Summary:
1. Move ATen threadpool & open registration mechanism to C10
2. Move the `global_work_queue` to use this open registration mechanism, to allow users to substitute in their own
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17788
Reviewed By: zdevito
Differential Revision: D14379707
Pulled By: jamesr66a
fbshipit-source-id: 949662d0024875abf09907d97db927f160c54d45
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/15041
Adding an alternative implementation of a task graph based on TBB
Reviewed By: dmudiger
Differential Revision: D13412517
fbshipit-source-id: f5efedd680bbe0072bf38d504e5682ab51dd630f
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/12295
Add ability to use custom implementations of thread pool instead of TaskThreadPool
Reviewed By: yinghai
Differential Revision: D10046685
fbshipit-source-id: 95da8c938b8e60b728484c520319b09b0c87ff11
Summary:
This does 6 things:
- add c10/util/Registry.h as the unified registry util
- cleaned up some APIs such as export condition
- fully remove aten/core/registry.h
- fully remove caffe2/core/registry.h
- remove a bogus aten/registry.h
- unifying all macros
- set up registry testing in c10
Also, an important note that we used to mark the templated Registry class as EXPORT - this should not happen, because one should almost never export a template class. This PR fixes that.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/12077
Reviewed By: ezyang
Differential Revision: D10050771
Pulled By: Yangqing
fbshipit-source-id: 417b249b49fed6a67956e7c6b6d22374bcee24cf
Summary:
TSIA. Right now we should basically use C10_EXPORT and C10_IMPORT for explicitly marking dllexport and dllimport, as a continued effort of the C10 unification.
This is a codemod by mechanically doing the following change:
CAFFE2_{EXPORT,IMPORT} -> C10_{EXPORT,IMPORT}
AT_CORE_{EXPORT,IMPORT} -> C10_{EXPORT,IMPORT}
Pull Request resolved: https://github.com/pytorch/pytorch/pull/12019
Reviewed By: ezyang, teng-li
Differential Revision: D10016276
Pulled By: Yangqing
fbshipit-source-id: a420d62c43d1110105fc88f9e9076e28a3203164
Summary:
Properly annotated all apis for cpu front. Checked with cmake using
cmake -DUSE_ATEN=ON -DUSE_CUDA=OFF -DBUILD_ATEN=ON
and resulting libcaffe2.so has about 11k symbols.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/10504
Reviewed By: ezyang
Differential Revision: D9316491
Pulled By: Yangqing
fbshipit-source-id: 215659abf350af7032e9a4b0f28a856babab2454
* add opencl + fpga context
adds an opencl context inside caffe2/fb which can be used for fpga access
* [Caffe2] Force tensor inference checks to be triggered during testing
We've started to rely on TensorInference functions more for different analysis. This diff ensures that the TensorInference function's result matches what is expected from the definition of the operator.
* Enable building //caffe2:torch with @mode/opt
In @mode/opt, python runs out of a PAR, which breaks a lot of
assumptions in the code about where templates/ folders live relative
to __file__. Rather than introduce hacks with parutil, I simply turn
template_path into a parameter for all the relevant functions and
thread it through from the top level.
* [Caffe2] Fix cost models for DotProduct and Div. Update Tensor Inference for dot product
As title. DotProduct states that output is a 1-D tensor (https://caffe2.ai/docs/operators-catalogue.html#dotproduct) though code suggests it is either 0- or 1-D depending on inputs. TensorInference defined to support implementation.
* [SG-MoE] Add an option to make the experts NOT as components
* [nomnigraph] Rename and fixup convertToNeuralNetOperator API
This will make things a bit cleaner
* no longer symlink THNN.h and THCUNN.h
* forced decoder network (onnx export)
Closes https://github.com/pytorch/translate/pull/95
Add networks in ensemble_export.py to create a forced decoding network from PyTorch NMT checkpoints. This network takes an arbitrary numberized (source, target) pair and returns the model score for the translation, including penalties.
Vocabulary reduction networks are also supported, but note that target indices which are not in the possible_translation_tokens generated for the source input will be trea
* Revert schema change to fix production models
Revert schema change to fix production models
* MockLogDeviceReader - rebase on FIX
# Goal
1), Build a make_mock_log_device_reader using make_mock_reader
2), Replace the real log_device_reader here: https://fburl.com/raihwf1p
# Log by D8151734
Real log_device_reader:
```
I0529 20:29:05.373108 954994 tensor.h:839] Tensor print_net/log of type std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> >. Dims: (): read_net/ParseOpenTrainingRow:0
I0529 20:29:05.373244 954994 tensor.h:839] Tensor read_net/ParseOpenTrainin
* [C2/D2][1/n]: Nonnegative-Constrained Optimization -- log barrier
implement log barrier as a regularization method
* Add teacher weight screening.
Add teacher weight sceening according to teacher labels. If teacher label is zero, we do not use the distill loss in the objective function.
* Add NormalizerContext
See task for more detail. This implementation is a copy of what exists for RegularizerContext except for how the parameters are defined in the model_definition thrift file.
I'll try an alternative implementation which overrides the default arguments of functions instead like for argscopes in tensorflow.
https://github.com/pytorch/pytorch/compare/master...MaximeBoucher:update-from-facebook-0939578c068c?expand=1
* Adding cosine similarity option in dot processor
Add pairwise cosine similarity option in dot product.
Add an option to concate dot product and cosine similarity.
Add test cases.
* [nomnigraph][redo] Concat elim for sparseNN
Same as D7962948, which was reverted because Operator Schema was not
defined
* [pytorch] Revert pytorch/pytorch#7918 'Release GIL when copying to shared memory', breaks ASAN
Revert this pytorch diff that breaks ASAN when running Filament in dev mode; in opt mode it gives "bad file descriptor" errors. Looks like a race when copying tensors to shared memory in multiple mp.Queue's (which spawn separate threads).
https://github.com/pytorch/pytorch/pull/7918/files
* [nomnigraph][mobile] Enable nomnigraph by default, use -Oz on nomnigraph related code to reduce code size
enables nomnigraph and reduces codesize
* [Warmup] Allow both offline incremental training and online training
Change plan name on saving side and reading side to support both training type
This diff depends on D8128530 and D8168651.
* Revert D7802642: [Warmup] Allow both offline incremental training and online training
This reverts commit afc213cf9b36cecf75333a788391c4d09f4afccc
@bypass-lint
An infra SEV is better than not reverting this diff.
If you copy this password, see you in SEV Review!
@cause_a_sev_many_files
* Add legacy grad logic to fix div op on old graphs.
Add legacy grad logic to fix div op on old graphs.
* Correctly propagate operator failures
Propagate errors from operators that throw exceptions and return false
* Revert D8374829: [caffe2][nomnigraph][redo] Concat elim for sparseNN
This reverts commit 6dda028c463e54bb5c32188bbbe9202107e188a5
@bypass-lint
An infra SEV is better than not reverting this diff.
If you copy this password, see you in SEV Review!
@cause_a_sev_many_files
* [Caffe2] Added extra_info to core.DeviceOption(), enforced extra_info to be inherited in scope.DeviceScope
extra_info is a newly defined field in DeviceOption proto. This diff added extra_info to the core.DeviceOption(). And, In scope.DeviceScope(), this diff enforce the new scope to inherit the extra_info from old scope.
* [opt] hgdirsync wasn't enabled, merge diverged code
Here's the damage, P59732616 basically xplat was left behind but had
the change from assert to CAFFE_ENFORCE
* OMP parallelism over RoIs for RoIAlign op
Simpler to parallelize over RoIs. Shouldn't affect other uses as it relies on
the number of OMP threads set during startup.
PR: https://github.com/pytorch/pytorch/pull/8562
* Use int64_t for shape in FillOps
to avoid overflow of int32
* Implement Rotated RoIAlign op
Based on Rotated RPNs as explained in https://arxiv.org/abs/1703.01086.
The idea is simple - orientation/angle is added as an RPN
anchor parameter and then the angle is further regressed similar to bbox
coords. There are some additional changes related to NMS and IoU, but besides
that it's a direct extension to Faster-RCNN. Further details in https://fb.quip.com/sZHlA1iMfWPZ.
RoIs are represented in [center_x, center_y, width, height, angle] format.
`angle` repre
* Rotated RoIAlign op CUDA forward implementation
CUDA forward impl for D8415490
* RoIAlignRotated op CUDA backward pass implementation
TSIA
* All remaining fixes to eliminate process_github.sh
Most of this diff has already been reviewed separately, except for the parts relating to _thnn/utils.py and _utils._internal.py
remove skipIf(True, 'Fbcode') line from process_github.sh
replace sed of cpp file with #ifdef to control cudnnDestroy use
undo sync-time deletion of .gitattributes, remove process_github.sh
switch to using _utils._internal rather than try-import-except
This diff also fixes the open-source bug where rebuilds have
* Back out "Revert D7802642: [Warmup] Allow both offline incremental training and online training"
Original commit changeset: 7707d2efe60e The original diff is backout becuase the online trainer package is backed out. This code would only work with new online trainer package
* [easy] improve error log in adagrad op
as title
* re-allow use of thnn_h_path
This fixes cffi usage in OSS
* [4/4] [tum] paralyzing layerNorm for GPU full sync
as title
* add compile=False to pytorch tests, remove hack with pyc
* Add shape and type inference for RowWiseArgMax operator
See title
* Revert D8515341: Back out "Revert D7802642: [Warmup] Allow both offline incremental training and online training"
This reverts commit 78167eeef0af16b60f72c82f9dcdda9b41b4dcbd
@bypass-lint
An infra SEV is better than not reverting this diff.
If you copy this password, see you in SEV Review!
@cause_a_sev_many_files
* [fix-flaky-test] mock_hive_reader_test flaky, because GlobalCounter collects local counts intervally
# Problem
`MockHiveReader` uses `GlobalCounter` to limit `max_examples`.
GlobalCounter on server node collect local counts from worker nodes every 1 sec.
This 1 sec delay makes it impossible to limit exactly to the `max_examples`, it will definitely exceed `max_examples`.
# Plan
Given,
```
Expected num_examples = max_examples + num_examples/sec (Read Speed) x 1 sec (GlobalCounter Sync Int
* [Caffe2] Fix FCGradient cost inference. Prevent overflow in cost inference
FCGradient missed a factor 2 in the `num_outputs == 3` case. Overflow was occurring with flop calculation for FC. Changed types to `uint64_t` to prevent future problems.
* Fix binary ops with empty inputs
Fix binary ops with empty inputs
* Support the filling of input blob with provided data
as title for Biz Integrity case
* Back out "Revert D8515341: Back out "Revert D7802642: [Warmup] Allow both offline incremental training and online training""
Original commit changeset: 30c55dd38816 Original diff is reverted due to introducing bad integration test. Fixed the integration test.
* [c2][easy] improve pack ops error loggings
as desc.
* Add ShapeTypeInference for LpNorm operator
As desc
* Shard test_nn to reduce runtime for each test target
Closes https://github.com/pytorch/pytorch/pull/8793
The current test_nn would time out and be disabled in GreenWarden, and we need to have an option to split it up in order to pass the stress test. Right now GreenWarden roughly allows running 100 test cases in test_nn before timing out, and here we have an option to divide test_nn into 30 shards (with ~40 tests in each shard) to allow for some test suite growth in the future.
* Change default caffe2_streams_per_gpu to 1
* Remove IN_SANDCASTLE from common.py and test_nn.py
We prefer to disable the failing tests through Sandcastle UI instead.
* Add a new class for an updated prof_dag.proto
This diff contains:
- An updated prof_dag.proto that contains blob profiles.
- A class to deserialize this information (serialization is in a follow up diff)
- Update to separate profiling information from NeuralNet (and use it as part of the class above).
- Unit tests
* Lambdarank for SparseNN
This diff adds a lambda_rank_layer for SparseNN.
changes include
1) Adds support for multi sessions in c2 op
2) Adds support for two different loss functions in c2 op
3) Unit tests for op
* Revert D8586950: Back out "Revert D8515341: Back out "Revert D7802642: [Warmup] Allow both offline incremental training and online training""
This reverts commit 012220ed63eccc35659a57b31d16a3625da6317b
@bypass-lint
An infra SEV is better than not reverting this diff.
If you copy this password, see you in SEV Review!
@cause_a_sev_many_files
* [easy] A few fixups to multithread predictor benchmark
(1) support perf on T6 server
(2) remove dead code
* fix a bug about the map size
as title
* Fix reduce sum on in-place case.
Fix reduce sum on in-place case.
* [Warmup] Reland reverted diff Allow both offline incremental training and online training
Closes https://github.com/pytorch/pytorch/pull/8827
fix net transform integration test. Allow offline and online trainer to coexist D7802642.
* Add StoreHandlerNotAvailableException
Add an exception for a store that is not available or has been
deleted.
* Use exception handling for fault tolerance, missing KV store
Remove status blobs to communication ops so that exceptions propagate on
failure.
* [C2/D2][2/n]: Nonnegative-Constrained Optimization -- bounded grad proj
for simple bounded constrained optimization, incl non-negative box constraints.
* [GanH]: Adaptive Weighting with More Estimations
With implemented postivity optimization, we now learn adaptive weights with different
parameterizations.
This improves parameter estimation and training stability.
* Revert some changes for landing
* Remove AutoNoGIL in StorageSharing
* Temporarily disable net_tests
* Revert "[Caffe2] Force tensor inference checks to be triggered during testing"
This reverts commit 67ef05c22b2f71b4a489695384932f968384a2a4.
* Revert "Fix reduce sum on in-place case."
This reverts commit 6cb8a8e1b3db7b6d20941b0053e3f3836068eb64.
* Revert "Revert "Fix reduce sum on in-place case.""
This reverts commit 130a257c0893dc09f4bd6e6a45d112261807fd2c.
* [mpscnn] MPSCNNChannelShuffle
att
* [Easy] Adding tags as an argument to the functional layer
Without it "tags" would be added as an argument to the operator.
The change here is based on the assumption that there is no operator that takes "tags" as an argument.
* Fix locally_connected_op schema check.
Fix locally_connected_op schema check.
* [C2] Add TypeAndShape inference for few more operators
As desc
* [c2] Shape inference should support 0 as dimension
Tensors can have 0 in their dimension.
* Make MockHiveReader loop over and support max_examples
Replace DatasetReader with RandomDatasetReader.
So that Mock Hive Reader can simulate a large data input using a small sample file as source.
* Utility function to wipe cache between benchmark runs
Caffe2 benchmark does not wipe out cache between runs, and this potentially creates an unrealistically optimistic picture of performance. This diff adds utility function to wipe out the cache.
* Allow caffe2 GlobalInit to be invoked multiple times
Allow caffe2 GlobalInit to be invoked multiple times. Will re-parse gflags and update logging levels on successive invocations, but will not re-run init functions or perform other one-time initialization.
* Add Caffe2 GlobalInitIsCalledGuard to base net and operator classes
Warn if caffe2's GlobalInit function has not been invoked before creating an operator or net object. This is based on discussion here: https://fb.quip.com/kqGIAbmK7vNG
* Rethrow current exception on failure
Rethrow current exception instead of copy constructing a new one on op failure.
* Make `clone()` return subclass of List/Struct
`clone()` is not working correctly when we subclass those classes
* Wipe the cache before the net run
the util function is copied from D7409424
will rebase once D7409424 is landed.
* [Caffe2] [Mobile] Support utils/cast.h::GetCastDataType with LITE_PROTO builds
* Correct includes
async_polling include -> async_base include
* Prepare execution flags for executor migration
Making async_scheduling aware of underlying net type to prepare for executor
migration
* Add operator level observers into async executor
Adding operator level observers into RunAsync operators' calls
* Cleanup TEST_Benchmark
Remove duplicate code and provide default implementation in NetBase
* [C2] Fix type and shape inference for binary comparison ops
As desc.
* Add GlobalInit to predictor to ensure initialization is always done before prediction
FACEBOOK:
Redo D7651453 the correct way.
Now use a static variable for the arguments passed to GLog
* Remove spammy log message
This method is currently used in various places inside Caffe itself.
* Disable events for operators inside a chain
We don't need to use events in operators within a chain because the chain is
always scheduled on a single stream, keeping only first and last event for
scheduling purposes
* Ensure correct finish run order
In rare cases we might call finishRun and trigger net's destruction while
another worker is still holding shared_ptr to a thread pool, that can cause
thread pool destruction from within a worker thread in case no other nets are
using the pool. This diff fixes the order of calling finishRun and also changes
pool() to return raw pointer to keep pool's ownership within the net
* Reduce unnecessary polling
Make sure we don't waste CPU by polling operators that we can set an efficient
callbacks on
* Squash commit of syncing 9506eeb from github to fbcode
Patch xplat buck fix
add virtual destructor to OptimizationPass
add virtual destructor to OptimizationPass
build fixes for sync
build fixes for sync
* Fix net tracing
Fix net tracing from async_scheduling
* Fix logging
* [bootcamp] Improve "Shape" operator to support axes specification
To improve .shape operator of Caffe2 to support x.shape(tensor, axes), which takes an optional int array "axes" as input. For example, x.shape(tensor, [1, 0]) will return the dimension for axis 1 and 0 following the specified order. For current version, "axes" input allows duplications and can have arbitrary length.
* Back out "Add barrier net that runs before training nets"
Original commit changeset: b373fdc9c30f. Need additional changes to some callers to support barrier failures.
* Change warning to verbose log to reduce log spam
The `LOG(WARNING)` was a bit spammy for regular use so lets just make it a `VLOG`.
* Extract the shared code from different caffe2_benchmark binaries
The OSS benchmark and Internal benchmark will share most functions in the benchmark.
* Support MFR in sequence training
As titled.
* Make knowledge distillation work with using logged prediction feature as teacher label.
1) Add loading raw dense feature as teacher label.
2) Optional calibration function for teacher label
3) Add teacher label into generic unit test
4) Deprecated TTSN workflow version using feature_options to config teacher label
* [C2/CUDA]: unjoined cross entropy sigmoid
as desc
* Add async_scheduling executor into deferrable_net_exec_test
Add async_scheduling into tests and fix some exception cases
* Fix Event disabled error
When disabling event in RNN ops make sure we don't call Finish on disabled
event from op's RunAsync
* cuda ensure cpu output op can handle both TensorCPU and TensorCUDA
as desc.
* [C2 Core] Infer input device option in C2 hypothesis_test checkers
Improve how we default input blob device options.
Previously it defaults as where op lives but it is not necessarily the case.
For example:
CopyCPUToGPU
* [C2 Op]SplitByLengthsOp CPU/GPU implementation
[C2 Op]SplitByLengthsOp CPU/GPU implementation
* fix undefined symbol error
not sure why we're getting undefined symbol even with link_whole = True
Need to figure out why but need this workaround for now
* Add tools in DAIPlayground platform to help debugging models
Add additional tools to allow Plauground override individual method defined in AnyExp. This will allow user to create module that specificly change certain default method behavior. An example included in this diff is deactivating test model and checkpointing. When debugging any model problems, switching off components helps me quickly narrow down the location of the bug. The technique is extensively used in task T27038712 (Steady memory increase in EDPM, eventually resulting in gloo/cuda.cu:34: out of memory)
* add shape and type inference for int8 conversion operator
* Fix flaky test for group_norm
Fix flaky test for group_norm
* Fix group_norm_op_test flaky
Fix group_norm_op_test flaky
* Implementation of composite learning rate policy
In many state-of-the-arts deep learning works, people use a simple trick to
schedule the learning rate: use a fixed learning rate until error plateaus
and then switch to a different fixed learning rate, and so on. In this diff,
we implemented a simple version of the composite learning rate. The user gives
a set of learning rates policies and corresponding iteration nums, and the
optimizer will change the learning rate policy based on the number of iterations so far.
For example, the user give two learning rate policies, one is FixedLearningRate
and PolyLearningRate, with an iteration number of 1k. Then the first 1k iteration,
we use FixedLearningRate. For the following iterations, we use PolyLearningRate.
* Split two use cases of CachedReader into two classes, DBFileReader and CachedReader
# Use Cases:
1). input: DB file -> output: DatasetReader.
Use DBFileReader.
2). input: Reader -> build cache DB file -> output: DatasetReader.
Use CachedReader.
# Changes to CachedReader:
1). Move db_path to the constructor.
Because in mock reader. cache will always be built ahead.
# Changes to tests:
1). Make a separate TestCase class for CachedReader and DBFileReader.
2). Make it possible to add more test functions by adding setUp, tearDown and _make_temp_path.
3). Make delete db_path more general. `db_path` could be a file for `log_file_db`, but could also be a directory for `leveldb`.
* Back out "On Mobile phones, call GlobalInit with no arguments in predictor in case we need to perform initialization"
Original commit changeset: 4489c6133f11
* Fix LARS bug
Fixed a bug in the LARS implementation which caused all subsequent blobs not using LARS to have the LARS learning rate multiplier applied to them.
* [tum] support sparse init & add uniformFill option
as title
* Propagate exception for async nets
Capture the exception when an exception is thrown in async nets and re-throw it after wait(). This allows exceptions to be propagated up to the caller.
This diff was a part of D7752068. We split the diff so that C2 core files changes are in a separate diff.
* Automatic update of fbcode/onnx to 69894f207dfcd72d1e70497d387201cec327efbc
Previous import was 403ccfbd0161c38f0834413d790bad0874afbf9a
Included changes:
- **[69894f2](https://github.com/onnx/onnx/commit/69894f2)**: Use op schema.all tensor types in random like definitions (#865) <Scott McKay>
- **[b9d6b90](https://github.com/onnx/onnx/commit/b9d6b90)**: Clarify random like operators (#846) <Scott McKay>
- **[fc6b5fb](https://github.com/onnx/onnx/commit/fc6b5fb)**: Refactor shape inference implementation (#855) <anderspapitto>
- **[b7d8dc8](https://github.com/onnx/onnx/commit/b7d8dc8)**: fix cmake warning message (#863) <Eric S. Yu>
- **[f585c5d](https://github.com/onnx/onnx/commit/f585c5d)**: add pytorch-operator test for tile (#831) <Wenhao Hu>
- **[993fe70](https://github.com/onnx/onnx/commit/993fe70)**: add install step (#832) <Eric S. Yu>
- **[68bc26c](https://github.com/onnx/onnx/commit/68bc26c)**: add type inference for traditional ml ops except classifier ops. (#857) <Ke Zhang>
- **[9cc0cda](https://github.com/onnx/onnx/commit/9cc0cda)**: fix string representation of scalar types (#858) <G. Ramalingam>
- **[1078925](https://github.com/onnx/onnx/commit/1078925)**: fix y in pow test case to scalar (#852) <Wenhao Hu>
- **[c66fb6f](https://github.com/onnx/onnx/commit/c66fb6f)**: Add some math function shape inference (#845) <anderspapitto>
- **[ff667d1](https://github.com/onnx/onnx/commit/ff667d1)**: Refactor return type and docs for ONNXIFI_BACKEND_DIRECTX_ID (#853) <Marat Dukhan>
- **[11c6876](https://github.com/onnx/onnx/commit/11c6876)**: clear initializer names when clear initializer (#849) <Wenhao Hu>
- **[73c34ae](https://github.com/onnx/onnx/commit/73c34ae)**: Clarify FeatureVectorizer description. (#843) <Scott McKay>
- **[1befb9b](https://github.com/onnx/onnx/commit/1befb9b)**: Remove useless text in docs (#850) <Lu Fang>
- **[e84788f](https://github.com/onnx/onnx/commit/e84788f)**: Fix SELU attributes' default values (#839) <Lu Fang>
- **[ebac046](https://github.com/onnx/onnx/commit/ebac046)**: Add tile test case (#823) <Wenhao Hu>
- **[8b7a925](https://github.com/onnx/onnx/commit/8b7a925)**: a few more shape inference functions (#772) <anderspapitto>
- **[9718f42](https://github.com/onnx/onnx/commit/9718f42)**: Make the coefficient non optional for LinearClassifier (#836) <Jaliya Ekanayake>
- **[ef083d0](https://github.com/onnx/onnx/commit/ef083d0)**: Add save_tensor and load_tensor functions for Protos (#770) <Lu Fang>
- **[45ceb55](https://github.com/onnx/onnx/commit/45ceb55)**: Check if CMAKE_BUILD_TYPE set before project(). (#812) <Sergii Dymchenko>
- **[4b3d2b0](https://github.com/onnx/onnx/commit/4b3d2b0)**: [WIP] reenable shape inference tests (#834) <anderspapitto>
- **[22d17ee](https://github.com/onnx/onnx/commit/22d17ee)**: RNN tests: LSTM, GRU, SimpleRNN (#739) <Peyman Manikashani>
- **[de65b95](https://github.com/onnx/onnx/commit/de65b95)**: dimension denotation (#443) <Tian Jin>
- **[eccc76e](https://github.com/onnx/onnx/commit/eccc76e)**: fix field number issue in onnx operator proto and enable its build (#829) <Ke Zhang>
- **[d582beb](https://github.com/onnx/onnx/commit/d582beb)**: disable shape inference test to unbreak ci (#830) <Lu Fang>
- **[485b787](https://github.com/onnx/onnx/commit/485b787)**: function proto for composite op. (#802) <Ke Zhang>
- **[cd58928](https://github.com/onnx/onnx/commit/cd58928)**: specify defaults for attributes of Affine op (#820) <G. Ramalingam>
- **[7ee2cf9](https://github.com/onnx/onnx/commit/7ee2cf9)**: merge the dummy backend back into the main one (#743) <anderspapitto>
- **[1c03a5a](https://github.com/onnx/onnx/commit/1c03a5a)**: [Proposal] ONNX Interface for Framework Integration (previously ONNX Backend API) header and docs (#551) <Marat Dukhan>
- **[3769a98](https://github.com/onnx/onnx/commit/3769a98)**: Rename real model test case from VGG-16 to ZFNet (#821) <Lu Fang>
* [C2]ReluN Op
relu n op.
tf reference: https://www.tensorflow.org/api_docs/python/tf/nn/relu6
* Call destructor when assigning a blob value
* Add executor overrides
Add executor overrides flag to enable migration to async_scheduling executor
* Add barrier net that runs before training nets - attempt #2
Add a synchonize barrier net that is run before training nets. With this net, shards that are faster will wait for other shards before start training. This reduce chances of the faster shards timing out during GLOO AllReduce.
Removed explicit data_parallel_model.py.synchronize call in holmes workflow.
This change was landed previously but caused errors for some EDPM workflows - See https://fb.facebook.com/groups/1426530000692545/permalink/1906766366002237/ - because EDPM assumes any call to CreateOrCloneCommonWorld and Gloo ops are wrapped in exception handlers but in this case exception thrown in the barrier init net is not handled.
To address this issue, we add _CreateOrCloneCommonWorld to the param_init_net instead of a new barrier init net. Since errors for param_init_net run is handled gracefully and re-rendezvous, it should fixes the problem.
* Handle empty nets in async_scheduling
Make sure we don't get stuck on empty nets
* use CUDA_ARCH for conditional compile
* [C2 fix] infer function for ensure_cpu_output_op
* Update group_norm test to reduce flaky test
* Fix lr_multiplier for GPU
* Track checkpoint performance in scuba
As title.
* [C2/CUDA]: fix cross entropy sigmoid with logits
when adding log_d_trick, I forgot to add it to the cuda impl; this diff fixes
it.
* Back out "[caffe2] Unregister MKL fallbacks for NCHW conversions"
Original commit changeset: 8918dd40205a
Will land after @jongsoo's diff https://phabricator.intern.facebook.com/D7596315 lands
* [Easy][C2] Don't add blob to external outputs from output_record if it's already external output
As desc.
* On Mobile phones, call GlobalInit with no arguments in predictor in case we need to perform initialization
FACEBOOK:
The QPL logger needs the initialization code. In the past, the initialization code is put in the pipeline calling Caffe2. However, those places become obsolete quickly, as the product teams change places to call Caffe2 from time to time. We also need to track which teams use Caffe2 so that we can put the initialization code there.
With this diff, the initialization code is put in the predictor constructor, only enabled for mobile phones. This way, we can always enable QPL logging.
Once we do this, we can check how many times Caffe2 inference is called in production, and which models are more popular in production. This way, we can prioritize our effort supporting those models.
Will clean up the old code calling the init in the product in a separate diff.
* add padding op for sparse length tensor
to pad length-based sparse tensor with padding_value
* Add conv_op with cudaconvnet engine
Add conv_op with cudaconvnet engine
* [numa] Fix simple NUMA copy benchmark
Move XavierFill into init_net and also compute BW
* call roundf (device function) instead of round (host function)
* [caffe2_benchmark][observer] Make caffe2_benchmark use its own observer
1. Add ClearGlobalNetObservers()
2. Make caffe2_benchmark use its own observer and observer_reporter
* [detectron] Use roundf instead of round in the detectron module ops
* allow K larger than number of elements in top k op
one use case is to use this op together with PackSegments for sparse tensors, where the number of elements in each slice is not statistically defined.
* add ChannelShuffle DNNLOWP op
* fixup math_cpu.cc break
Summary:
There is a lot of bussiness logic around various events in
the base net class. SimpleNet doesn't have to handle those (checked
with ilia-cher). Normally these should be no events registered for
simple nets, but we can have some issues where they will be added, so
its less error prone to just have a SimpleNet::Run pure. And then we
also avoid extra virtual calls / empty vector iterations.
Reviewed By: ilia-cher
Differential Revision: D6551440
fbshipit-source-id: c97a732a00bb36eed49d35e727156ce94225a08b
Summary:
Async executor based on async_polling (D5985110):
- Tasks scheduling other tasks, using polling only when necessary (e.g.
CUDA->CPU case)
- Fully async, i.e. RunAsync immediately returns
Reviewed By: azzolini
Differential Revision: D6281681
fbshipit-source-id: 06e3723e1424ffab652c38ca7b279cf76e43fa44
Summary:
Implementation of polling async net executor.
Notes:
- New net executor async_polling - schedules CPU and GPU ops asynchronously, uses single polling thread
- Events: update to Caffe2 events to support async CPU events, adding new methods:
Query() - non-blocking checking of event states: INITIALIZED -> RECORDED -> SUCCESS/FAILED
ErrorMessage() - when operation runs asynchronously and fails calling this on event will give error message
- Tasks: using existing DAGNet's algorithm to compute CPU and GPU chains, a separate task for each chain
- Polling: using single thread to query state of events - for CPU tasks atomically queries task state, for GPU task - uses cudaEventQuery; using Event
- Scheduling of CPU ops: using global thread pools
- Scheduling of GPU ops: using GPU thread pool per GPU device
Reviewed By: dzhulgakov
Differential Revision: D5985110
fbshipit-source-id: a9de7fcbb71d046a3aa1b573072b89a65dfeee8c
Summary: observer framework can now be used in python + a small writeup of how to use it. this is D6035393 with a fix for ct-scan
Reviewed By: salexspb
Differential Revision: D6066380
fbshipit-source-id: 896c4c580d4387240b81ac2dbbc43db51d4bfeb9
Summary: observer framework can now be used in python + a small writeup of how to use it
Reviewed By: sf-wind
Differential Revision: D6035393
fbshipit-source-id: 4563cf0203095fa979bb2160621cd16dd22ff830
Summary: observer framework can now be used in python + a small writeup of how to use it
Reviewed By: salexspb
Differential Revision: D5905002
fbshipit-source-id: e40ec24a55e08fb73beea9b4f3b68e71fc66ffb1
Summary:
Right now, each net implements 2 functions: Run() and RunAsync(). The (loose) abstraction is:
* Run(): run the network in a synchronous way. The call is synchronous.
* RunAsync(): run the network *still synchronously*, but potentially use asynchronous scheduling of the underlying operators.
As one can see, this is highly confusing: RunAsync() is actually a sync call, and the semantics it tries to implement should actually be done by a different net type. For example, DAGNet and AsyncDAGNet both implement the Run() function, and under the hood one uses sync scheduling and one uses async scheduling. Currently, the only user of the RunAsync() function is in SimpleNet::RunAsync(). The only call site is in recurrent_net_op.
Instead, the operator implements the two Run() and RunAsync() functions as follows:
* Run(): run the operator in a synchronous way. aka doing FinishDeviceComputation().
* RunAsync(): run the operator in an asynchronous way if possible (i.e. still sync in CPU, but async in cuda), records the action in the event_, and return immediately.
Semantically, Run() is equal to RunAsync() followed by event().Finish().
As a result, we propose in diff D5812854 to change the network interface similar to the operator interface, and explicitly raise RunAsync() as a first class citizen of the net interface. Specifically, whether a net can run asynchronously is now determined by the
* Adding a SupportsAsync() function that determines if a net supports async execution or not.
* Run(): run the net in a synchronous way.
* RunAsync(): if SupportsAsync() is false, same as Run(). if SupportsAsync() is true, run the operator in an asynchronous way, with the scheduling algorithm determined by the implementation itself. Then, record all outstanding events in the events_ field, and return immediately.
Semantically, Run() is equal to RunAsync, and call event.Finish() for all the events. This is actually the implementation and Run() is no longer a virtual function, RunAsync() is: all sub classes of NetBase shall implement SupportsAsync() and RunAsync() now.
**Why SupportsAsync()?**
This is a design idea that probably needs iterating. Basically, the idea is that RunAsync() is the main entry for the net execution, and it's actually like RunAsyncIfTheNetSupportsIt().
In theory, Run() is basically a wrapper on top of RunAsync() to reduce code duplication: if a net type does not support RunAsync(), its RunAsync() implementation simply is sync (see e.g. SimpleNet) and the Run() to RunAsync() lowering is a no-op (with the only overhead being a nested function call).
I exposed the SupportsAsync() function just in case some caller wants to explicitly check whether an instantiated net supports async call or not - for example, a caller may want to make sure that it is actually running a net asynchronously, in which case SupportsAsync() is the place to query.
Reviewed By: dzhulgakov
Differential Revision: D5812854
fbshipit-source-id: 916b38fded0eb14439f340ab254a034ac5a9a465
Summary: SimpleNet and DAGNetBase are the only two direct subclasses of NetBase. This feature has already been applied to SimpleNet before, with this diff all nets should be covered.
Reviewed By: dzhulgakov
Differential Revision: D5475498
fbshipit-source-id: 339edac31d008ec1e4630d93d2e27d0f518f4ebb
Summary: This uses `clang-tidy` to comment out unused parameters (in functions, methods and lambdas) in fbcode. Cases that the tool failed to handle are fixed manually.
Reviewed By: igorsugak
Differential Revision: D5454343
fbshipit-source-id: 5dee339b4334e25e963891b519a5aa81fbf627b2
Summary: Constructor should extract everything needed from NetDef instead of keeping it for usage after construction.
Reviewed By: akyrola
Differential Revision: D5469095
fbshipit-source-id: 288ea3243d85061ba9c018d2aef3b4d97485dd00
Summary: As Part 1 in reducing the size of operator objects, this removes the outside access to def() and moves debug-uses under a new debug_def() function. Next phase will be by jbai to remove all access from subclasses to def().
Reviewed By: Yangqing
Differential Revision: D5393893
fbshipit-source-id: 7301cff4138dce620b49f6c4db315df85fee7266
Summary: Fixed a lot of issues that salexspb brought up, and templates on NetBase which basically adds compatibility for DAGNetBase. This will be useful for Fei's future work.
Reviewed By: salexspb
Differential Revision: D5272352
fbshipit-source-id: b5ffe1d6fb0566dc1bfad9041c129a3ab7f6d93a
Summary:
This brings back DAGNet up to parity with SimpleNet, where
execution stops as expected after an operator fails. For the DAGNet
it's more involved, since we have to deal with all worker threads
stopping execution. Because the job queue may still hold an arbitrary
number of chains to execute, this diff explicitly closes it down,
waits for all workers to terminate, and resets the job queue, upon
seeing a failure.
Reviewed By: akyrola
Differential Revision: D5232955
fbshipit-source-id: 4dac3c3ed6e5c2ebd07473b0f8be2b02c28978e9
Summary: Contains the ObserverBase class and some unittests.
Reviewed By: bwasti, pietern
Differential Revision: D5099367
fbshipit-source-id: fabde126d3281729dfc772d63dbf363e5d649319
Summary:
This is from discussion with dzhulgakov : as a step towards revisiting the
core.Net autonaming, we will first guard against accidental overwrites of
existing networks in the workspace.
ajtulloch since we are doing Predictors in mobile, this should be safe right?
azzolini - I assume this would be safe, but would love to get your approval.
akyrola - would this hurt xray?
Reviewed By: dzhulgakov
Differential Revision: D4897725
fbshipit-source-id: aa41271927ad6671f07a53b9505283623f8c49e5
Summary:
RecurrentNetOp created workspaces at every run, which is very wasteful, as it had to also recreate the stepnets (forward and backward!).
.
Reviewed By: salexspb
Differential Revision: D4704547
fbshipit-source-id: 460028d912d6a735448c445cb83c0c4d03286351
Summary:
In the past we have moved most of the CHECKs to CAFFE_ENFORCE (exceptions).
However, we kept the name "*_CHECK" for cuda calls, and that caused some
confusion especially in the destructor calls: while our destructors are not
written to handle exceptions, these CUDA_CHECKs could actually throw some
exceptions.
As a result, this diff
(1) Renames all cuda related "*_CHECK" to "*_ENFORCE"
(2) Explicitly marked the destructor of core Caffe2 classes as noexcept
(3) Added proper, really-CHECK cuda check macros, and used those in the
corresponding destructors.
This should not change any of existing functionality.
Reviewed By: dzhulgakov
Differential Revision: D4656368
fbshipit-source-id: 32e3056e66c0400156c5ca0187b6151cf3d52404
Summary:
This diff introduces a new net type 'singlethread_async' which is based on my investigation of DPER/hogwild MLP bottlenecks.
It only uses one CPU thread, but multiple GPUs on each GPU. This is implemented by having each Net to submit their list of operators to
a central GPU-specific executor queue and a thread that executes them asynchronously. This executor takes all tasks in the queue and executes them on separate cuda streams and then waits them in the end. This solution can achieve >95% GPU utilization on 8 GPUs when sufficient amount of workers is used.
FYI: I also tried fancier solution such as using cudaStreamCallbacks(), but they did not have as good performance.
Improved the dper bench by adding the MomentumSGDUpdate operations and adding speed test capabilities. During my testing I also noticed that the startup costs for inizialing CUDA streams and contexts are high, so it is important to do a warm up.
Reviewed By: Yangqing
Differential Revision: D4553941
fbshipit-source-id: bb00524bef653d75de026dd64097b8d9b7a0acb3
Summary:
Matt uyt reported (1) a very infrequent assertion failure at net.cc worker function. This was caused because an operator, that was not a chain, was scheduled in the job queue. This was possibly to happen since our DAGnet operator graph is graph of operators, and not chains. The dependency prunign that I introduced last week exposed this problem since it removed some "middle-to-chain" dependencies when computing the chains. (It is bit hard to explain).
This diff attempts to fix the problem by only allowing scheduling of chains. I added, in addition, extra check to confirm that all parents of all nodes were indeed executed before starting next roud. This adds additional safety and breakpoint to see if there are still problems.
I also fixed a bug in the operator graph pruning that made pruning less effective.
(1) Matt's report:
https://www.prod.facebook.com/groups/1405155842844877/permalink/1639428779417581/
Reviewed By: dzhulgakov
Differential Revision: D4531424
fbshipit-source-id: 80fa7def6e8aff6910ebf0d9d5fef15ff20e0aec
Summary:
We have noticed that the number of chains computed is usually much larger than necessary, when there is a backward pass. For example having a network of 5 FCs with gradient operators (but no parameter updates) should yield only one chain, but instead over 20 were created. After adding parameter updates, the forward pass still should remain one chain, while the backward pass will be splintered.
Analysis showed that the problem was the dependices from forward ops to the gradient computation. But these are redundant since the gradient op is already dependent from the op via the full path over ops. Example:
fc1 -> fc2 ---> fc3 --> loss
| | | |
fc1grad <- fc2grad <- fc3grad <-
Here fc1 and fc1 grad have a direct dependency, but indirect dependency via fc2->fc3->[...]->fc1grad already covers that dependency.
To fix this, I added a pruning step prior to the chain computation. The chain computation is done on the pruned tree, but I do not modify the runtime chains for safety.
Pruning is based on following logic:
- if one of my direct parents is ancestor via an another traversar, I can remove the direct dependency
Pruning is extremely fast, linear in the number of dependencies.
Reviewed By: dzhulgakov
Differential Revision: D4500293
fbshipit-source-id: 0994ae6775c53378ea1e0074365cef041764a1b4
Summary:
This is a bit large diff, sorry about it. It includes basic shape and type inference functionality, based on YQ's Schema scaffolding. I added some helper functions to make it easier to write simple translations.
Bigger refactoring was needed for ConvPoolBase so that we could use the shape inference already there in the schema.
I annotated enough operators to be able to infer forward-pass of shapes for basic convnet, and added test for that. I intend to bootcamp some annotations and annotate enough to handle Resnets fully. Need to think about gradients, if they could be annotated in an easier way.
Only shapes are now exposed to Python, types will follow later. Also the inference is not called yet anywhere but unit test.
Also I am not sure if everything is in the best location in the code, but shouldn't be hard to move stuff around.
Reviewed By: dzhulgakov
Differential Revision: D4436818
fbshipit-source-id: eebee5937ccc9ac09c245465302388a1fae6933c
