is reduced from 1.11 sec to 0.97 sec. A 16.32% improvement. * For microbenchmarks. Here are the improvement. Benchmark Base (ns) New (ns) Improvement ------------------------------------------------------------------ BM_ConvFloatFwdGPU_conv0 235083 230767 +1.8% BM_ConvFloatFwdGPU_conv1 1075211 1107800 -3.0% BM_ConvFloatFwdGPU_conv2 1015770 1044204 -2.8% BM_ConvFloatFwdGPU_conv3 1338677 1333840 +0.4% BM_ConvFloatFwdGPU_conv4 1724488 1727875 -0.2% BM_ConvFloatFwdGPU_conv5 1504656 1531122 -1.8% BM_ConvFloatFwdGPU_conv6 1912314 1921835 -0.5% BM_ConvFloatFwdGPU_conv7 833252 812320 +2.5% BM_ConvFloatFwdGPU_conv8 704842 742914 -5.4% BM_ConvFloatFwdGPU_conv9 1181595 1153533 +2.4% BM_ConvFloatFwdGPU_conv10 1626990 1602748 +1.5% BM_ConvFloatFwdGPU_conv11 1266993 1334205 -5.3% BM_ConvFloatFwdGPU_conv12 778462 767860 +1.4% BM_ConvFloatFwdGPU_conv13 3850331 2107377 +45.3% BM_ConvFloatFwdGPU_conv14 4126061 4180073 -1.3% BM_ConvFloatFwdGPU_conv15 678327 675829 +0.4% BM_ConvFloatFwdGPU_conv16 1337845 1324671 +1.0% BM_ConvFloatFwdGPU_conv17 1605443 1609892 -0.3% BM_ConvFloatFwdGPU_conv18 1501101 1504725 -0.2% BM_ConvFloatFwdGPU_conv19 1591419 1465860 +7.9% BM_ConvFloatFwdGPU_conv20 3978635 4008382 -0.7% BM_ConvFloatFwdGPU_conv21 1512956 1491781 +1.4% BM_ConvFloatFwdGPU_conv22 1512534 1492847 +1.3% BM_ConvFloatFwdGPU_conv23 4250634 2449213 +42.4% BM_ConvFloatFwdGPU_conv24 1252755 1250047 +0.2% BM_ConvFloatFwdGPU_conv25 3771888 3727033 +1.2% BM_ConvFloatFwdGPU_conv26 1176322 1188693 -1.1% BM_ConvFloatFwdGPU_conv27 1190219 1166078 +2.0% BM_ConvFloatFwdGPU_conv28 1736335 1738561 -0.1% BM_ConvFloatFwdGPU_conv29 2470491 2526576 -2.3% BM_ConvFloatFwdGPU_conv30 880584 845829 +3.9% BM_ConvFloatFwdGPU_conv31 950092 971105 -2.2% BM_ConvFloatFwdGPU_conv32 1968954 1987465 -0.9% BM_ConvFloatFwdGPU_conv33 918658 900723 +2.0% BM_ConvFloatFwdGPU_conv34 1458851 1462193 -0.2% BM_ConvFloatFwdGPU_conv35 687912 667973 +2.9% BM_ConvFloatFwdGPU_conv36 2475984 2447614 +1.1% BM_ConvFloatFwdGPU_conv37 691891 693028 -0.2% BM_ConvFloatFwdGPU_conv38 1024818 1028796 -0.4% BM_ConvFloatFwdGPU_conv39 792712 804433 -1.5% BM_ConvFloatFwdGPU_conv40 2866529 2831128 +1.2% BM_ConvFloatFwdGPU_conv41 825719 821005 +0.6% BM_ConvFloatFwdGPU_conv42 6178692 6055745 +2.0% BM_ConvFloatFwdGPU_conv43 1770585 1758200 +0.7% BM_ConvFloatFwdGPU_conv44 1101019 1121091 -1.8% BM_ConvFloatFwdGPU_conv45 953025 974867 -2.3% BM_ConvFloatFwdGPU_conv46 1976171 1907614 +3.5% BM_ConvFloatFwdGPU_conv47 926263 930521 -0.5% BM_ConvFloatFwdGPU_conv48 2486172 2451860 +1.4% BM_ConvFloatFwdGPU_conv49 612463 619752 -1.2% BM_ConvFloatFwdGPU_conv50 669415 688190 -2.8% BM_ConvFloatFwdGPU_conv51 669922 642478 +4.1% BM_ConvFloatFwdGPU_conv52 13375846 13288659 +0.7% BM_ConvFloatFwdGPU_conv53 1165725 1180657 -1.3% BM_ConvFloatFwdGPU_conv54 8067519 7854240 +2.6% BM_ConvFloatBkInGPU_conv0 211400 182926 +13.5% BM_ConvFloatBkFilterGPU_conv0 202238 167241 +17.3% BM_ConvFloatBkInGPU_conv1 1498547 1037392 +30.8% BM_ConvFloatBkFilterGPU_conv1 781429 754140 +3.5% BM_ConvFloatBkInGPU_conv2 1542272 970764 +37.1% BM_ConvFloatBkFilterGPU_conv2 827241 812901 +1.7% BM_ConvFloatBkInGPU_conv3 554840 575366 -3.7% BM_ConvFloatBkFilterGPU_conv3 1390099 1297130 +6.7% BM_ConvFloatBkInGPU_conv4 2679557 2629239 +1.9% BM_ConvFloatBkFilterGPU_conv4 2391921 2415243 -1.0% BM_ConvFloatBkInGPU_conv5 775991 819557 -5.6% BM_ConvFloatBkFilterGPU_conv5 1518350 1555425 -2.4% BM_ConvFloatBkInGPU_conv6 1164620 1124455 +3.4% BM_ConvFloatBkFilterGPU_conv6 1886563 1878187 +0.4% BM_ConvFloatBkInGPU_conv7 1014010 997501 +1.6% BM_ConvFloatBkFilterGPU_conv7 836499 818683 +2.1% BM_ConvFloatBkInGPU_conv8 1096261 976079 +11.0% BM_ConvFloatBkFilterGPU_conv8 819271 809679 +1.2% BM_ConvFloatBkInGPU_conv9 638050 599533 +6.0% BM_ConvFloatBkFilterGPU_conv9 1204812 1178843 +2.2% BM_ConvFloatBkInGPU_conv10 1158430 1223196 -5.6% BM_ConvFloatBkFilterGPU_conv10 1732046 1718558 +0.8% BM_ConvFloatBkInGPU_conv11 940582 890771 +5.3% BM_ConvFloatBkFilterGPU_conv11 1538670 1436865 +6.6% BM_ConvFloatBkInGPU_conv12 6819839 960485 +85.9% BM_ConvFloatBkFilterGPU_conv12 686978 730785 -6.4% BM_ConvFloatBkInGPU_conv13 2193316 2206764 -0.6% BM_ConvFloatBkFilterGPU_conv13 3938868 2091134 +46.9% BM_ConvFloatBkInGPU_conv14 2035871 2138318 -5.0% BM_ConvFloatBkFilterGPU_conv14 4029626 4033444 -0.1% BM_ConvFloatBkInGPU_conv15 6997156 890109 +87.3% BM_ConvFloatBkFilterGPU_conv15 740402 701366 +5.3% BM_ConvFloatBkInGPU_conv16 1424744 1406938 +1.2% BM_ConvFloatBkFilterGPU_conv16 1671854 1462868 +12.5% BM_ConvFloatBkInGPU_conv17 2700862 1992674 +26.2% BM_ConvFloatBkFilterGPU_conv17 1305656 1322830 -1.3% BM_ConvFloatBkInGPU_conv18 2957025 1864698 +36.9% BM_ConvFloatBkFilterGPU_conv18 1225843 1221011 +0.4% BM_ConvFloatBkInGPU_conv19 2983442 1838917 +38.4% BM_ConvFloatBkFilterGPU_conv19 1143908 1181473 -3.3% BM_ConvFloatBkInGPU_conv20 1746891 1792048 -2.6% BM_ConvFloatBkFilterGPU_conv20 3858859 3947101 -2.3% BM_ConvFloatBkInGPU_conv21 1049381 1057465 -0.8% BM_ConvFloatBkFilterGPU_conv21 1960184 1963597 -0.2% BM_ConvFloatBkInGPU_conv22 2709485 1962671 +27.6% BM_ConvFloatBkFilterGPU_conv22 1347473 1337113 +0.8% BM_ConvFloatBkInGPU_conv23 2488277 2444806 +1.7% BM_ConvFloatBkFilterGPU_conv23 2393383 2361463 +1.3% BM_ConvFloatBkInGPU_conv24 2317770 1555267 +32.9% BM_ConvFloatBkFilterGPU_conv24 1005172 987688 +1.7% BM_ConvFloatBkInGPU_conv25 1282727 1313422 -2.4% BM_ConvFloatBkFilterGPU_conv25 3467895 3520604 -1.5% BM_ConvFloatBkInGPU_conv26 931302 887955 +4.7% BM_ConvFloatBkFilterGPU_conv26 1413088 1348387 +4.6% BM_ConvFloatBkInGPU_conv27 2285721 1501425 +34.3% BM_ConvFloatBkFilterGPU_conv27 1209520 1168316 +3.4% BM_ConvFloatBkInGPU_conv28 2157998 2157376 +0.0% BM_ConvFloatBkFilterGPU_conv28 3074795 1853044 +39.7% BM_ConvFloatBkInGPU_conv29 1144831 1075297 +6.1% BM_ConvFloatBkFilterGPU_conv29 2340646 2340184 +0.0% BM_ConvFloatBkInGPU_conv30 858060 837645 +2.4% BM_ConvFloatBkFilterGPU_conv30 1315830 1353214 -2.8% BM_ConvFloatBkInGPU_conv31 1224674 1128456 +7.9% BM_ConvFloatBkFilterGPU_conv31 707870 726953 -2.7% BM_ConvFloatBkInGPU_conv32 996074 1014381 -1.8% BM_ConvFloatBkFilterGPU_conv32 2107132 2063072 +2.1% BM_ConvFloatBkInGPU_conv33 1223802 1110516 +9.3% BM_ConvFloatBkFilterGPU_conv33 912262 862748 +5.4% BM_ConvFloatBkInGPU_conv34 1466738 1551351 -5.8% BM_ConvFloatBkFilterGPU_conv34 1974404 1923035 +2.6% BM_ConvFloatBkInGPU_conv35 922659 939845 -1.9% BM_ConvFloatBkFilterGPU_conv35 612561 568035 +7.3% BM_ConvFloatBkInGPU_conv36 908895 895344 +1.5% BM_ConvFloatBkFilterGPU_conv36 2953348 2899676 +1.8% BM_ConvFloatBkInGPU_conv37 938952 892167 +5.0% BM_ConvFloatBkFilterGPU_conv37 577438 569059 +1.5% BM_ConvFloatBkInGPU_conv38 1138055 1096089 +3.7% BM_ConvFloatBkFilterGPU_conv38 1011368 1008415 +0.3% BM_ConvFloatBkInGPU_conv39 668144 673298 -0.8% BM_ConvFloatBkFilterGPU_conv39 1358847 1298098 +4.5% BM_ConvFloatBkInGPU_conv40 1380139 1331826 +3.5% BM_ConvFloatBkFilterGPU_conv40 3541527 3069069 +13.3% BM_ConvFloatBkInGPU_conv41 1638383 1595251 +2.6% BM_ConvFloatBkFilterGPU_conv41 1005443 987946 +1.7% BM_ConvFloatBkInGPU_conv42 17024559 10725787 +37.0% BM_ConvFloatBkFilterGPU_conv42 6567765 6515355 +0.8% BM_ConvFloatBkInGPU_conv43 1780598 1708543 +4.0% BM_ConvFloatBkFilterGPU_conv43 2356016 2281999 +3.1% BM_ConvFloatBkInGPU_conv44 931335 971200 -4.3% BM_ConvFloatBkFilterGPU_conv44 1346236 1339928 +0.5% BM_ConvFloatBkInGPU_conv45 610336 608156 +0.4% BM_ConvFloatBkFilterGPU_conv45 1247724 1208773 +3.1% BM_ConvFloatBkInGPU_conv46 3368269 2161475 +35.8% BM_ConvFloatBkFilterGPU_conv46 2161988 2140970 +1.0% BM_ConvFloatBkInGPU_conv47 500600 549664 -9.8% BM_ConvFloatBkFilterGPU_conv47 1239103 1201332 +3.0% BM_ConvFloatBkInGPU_conv48 2505748 2487250 +0.7% BM_ConvFloatBkFilterGPU_conv48 3181887 3196408 -0.5% BM_ConvFloatBkInGPU_conv49 654636 752578 -15.0% BM_ConvFloatBkFilterGPU_conv49 614054 640264 -4.3% BM_ConvFloatBkInGPU_conv50 1046576 1022585 +2.3% BM_ConvFloatBkFilterGPU_conv50 928998 884173 +4.8% BM_ConvFloatBkInGPU_conv51 831912 805962 +3.1% BM_ConvFloatBkFilterGPU_conv51 833301 890314 -6.8% BM_ConvFloatBkInGPU_conv52 13575989 13244294 +2.4% BM_ConvFloatBkFilterGPU_conv52 26960865 14528291 +46.1% BM_ConvFloatBkInGPU_conv53 1212746 1193415 +1.6% BM_ConvFloatBkFilterGPU_conv53 1617787 1599532 +1.1% BM_ConvFloatBkInGPU_conv54 7143853 9979079 -39.7% BM_ConvFloatBkFilterGPU_conv54 7261642 9516172 -31.0% BM_ConvFloatBkFGPU_128_128_128_3_96_11_11 65676662 67077190 -2.1% BM_ConvFloatBkFGPU_128_64_64_64_128_9_9 30481444 30479591 +0.0% BM_ConvFloatBkFGPU_128_32_32_128_128_9_9 9184052 9309441 -1.4% BM_ConvFloatBkFGPU_128_16_16_128_128_7_7 1783974 1728034 +3.1% BM_ConvFloatBkFGPU_128_13_13_384_384_3_3 9793620 9728267 +0.7% BM_ConvFloatDepthwiseFwdGPU_conv0 2485053 2423786 +2.5% BM_ConvFloatDepthwiseFwdGPU_conv1 9232311 9385025 -1.7% BM_ConvFloatDepthwiseFwdGPU_conv2 3951763 4081355 -3.3% BM_ConvFloatDepthwiseFwdGPU_conv3 1072853 1075711 -0.3% BM_ConvFloatDepthwiseFwdGPU_conv4 857038 834950 +2.6% BM_ConvFloatDepthwiseFwdGPU_conv5 849175 851622 -0.3% BM_ConvFloatDepthwiseFwdGPU_conv6 492490 463820 +5.8% BM_ConvFloatDepthwiseFwdGPU_conv7 699378 715631 -2.3% BM_ConvFloatDepthwiseFwdGPU_conv8 655144 622416 +5.0% BM_ConvFloatDepthwiseBkInGPU_conv0 2521530 2564153 -1.7% BM_ConvFloatDepthwiseBkFilterGPU_conv0 65102549 65026603 +0.1% BM_ConvFloatDepthwiseBkInGPU_conv1 9572755 11527412 -20.4% BM_ConvFloatDepthwiseBkFilterGPU_conv1 67341583 66854785 +0.7% BM_ConvFloatDepthwiseBkInGPU_conv2 5038497 5191688 -3.0% BM_ConvFloatDepthwiseBkFilterGPU_conv2 17959694 18075388 -0.6% BM_ConvFloatDepthwiseBkInGPU_conv3 2733609 2583003 +5.5% BM_ConvFloatDepthwiseBkFilterGPU_conv3 5430360 5479402 -0.9% BM_ConvFloatDepthwiseBkInGPU_conv4 895653 933757 -4.3% BM_ConvFloatDepthwiseBkFilterGPU_conv4 8054025 8008655 +0.6% BM_ConvFloatDepthwiseBkInGPU_conv5 859509 818970 +4.7% BM_ConvFloatDepthwiseBkFilterGPU_conv5 3151537 3052507 +3.1% BM_ConvFloatDepthwiseBkInGPU_conv6 482079 474622 +1.5% BM_ConvFloatDepthwiseBkFilterGPU_conv6 1040890 1048557 -0.7% BM_ConvFloatDepthwiseBkInGPU_conv7 920856 905214 +1.7% BM_ConvFloatDepthwiseBkFilterGPU_conv7 16564049 16556264 +0.0% BM_ConvFloatDepthwiseBkInGPU_conv8 862814 864988 -0.3% BM_ConvFloatDepthwiseBkFilterGPU_conv8 16168442 16169527 -0.0% Change: 123168911
* Update generated Python Op docs.Gravatar A. Unique TensorFlower2016-05-24
| | | | Change: 123168816
* Update ops-related pbtxt files.Gravatar A. Unique TensorFlower2016-05-24
| | | | Change: 123168685
* Merge changes from github.Gravatar Vijay Vasudevan2016-05-24
| | | | Change: 123167405
* Run update_op in every step of evaluation. Calculate evaluate metric at log ↵Gravatar Mustafa Ispir2016-05-24
| | | | | | steps and at final step. Change: 123167151
* Summary ops should run on only Chef not on all workers.Gravatar Mustafa Ispir2016-05-24
| | | | Change: 123163672
* Register GPU device OpKernel for Mod, for int32 type.Gravatar Manjunath Kudlur2016-05-24
| | | | Change: 123160537
* Add support for arbitrarily nested tuples for RNN state.Gravatar Eugene Brevdo2016-05-24
| | | | | Also fixed a bug in the RNN unit tests. Change: 123150781
* Remove heuristic caps on parallelism that should now be handled by cost model.Gravatar A. Unique TensorFlower2016-05-24
| | | | | | | Adjust cost model for FloatToBFloat16 and BFloat16ToFloat. They do not take 100 cycles per element. This cl is a companion to cl/122779011, which makes the caps effective again, even with the nonblocking threadpool. Change: 123144919
* Enforce max_parallelism in work sharder by reverting to old sharding code if ↵Gravatar A. Unique TensorFlower2016-05-24
| | | | | | max_parallelism is less than number of workers in the pool. Change: 123139558
* Update generated Python Op docs.Gravatar A. Unique TensorFlower2016-05-24
| | | | Change: 123137273
* Update ops-related pbtxt files.Gravatar A. Unique TensorFlower2016-05-24
| | | | Change: 123136978
* Add sparse_softmax() op.Gravatar Zongheng Yang2016-05-24
| | | | Change: 123133067
* Update generated Python Op docs.Gravatar A. Unique TensorFlower2016-05-24
| | | | Change: 123133051
* - Allowing sparse_concat to handle sparse tensors with different sizes in ↵Gravatar A. Unique TensorFlower2016-05-24
| | | | | | | | the non- concating dimensions. Default behavior unchanged. - Adding a sparse_reset_shape which can be used to set the shape of a SparseTensor to its tight bounding box or set to a new shape with each dimension equal or larger. Change: 123130219
* Added losses_collections to allow to collect losses into collections.Gravatar A. Unique TensorFlower2016-05-24
| | | | Change: 123128114
* Add DataFrame.to_input_fn() to provide data to Estimator.Gravatar David Soergel2016-05-24
| | | | Change: 123127104
* Use running mean as a shift value to stabilize moment computation in layers.Gravatar Vincent Vanhoucke2016-05-24
| | | | Change: 123125587
* Set upper limit on the number of the number dense features to ↵Gravatar A. Unique TensorFlower2016-05-24
| | | | | | std::numeric_limits<int32>::max(). Change: 123124522
* Update ops-related pbtxt files.Gravatar A. Unique TensorFlower2016-05-24
| | | | Change: 123122170
* Adds streaming AUC for binary classification in dnnlinearcombined models.Gravatar A. Unique TensorFlower2016-05-24
| | | | Change: 123122027
* Register a new SparseDenseCwiseAdd op and add its Python sugar.Gravatar Zongheng Yang2016-05-24
| | | | | | | It has the special semantics in that not all dense-side values participate in the calculation. Hence, we do not override the Python "+" (__add__) operator. Change: 123118195
* Update generated Python Op docs.Gravatar A. Unique TensorFlower2016-05-24
| | | | Change: 123112607
* Update ops-related pbtxt files.Gravatar A. Unique TensorFlower2016-05-24
| | | | Change: 123112375
* Make SummarizeImageOp work with fp16 images.Gravatar A. Unique TensorFlower2016-05-24
| | | | Change: 123110286
* Documentation fix for tf.nn.sufficient_statistics().Gravatar A. Unique TensorFlower2016-05-24
| | | | Change: 123109819
* Update generated Python Op docs.Gravatar A. Unique TensorFlower2016-05-24
| | | | Change: 123108585
* Add max shards parameter to variable_axis_size_partitioner.Gravatar A. Unique TensorFlower2016-05-24
| | | | Change: 123107048
* Update generated Python Op docs.Gravatar A. Unique TensorFlower2016-05-24
| | | | Change: 123101962
* Update ops-related pbtxt files.Gravatar A. Unique TensorFlower2016-05-24
| | | | Change: 123101786
* Add fp16 support to the DrawBoundingBoxes op; used for e.g. Inception.Gravatar A. Unique TensorFlower2016-05-24
| | | | | | | (The bounding box is still in fp32, since it most likely comes from ExampleReader, which doesn't support fp16. In any case, it's much less data.) Change: 123100239
* Update generated Python Op docs.Gravatar A. Unique TensorFlower2016-05-24
| | | | Change: 123096337
* Allow callers to provide a shift value in tf.nn.moments().Gravatar A. Unique TensorFlower2016-05-24
| | | | Change: 123095477
* Update generated Python Op docs.Gravatar A. Unique TensorFlower2016-05-24
| | | | Change: 123081047
* Update ops-related pbtxt files.Gravatar A. Unique TensorFlower2016-05-24
| | | | Change: 123080651
* Add a strong keyed hash function based on highwayhash's siphash.Gravatar Yutaka Leon2016-05-24
| | | | | Add string_to_hash_bucket_strong to assign hash buckets using the strong keyed hash function. Change: 123080459
* Speeds up Softmax by up to 43%, by changing "/ sum" to "* (1/sum)".Gravatar Zongheng Yang2016-05-23
| | | | | | | | | | | | | | | | | | | | | | | Benchmarked using third_party/tensorflow/core/kernels:nn_ops_test. Wall time improves 10-43%: Benchmark Base (ns) New (ns) Improvement ------------------------------------------------------------------ BM_ImageNetSoftmaxFwd_32_1008_1 713325 620705 +13.0% BM_ImageNetSoftmaxFwd_128_1008_1 3097766 2782433 +10.2% BM_ImageNetSoftmaxFwd_32_1008_4 1254561 703238 +43.9% BM_ImageNetSoftmaxFwd_128_1008_4 3225011 2543525 +21.1% CPU time improves 4-17%: Benchmark Base (ns) New (ns) Improvement ------------------------------------------------------------------ BM_ImageNetSoftmaxFwd_32_1008_1 711375 618729 +13.0% BM_ImageNetSoftmaxFwd_128_1008_1 3087158 2779777 +10.0% BM_ImageNetSoftmaxFwd_32_1008_4 959016 795579 +17.0% BM_ImageNetSoftmaxFwd_128_1008_4 3774543 3591573 +4.8% Change: 123074430
* Inflow: establish Column, Transform, and DataFrame abstractions.Gravatar David Soergel2016-05-23
| | | | | Part of a series of CLs setting up the minimal Inflow. Change: 123072135
generated by cgit v1.2.3 (git 2.39.1) at 2024-10-15 12:54:49 +0000