Functional bias_add #857

szagoruyko · 2017-02-26T13:27:59Z

This is occasionally very useful to avoid slow expands. Added doc for F.linear too

Depends on #851

soumith · 2017-02-26T13:54:59Z

why have a separate function called bias_add? it's just an expand over the mini-batch + add
I'm not sure I see the value, or want to add this into core.

szagoruyko · 2017-02-26T13:57:47Z

@soumith expands are very slow on GPU, that's why we use sger in F.linear and not expand, for example.

fmassa · 2017-02-26T14:44:03Z

While I see that this is more efficient and expand + add, I think that we will (soon?) be adding automatic broadcast to torch tensors, so this might look weird. I wonder if there is a way to optimize the pointwise operations in THC for the stride 0 case... that might be the cleanest solution.

szagoruyko · 2017-02-26T14:54:15Z

@fmassa naive expand-broadcasting would prob be very slow, I guess it could be made to recognize CUBLAS ops and use them?

fmassa · 2017-02-26T15:34:54Z

@szagoruyko a very dirty hack for this specific case would be to add the code in this PR in here, by asserting the dimensions of both tensors and that the stride of the 2nd tensor is 0, and calling the underlying cuBLAS routine.
But I wonder if we can do something in the pointwise kernels themselves...

apaszke · 2017-02-26T18:11:23Z

@fmassa I think it makes sense, because it won't be covered by the automatic broadcasting, and will allow for doing that in an efficient way. Broadcast can only unsqueeze additional dimensions at the front of the tensor, while in this case you actually want to unsqueeze one at the front and N-2 at the back.

ngimel · 2017-02-26T20:22:27Z

I must be missing something - why are expands slow on the GPU? Those are CPU operations setting the necessary dimensions and strides, without operations on the underlying GPU data. I was experimenting recently with replacing creating add_buffer and addr_ with expand and add for Linear, and while cublas ger_kernel is usually faster than torch's pointwiseApply2, overall time is on par or slightly less (because in the first case, input.new(1).resize().fill_ translates into cudaMemset, a bunch of cudaGetDevice/cudaSetDevice (why so many?), launching pointwiseApply1 kernel to fill, and only then launching cublas ger_kernel). There is no reason why torch's pointwiseApply kernels should be much slower than cublas - they can be improved by adding unrolls and handling 0 strides more efficiently, except bias_add is rarely a bottleneck, so no one is doing it.
cudnn has addTensor and opTensor that handle addition with broadcasting, but they are a pain to use and have strange performance holes for some sizes.

apaszke · 2017-03-24T19:36:07Z

I'm not sure if it's the right way to use biases on the GPU. I just ran a quick benchmark and it seems that addr_ is 20x slower than expand_as + add_. Here's the script (requires Python>=3.3):

import torch
import time                             
                                    
x = torch.randn(4000).cuda()        
y = torch.zeros(128, 4000).cuda()                            
                    
s = time.perf_counter()             
for i in range(1000):               
    add_buffer = x.new(128).fill_(1)
    y.addr_(add_buffer, x)          
torch.cuda.synchronize()            
print(time.perf_counter() - s)      
                                    
s = time.perf_counter()             
for i in range(1000):               
    y.add_(x.expand_as(y))          
torch.cuda.synchronize()            
print(time.perf_counter() - s)

And the results are:

0.23105818685144186
0.01973715308122337

ngimel · 2017-03-24T19:40:24Z

Depends on the sizes. Also, once extra overhead of a dozen getDevice/setDevice for every pointwise kernel is taken care of, addr_ can be better. But overall you are right, for medium to small sizes, expand_as is faster than addr_ (expand_as + add is a single pointwise apply kernel, addr is create tensor (for some reason there is memcopy there), fill (overhead of starting a pointwise kernel), call cublas). Even though cublas ger kernel is usually faster than torch's pointwiseApply, when all overheads are taken into account expand_as + add wins.

ngimel · 2017-03-24T19:47:17Z

For backwards though expand_as becomes a sum, and there were some issues with the sum speed on CPU? (GPU is Ok there, though once again torch's reduce kernel is slower than cublas gemv, but overheads are less, for a net win).

apaszke · 2017-03-24T19:48:04Z

Actually I just realized that I forgot to synchronize after creating the input, so the addr time was skewed by the CPU copy. This is a new script, and it seems that addr is consistently faster 1.4x faster than expand + add 😕

import torch                                                             
import time                                                              
                                                                         
batch_size = 32                                                          
features = 128                                                           
                                                                         
for batch_size in (2 ** i for i in range(5, 11)):                        
    for features in (2 ** i for i in range(5, 11)):                      
        x = torch.randn(features).cuda()                                 
        y = torch.zeros(batch_size, features).cuda()                     
        torch.cuda.synchronize()                                         
                                                                         
        s = time.perf_counter()                                          
        for i in range(1000):                                            
            add_buffer = x.new(batch_size).fill_(1)                      
            y.addr_(add_buffer, x)                                       
        torch.cuda.synchronize()                                         
        addr_t = time.perf_counter() - s                                 
                                                                         
        s = time.perf_counter()                                          
        for i in range(1000):                                            
            y.add_(x.expand_as(y))                                       
        torch.cuda.synchronize()                                         
        expand_t = time.perf_counter() - s                               
        print('{}\t{}\t{:.4f}\t{:.4f}\t{:.2f}'.format(                   
              batch_size, features, addr_t, expand_t, addr_t / expand_t))

ngimel · 2017-03-24T19:55:49Z

Ha! In linear add bias is implemented as

add_buffer=input.new(1).resize_(input.size(0)).fill_(1)

https://github.com/pytorch/pytorch/blob/master/torch/nn/_functions/linear.py#L13
and when I modify your script to do this, addr is 20% slower than expand_as.
I guess linear should be fixed.

soumith · 2017-03-28T22:16:35Z

as discussed offline, this PR is kinda redundant. The speed benefits are negative (unlike what we originally thought).
So closing this out.

… ops except classifier ops. (#857) onnx/onnx@68bc26c

… ops except classifier ops. (pytorch#857) onnx/onnx@68bc26c

…27efbc Previous import was 403ccfbd0161c38f0834413d790bad0874afbf9a Included changes: - **[69894f2](onnx/onnx@69894f2)**: Use op schema.all tensor types in random like definitions (pytorch#865) <Scott McKay> - **[b9d6b90](onnx/onnx@b9d6b90)**: Clarify random like operators (pytorch#846) <Scott McKay> - **[fc6b5fb](onnx/onnx@fc6b5fb)**: Refactor shape inference implementation (pytorch#855) <anderspapitto> - **[b7d8dc8](onnx/onnx@b7d8dc8)**: fix cmake warning message (pytorch#863) <Eric S. Yu> - **[f585c5d](onnx/onnx@f585c5d)**: add pytorch-operator test for tile (pytorch#831) <Wenhao Hu> - **[993fe70](onnx/onnx@993fe70)**: add install step (pytorch#832) <Eric S. Yu> - **[68bc26c](onnx/onnx@68bc26c)**: add type inference for traditional ml ops except classifier ops. (pytorch#857) <Ke Zhang> - **[9cc0cda](onnx/onnx@9cc0cda)**: fix string representation of scalar types (pytorch#858) <G. Ramalingam> - **[1078925](onnx/onnx@1078925)**: fix y in pow test case to scalar (pytorch#852) <Wenhao Hu> - **[c66fb6f](onnx/onnx@c66fb6f)**: Add some math function shape inference (pytorch#845) <anderspapitto> - **[ff667d1](onnx/onnx@ff667d1)**: Refactor return type and docs for ONNXIFI_BACKEND_DIRECTX_ID (pytorch#853) <Marat Dukhan> - **[11c6876](onnx/onnx@11c6876)**: clear initializer names when clear initializer (pytorch#849) <Wenhao Hu> - **[73c34ae](onnx/onnx@73c34ae)**: Clarify FeatureVectorizer description. (pytorch#843) <Scott McKay> - **[1befb9b](onnx/onnx@1befb9b)**: Remove useless text in docs (pytorch#850) <Lu Fang> - **[e84788f](onnx/onnx@e84788f)**: Fix SELU attributes' default values (pytorch#839) <Lu Fang> - **[ebac046](onnx/onnx@ebac046)**: Add tile test case (pytorch#823) <Wenhao Hu> - **[8b7a925](onnx/onnx@8b7a925)**: a few more shape inference functions (pytorch#772) <anderspapitto> - **[9718f42](onnx/onnx@9718f42)**: Make the coefficient non optional for LinearClassifier (pytorch#836) <Jaliya Ekanayake> - **[ef083d0](onnx/onnx@ef083d0)**: Add save_tensor and load_tensor functions for Protos (pytorch#770) <Lu Fang> - **[45ceb55](onnx/onnx@45ceb55)**: Check if CMAKE_BUILD_TYPE set before project(). (pytorch#812) <Sergii Dymchenko> - **[4b3d2b0](onnx/onnx@4b3d2b0)**: [WIP] reenable shape inference tests (pytorch#834) <anderspapitto> - **[22d17ee](onnx/onnx@22d17ee)**: RNN tests: LSTM, GRU, SimpleRNN (pytorch#739) <Peyman Manikashani> - **[de65b95](onnx/onnx@de65b95)**: dimension denotation (pytorch#443) <Tian Jin> - **[eccc76e](onnx/onnx@eccc76e)**: fix field number issue in onnx operator proto and enable its build (pytorch#829) <Ke Zhang> - **[d582beb](onnx/onnx@d582beb)**: disable shape inference test to unbreak ci (pytorch#830) <Lu Fang> - **[485b787](onnx/onnx@485b787)**: function proto for composite op. (pytorch#802) <Ke Zhang> - **[cd58928](onnx/onnx@cd58928)**: specify defaults for attributes of Affine op (pytorch#820) <G. Ramalingam> - **[7ee2cf9](onnx/onnx@7ee2cf9)**: merge the dummy backend back into the main one (pytorch#743) <anderspapitto> - **[1c03a5a](onnx/onnx@1c03a5a)**: [Proposal] ONNX Interface for Framework Integration (previously ONNX Backend API) header and docs (pytorch#551) <Marat Dukhan> - **[3769a98](onnx/onnx@3769a98)**: Rename real model test case from VGG-16 to ZFNet (pytorch#821) <Lu Fang>

* [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](onnx/onnx@69894f2)**: Use op schema.all tensor types in random like definitions (#865) <Scott McKay> - **[b9d6b90](onnx/onnx@b9d6b90)**: Clarify random like operators (#846) <Scott McKay> - **[fc6b5fb](onnx/onnx@fc6b5fb)**: Refactor shape inference implementation (#855) <anderspapitto> - **[b7d8dc8](onnx/onnx@b7d8dc8)**: fix cmake warning message (#863) <Eric S. Yu> - **[f585c5d](onnx/onnx@f585c5d)**: add pytorch-operator test for tile (#831) <Wenhao Hu> - **[993fe70](onnx/onnx@993fe70)**: add install step (#832) <Eric S. Yu> - **[68bc26c](onnx/onnx@68bc26c)**: add type inference for traditional ml ops except classifier ops. (#857) <Ke Zhang> - **[9cc0cda](onnx/onnx@9cc0cda)**: fix string representation of scalar types (#858) <G. Ramalingam> - **[1078925](onnx/onnx@1078925)**: fix y in pow test case to scalar (#852) <Wenhao Hu> - **[c66fb6f](onnx/onnx@c66fb6f)**: Add some math function shape inference (#845) <anderspapitto> - **[ff667d1](onnx/onnx@ff667d1)**: Refactor return type and docs for ONNXIFI_BACKEND_DIRECTX_ID (#853) <Marat Dukhan> - **[11c6876](onnx/onnx@11c6876)**: clear initializer names when clear initializer (#849) <Wenhao Hu> - **[73c34ae](onnx/onnx@73c34ae)**: Clarify FeatureVectorizer description. (#843) <Scott McKay> - **[1befb9b](onnx/onnx@1befb9b)**: Remove useless text in docs (#850) <Lu Fang> - **[e84788f](onnx/onnx@e84788f)**: Fix SELU attributes' default values (#839) <Lu Fang> - **[ebac046](onnx/onnx@ebac046)**: Add tile test case (#823) <Wenhao Hu> - **[8b7a925](onnx/onnx@8b7a925)**: a few more shape inference functions (#772) <anderspapitto> - **[9718f42](onnx/onnx@9718f42)**: Make the coefficient non optional for LinearClassifier (#836) <Jaliya Ekanayake> - **[ef083d0](onnx/onnx@ef083d0)**: Add save_tensor and load_tensor functions for Protos (#770) <Lu Fang> - **[45ceb55](onnx/onnx@45ceb55)**: Check if CMAKE_BUILD_TYPE set before project(). (#812) <Sergii Dymchenko> - **[4b3d2b0](onnx/onnx@4b3d2b0)**: [WIP] reenable shape inference tests (#834) <anderspapitto> - **[22d17ee](onnx/onnx@22d17ee)**: RNN tests: LSTM, GRU, SimpleRNN (#739) <Peyman Manikashani> - **[de65b95](onnx/onnx@de65b95)**: dimension denotation (#443) <Tian Jin> - **[eccc76e](onnx/onnx@eccc76e)**: fix field number issue in onnx operator proto and enable its build (#829) <Ke Zhang> - **[d582beb](onnx/onnx@d582beb)**: disable shape inference test to unbreak ci (#830) <Lu Fang> - **[485b787](onnx/onnx@485b787)**: function proto for composite op. (#802) <Ke Zhang> - **[cd58928](onnx/onnx@cd58928)**: specify defaults for attributes of Affine op (#820) <G. Ramalingam> - **[7ee2cf9](onnx/onnx@7ee2cf9)**: merge the dummy backend back into the main one (#743) <anderspapitto> - **[1c03a5a](onnx/onnx@1c03a5a)**: [Proposal] ONNX Interface for Framework Integration (previously ONNX Backend API) header and docs (#551) <Marat Dukhan> - **[3769a98](onnx/onnx@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

… ops except classifier ops. (pytorch#857) onnx/onnx@68bc26c

* [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](onnx/onnx@69894f2)**: Use op schema.all tensor types in random like definitions (pytorch#865) <Scott McKay> - **[b9d6b90](onnx/onnx@b9d6b90)**: Clarify random like operators (pytorch#846) <Scott McKay> - **[fc6b5fb](onnx/onnx@fc6b5fb)**: Refactor shape inference implementation (pytorch#855) <anderspapitto> - **[b7d8dc8](onnx/onnx@b7d8dc8)**: fix cmake warning message (pytorch#863) <Eric S. Yu> - **[f585c5d](onnx/onnx@f585c5d)**: add pytorch-operator test for tile (pytorch#831) <Wenhao Hu> - **[993fe70](onnx/onnx@993fe70)**: add install step (pytorch#832) <Eric S. Yu> - **[68bc26c](onnx/onnx@68bc26c)**: add type inference for traditional ml ops except classifier ops. (pytorch#857) <Ke Zhang> - **[9cc0cda](onnx/onnx@9cc0cda)**: fix string representation of scalar types (pytorch#858) <G. Ramalingam> - **[1078925](onnx/onnx@1078925)**: fix y in pow test case to scalar (pytorch#852) <Wenhao Hu> - **[c66fb6f](onnx/onnx@c66fb6f)**: Add some math function shape inference (pytorch#845) <anderspapitto> - **[ff667d1](onnx/onnx@ff667d1)**: Refactor return type and docs for ONNXIFI_BACKEND_DIRECTX_ID (pytorch#853) <Marat Dukhan> - **[11c6876](onnx/onnx@11c6876)**: clear initializer names when clear initializer (pytorch#849) <Wenhao Hu> - **[73c34ae](onnx/onnx@73c34ae)**: Clarify FeatureVectorizer description. (pytorch#843) <Scott McKay> - **[1befb9b](onnx/onnx@1befb9b)**: Remove useless text in docs (pytorch#850) <Lu Fang> - **[e84788f](onnx/onnx@e84788f)**: Fix SELU attributes' default values (pytorch#839) <Lu Fang> - **[ebac046](onnx/onnx@ebac046)**: Add tile test case (pytorch#823) <Wenhao Hu> - **[8b7a925](onnx/onnx@8b7a925)**: a few more shape inference functions (pytorch#772) <anderspapitto> - **[9718f42](onnx/onnx@9718f42)**: Make the coefficient non optional for LinearClassifier (pytorch#836) <Jaliya Ekanayake> - **[ef083d0](onnx/onnx@ef083d0)**: Add save_tensor and load_tensor functions for Protos (pytorch#770) <Lu Fang> - **[45ceb55](onnx/onnx@45ceb55)**: Check if CMAKE_BUILD_TYPE set before project(). (pytorch#812) <Sergii Dymchenko> - **[4b3d2b0](onnx/onnx@4b3d2b0)**: [WIP] reenable shape inference tests (pytorch#834) <anderspapitto> - **[22d17ee](onnx/onnx@22d17ee)**: RNN tests: LSTM, GRU, SimpleRNN (pytorch#739) <Peyman Manikashani> - **[de65b95](onnx/onnx@de65b95)**: dimension denotation (pytorch#443) <Tian Jin> - **[eccc76e](onnx/onnx@eccc76e)**: fix field number issue in onnx operator proto and enable its build (pytorch#829) <Ke Zhang> - **[d582beb](onnx/onnx@d582beb)**: disable shape inference test to unbreak ci (pytorch#830) <Lu Fang> - **[485b787](onnx/onnx@485b787)**: function proto for composite op. (pytorch#802) <Ke Zhang> - **[cd58928](onnx/onnx@cd58928)**: specify defaults for attributes of Affine op (pytorch#820) <G. Ramalingam> - **[7ee2cf9](onnx/onnx@7ee2cf9)**: merge the dummy backend back into the main one (pytorch#743) <anderspapitto> - **[1c03a5a](onnx/onnx@1c03a5a)**: [Proposal] ONNX Interface for Framework Integration (previously ONNX Backend API) header and docs (pytorch#551) <Marat Dukhan> - **[3769a98](onnx/onnx@3769a98)**: Rename real model test case from VGG-16 to ZFNet (pytorch#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 pytorch#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

* Remove the use of `Tensor` and `LongTensor` * Correct formatting

szagoruyko added 2 commits February 26, 2017 14:38

bias_add

b890ba3

bias_add gradcheck

c4596ad

szagoruyko force-pushed the bias_add branch from b7e8c3a to c4596ad Compare February 26, 2017 13:41

fix lint

83850ea

apaszke added the requested-changes label Mar 13, 2017

soumith closed this Mar 28, 2017

apaszke reopened this Mar 29, 2017

apaszke removed the requested changes label Apr 9, 2017

apaszke closed this Apr 21, 2017

onnxbot added a commit that referenced this pull request May 1, 2018

[auto] Update onnx to 68bc26c - add type inference for traditional ml…

cdc6d10

… ops except classifier ops. (#857) onnx/onnx@68bc26c

Jorghi12 pushed a commit to wsttiger/pytorch that referenced this pull request May 10, 2018

[auto] Update onnx to 68bc26c - add type inference for traditional ml…

57199c9

… ops except classifier ops. (pytorch#857) onnx/onnx@68bc26c

weiyangfb pushed a commit to weiyangfb/pytorch that referenced this pull request Jun 11, 2018

[auto] Update onnx to 68bc26c - add type inference for traditional ml…

ac5f1a7

… ops except classifier ops. (pytorch#857) onnx/onnx@68bc26c

mrshenli pushed a commit to mrshenli/pytorch that referenced this pull request Apr 11, 2020

Remove the use of Tensor and LongTensor (pytorch#857)

af5bf50

* Remove the use of `Tensor` and `LongTensor` * Correct formatting

jjsjann123 pushed a commit to jjsjann123/pytorch that referenced this pull request May 19, 2021

Avoid random failure (pytorch#857)

9f1af6f

zhuhong61 pushed a commit to zhuhong61/pytorch that referenced this pull request Jun 8, 2022

upgrade to cu11.1.1 (pytorch#857)

18ad7fc

Functional bias_add #857

Functional bias_add #857

Uh oh!

Conversation

szagoruyko commented Feb 26, 2017

Uh oh!

soumith commented Feb 26, 2017 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

szagoruyko commented Feb 26, 2017

Uh oh!

fmassa commented Feb 26, 2017

Uh oh!

szagoruyko commented Feb 26, 2017

Uh oh!

fmassa commented Feb 26, 2017

Uh oh!

apaszke commented Feb 26, 2017

Uh oh!

ngimel commented Feb 26, 2017

Uh oh!

apaszke commented Mar 24, 2017 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

ngimel commented Mar 24, 2017

Uh oh!

ngimel commented Mar 24, 2017

Uh oh!

apaszke commented Mar 24, 2017

Uh oh!

ngimel commented Mar 24, 2017

Uh oh!

soumith commented Mar 28, 2017

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

5 participants

soumith commented Feb 26, 2017 •

edited

Loading

apaszke commented Mar 24, 2017 •

edited

Loading