58c8746e
fix(opr): fix fast-run error in cuda by
2021-03-16 13:35:36 +0800
ba2ad46e
feat(gopt): add deconv nchw4 int8 opt pass, add deconv nchw int8 by
2021-03-08 13:49:43 +0800
5d350fc8
feat(dnn/cuda): add deconv int8 and fix cutlass conv wrapper base on modify cutlass 2.4 by
2021-03-05 17:40:25 +0800
4917534b
feat(mge/optimizer): save state's numpy value by default in `state_dict` by
2021-03-16 13:38:41 +0800
84f990a0
fix(mge/tensor): fix `__repr__` for Tensor's subclass by
2021-03-16 11:29:06 +0800
cdbb4a20
fix(mge/tensor): fix tensor's serialization behavior by
2021-03-12 18:06:50 +0800
9da26407
feat(ci): and model to test algo policy compatible by
2021-03-11 17:00:05 +0800
e19b9af1
feat(imperative): add bit combined enum to python C extension by
2021-03-09 17:53:06 +0800
a3ea1f15
feat(mgb/opr): add fast profile and combined Execution strategy by
2020-11-30 21:29:01 +0800
80f00643
fix(cuda): show more when cuda env check fail add cublas version check when build whl by
2021-03-15 14:33:33 +0800
89e52281
fix(build/lite): fix lite rc4_encryptor build failed at new ld env when cuda enable: by
2021-03-12 18:22:11 +0800
992a90bb
docs(mge/quantization): add docstring for Observer by
2021-03-16 18:59:23 +0800
c45f1eb2
fix(mge/tools): improve `module_visualize` result's robustness and beauty by
2021-03-16 16:27:38 +0800
786c36ff
fix(mge/tools): rename `net_stats` in function and examples to match file name by
2021-03-16 15:28:11 +0800
9c90ce8c
fix(mge/quantization): fix `quantize` and `quantize_qat`'s `set_expand_structure` arguments by
2021-03-16 15:15:51 +0800
c8697a70
feat(imperative/src): python wrapper for cambricon and atlas runtime opr by
2021-03-02 13:29:12 +0800
8cfed4a1
refactor(functional/nn): remove nn.svd by
2021-03-15 15:28:06 +0800
e6b06914
feat(mge/tensor): support non-Tensor value in `_reset` and remove depreciated tests by
2021-03-12 17:20:30 +0800
5a38ad39
feat(mge/utils): add get/set_expand_structure to deal with complex key by
2021-03-11 19:17:07 +0800
fad5bc74
feat(mge/tools): add load_network_and_run for megengine by
2021-03-04 17:24:45 +0800
3f2e15f5
fix(mge/logger): fix set_mgb_log_level by
2021-03-11 15:18:44 +0800
594fa722
feat(mge/graph): add `modify_opr_algo_strategy_inplace` for fast-run by
2021-03-11 15:23:18 +0800
813be72c
feat(mge/utils): refactor GraphInference and add more options by
2021-03-04 17:20:33 +0800
d970b85d
feat(functional): add conv3d by
2021-03-09 19:45:46 +0800
46cad4d3
feat(functional/ops): add _assert_equal by
2021-03-11 13:42:13 +0800
585aa561
fix(functional): replace usage of deprecated get_conv_execution_strategy() by
2021-03-11 13:41:33 +0800
dcfb6a53
refactor(mge/functional): move functional api by
2021-03-02 12:00:12 +0800
bfb9def9
test(mge/test): reorganize utils test files by
2021-03-02 11:30:00 +0800
53075cd3
feat(mge/experimental): add visualization and net stats for python graph by
2021-02-25 11:17:34 +0800
ae3123b3
feat(mge): add python graph for mgb graph editing by
2021-02-05 16:43:25 +0800
c82d8875
fix(dnn/cuda): add cuda nchw int8 conv impl with nchw4 to fix cu111 compatibility by
2021-03-10 17:58:08 +0800
0d096855
fix(imperative/quantization): add conditional when passing parameters name by
2021-03-08 14:25:58 +0800
652ec9f2
fix(mgb/dnn): fix backward computation of tqt by
2021-03-08 14:07:08 +0800
27638461
fix(mge/module): support list/dict/tuple in module __repr__ by
2021-03-04 14:36:28 +0800
07826f5e
test(dataloder): fix dataloader multiprocess test by
2021-03-04 21:13:24 +0800
fe9c6e26
feat(imperative/opr): deprecate resize op and make it as a special case of interpolate by
2021-03-09 16:48:00 +0800
798ae5e5
fix(imperative/opr): close SCALAR_IDX warning of IndexingMultiAxisVec for proxy_graph by
2021-03-09 14:10:21 +0800
1670a8cf
docs(mge/quantization): fix wrong unexpected unindent in docstring by
2021-03-09 15:12:37 +0800
365e80b5
fix(build): fix cu111 cudnn rpath by
2021-03-11 15:46:24 +0800
c89c6735
feat(lite): add windows lite whl scripts, enable megenginelite pytest and enable cpp test in ci by
2021-03-06 21:52:11 +0800
4c6484b7
feat(lite): add macos lite whl scripts, enable megenginelite pytest and enable cpp test in ci by
2021-03-05 10:38:38 +0800
ef079b84
feat(lite): add linux lite whl scripts, enable megenginelite pytest/cmake gtest in ci by
2021-03-02 19:26:29 +0800
3fab21ed
fix(lite): fix BUILD and misc by
2021-02-25 16:27:44 +0800
65c3159c
feat(lite): add version interface by
2021-02-23 13:42:53 +0800
af60ff97
feat(lite): add cmake build for lite by
2021-01-18 19:00:30 +0800
1ee331dd
fix(git): fix ci resource lfs, caused by f5993 by
2021-03-11 08:58:37 +0800
f2b42bf0
chore(dotprod): add arm dotprod attribute for easy use by
2021-02-07 10:02:26 +0800
fa4bf168
feat(mge/functional): add repeat and tile opr by
2020-09-30 22:54:17 +0800
c33a7173
feat(dnn): repalce is_reproducible with algo attribute in opencl, cpu, rocm and cuda by
2020-11-30 21:32:26 +0800
249a116b
fix(linux/whl): fix megengine whl build faild at manylinux2014 as > cuda11 cudnn8 trt7 have some dlopen libs, auditwheel can not handle correctly, also at lastest auditwheel, NEW LIBS dirs change from package data dir to package dir(eg, megengine to MegEngine), which lead to copy manually libs change rpath hard with auditwheel rpath, so do not auditwheel now! by
2021-03-03 13:55:07 +0800
0c37a588
fix(mge/functional): fix F.ones when input is a tensor of scalar type by
2021-02-24 11:39:16 +0800
b0944dc7
fix(mgb): fix tx1 compile by
2021-03-01 23:42:12 +0800
ce88e6c4
feat(mge/quantization): use extra act_fakequant to decide whether to do bias fakequant by
2021-02-25 17:12:22 +0800
1d7dd001
feat(mge/quantization): add QParams and QuantDtypeMeta for quantization data structure by
2021-02-09 18:06:55 +0800
4130dcd3
fix(mge/quantization): fix QATModule filter in `reset_qconfig` and `hook_qat_module` by
2021-02-22 15:38:54 +0800
97beae2f
fix(megdnn): fix megdnn benchmark testcase by
2021-03-02 17:39:53 +0800
1d4baa42
fix(imperative/dtype): fix is_dtype_equal by
2021-03-02 17:24:42 +0800
0035efa6
fix(mge): limit task queue size by
2021-02-24 15:26:12 +0800
5a1f9134
chore(imperative): remove unnecessary function template by
2021-02-26 19:38:52 +0800
2de2222e
feat(dnn/cuda): add cutlass batched gemv kernel for matmul operator by
2021-01-19 16:22:45 +0800
973d2a0a
feat(dnn/cuda): add cutlass matmul using split k parallel by
2021-01-13 16:42:34 +0800
03c921f7
feat(dnn/cuda): add cutlass matmul impls by
2021-01-13 13:29:26 +0800
886e7c6e
refactor(mge/tensor): support scalar in detach and grad by
2021-02-04 13:52:52 +0800
a1ca50c9
feat(mge/quantization): add name for quantized module by
2021-02-25 13:32:54 +0800
d0f70a44
fix(imperative/utils): module parameters' name do not have scope after dumping by
2021-02-25 11:07:42 +0800
b9bbf802
fix(mgb/opr): fix string append error by
2021-02-26 19:02:09 +0800
47bdb4ef
fix(mgb/core): add MGB_HAVE_THREAD macro to thread_local by
2021-02-07 12:59:46 +0800
5b62acfa
feat(dnn/armv7): add new matmul strategy k8x8x4 by
2021-01-27 16:59:55 +0800
ad87f78a
chore(imperative): refine tblgen for generating op name by
2021-02-26 21:25:36 +0800
4b98c721
fix(mge/sublinear): change no virtual grad warning to log_debug by
2021-02-26 16:50:06 +0800
51868533
fix(mgb/gopt): fix opt pass elementwise operation shape issue at tranform to NCHW4 by
2021-02-26 16:13:14 +0800
0a191f8d
fix(misc): some fixes to make faster_rcnn run in no-symbolic-trace by
2020-12-03 19:44:47 +0800
f95bb7b7
perf(mge/profiler): reduce profiler overhead by
2021-02-04 21:29:13 +0800
9cc732f8
fix(opencl): fix opencl search algo negative stride support by
2021-02-04 15:01:28 +0800
4f3875eb
docs(mge): add missing docstring and fix sphinx build warnings by
2021-02-22 10:57:58 +0800
98d6ba7e
refactor(mge/functional): move matinv from nn to math by
2021-02-18 14:56:14 +0800
17323dbd
feat(dist): collect return values in dist.launcher by
2021-02-26 14:26:21 +0800
5b697c71
chore(imperative): remove unused reference to megbrain_graph by
2021-02-22 19:48:13 +0800
f0170b9c
fix(whl): opt whl version by
2021-02-25 14:26:02 +0800
47db29aa
fix(mge/module): add kwargs param for all modules by
2021-02-07 18:09:24 +0800
6fb19b66
feat(imperative/src): name operators automatically when tracing by
2021-01-15 15:08:26 +0800
09de5a07
feat(mgb/serialization): be able to serialize operator names by
2021-01-28 22:09:14 +0800
bb8f2928
refactor(mge/func): change `conv_execution_strategy` to `execution_strategy` by
2021-01-22 18:00:25 +0800
914af286
feat(cmake/arm/cuda): support cmake arm with cuda, * support arm-host and x86-cross-arm * fix trt myelib cmake build issue at 'trt copy env'(about 'copy env', Please refs cb92123f) by
2021-02-03 13:40:11 +0800
096d340c
fix(whl): fix whl version, fix create manylinux2014 image by
2021-02-23 10:53:25 +0800
f2111248
feat(mge/trace): support dict return value processing in trace by
2021-02-01 10:24:02 +0800
cbff4d7c
fix(imperative/ops): fix infer_output_attrs_fallible for reshape by
2021-01-29 14:57:24 +0800
d1be3127
fix(mge/quantization): fix tqt load and convert issue and observer calculate params issue by
2021-02-02 12:42:57 +0800
30d6b4f6
fix(mge): fix scalar parameter change to 1-dim parameter after backward and optimize by
2021-02-02 12:35:09 +0800
cf27dd64
fix(cuda): use cudnn8.0.4 as cu111 default libs by
2021-02-07 22:15:39 +0800
649e4dd7
test(cuda): fix test for cu111 by
2021-02-05 23:31:00 +0800
c69359d0
fix(dnn/cuda): disable cudnn conv_bias kernels for NCHW4_NCHW tensor format by
2021-02-08 16:42:44 +0800
2e4b9a42
fix(mgb/gopt): fix folding conv dimshuffle opt pass by
2021-02-05 14:23:25 +0800
d62dabe5
fix(mgb): fix matmul model compat in flatbuffer by
2021-02-06 01:09:26 +0800
05692d05
refactor(functional): move matmul dot svd to math pkg by
2021-01-28 20:03:25 +0800
02df634d
fix(mge/quantization): set ``q_dict`` as an instance property by
2021-02-03 18:21:12 +0800
193e77d4
misc(megdnn): misc by
2021-02-05 21:48:02 +0800
b6376c83
fix(mgb/comp_node): use extra physical device to decide whether reuse existed CompNodeImpl by
2021-02-01 15:17:18 +0800
ccbc6761
Revert "fix(whl): append buildid to wheel version" by
2021-02-05 15:13:06 +0800
0e3a6329
build(cuda): support cu111 build by
2021-02-05 10:38:18 +0800