Add numerically stable BCELoss which takes logits as input #1792

alykhantejani · 2017-06-13T17:16:25Z

Add numerically stable BCELoss which takes logits as input and combines the sigmoid and BCELoss in one layer. This is in response to issue #751

…nd num input planes

…stable_BCELoss

…nd num input planes

soumith · 2017-06-20T02:05:57Z

thanks man!

esube · 2017-06-26T17:50:03Z

@alykhantejani @soumith
I just saw this commit. Thanks for implementing this tf like BCE with logits!

I noticed that the function repeats the weight vector to match target if it is provided. However, when the weight vector is the same shape as the target vector to begin with (as is the case in multi-label classification), the check line fails as it tries to expand the weight vector. So, it would be nice to check if the weight is already the same shape as the target before trying to expand it. This will solve some issues related to mult-label sample by sample weighting rather than class-based weighting.

alykhantejani · 2017-06-27T15:34:12Z

@esube can you provide an example of the failure case (in code)?

This worked fine for me:

output = torch.autograd.Variable(torch.Tensor(2, 10).uniform_())
target = torch.autograd.Variable(torch.Tensor(2, 10).uniform_())
weigth = torch.Tensor(10).uniform_()
loss_fn = nn.BCEWithLogitsLoss(weight)
loss = loss_fn(output, target)

esube · 2017-06-27T18:50:56Z

@alykhantejani
The following:

output = torch.autograd.Variable(torch.Tensor(2, 10).uniform_())
target = torch.autograd.Variable(torch.Tensor(2, 10).uniform_())
weight = torch.Tensor(2,10).uniform_()
loss_fn = nn.BCEWithLogitsLoss(weight)
loss = loss_fn(output, target)

fails with error code:

Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "python2.7/site-packages/torch/nn/modules/module.py", line 206, in __call__
    result = self.forward(*input, **kwargs)
  File "python2.7/site-packages/torch/nn/modules/loss.py", line 232, in forward
    return F.binary_cross_entropy_with_logits(input, target, Variable(self.weight), self.size_average)
  File "python2.7/site-packages/torch/nn/functional.py", line 659, in binary_cross_entropy_with_logits
    weight = weight.view(1, target.size(1)).expand_as(target)
  File "python2.7/site-packages/torch/autograd/variable.py", line 479, in view
    return View.apply(self, sizes)
  File "python2.7/site-packages/torch/autograd/_functions/tensor.py", line 96, in forward
    result = i.view(*sizes)
RuntimeError: size '[1 x 10]' is invalid for input of with 20 elements at pytorch/torch/lib/TH/THStorage.c:41

In multi-label (not multi-class) classification tasks, there are conditions that you want 2D weight for 2D target vector to weigh each label of each sample separately such as in the case of extreme imbalance in positive and negative examples of multi-label learning.

Note: I am not saying this loss function would work out of the box even if you fix the check as the weight is passed to the initializer than the forward function in the pytorch loss functions paradigm (the weight multiplication would fail). However, I locally have the function that it accepts batch of weights in the forward function just like the targets and perform the weight multiplication. I have this locally working and if needed can do a PR.

alykhantejani · 2017-06-27T18:59:44Z

@esube Ah ok, gotcha. ok I can add support for this. Out of interest, how are you currently doing this with the normal BCELoss + Sigmoid?

esube · 2017-06-27T19:08:33Z

@alykhantejani
Sorry for the edited response above. I have changed the class and the function so that the forward accepts three inputs as I described above and I have it working locally. It would be nice to have this either in your PR as correction or on a separate PR if you think it is completely different.

class WBCELoss(nn.Module):
    def __init__(self, size_average=True):
        super(WBCELoss, self).__init__()
        self.size_average = size_average

    def forward(self, input, target, weight=None):
        return binary_cross_entropy_with_logits(input, target, weight, self.size_average)

def binary_cross_entropy_with_logits(input, target, weight=None, size_average=True):
   
    if weight is not None and weight.dim() != target.dim() and target.dim() != 1:
        weight = weight.view(1, target.size(1)).expand_as(target)

    neg_abs = - input.abs()
    loss = input.clamp(min=0) - input * target + (1 + neg_abs.exp()).log()

    if weight is not None:
        loss = loss * weight

    if size_average:
        return loss.mean()
    else:
        return loss.sum()

Note: this expects the dataset loader to supply a minibatch of input, target and weight tensors to the loss function.

alykhantejani · 2017-06-27T19:18:51Z

Should be fairly easy to amend the binary_cross_entropy_with_logits to do this (I can do this tomorrow). I was just curious how you achieve this behaviour with the old BCELoss + Sigmoid (not the new BCELossWithLogits), because that too does the same resize, see here: https://github.com/pytorch/pytorch/blob/master/torch/nn/_functions/thnn/loss.py#L15

esube · 2017-06-27T19:56:20Z

@alykhantejani Thanks for updating it to accommodate this. I had to do these modifications with the regular BCELoss + Sigmoid as well. In short most losses are written without multi-label weighting in mind.

Note: that adding this support shouldn't break the support for class-based weighting so you might want to have two weights (class weight in the constructor as ussual and sample weight in the forward function and multiply with both).

alykhantejani · 2017-07-24T20:04:26Z

@Jiaming-Liu this should be fixed in #2195

F0REacH · 2017-07-26T18:34:50Z

@alykhantejani Fix from #2195 gives some unexpected behavior:
input, target = torch.Tensor([-100.0]), torch.Tensor([0]) print(F.binary_cross_entropy(torch.sigmoid(input), torch.Tensor(target)))

Variable containing:
1.00000e-12 *
-1.0001
[torch.FloatTensor of size 1]

print(F.binary_cross_entropy_with_logits(input, torch.Tensor(target)))

inf

neg_abs = - input.abs() loss = input.clamp(min=0) - input * target + (1 + neg_abs.exp()).log() print(loss)

0
[torch.FloatTensor of size 1]

Jiaming-Liu · 2017-07-26T21:45:28Z

https://github.com/Jiaming-Liu/pytorch-misc/blob/master/test_binary_cross_entropy.ipynb

This might be helpful for debugging the loss. Just change loss definition in cell [2] and this should work.

https://github.com/Jiaming-Liu/pytorch-misc/blob/master/sigmoid_with_binary_cross_entropy.py

This is my personal fix of BCE+logit issue.

alykhantejani · 2017-07-26T21:55:08Z

@F0REacH thanks for reporting. This is a subtle bug that only occurs with large values that cause the exp to overflow. It's a very easy fix, but I will also add a test, because this is just a plain bug. Sending a PR in a min.

alykhantejani · 2017-07-26T22:06:33Z

@Jiaming-Liu Thanks. I found the issue, I forgot to do (-input).clamp(min=0) i.e. I forgot the - sign.

This only shows a error when exp(x) overflows, and the tests were run with DoubleTensors so this didn't happen. Sending a PR now

…19bdc7 Summary: Previous import was bfa8b335ab6d1ed7b688dc2ec96421a3fe9e644c Included changes: - **[822d8df](onnx/onnx@822d8df)**: allow removed experimental ops in the checker for now (pytorch#1792) <Lu Fang> Differential Revision: D13970103 fbshipit-source-id: 5174767578cb44caa4cf4b093ce57e605ab8df1e

…19bdc7 (#16791) Summary: Pull Request resolved: #16791 Previous import was bfa8b335ab6d1ed7b688dc2ec96421a3fe9e644c Included changes: - **[822d8df](onnx/onnx@822d8df)**: allow removed experimental ops in the checker for now (#1792) <Lu Fang> Reviewed By: MisterTea Differential Revision: D13970103 fbshipit-source-id: 5feaaa6c65ba10901eeba0b63724d7e451e9b642

Syncing nvfuser devel branch to upstream master. https://github.com/csarofeen/pytorch/ Code changes includes: - codegen improvements: 1. Indexing refactor -> Remove reference tensor in predicate indexing logic 2. MMA Rfactor support for cross-warp and cross-CTA split on K dimension 3. Grouping grid allreduces across iterations 4. Swizzle op formulation for non-affine swizzles 5. Use scheduler_utils to cache inputs and outputs in schedulePointwise - scheduler refactor 1. New compute at interface - transformation propagation refactor on MaxInfoSpanningTree 1. Added sibling path that is required to generate consistent replay for some cases where `MaxInfoSpanningTree` is used with a selector. 2. Optimization to skip Transform propagator 3. SpanningTreePrinter for debugging - parser update 1. Fixes `div` 2. Added `_to_copy` 3. Broadcast in dim with expand to support expanding to concrete size 4. Dropout prob extremal patch - executor patch on caching strides for output allocation Squashed commits to WAR github API Commits that's actually in this PR from the devel branch: ``` 3b87896 Fix allocation of work buffers and `fused_reduction::ParallelReduce` with unswitch (#1818) 4cae122 schedulePointwise cleanup: - computeAt + InlinePropagator (#1815) 3df9742 Use scheduler_utils to cache inputs and outputs in schedulePointwise (#1811) 03180aa improve broadcast resolution (#1792) bee6c69 bug fix (#1819) 4413c8f Support PYTORCH_NVFUSER_DUMP=transform_propagator (#1812) de6b7ca Fix negative position in InlinePropagator (#1813) 10a996c Remove redundant check in schedulePointwise (#1810) acd5ed4 Swizzle op formulation for non-affine swizzles (#1441) 3ed8330 Kernel args patch to show zero_init buffer (#1809) 037a75a Dropout prob extremal patch (#1804) 282c429 spam nvrtc options (#1783) 3ba6a5f Broadcast in dim with expand (#1794) fd4be12 remove dead indexing code (#1806) fa4e6a4 Check siblings in getMaxPosAll (#1805) 025c840 Grouping grid allreduces across iterations (#1755) 37c579e Temporarily disable test requring large shared memory. (#1802) 5f375d0 More cleanup on InlinePropagator (#1800) 8d384da Indexing refactor stage 2 : Remove reference tensor in predicate indexing logic (#1784) f008140 MMA Rfactor support for cross-warp and cross-CTA split on K dimension (#1554) 76b3cca Add parsing support for `_to_copy` to handle AMP casts. (#1756) ef04f6c Coding style cleanups (#1798) 38c7f3c InlinePropagator please don't replay (#1797) 3f2c263 validateDomain in TransformPropagator (#1796) c077085 Use TransformPropagatorWithCheck in many tests (#1795) d0d0908 Some further cleanup for the new computeAt interface (#1793) 45f5203 Fix TransformReplay::getMatchedLeafPosWithoutReplay* (#1791) 28cbaf9 New compute at interface (#1743) 635ebfc Add SpanningTreePrinter (#1786) 59f3c32 Output allocate patch (#1790) fe93bf5 Transform propagator skip replay when possible (#1782) ebf23a5 Fix isIntegralType error msg (#1789) 0c82ecf Disable register reuse across serial broadcast ops (#1787) 33a824d Adding sibling path for MaxInfoSpanningTree (#1776) 86f46aa Fix div(Val, TensorView) (#1778) d3de227 Fix FusionMaxRootDomainInfoSpanningTreePrintTwice_CUDA (#1781) ecc7a87 Extend mma dimension and layout checking to support strided batched matmul and tensor contractions (#1761) ``` [ghstack-poisoned]

Syncing nvfuser devel branch to upstream master. https://github.com/csarofeen/pytorch/ Code changes includes: - codegen improvements: 1. Indexing refactor -> Remove reference tensor in predicate indexing logic 2. MMA Rfactor support for cross-warp and cross-CTA split on K dimension 3. Grouping grid allreduces across iterations 4. Swizzle op formulation for non-affine swizzles 5. Use scheduler_utils to cache inputs and outputs in schedulePointwise - scheduler refactor 1. New compute at interface - transformation propagation refactor on MaxInfoSpanningTree 1. Added sibling path that is required to generate consistent replay for some cases where `MaxInfoSpanningTree` is used with a selector. 2. Optimization to skip Transform propagator 3. SpanningTreePrinter for debugging - parser update 1. Fixes `div` 2. Added `_to_copy` 3. Broadcast in dim with expand to support expanding to concrete size 4. Dropout prob extremal patch - executor patch on caching strides for output allocation Squashed commits to WAR github API Commits that's actually in this PR from the devel branch: ``` 3b87896 Fix allocation of work buffers and `fused_reduction::ParallelReduce` with unswitch (#1818) 4cae122 schedulePointwise cleanup: - computeAt + InlinePropagator (#1815) 3df9742 Use scheduler_utils to cache inputs and outputs in schedulePointwise (#1811) 03180aa improve broadcast resolution (#1792) bee6c69 bug fix (#1819) 4413c8f Support PYTORCH_NVFUSER_DUMP=transform_propagator (#1812) de6b7ca Fix negative position in InlinePropagator (#1813) 10a996c Remove redundant check in schedulePointwise (#1810) acd5ed4 Swizzle op formulation for non-affine swizzles (#1441) 3ed8330 Kernel args patch to show zero_init buffer (#1809) 037a75a Dropout prob extremal patch (#1804) 282c429 spam nvrtc options (#1783) 3ba6a5f Broadcast in dim with expand (#1794) fd4be12 remove dead indexing code (#1806) fa4e6a4 Check siblings in getMaxPosAll (#1805) 025c840 Grouping grid allreduces across iterations (#1755) 37c579e Temporarily disable test requring large shared memory. (#1802) 5f375d0 More cleanup on InlinePropagator (#1800) 8d384da Indexing refactor stage 2 : Remove reference tensor in predicate indexing logic (#1784) f008140 MMA Rfactor support for cross-warp and cross-CTA split on K dimension (#1554) 76b3cca Add parsing support for `_to_copy` to handle AMP casts. (#1756) ef04f6c Coding style cleanups (#1798) 38c7f3c InlinePropagator please don't replay (#1797) 3f2c263 validateDomain in TransformPropagator (#1796) c077085 Use TransformPropagatorWithCheck in many tests (#1795) d0d0908 Some further cleanup for the new computeAt interface (#1793) 45f5203 Fix TransformReplay::getMatchedLeafPosWithoutReplay* (#1791) 28cbaf9 New compute at interface (#1743) 635ebfc Add SpanningTreePrinter (#1786) 59f3c32 Output allocate patch (#1790) fe93bf5 Transform propagator skip replay when possible (#1782) ebf23a5 Fix isIntegralType error msg (#1789) 0c82ecf Disable register reuse across serial broadcast ops (#1787) 33a824d Adding sibling path for MaxInfoSpanningTree (#1776) 86f46aa Fix div(Val, TensorView) (#1778) d3de227 Fix FusionMaxRootDomainInfoSpanningTreePrintTwice_CUDA (#1781) ecc7a87 Extend mma dimension and layout checking to support strided batched matmul and tensor contractions (#1761) ``` ghstack-source-id: f24793f Pull Request resolved: #81861

Syncing nvfuser devel branch to upstream master. https://github.com/csarofeen/pytorch/ Code changes includes: - codegen improvements: 1. Indexing refactor -> Remove reference tensor in predicate indexing logic 2. MMA Rfactor support for cross-warp and cross-CTA split on K dimension 3. Grouping grid allreduces across iterations 4. Swizzle op formulation for non-affine swizzles 5. Use scheduler_utils to cache inputs and outputs in schedulePointwise - scheduler refactor 1. New compute at interface - transformation propagation refactor on MaxInfoSpanningTree 1. Added sibling path that is required to generate consistent replay for some cases where `MaxInfoSpanningTree` is used with a selector. 2. Optimization to skip Transform propagator 3. SpanningTreePrinter for debugging - parser update 1. Fixes `div` 2. Added `_to_copy` 3. Broadcast in dim with expand to support expanding to concrete size 4. Dropout prob extremal patch - executor patch on caching strides for output allocation Squashed commits to WAR github API Commits that's actually in this PR from the devel branch: ``` 3b87896 Fix allocation of work buffers and `fused_reduction::ParallelReduce` with unswitch (#1818) 4cae122 schedulePointwise cleanup: - computeAt + InlinePropagator (#1815) 3df9742 Use scheduler_utils to cache inputs and outputs in schedulePointwise (#1811) 03180aa improve broadcast resolution (#1792) bee6c69 bug fix (#1819) 4413c8f Support PYTORCH_NVFUSER_DUMP=transform_propagator (#1812) de6b7ca Fix negative position in InlinePropagator (#1813) 10a996c Remove redundant check in schedulePointwise (#1810) acd5ed4 Swizzle op formulation for non-affine swizzles (#1441) 3ed8330 Kernel args patch to show zero_init buffer (#1809) 037a75a Dropout prob extremal patch (#1804) 282c429 spam nvrtc options (#1783) 3ba6a5f Broadcast in dim with expand (#1794) fd4be12 remove dead indexing code (#1806) fa4e6a4 Check siblings in getMaxPosAll (#1805) 025c840 Grouping grid allreduces across iterations (#1755) 37c579e Temporarily disable test requring large shared memory. (#1802) 5f375d0 More cleanup on InlinePropagator (#1800) 8d384da Indexing refactor stage 2 : Remove reference tensor in predicate indexing logic (#1784) f008140 MMA Rfactor support for cross-warp and cross-CTA split on K dimension (#1554) 76b3cca Add parsing support for `_to_copy` to handle AMP casts. (#1756) ef04f6c Coding style cleanups (#1798) 38c7f3c InlinePropagator please don't replay (#1797) 3f2c263 validateDomain in TransformPropagator (#1796) c077085 Use TransformPropagatorWithCheck in many tests (#1795) d0d0908 Some further cleanup for the new computeAt interface (#1793) 45f5203 Fix TransformReplay::getMatchedLeafPosWithoutReplay* (#1791) 28cbaf9 New compute at interface (#1743) 635ebfc Add SpanningTreePrinter (#1786) 59f3c32 Output allocate patch (#1790) fe93bf5 Transform propagator skip replay when possible (#1782) ebf23a5 Fix isIntegralType error msg (#1789) 0c82ecf Disable register reuse across serial broadcast ops (#1787) 33a824d Adding sibling path for MaxInfoSpanningTree (#1776) 86f46aa Fix div(Val, TensorView) (#1778) d3de227 Fix FusionMaxRootDomainInfoSpanningTreePrintTwice_CUDA (#1781) ecc7a87 Extend mma dimension and layout checking to support strided batched matmul and tensor contractions (#1761) ``` RUN_TORCHBENCH: nvfuser Differential Revision: [D38043938](https://our.internmc.facebook.com/intern/diff/D38043938) [ghstack-poisoned]

Syncing nvfuser devel branch to upstream master. https://github.com/csarofeen/pytorch/ Code changes includes: - codegen improvements: 1. Indexing refactor -> Remove reference tensor in predicate indexing logic 2. MMA Rfactor support for cross-warp and cross-CTA split on K dimension 3. Grouping grid allreduces across iterations 4. Swizzle op formulation for non-affine swizzles 5. Use scheduler_utils to cache inputs and outputs in schedulePointwise - scheduler refactor 1. New compute at interface - transformation propagation refactor on MaxInfoSpanningTree 1. Added sibling path that is required to generate consistent replay for some cases where `MaxInfoSpanningTree` is used with a selector. 2. Optimization to skip Transform propagator 3. SpanningTreePrinter for debugging - parser update 1. Fixes `div` 2. Added `_to_copy` 3. Broadcast in dim with expand to support expanding to concrete size 4. Dropout prob extremal patch - executor patch on caching strides for output allocation Squashed commits to WAR github API Commits that's actually in this PR from the devel branch: ``` 3b87896 Fix allocation of work buffers and `fused_reduction::ParallelReduce` with unswitch (#1818) 4cae122 schedulePointwise cleanup: - computeAt + InlinePropagator (#1815) 3df9742 Use scheduler_utils to cache inputs and outputs in schedulePointwise (#1811) 03180aa improve broadcast resolution (#1792) bee6c69 bug fix (#1819) 4413c8f Support PYTORCH_NVFUSER_DUMP=transform_propagator (#1812) de6b7ca Fix negative position in InlinePropagator (#1813) 10a996c Remove redundant check in schedulePointwise (#1810) acd5ed4 Swizzle op formulation for non-affine swizzles (#1441) 3ed8330 Kernel args patch to show zero_init buffer (#1809) 037a75a Dropout prob extremal patch (#1804) 282c429 spam nvrtc options (#1783) 3ba6a5f Broadcast in dim with expand (#1794) fd4be12 remove dead indexing code (#1806) fa4e6a4 Check siblings in getMaxPosAll (#1805) 025c840 Grouping grid allreduces across iterations (#1755) 37c579e Temporarily disable test requring large shared memory. (#1802) 5f375d0 More cleanup on InlinePropagator (#1800) 8d384da Indexing refactor stage 2 : Remove reference tensor in predicate indexing logic (#1784) f008140 MMA Rfactor support for cross-warp and cross-CTA split on K dimension (#1554) 76b3cca Add parsing support for `_to_copy` to handle AMP casts. (#1756) ef04f6c Coding style cleanups (#1798) 38c7f3c InlinePropagator please don't replay (#1797) 3f2c263 validateDomain in TransformPropagator (#1796) c077085 Use TransformPropagatorWithCheck in many tests (#1795) d0d0908 Some further cleanup for the new computeAt interface (#1793) 45f5203 Fix TransformReplay::getMatchedLeafPosWithoutReplay* (#1791) 28cbaf9 New compute at interface (#1743) 635ebfc Add SpanningTreePrinter (#1786) 59f3c32 Output allocate patch (#1790) fe93bf5 Transform propagator skip replay when possible (#1782) ebf23a5 Fix isIntegralType error msg (#1789) 0c82ecf Disable register reuse across serial broadcast ops (#1787) 33a824d Adding sibling path for MaxInfoSpanningTree (#1776) 86f46aa Fix div(Val, TensorView) (#1778) d3de227 Fix FusionMaxRootDomainInfoSpanningTreePrintTwice_CUDA (#1781) ecc7a87 Extend mma dimension and layout checking to support strided batched matmul and tensor contractions (#1761) ``` RUN_TORCHBENCH: nvfuser ghstack-source-id: cfd5278 Pull Request resolved: #81861

Syncing nvfuser devel branch to upstream master. https://github.com/csarofeen/pytorch/ Code changes includes: - codegen improvements: 1. Indexing refactor -> Remove reference tensor in predicate indexing logic 2. MMA Rfactor support for cross-warp and cross-CTA split on K dimension 3. Grouping grid allreduces across iterations 4. Swizzle op formulation for non-affine swizzles 5. Use scheduler_utils to cache inputs and outputs in schedulePointwise - scheduler refactor 1. New compute at interface - transformation propagation refactor on MaxInfoSpanningTree 1. Added sibling path that is required to generate consistent replay for some cases where `MaxInfoSpanningTree` is used with a selector. 2. Optimization to skip Transform propagator 3. SpanningTreePrinter for debugging - parser update 1. Fixes `div` 2. Added `_to_copy` 3. Broadcast in dim with expand to support expanding to concrete size 4. Dropout prob extremal patch - executor patch on caching strides for output allocation Squashed commits to WAR github API Commits that's actually in this PR from the devel branch: ``` 3b87896 Fix allocation of work buffers and `fused_reduction::ParallelReduce` with unswitch (#1818) 4cae122 schedulePointwise cleanup: - computeAt + InlinePropagator (#1815) 3df9742 Use scheduler_utils to cache inputs and outputs in schedulePointwise (#1811) 03180aa improve broadcast resolution (#1792) bee6c69 bug fix (#1819) 4413c8f Support PYTORCH_NVFUSER_DUMP=transform_propagator (#1812) de6b7ca Fix negative position in InlinePropagator (#1813) 10a996c Remove redundant check in schedulePointwise (#1810) acd5ed4 Swizzle op formulation for non-affine swizzles (#1441) 3ed8330 Kernel args patch to show zero_init buffer (#1809) 037a75a Dropout prob extremal patch (#1804) 282c429 spam nvrtc options (#1783) 3ba6a5f Broadcast in dim with expand (#1794) fd4be12 remove dead indexing code (#1806) fa4e6a4 Check siblings in getMaxPosAll (#1805) 025c840 Grouping grid allreduces across iterations (#1755) 37c579e Temporarily disable test requring large shared memory. (#1802) 5f375d0 More cleanup on InlinePropagator (#1800) 8d384da Indexing refactor stage 2 : Remove reference tensor in predicate indexing logic (#1784) f008140 MMA Rfactor support for cross-warp and cross-CTA split on K dimension (#1554) 76b3cca Add parsing support for `_to_copy` to handle AMP casts. (#1756) ef04f6c Coding style cleanups (#1798) 38c7f3c InlinePropagator please don't replay (#1797) 3f2c263 validateDomain in TransformPropagator (#1796) c077085 Use TransformPropagatorWithCheck in many tests (#1795) d0d0908 Some further cleanup for the new computeAt interface (#1793) 45f5203 Fix TransformReplay::getMatchedLeafPosWithoutReplay* (#1791) 28cbaf9 New compute at interface (#1743) 635ebfc Add SpanningTreePrinter (#1786) 59f3c32 Output allocate patch (#1790) fe93bf5 Transform propagator skip replay when possible (#1782) ebf23a5 Fix isIntegralType error msg (#1789) 0c82ecf Disable register reuse across serial broadcast ops (#1787) 33a824d Adding sibling path for MaxInfoSpanningTree (#1776) 86f46aa Fix div(Val, TensorView) (#1778) d3de227 Fix FusionMaxRootDomainInfoSpanningTreePrintTwice_CUDA (#1781) ecc7a87 Extend mma dimension and layout checking to support strided batched matmul and tensor contractions (#1761) ``` RUN_TORCHBENCH: nvfuser Differential Revision: [D38043938](https://our.internmc.facebook.com/intern/diff/D38043938) [ghstack-poisoned]

Syncing nvfuser devel branch to upstream master. https://github.com/csarofeen/pytorch/ Code changes includes: - codegen improvements: 1. Indexing refactor -> Remove reference tensor in predicate indexing logic 2. MMA Rfactor support for cross-warp and cross-CTA split on K dimension 3. Grouping grid allreduces across iterations 4. Swizzle op formulation for non-affine swizzles 5. Use scheduler_utils to cache inputs and outputs in schedulePointwise - scheduler refactor 1. New compute at interface - transformation propagation refactor on MaxInfoSpanningTree 1. Added sibling path that is required to generate consistent replay for some cases where `MaxInfoSpanningTree` is used with a selector. 2. Optimization to skip Transform propagator 3. SpanningTreePrinter for debugging - parser update 1. Fixes `div` 2. Added `_to_copy` 3. Broadcast in dim with expand to support expanding to concrete size 4. Dropout prob extremal patch - executor patch on caching strides for output allocation Squashed commits to WAR github API Commits that's actually in this PR from the devel branch: ``` 3b87896 Fix allocation of work buffers and `fused_reduction::ParallelReduce` with unswitch (#1818) 4cae122 schedulePointwise cleanup: - computeAt + InlinePropagator (#1815) 3df9742 Use scheduler_utils to cache inputs and outputs in schedulePointwise (#1811) 03180aa improve broadcast resolution (#1792) bee6c69 bug fix (#1819) 4413c8f Support PYTORCH_NVFUSER_DUMP=transform_propagator (#1812) de6b7ca Fix negative position in InlinePropagator (#1813) 10a996c Remove redundant check in schedulePointwise (#1810) acd5ed4 Swizzle op formulation for non-affine swizzles (#1441) 3ed8330 Kernel args patch to show zero_init buffer (#1809) 037a75a Dropout prob extremal patch (#1804) 282c429 spam nvrtc options (#1783) 3ba6a5f Broadcast in dim with expand (#1794) fd4be12 remove dead indexing code (#1806) fa4e6a4 Check siblings in getMaxPosAll (#1805) 025c840 Grouping grid allreduces across iterations (#1755) 37c579e Temporarily disable test requring large shared memory. (#1802) 5f375d0 More cleanup on InlinePropagator (#1800) 8d384da Indexing refactor stage 2 : Remove reference tensor in predicate indexing logic (#1784) f008140 MMA Rfactor support for cross-warp and cross-CTA split on K dimension (#1554) 76b3cca Add parsing support for `_to_copy` to handle AMP casts. (#1756) ef04f6c Coding style cleanups (#1798) 38c7f3c InlinePropagator please don't replay (#1797) 3f2c263 validateDomain in TransformPropagator (#1796) c077085 Use TransformPropagatorWithCheck in many tests (#1795) d0d0908 Some further cleanup for the new computeAt interface (#1793) 45f5203 Fix TransformReplay::getMatchedLeafPosWithoutReplay* (#1791) 28cbaf9 New compute at interface (#1743) 635ebfc Add SpanningTreePrinter (#1786) 59f3c32 Output allocate patch (#1790) fe93bf5 Transform propagator skip replay when possible (#1782) ebf23a5 Fix isIntegralType error msg (#1789) 0c82ecf Disable register reuse across serial broadcast ops (#1787) 33a824d Adding sibling path for MaxInfoSpanningTree (#1776) 86f46aa Fix div(Val, TensorView) (#1778) d3de227 Fix FusionMaxRootDomainInfoSpanningTreePrintTwice_CUDA (#1781) ecc7a87 Extend mma dimension and layout checking to support strided batched matmul and tensor contractions (#1761) ``` RUN_TORCHBENCH: nvfuser ghstack-source-id: 93c6b1e Pull Request resolved: #81861

Syncing nvfuser devel branch to upstream master. https://github.com/csarofeen/pytorch/ Code changes includes: - codegen improvements: 1. Indexing refactor -> Remove reference tensor in predicate indexing logic 2. MMA Rfactor support for cross-warp and cross-CTA split on K dimension 3. Grouping grid allreduces across iterations 4. Swizzle op formulation for non-affine swizzles 5. Use scheduler_utils to cache inputs and outputs in schedulePointwise - scheduler refactor 1. New compute at interface - transformation propagation refactor on MaxInfoSpanningTree 1. Added sibling path that is required to generate consistent replay for some cases where `MaxInfoSpanningTree` is used with a selector. 2. Optimization to skip Transform propagator 3. SpanningTreePrinter for debugging - parser update 1. Fixes `div` 2. Added `_to_copy` 3. Broadcast in dim with expand to support expanding to concrete size 4. Dropout prob extremal patch - executor patch on caching strides for output allocation Squashed commits to WAR github API Commits that's actually in this PR from the devel branch: ``` 3b87896 Fix allocation of work buffers and `fused_reduction::ParallelReduce` with unswitch (#1818) 4cae122 schedulePointwise cleanup: - computeAt + InlinePropagator (#1815) 3df9742 Use scheduler_utils to cache inputs and outputs in schedulePointwise (#1811) 03180aa improve broadcast resolution (#1792) bee6c69 bug fix (#1819) 4413c8f Support PYTORCH_NVFUSER_DUMP=transform_propagator (#1812) de6b7ca Fix negative position in InlinePropagator (#1813) 10a996c Remove redundant check in schedulePointwise (#1810) acd5ed4 Swizzle op formulation for non-affine swizzles (#1441) 3ed8330 Kernel args patch to show zero_init buffer (#1809) 037a75a Dropout prob extremal patch (#1804) 282c429 spam nvrtc options (#1783) 3ba6a5f Broadcast in dim with expand (#1794) fd4be12 remove dead indexing code (#1806) fa4e6a4 Check siblings in getMaxPosAll (#1805) 025c840 Grouping grid allreduces across iterations (#1755) 37c579e Temporarily disable test requring large shared memory. (#1802) 5f375d0 More cleanup on InlinePropagator (#1800) 8d384da Indexing refactor stage 2 : Remove reference tensor in predicate indexing logic (#1784) f008140 MMA Rfactor support for cross-warp and cross-CTA split on K dimension (#1554) 76b3cca Add parsing support for `_to_copy` to handle AMP casts. (#1756) ef04f6c Coding style cleanups (#1798) 38c7f3c InlinePropagator please don't replay (#1797) 3f2c263 validateDomain in TransformPropagator (#1796) c077085 Use TransformPropagatorWithCheck in many tests (#1795) d0d0908 Some further cleanup for the new computeAt interface (#1793) 45f5203 Fix TransformReplay::getMatchedLeafPosWithoutReplay* (#1791) 28cbaf9 New compute at interface (#1743) 635ebfc Add SpanningTreePrinter (#1786) 59f3c32 Output allocate patch (#1790) fe93bf5 Transform propagator skip replay when possible (#1782) ebf23a5 Fix isIntegralType error msg (#1789) 0c82ecf Disable register reuse across serial broadcast ops (#1787) 33a824d Adding sibling path for MaxInfoSpanningTree (#1776) 86f46aa Fix div(Val, TensorView) (#1778) d3de227 Fix FusionMaxRootDomainInfoSpanningTreePrintTwice_CUDA (#1781) ecc7a87 Extend mma dimension and layout checking to support strided batched matmul and tensor contractions (#1761) ``` RUN_TORCHBENCH: nvfuser Differential Revision: [D38043938](https://our.internmc.facebook.com/intern/diff/D38043938) [ghstack-poisoned]

Syncing nvfuser devel branch to upstream master. https://github.com/csarofeen/pytorch/ Code changes includes: - codegen improvements: 1. Indexing refactor -> Remove reference tensor in predicate indexing logic 2. MMA Rfactor support for cross-warp and cross-CTA split on K dimension 3. Grouping grid allreduces across iterations 4. Swizzle op formulation for non-affine swizzles 5. Use scheduler_utils to cache inputs and outputs in schedulePointwise - scheduler refactor 1. New compute at interface - transformation propagation refactor on MaxInfoSpanningTree 1. Added sibling path that is required to generate consistent replay for some cases where `MaxInfoSpanningTree` is used with a selector. 2. Optimization to skip Transform propagator 3. SpanningTreePrinter for debugging - parser update 1. Fixes `div` 2. Added `_to_copy` 3. Broadcast in dim with expand to support expanding to concrete size 4. Dropout prob extremal patch - executor patch on caching strides for output allocation Squashed commits to WAR github API Commits that's actually in this PR from the devel branch: ``` 3b87896 Fix allocation of work buffers and `fused_reduction::ParallelReduce` with unswitch (#1818) 4cae122 schedulePointwise cleanup: - computeAt + InlinePropagator (#1815) 3df9742 Use scheduler_utils to cache inputs and outputs in schedulePointwise (#1811) 03180aa improve broadcast resolution (#1792) bee6c69 bug fix (#1819) 4413c8f Support PYTORCH_NVFUSER_DUMP=transform_propagator (#1812) de6b7ca Fix negative position in InlinePropagator (#1813) 10a996c Remove redundant check in schedulePointwise (#1810) acd5ed4 Swizzle op formulation for non-affine swizzles (#1441) 3ed8330 Kernel args patch to show zero_init buffer (#1809) 037a75a Dropout prob extremal patch (#1804) 282c429 spam nvrtc options (#1783) 3ba6a5f Broadcast in dim with expand (#1794) fd4be12 remove dead indexing code (#1806) fa4e6a4 Check siblings in getMaxPosAll (#1805) 025c840 Grouping grid allreduces across iterations (#1755) 37c579e Temporarily disable test requring large shared memory. (#1802) 5f375d0 More cleanup on InlinePropagator (#1800) 8d384da Indexing refactor stage 2 : Remove reference tensor in predicate indexing logic (#1784) f008140 MMA Rfactor support for cross-warp and cross-CTA split on K dimension (#1554) 76b3cca Add parsing support for `_to_copy` to handle AMP casts. (#1756) ef04f6c Coding style cleanups (#1798) 38c7f3c InlinePropagator please don't replay (#1797) 3f2c263 validateDomain in TransformPropagator (#1796) c077085 Use TransformPropagatorWithCheck in many tests (#1795) d0d0908 Some further cleanup for the new computeAt interface (#1793) 45f5203 Fix TransformReplay::getMatchedLeafPosWithoutReplay* (#1791) 28cbaf9 New compute at interface (#1743) 635ebfc Add SpanningTreePrinter (#1786) 59f3c32 Output allocate patch (#1790) fe93bf5 Transform propagator skip replay when possible (#1782) ebf23a5 Fix isIntegralType error msg (#1789) 0c82ecf Disable register reuse across serial broadcast ops (#1787) 33a824d Adding sibling path for MaxInfoSpanningTree (#1776) 86f46aa Fix div(Val, TensorView) (#1778) d3de227 Fix FusionMaxRootDomainInfoSpanningTreePrintTwice_CUDA (#1781) ecc7a87 Extend mma dimension and layout checking to support strided batched matmul and tensor contractions (#1761) ``` RUN_TORCHBENCH: nvfuser ghstack-source-id: a74f653 Pull Request resolved: #81861

Syncing nvfuser devel branch to upstream master. https://github.com/csarofeen/pytorch/ Code changes includes: - codegen improvements: 1. Indexing refactor -> Remove reference tensor in predicate indexing logic 2. MMA Rfactor support for cross-warp and cross-CTA split on K dimension 3. Grouping grid allreduces across iterations 4. Swizzle op formulation for non-affine swizzles 5. Use scheduler_utils to cache inputs and outputs in schedulePointwise - scheduler refactor 1. New compute at interface - transformation propagation refactor on MaxInfoSpanningTree 1. Added sibling path that is required to generate consistent replay for some cases where `MaxInfoSpanningTree` is used with a selector. 2. Optimization to skip Transform propagator 3. SpanningTreePrinter for debugging - parser update 1. Fixes `div` 2. Added `_to_copy` 3. Broadcast in dim with expand to support expanding to concrete size 4. Dropout prob extremal patch - executor patch on caching strides for output allocation Squashed commits to WAR github API Commits that's actually in this PR from the devel branch: ``` 3b87896 Fix allocation of work buffers and `fused_reduction::ParallelReduce` with unswitch (#1818) 4cae122 schedulePointwise cleanup: - computeAt + InlinePropagator (#1815) 3df9742 Use scheduler_utils to cache inputs and outputs in schedulePointwise (#1811) 03180aa improve broadcast resolution (#1792) bee6c69 bug fix (#1819) 4413c8f Support PYTORCH_NVFUSER_DUMP=transform_propagator (#1812) de6b7ca Fix negative position in InlinePropagator (#1813) 10a996c Remove redundant check in schedulePointwise (#1810) acd5ed4 Swizzle op formulation for non-affine swizzles (#1441) 3ed8330 Kernel args patch to show zero_init buffer (#1809) 037a75a Dropout prob extremal patch (#1804) 282c429 spam nvrtc options (#1783) 3ba6a5f Broadcast in dim with expand (#1794) fd4be12 remove dead indexing code (#1806) fa4e6a4 Check siblings in getMaxPosAll (#1805) 025c840 Grouping grid allreduces across iterations (#1755) 37c579e Temporarily disable test requring large shared memory. (#1802) 5f375d0 More cleanup on InlinePropagator (#1800) 8d384da Indexing refactor stage 2 : Remove reference tensor in predicate indexing logic (#1784) f008140 MMA Rfactor support for cross-warp and cross-CTA split on K dimension (#1554) 76b3cca Add parsing support for `_to_copy` to handle AMP casts. (#1756) ef04f6c Coding style cleanups (#1798) 38c7f3c InlinePropagator please don't replay (#1797) 3f2c263 validateDomain in TransformPropagator (#1796) c077085 Use TransformPropagatorWithCheck in many tests (#1795) d0d0908 Some further cleanup for the new computeAt interface (#1793) 45f5203 Fix TransformReplay::getMatchedLeafPosWithoutReplay* (#1791) 28cbaf9 New compute at interface (#1743) 635ebfc Add SpanningTreePrinter (#1786) 59f3c32 Output allocate patch (#1790) fe93bf5 Transform propagator skip replay when possible (#1782) ebf23a5 Fix isIntegralType error msg (#1789) 0c82ecf Disable register reuse across serial broadcast ops (#1787) 33a824d Adding sibling path for MaxInfoSpanningTree (#1776) 86f46aa Fix div(Val, TensorView) (#1778) d3de227 Fix FusionMaxRootDomainInfoSpanningTreePrintTwice_CUDA (#1781) ecc7a87 Extend mma dimension and layout checking to support strided batched matmul and tensor contractions (#1761) ``` RUN_TORCHBENCH: nvfuser Differential Revision: [D38043938](https://our.internmc.facebook.com/intern/diff/D38043938) Pull Request resolved: #81861 Approved by: https://github.com/davidberard98

Summary: Pull Request resolved: #81861 Syncing nvfuser devel branch to upstream master. https://github.com/csarofeen/pytorch/ Code changes includes: - codegen improvements: 1. Indexing refactor -> Remove reference tensor in predicate indexing logic 2. MMA Rfactor support for cross-warp and cross-CTA split on K dimension 3. Grouping grid allreduces across iterations 4. Swizzle op formulation for non-affine swizzles 5. Use scheduler_utils to cache inputs and outputs in schedulePointwise - scheduler refactor 1. New compute at interface - transformation propagation refactor on MaxInfoSpanningTree 1. Added sibling path that is required to generate consistent replay for some cases where `MaxInfoSpanningTree` is used with a selector. 2. Optimization to skip Transform propagator 3. SpanningTreePrinter for debugging - parser update 1. Fixes `div` 2. Added `_to_copy` 3. Broadcast in dim with expand to support expanding to concrete size 4. Dropout prob extremal patch - executor patch on caching strides for output allocation Squashed commits to WAR github API Commits that's actually in this PR from the devel branch: ``` 3b87896 Fix allocation of work buffers and `fused_reduction::ParallelReduce` with unswitch (#1818) 4cae122 schedulePointwise cleanup: - computeAt + InlinePropagator (#1815) 3df9742 Use scheduler_utils to cache inputs and outputs in schedulePointwise (#1811) 03180aa improve broadcast resolution (#1792) bee6c69 bug fix (#1819) 4413c8f Support PYTORCH_NVFUSER_DUMP=transform_propagator (#1812) de6b7ca Fix negative position in InlinePropagator (#1813) 10a996c Remove redundant check in schedulePointwise (#1810) acd5ed4 Swizzle op formulation for non-affine swizzles (#1441) 3ed8330 Kernel args patch to show zero_init buffer (#1809) 037a75a Dropout prob extremal patch (#1804) 282c429 spam nvrtc options (#1783) 3ba6a5f Broadcast in dim with expand (#1794) fd4be12 remove dead indexing code (#1806) fa4e6a4 Check siblings in getMaxPosAll (#1805) 025c840 Grouping grid allreduces across iterations (#1755) 37c579e Temporarily disable test requring large shared memory. (#1802) 5f375d0 More cleanup on InlinePropagator (#1800) 8d384da Indexing refactor stage 2 : Remove reference tensor in predicate indexing logic (#1784) f008140 MMA Rfactor support for cross-warp and cross-CTA split on K dimension (#1554) 76b3cca Add parsing support for `_to_copy` to handle AMP casts. (#1756) ef04f6c Coding style cleanups (#1798) 38c7f3c InlinePropagator please don't replay (#1797) 3f2c263 validateDomain in TransformPropagator (#1796) c077085 Use TransformPropagatorWithCheck in many tests (#1795) d0d0908 Some further cleanup for the new computeAt interface (#1793) 45f5203 Fix TransformReplay::getMatchedLeafPosWithoutReplay* (#1791) 28cbaf9 New compute at interface (#1743) 635ebfc Add SpanningTreePrinter (#1786) 59f3c32 Output allocate patch (#1790) fe93bf5 Transform propagator skip replay when possible (#1782) ebf23a5 Fix isIntegralType error msg (#1789) 0c82ecf Disable register reuse across serial broadcast ops (#1787) 33a824d Adding sibling path for MaxInfoSpanningTree (#1776) 86f46aa Fix div(Val, TensorView) (#1778) d3de227 Fix FusionMaxRootDomainInfoSpanningTreePrintTwice_CUDA (#1781) ecc7a87 Extend mma dimension and layout checking to support strided batched matmul and tensor contractions (#1761) ``` RUN_TORCHBENCH: nvfuser Test Plan: Imported from OSS Reviewed By: samdow Differential Revision: D38043938 Pulled By: davidberard98 fbshipit-source-id: b94245f83dab6faee31e0c154d3b969bddeb3d47

* Enable tests previously disabled due to an aliasing bug The bug was fixed by pytorch#1792

alykhantejani added 7 commits June 5, 2017 12:10

added checks to cudnn Convolution for stride, dilation, kernel size a…

7efd8d7

…nd num input planes

add convolution_shape_checks for cudnn convs

8abca10

torch.manual_seed now seeds all cuda devices too

cb69016

Merge branch 'master' into torch_manual_seed_to_seed_cuda_devices

67a762a

Merge branch 'master' of https://github.com/pytorch/pytorch into add_…

8678147

…stable_BCELoss

added stubs

9aabde9

added BCEWithLogitsLoss

45d47cc

alykhantejani closed this Jun 13, 2017

alykhantejani added 7 commits June 13, 2017 18:43

added checks to cudnn Convolution for stride, dilation, kernel size a…

0978510

…nd num input planes

add convolution_shape_checks for cudnn convs

cd93799

added stubs

08bd291

added BCEWithLogitsLoss

baf77d4

add stable BCEWithLogitsLoss and add new_criterion_tests for nn

38deb87

add stable BCEWithLogitsLoss and add new_criterion_tests for nn

a496fc6

remove tests from legacy_nn

ded437b

alykhantejani reopened this Jun 13, 2017

alykhantejani added 3 commits June 14, 2017 09:54

Merge branch 'master' into add_stable_BCELoss

96bbc5d

merge upstream

1ee3129

merge upstream

189b32f

soumith approved these changes Jun 20, 2017

View reviewed changes

soumith merged commit 67968cb into pytorch:master Jun 20, 2017

alykhantejani deleted the add_stable_BCELoss branch June 27, 2017 16:00

houseroad mentioned this pull request Feb 6, 2019

Automatic update of fbcode/onnx to 822d8df0a2a32233c6022f50a158817a0f19bdc7 #16791

Closed

ezyang added the open source label Jun 24, 2019

ddehueck mentioned this pull request May 15, 2020

Is bce_w_logits equivalent to the pytorch BCEWithLogitsLoss? ddehueck/jax-loss#4

Closed

jjsjann123 pushed a commit to jjsjann123/pytorch that referenced this pull request Jul 15, 2022

improve broadcast resolution (pytorch#1792)

03180aa

jjsjann123 pushed a commit to jjsjann123/pytorch that referenced this pull request Oct 4, 2022

Enable tests previously disabled due to an aliasing bug (pytorch#2005)

45045cd

* Enable tests previously disabled due to an aliasing bug The bug was fixed by pytorch#1792

Add numerically stable BCELoss which takes logits as input #1792

Add numerically stable BCELoss which takes logits as input #1792

Uh oh!

Conversation

alykhantejani commented Jun 13, 2017

Uh oh!

soumith commented Jun 20, 2017

Uh oh!

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

Uh oh!

Uh oh!

alykhantejani commented Jun 27, 2017

Uh oh!

esube commented Jun 27, 2017 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

alykhantejani commented Jun 27, 2017

Uh oh!

esube commented Jun 27, 2017 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

alykhantejani commented Jun 27, 2017

Uh oh!

esube commented Jun 27, 2017 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

alykhantejani commented Jul 24, 2017

Uh oh!

F0REacH commented Jul 26, 2017

Uh oh!

Jiaming-Liu commented Jul 26, 2017 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

alykhantejani commented Jul 26, 2017

Uh oh!

alykhantejani commented Jul 26, 2017

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

7 participants

esube commented Jun 26, 2017 •

edited

Loading

esube commented Jun 27, 2017 •

edited

Loading

esube commented Jun 27, 2017 •

edited

Loading

esube commented Jun 27, 2017 •

edited

Loading

Jiaming-Liu commented Jul 26, 2017 •

edited

Loading