* [fix] Re-enable events in RNN ops
We have earlier added event disabling in RNN ops as back then we didn't use
events, with current use cases this is no longer true
(https://fburl.com/8vd0lp8y)
* use ops with cude impl
* Revert D7729695: [caffe2][fix] Re-enable events in RNN ops
This reverts commit 4b215c7496fb724656ff4c776933a15bdbbcde5e
@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
* [observer] Clean up observer_config.h
#accept2ship
* [1/n] Refactor dataio_test.py
Replace code duplication with a common function
* Add barrier net that runs before training nets
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. Similar change in speech/asr_training workflow will come in another diff.
* Support the dnnlowp backend in caffe2_benchmark
This is for SHARE operator latency evaluation
* Migrate integral_image_op to main caffe2
migrate integral_image_op(GPU version) given by https://fburl.com/yvqezigi
to caffe2/caffe2/operators and implement its CPU version. Write up a test
using the hypothesis_test mechanism
* [pos_disc, fbcode] Implement unjoined lr loss
As explained in https://our.intern.facebook.com/intern/wiki/Model_Based_Calibration/, when the dataset is an joined data set, where labels might change later, we need to use unjoined logloss.
The implementation is almost the same as in Sigrid (https://fburl.com/1trngsls), where
loss = y (log(p) - log(1-p)) + (1-y)(log(1-p)) = xy - (1-y)x - (1-y)log(1+exp(-x))
For x < 0, to ensure stability and avoid overflow, we reformulate the above exp as
loss = xy - (1-y)x - (1-y)x + (1-y)log(1+exp(x)) = xy + (1-y)log(1+exp(x))
Then the final expression becomes
loss = xy + (y - 1) x (x >= 0) - (1 - y) log(1 + exp(x - 2 x (x >= 0)))
where y is the true label, x is the dot product and p = logistic(x).
This kind of implementation is align with the current implementation of the original cross entropy in
https://phabricator.intern.facebook.com/diffusion/FBS/browse/master/fbcode/caffe2/caffe2/operators/cross_entropy_op.cc;0bae3b5d0f825897c5e0dd0ff10f489d7271bf25$7-13
* Keep the array to fix the conflict
* [C2] Compute Adagrad effective LR
The AdagradWithLR op outputs an extra blob which is contains the average effective learning rate across all weights in this blob.
* Open-source extractMetaNetDef & runGlobalInitialization, add new Predictor constructor from db file, and add run_map_outputs
1. Open-source extractMetaNetDef and runGlobalInitialization, for use in
2. new Predictor constructor from db file.
3. Add new run function that returns outputs as TensorMap
* Disable eigen cpu
Disable eigen cpu in transpose and reduce
* Introduce request_only/object_only property of ModelLayer
by default this is False
* A simple TC Caffe2 benchmark
We can run tunner, get MappingOptions and then use them to
compare against cuBLAS
currently broken due to LLVM issues. How to run:
hg checkout eec1ab31b59c03b8deded1c755a9abaf8c45be01
add D7401202
add D7434625
add D7506031
add D7540728
buck run @mode/dev-nosan tc/tc/benchmarks_python:caffe2_benchmark
* Move Caffe2 feature_maps_ops to open source
Need feature maps operators in open source project facebookresearch/BlueWhale
* Manually fix the conflicts in channel shuffle op
* Fix the inconsistency between different gh and fbcode
* Skip Adagrad GPU Test (Because some gpu implementation is missing)
* Fix another test to make sure it won't run on gpu when implementation is not available yet
* [GanH][Easy]: Add assertion to adaptive weighting layer
0 weight causes numeric instability and exploding ne
* [Easy] Add cast op before computing norm in diagnose options
As LpNorm only takes floats we add a manual casting here.
* Introduce a new caching device allocator
`cudaMalloc` and `cudaFree` calls are slow, and become slower the
more GPUs there are. Essentially, they grab a host-wide (not device-wide) lock
because GPU memory is transparently shared across all GPUs. Normally, this
isn't much of a concern since workloads allocate memory upfront, and reuse it
during later computation.
However, under some computation models (specifically, memory conserving
approaches like checkpoint-and-recompute, see
https://medium.com/@yaroslavvb/fitting-larger-networks-into-memory-583e3c758ff9)
this assumption is no longer true. In these situations, `cudaMalloc` and
`cudaFree` are common and frequent. Furthermore, in data parallel contexts,
these calls happen at nearly the same time from all GPUs worsening lock
contention.
A common solution to this problem is to add a custom allocator. In fact,
nVIDIA provides one out of the box: CUB, which Caffe2 already supports.
Unfortunately, the CUB allocator suffers from very high fragmentation. This is
primarily because it is a "buddy" allocator which neither splits nor merges
free cached blocks. Study
https://github.com/NVlabs/cub/blob/1.8.0/cub/util_allocator.cuh#L357 if you
want to convince yourself.
This diff adapts a caching allocator from the Torch codebase
https://github.com/torch/cutorch/blob/master/lib/THC/THCCachingAllocator.cpp
which does splitting and merging and ends up working really well, at least for
workloads like the checkpoint-and-recompute computation models noted above.
I simplified the implementation a little bit, made it a bit more C++-like. I
also removed a bunch of stream synchronization primitives for this diff. I
plan to add them back in subsequent diffs.
* Report reader progress in fblearner workflows
Integrate with fblearner progress reporting API and add support to report training progress from reader nodes.
If reader is constructed with batch limits, report based on finished batch vs total batch. The finished batch may be more than total batch because we evaludate if we should stop processing everytime we dequeue a split.
If no limit for the reader, report based on finished splits (Hive files) vs total splits. This is fairly accurate.
* [GanH][Diagnose]: fix plotting
1. ganh diagnose needs to set plot options
2. modifier's blob name is used for metric field can need to be fixed before
generating net
* Automatic update of fbcode/onnx to 985af3f5a0f7e7d29bc0ee6b13047e7ead9c90c8
* Make CompositeReader stops as soon as one reader finishes
Previously, CompositeReader calls all readers before stopping. It results in flaky test since the last batch may be read by different threads; resulting in dropped data.
* [dper] make sure loss is not nan
as desc.
* [rosetta2] [mobile-vision] Option to export NHWC order for RoIWarp/RoIAlign
Thanks for finding this @stzpz and @wangyanghan. Looks like NHWC is more
optimized. For OCR though it doesn't yet help since NHWC uses more mem b/w but
will soon become important.
* Intra-op parallel FC operator
Intra-op parallel FC operator
* [C2 Proto] extra info in device option
passing extra information in device option
design doc: https://fb.quip.com/yAiuAXkRXZGx
* Unregister MKL fallbacks for NCHW conversions
* Tracing for more executors
Modified Tracer to work with other executors and add more tracing
* Remove ShiftActivationDevices()
* Check for blob entry iff it is present
When processing the placeholders ops, ignore if the blob is not present in the blob_to_device.
* Internalize use of eigen tensor
Move use of eigen tensor out of the header file so we don't get template partial specialization errors when building other libraries.
* feature importance for transformed features.
* - Fix unused parameter warnings
The changes in this diff comments out unused parameters.
This will allow us to enable -Wunused-parameter as error.
#accept2ship
* add opencv dependencies to caffe2
The video input op requires additional opencv packages. This is to add them to
cmake so that it can build
* Add clip_by_value option in gradient clipping
Add clip_by_value option in gradient clipping
when the value is bigger than max or smaller than min, do the clip
* std::round compat
* fix unit test for sqrt op
From the error logging:
[idx, grad, grad_estimate] are:
[[ 146. 0.5 0.45776367]
[ 147. 0.5 0.45776367]
The gradient == 0.5 is correct, which means the SqrtOp and its gradient is doing right job. (Because y = sqrt(x), loss = y^2/2 = x/2, and then d(loss)/dx = 1/2 = 0.5; )
The test failed because of numerical problem of grad_estimate (in unit test). It can be because the step_size is small, and float precision is not high (when there are multiple elements in the tensor, we do sum(y^2) to compute loss)
This diff
- increase the step size, and also move the test cases to be further away from 0 (where sqrt(x) is not well defined) to be safe :)
- also clean up, and merge the test case for inplace Vs. non-inplace
Tested with:
`CAFFE2_HYPOTHESIS_PROFILE=debug ai_bt caffe2/caffe2/python/operator_test:elementwise_ops_test -- "test_sqrt"`
* CompositeReader & CompositeReaderBuilder
A new type of reader gluing multiple readers together.
* Back out "Revert D7394363: [GanH]: Log D Trick for Cross Entropy with Sigmoid"
Original commit changeset: 9325a4356dbe
* [dai][WIP] convert params to int8 on ps before sending to trainer
Add float->uint8 conversion in addition to float->fp16 conversion in model_saver.
* [easy] improve unit test for sparse length sum ops
as desc.
#accept2ship
* Update GitHub upstream to 771fcb3455
* move sparse hash unique ops to OOS and add unit tests
- move the SparseHash version to OOS, since 'sparsehash' is already deps of caffe2 OOS: https://fburl.com/arssw4n1
- The 'SparseHash' engine is also being used in OOS, so the SparseHash version shall be in OOS to reduce confusion: https://fburl.com/o5ea7ah2
- fix the CUDA UniqueOp for the case when batch is empty.
- add unit test
* group_norm_op for caffe2
This is the cuda op for Group Normalization (GN): https://arxiv.org/abs/1803.08494
This code implements GN in one op that computes Y=gamma * (X-mu) / sigma + beta and also its gradients. It is expected to have minimal memory consumption (similar to the BN op), without creating new blobs if GN were implemented as several ops (e.g., reshape, norm_mean/std, affine_channel).
* Resubmit D7405233: disappeared in D7464958
OOS publish causes the op missing -- however, test was still there
* [c2] add sparse hash engine for cuda unique op
The SparseHash version of UniqueOp copy input tensor to CPU, and make use of sparse hash map to get unique output, and then copy back to GPU.
* [dper][gpu] enable unit testing gpu trainer for sparse nn
to debug the GPU trainer using mock data in unit test.
make it easier to develop GPU trainer for new models.
* Reuse Gloo context for Synchronize() calls
Previously we were creating (and leaking) the Gloo context on each call to Synchronize(). Now only run the common world op and create the barrier net once, then run the barrier net on each Synchronize() call. Since timeout is associated with the Gloo context, assert that the timeout is fixed instead of trying to handle the complexity of multiple timeouts (and associated contexts).
* [GanH/WGAN][1/n]: add FC param clipping
as titled
* [mobile] minimizing changes between caffe2_benchmark and speed_benchmark
* [GanH]: enable diagnose within model
avoid finding blob names but to directly enable inside the model
* Add `net_transformer_fun` option to DPM
This callback allows for various transformations to be made to the
model after gradient operators have been added. The immediate motivation for
this is to allow transformations such has "checkpoint-and-recompute" which
allow trading off memory for additional compute.
Adding several callbacks like this has made DPM's API less than ideal at this
stage. However, I could not find any reasonable alternative.
* [DT] [33/n] Compile flow task groups
task groups need to compiled in order to pickle the object in fblearner. However I also changed the Job's compile function as creating new object is not necessary.
* Initial commit for sparse_normalize vectorization and benchmark
* [GanH]: LB Calibration for JSD
as titled
* Tracing event in async executor
Adding event tracing through TRACE_EVENT macro in async executor
* [Resubmit] D7409751 Reseting book-keeping blobs when the reservoir is reset
D7409751 got lost in D7464958
* Visualizing realtime weights values
we want to visualize the weights values as optimizer is iterating. This diff supports to visual the weights at an assigned index.
Currently, we assume the blob to be 2 dimensional.
* [GanH][Easy]: Fix Homotopy Weighting
apparantely, there was a bug in homotopy weight (alpha, beta) update
* [c2] move sparse hash unique op out of oss
so that oss do not need to depend on google hash map.
* Get rid of std::round as it's not supported on Android
* Revert changes on setup.py
* Skip shaky test on Dataio
* fix
Summary: ReaderWithTimeLimit() class to stop after a certain amount of time
Reviewed By: boryiingsu
Differential Revision: D6477623
fbshipit-source-id: 165874c9344b0c9c7e0b33e12e72e24c46669cb2
Summary:
Advantages of cloning the tasks/execution_steps at runtime:
- Less complexity on the python side: no need to clone nets and add prefixes to blob names
- Faster start-up: we had cases of complex plans that took up to 30min to be created.
- Better isolation: each task cloned at runtime has its own child workspace, preventing false sharing of blobs.
- Opens up possibility for dynamic scheduling: Number of threads per task can be increased on the fly, at runtime.
Reviewed By: dzhulgakov
Differential Revision: D5100730
fbshipit-source-id: 71b83193b135da4e6eaf2536d8fc266528e1fdcc
Summary: This diff fixes an issue with running the same reader in the same workspace multiple times. In order to achieve correct behavior of execution step we have to explicitly initialize should_stop_blob with False.
Reviewed By: kennyhorror
Differential Revision: D5224410
fbshipit-source-id: 4ad2740e187b62b0a1f5612ea3eef223dcc8a799
Summary:
Before we didn't propagate the 'out-of-data' signal if splits_per_epoch wasn't specified.
Right now it's a hacky fix (just reuse ReaderWithLimit). azzolini - any suggestions of more elegant solution? I can create an extra reader that just export "is empty" signal out.
Overall, I guess we need to turn global_queue into a more sustainable unittest that verifies all possible combinations - I'm still not sure it's correct :-\
Reviewed By: xianjiec
Differential Revision: D4665677
fbshipit-source-id: fe44d10ee82c3383145635e67dea1d9b666e061f
Summary:
Remove the use of `NextName` in layer model helper, so that the same function return `model_helper` that should construct identical `Net`, when under the same NameScope.
The `NextScopedBlob` should only take effect when there is real name conflicting, otherwise it returns ScopedBlobReference.
This is critical for parameter blobs. In long run, we need to be able to specify parameter blobs more explicitly. (kennyhorror is working on this). This solution works in short term for e.g., two tower sparse nn models.
Reviewed By: kennyhorror
Differential Revision: D4555423
fbshipit-source-id: 2c4b99a61392e5d51aa878f7346466a8f14be187
