mmr/tkDNN
Archived
1
0
Fork 0
Code Issues 46 Pull Requests 12 Releases 4 Activity
608 Commits 11 Branches 5 Tags
56da10df647bd5cc7984a0b16520e25c2e87dfe2
Commit Graph

111 Commits

Author SHA1 Message Date
Micaela Verucchi 56da10df64 (tkDNN): Support Resnet-101-AP-GeM for CUDNN, TRT to be fixed 
Signed-off-by: Micaela Verucchi <micaelaverucchi@gmail.com>
 2022-09-24 16:57:12 +02:00
Francesco Gatti 5e71b99265 YoloRT save bias, mask and clasesName into RT file 2022-03-30 20:46:51 +02:00
Harshvardhan Chandirasekar 40266a6c32 Fixed tensorrt8 branch to work jetpack 4.5 and tensorrt7 
Signed-off-by: perseusdg <f20180523@goa.bits-pilani.ac.in>
 2022-03-16 18:09:32 +05:30
Harshvardhan Chandirasekar afdad8e661 Moved batchnorm and test_monodepth2_new_format layer to dev 
Signed-off-by: perseusdg <harshvardhan.chandira@gmail.com>
 2022-01-24 21:48:21 +05:30
Harshvardhan Chandirasekar 6a133d8dec -Replaced ISliceLayer based paddings(reflection,constant and zero) with the IPluginV2 version for TensorRT >= 8.2.0 
-Fixed demo build issue on Windows

Signed-off-by: perseusdg <harshvardhan.chandira@gmail.com>
 2022-01-20 00:52:22 +05:30
Micaela Verucchi 907df27e07 Fix monodepth2, add demoDepth: 
- Fix monodepth2 network, now works with both cuDNN and tensorRT
- Substitute cuDNN ELU with tkDNN one
- add DepthNN class
- add demoDepth demo, now only works with monodepth2 net

Signed-off-by: Micaela Verucchi <micaelaverucchi@gmail.com>
               Francesco Gatti <gattifrancesco@hotmail.it>
 2022-01-18 21:34:16 -08:00
perseusdg 061bc79a69 Added BatchNorm to NetworkRT (conver_layer) 2022-01-10 22:48:16 +05:30
perseusdg 3b58fbcb82 Added ConstantPadding plugin for TensorRT < 8.2 2022-01-07 13:05:09 +05:30
perseusdg f189efcbb1 Reflection Padding native plugin fix ,forgot to added input_dim.c in the plugin creator 2022-01-07 05:41:26 +00:00
perseusdg b75cecb105 Added constant and zero padding with ISliceLayer,need to add them for tensorrt versions less than 8.2 using IPluginV2Ext instead of ISliceLayer since they dont seem to support reflect and zero 2022-01-04 09:03:04 +05:30
perseusdg ba022663f1 completed padding migrations from github 2022-01-03 19:49:32 +05:30
perseusdg dcf4054bc6 performance improvements for yolo based networks,significant reduction in inference time can be seen yolo4tiny ,yolo4 and minor reduction in inference time can be seen in yolo4_berkeley_f1 and yolo4_berkeley - tested with a batchsize of 1 and 2 and on gtx 1070,it is possible that the performance improvement is more signficant in newer hardware 2022-01-03 13:52:16 +05:30
perseusdg a8c98e3c31 Signed-off-by: perseusdg <43143075+perseusdg@users.noreply.github.com> 
removed all TRT8_DEPRACTED functions
 2021-11-19 22:40:52 +05:30
perseusdg 6f936096ae Added num of layers in net and netRT,suggestions from pull request 270 on main repo 2021-11-14 07:13:42 +05:30
perseusdg 367061fea2 fps fix 2021-11-13 18:11:46 +05:30
perseusdg d488c3bd17 remove duplicated lines 2021-11-13 02:12:50 +05:30
perseusdg ee5000ccca bug fixes for dla networks and ported optimization from different pull request 2021-11-13 02:00:49 +05:30
perseusdg 9e328c0daa Migrate tkDNN max pooling plugin creation to pluginRegistry from the default method 2021-11-10 20:09:57 +05:30
perseusdg c5e66c6bf6 TRT8 works with almost every nerual network now!!!!(including demo3d) 2021-10-28 23:35:37 +05:30
perseusdg 8c36dd0431 mobilenet works with trt7(IPluginV2IOExt) ,need to test it with trt8 2021-10-27 18:00:33 +05:30
perseusdg 03473743c4 tkDNN works with trt8!!,need to test int8 and mobilenet,dla_cnet (fps seems to be a bit low 350 on trt8 compared to 396 on trt7) 2021-09-07 22:00:28 +05:30
perseusdg de83ae5d25 update tensorrt8 branch 2021-08-30 19:04:26 +05:30
perseusdg 2ffe07057e Mnist works at the moment with trt8,others like yolo4tiny and mobilenet generate the engine files but crash after throwing nvifer1::CudaRuntimeError and when demo is being run ,it doesnt deserialize properly and crashes 2021-08-29 03:18:58 +05:30
perseusdg 3d8b1ac494 Yolo3Detection.cpp doesn't build yet,issues with dependecies of preprocessing and postprocessing on IPluginFactory 2021-08-21 22:29:36 +05:30
Micaela Verucchi 84ff978ccb Merge with master, all tests passed 
Signed-off-by: Micaela Verucchi <micaelaverucchi@gmail.com>
 2021-07-20 12:48:41 +02:00
Francesco Gatti ba8199a030 fix compile error on old tensorrt and remove useless prints 2021-05-16 22:33:45 +02:00
Francesco Gatti 6611a91201 merge perseusdg windows10 support 2021-04-24 18:55:26 +02:00
perseusdg 1de804f98d Code cleanup and readme fixes 2021-04-09 13:17:41 +05:30
perseusdg cc594f09ef merge from gitlab 2021-04-08 00:28:00 +05:30
hchandirasekar f12ec3c935 added yolo4x and yolo4-csp from the github repo and download file corrections 2021-03-26 12:18:46 +05:30
hchandirasekar e94e1f7622 tkdnn first patch for windows 2021-03-24 22:26:46 +05:30
Harshvardhan Chandirasekar 06787a931f minor migrations 2021-03-17 23:06:50 +05:30
Harshvardhan Chandirasekar 304ab49897 shared_ptr migrations 2021-03-17 22:26:26 +05:30
Micaela Verucchi adac8576b0 Add support to Scaled-YOLO4, update Yolov4x-mish (tested) 
Signed-off-by: Micaela Verucchi <micaelaverucchi@gmail.com>
 2021-01-22 17:57:42 +01:00
Micaela Verucchi 702791e41a Add support for yolov4x-mish. 
Changes:
- add parameters nms_kind, nms_thresh, new_coords to yolo layer and darknet parser
- added diou nms, new method to compute the BBs
- created test for yolov4x-mish called yolo4x

Tested, all tests work.
Problem to solve: little loss in mAP of yolo4x

Signed-off-by: Micaela Verucchi <micaelaverucchi@gmail.com>
 2020-11-23 11:29:32 +01:00
micaela fa2b3d26cb Fix typos (#107) 
Signed-off-by: micaela <micaelaverucchi@gmail.com>
 2020-09-11 09:13:59 +02:00
Micaela Verucchi fe2e4eae92 yolo4tiny works on tensorRT 🐬 🐬 🐬 🐬 
Signed-off-by: Micaela Verucchi <micaelaverucchi@gmail.com>
 2020-06-30 15:52:58 +02:00
Micaela Verucchi 082920f3f5 Shelfnet works, also visualization. Postprocessing need to be parallelized 
Signed-off-by: Micaela Verucchi <micaelaverucchi@gmail.com>
 2020-06-23 20:01:47 +02:00
Micaela Verucchi 94e558003d Shelfnet works on tensorRT (shortcut need to be fixed) 
Signed-off-by: Micaela Verucchi <micaelaverucchi@gmail.com>
 2020-06-23 12:50:24 +02:00
Micaela Verucchi 6fd261f628 Shelfnet works on cuDNN. To test everything else 
Signed-off-by: Micaela Verucchi <micaelaverucchi@gmail.com>
 2020-06-22 18:11:19 +02:00
Francesco Gatti cbfc8ea4f2 serialize fix 2020-06-14 13:01:45 +02:00
Francesco Gatti 62e4a3f779 memory release 2020-06-02 12:43:06 +02:00
Francesco Gatti d2e2669b6d darknet parse all net to be tested 2020-06-01 12:22:55 +02:00
Micaela Verucchi db0f8a4d99 Support yolov4 
Signed-off-by: Micaela Verucchi <micaelaverucchi@gmail.com>
 2020-04-28 20:35:30 +02:00
Francesco Gatti 3d940a9fa2 batch seems ok in yolo3_berkely 
layers to be checked:
DeformableConvRT
FlattenConcatRT
ReshapeRT
RouteRT (dont know why but seems working)
 2020-04-21 19:41:40 +00:00
Francesco Gatti 7c81c5a43c batch size > 1 2020-04-21 19:20:44 +02:00
Micaela Verucchi 2fbac7705d Move download of weights inside build folder 
Signed-off-by: Micaela Verucchi <micaelaverucchi@gmail.com>
 2020-04-17 22:36:25 +02:00
Micaela Verucchi 3502c5b676 Fix pooling, add return code in each test, add return code handling in test_all_tests script 
Signed-off-by: Micaela Verucchi <micaelaverucchi@gmail.com>
 2020-04-09 15:25:05 +02:00
Micaela Verucchi 6d1fda0c21 Compile on tx2 
Signed-off-by: Micaela Verucchi <micaelaverucchi@gmail.com>
 2020-04-08 16:23:08 +02:00
Davide Sapienza 26dcb7b8c9 Fix the tensorRT code to support versions prior to 6.0 
Signed-off-by: Davide Sapienza <sapienza.dav@gmail.com>
 2020-04-06 19:54:40 +02:00