Add IoU3D as a custom c++ op (CPU) #890

ianstenbit · 2022-10-07T01:17:30Z

No description provided.

ianstenbit · 2022-10-07T02:47:25Z

/gcbrun

ianstenbit · 2022-10-07T05:17:28Z

/gcbrun

ianstenbit · 2022-10-07T05:37:01Z

/gcbrun

ianstenbit · 2022-10-07T05:50:26Z

/gcbrun

ianstenbit · 2022-10-07T05:58:43Z

This is now working with CI across all of our platforms, and I've also manually tested this on MacOS.

Some todos are:

Move custom op away from IoU3D and into some _internal object (I think it's worth getting this committed even before we have a real op so that I can set up the internal build configuration necessary). Probably a ZeroOutTestLayer is appropriate.
Update README to show how to build from source
Update contributing guide to explain that building from source will be required to run custom op tests.

ianstenbit · 2022-10-07T06:04:44Z

@tanzhenyu

bhack · 2022-10-07T10:18:22Z

Are we sure that we want to go in a direction of limited HW compatibility for keras libraries considering also all the overhead of maintaining custom ops?

I supposed that the scope of keras-* was to have a better interaction with the compiler team and stress test the technology but to still express our needs through the TF/*HLO composability.

But I must also admit that recently, also on the "modern" and (M)HLO native JAX, something like custom ops is emerging again with google/jax#12632 (XLA's CustomCall)

/Cc @andrew-leaver @jpienaar @paynecl

bhack · 2022-10-07T13:34:03Z

keras_cv/losses/iou_3d_loss.py

@@ -0,0 +1,23 @@
+# Copyright 2022 The KerasCV Authors. All Rights Reserved.


What are the Google TF3d non custom-op performances of their 3d losses?
https://github.com/google-research/google-research/blob/master/tf3d/losses/box_prediction_losses.py

bhack · 2022-10-07T13:51:30Z

From https://github.com/google/jax/pull/12632/files#diff-b8191ca86b51eea58dc118d756de46521d77321338163145aea6b1ef64d126c5R77

JAX uses XLA to compile staged-out Python programs, which are represented with MHLO. MHLO offers common operations used in machine learning, such as the dot-product, exponential function, and so on. Unfortunately, sometimes these operations (and compositions of these operations) are not sufficiently expressive or performant and users look to external libraries (Triton, Numba) or hand-written code (CUDA, C/C++) to implement operations for their desired program.

So are we already here with Keras-* as we was at the same point for many years with TFAddons with the Custom ops proliferation (often with 1K lines of code x PR for components that required custom ops with CPU or CUDA only support)? /cc @seanpmorgan

ianstenbit · 2022-10-07T15:22:49Z

Hey @bhack -- thanks for the comments

This is for an existing internal use case which will require this custom op (the TF3D losses are insufficient) as well as a few other custom ops for 3D spatial data.

I also worry about package accessibility and I want to make sure that we package this in a way which does not add user friction for pip install and lets users who don't want the custom ops install without binaries. I'm not certain what the right solution to that is just yet.

Please let me know if you have any suggestions.

bhack · 2022-10-07T15:44:40Z

I also worry about package accessibility and I want to make sure that we package this in a way which does not add user friction for pip install and lets users who don't want the custom ops install without binaries. I'm not certain what the right solution to that is just yet.

I could understand that there was no time to address this subject and now, also on your side, you need just to "deliver" but I have tried to discuss this topic since June 2021 and earlier.

What will we do, once you have introduced the c++ infrastructure in the project, when contributors will start to push for their own custom kernels for performance, uncovered vectorized maps and the underline API design of native TF ops or for XLA bridges uncovered ops in TF openxla/stablehlo#216 (comment))?

Are you going to maintain a proliferation of user kernels excluding probably some HW vendor like e.g. AMD or Google CLOUD TPU?
And what about exporting to TFlite when you are going to do this soon or later on an inference component instead of a loss?

So you could also solve something urgent, for your internal delivery with this, but after 1 year and half I still not understand what the strategy is.

ianstenbit · 2022-10-07T17:48:03Z

@bhack I've updated the README and contributing guide to reflect a posture that we do not intend to invite user-contributed custom ops at this time, and that we only intend to include training-time CPU ops.

ianstenbit · 2022-10-07T17:48:25Z

/gcbrun

tanzhenyu

This is great! Thanks for pushing it forward.

.github/CONTRIBUTING.md

tanzhenyu · 2022-10-18T17:26:35Z

build_deps/build_pip_pkg.sh

+set -x
+
+PLATFORM="$(uname -s | tr 'A-Z' 'a-z')"
+function is_windows() {


Some high level question on this:

does this build script follow some other scripts? It would be nice to point out the reference so users can follow a pattern

do we really want to support windows and mac builds? This will put more burden on maintainence and doesn't give immediate benefit.

This is adapted from TFA. I'm adding a link to the original script and a brief blurb about what this does.

We don't necessarily need to ship binaries for Windows and Mac OS, but including this makes it a lot easier for people to install from source it they want. I'm not sure if it makes sense to offer pre-built wheels for these environments, as I agree that it does add maintenance cost for not a ton of benefit. But it doesn't cost much to support it in the setup script IMO

I suppose that these things are ported from Addons:
https://github.com/tensorflow/addons/blob/master/build_deps/build_pip_pkg.sh

The pain is to support those if TF drops pre-built wheels for them

tanzhenyu · 2022-10-18T17:27:28Z

build_deps/configure.py

+
+
+# Writes variables to bazelrc file
+def write(line):


Same question with build_pip_pkg.sh

Added a little doc blurb at the top -- is there anything else specifically you're looking for here?

tanzhenyu · 2022-10-18T17:29:35Z

build_deps/tf_dependency/BUILD.tpl

+
+cc_library(
+    name = "tf_header_lib",
+    hdrs = [":tf_header_include"],


the header file inclusion and libtensorflow_framework inclusion seems correct to me.
Though it'd be better to point out where we're relying this from. The main reason for this ask is that "tensorflow_framework" might not be a guaranteed binary in the long run, and we should know how to fix this once that becomes true.

I'm not sure I follow -- do you mean comments specifying that this is enabling linking to the installed TF version, or something else?

cloudbuild/README.md

tanzhenyu · 2022-10-18T17:32:34Z

keras_cv/custom_ops/__init__.py

+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.


I see a _pariwise_iou_op.so being included,

should we just have a single custom_op.so that includes all the targets?

if a user wants to change some ops implementation, maybe we should have a documentation on how to do that?

I'm not sure what the right answer is. I figured that for now we can have a specific .so object for this op, and when we add further ops we can decide whether to merge them into a single binary or keep them split. I've seen packages take both approaches, and they both seem reasonable to me (e.g. TFA uses separate binaries but lingvo uses one)

This PR includes docs for installing from source, which should cover the biggest difficulty of developing custom ops. I think we probably want to generally discourage contribution of custom ops, but if you think this is valuable I can write a blurb in CONTRIBUTING.md about the workflow of working on these ops.

yeah please do so, my main ask here is how users can build and update the *.so file once they change the implementation.
having a single .so object seems better to me, even tensorflow itself only contains very few .so.

Gotcha -- that is already covered in Tests that require custom ops in CONTRIBUTING.md. I've updated that file to refer to the instructions for building the binaries in the Contributing Custom Ops section

I've also renamed the .so to _keras_cv_custom_ops.so to indicate that later ops should be included in this same binary.

Sounds good!

tanzhenyu · 2022-10-18T17:34:29Z

keras_cv/losses/iou_3d_loss.py

+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ============================================================================
+"""IoU3D loss using a custom TF op."""


Hmm...this is indeed very interesting...

keras_cv/losses/iou_3d_loss_test.py

bhack

It could be nice if you could adapt our .devcontainer for C++

You could port copy the c++ entries from:
https://github.com/tensorflow/addons/blob/master/.devcontainer/devcontainer.json

…ute only major axis of pairwise IoU matrix

ianstenbit · 2022-10-18T21:36:04Z

/gcbrun

ianstenbit · 2022-10-18T22:24:54Z

/gcbrun

bhack · 2022-10-19T13:16:44Z

keras_cv/metrics/iou_3d.py

+    Note that this is implemented using a custom TensorFlow op. Initializing an
+    IoU3D object will attempt to load the binary for that op.
+
+    Boxes should have the format [center_x, center_y, center_z, dimension_x,


Can we document a little bit the rational of this format choice in our python API?

I don't have stats on the popularity in the datasets, but If it is not the most popular native format we need to extend our current 2d bbox converters to 3d.

E.g. pytorch3d # Assume inputs: boxes1 (M, 8, 3) and boxes2 (N, 8, 3)

Please, check also the unified format in one of the largest 3d bbox dataset:
https://github.com/facebookresearch/omni3d/blob/main/DATA.md#data-usage

I think it's likely that we'll want to add converters in the future. I've opened #947 for this and we can address it once we have a minimal 3D bbox pipeline working.

tanzhenyu

LGTM, one minor comment

keras_cv/metrics/iou_3d.py

tanzhenyu · 2022-10-20T14:27:40Z

Also I'm leaning towards only supporting Linux for now until we know better

ianstenbit · 2022-10-21T14:42:08Z

Also I'm leaning towards only supporting Linux for now until we know better

sgtm. I will leave the config code in for now for all platforms, but when we go to build our wheels we can build linux only for now (and other platforms can have a Python-only wheel)

ianstenbit · 2022-10-21T15:29:57Z

/gcbrun

ianstenbit · 2022-10-21T16:40:54Z

/gcbrun

ianstenbit · 2022-10-21T17:17:14Z

/gcbrun

* Initial custom ops testing * Switch to cpp14 * Add include_package_data flag * cpp17 * Configure CI tests * Formatting * Call configure before running bazel build for CI * Verbose failures for CI/Bazel * Switch to tf-cpu for test flow * Use build dep impls from TFA * Update cloudbuild workflow to build binaries from source * Update cloudbuild docks (updated docker image) * Update GPU tests to TF 2.10 * Update copyright and docstrings * Move loss to internal, update README and CONTRIBUTING.md * Update to IoU op ported from Lingvo * Remove cuda config * Make iou3d a real keras loss * Skip custom-ops tests in devcontainer * Review comments * Remove skipping of IoU test from devcontainer (it is now opt-in) * Fix environment variable for windows * Update environment variable setting * Use env in yaml for environment variable * Review comments * Make IoU a metric, not a loss. Also update Cpp implementation to compute only major axis of pairwise IoU matrix * typo * Make test case more demonstrative * Switch to a single binary, remove binary from git * Move IoU3D from metrics to ops * Switch to full matrix computation of iou * Formatting

ianstenbit added 11 commits October 6, 2022 17:47

Initial custom ops testing

6bcac82

Switch to cpp14

87bcb8e

Add include_package_data flag

2c87da0

cpp17

3112a03

Configure CI tests

3b92be3

Formatting

b26f9d7

Call configure before running bazel build for CI

99b3b92

Verbose failures for CI/Bazel

da958cf

Switch to tf-cpu for test flow

734f74d

Use build dep impls from TFA

a25796d

Update cloudbuild workflow to build binaries from source

2143176

Update cloudbuild docks (updated docker image)

3b10e4a

Update GPU tests to TF 2.10

3b707b1

Update copyright and docstrings

6085f7a

ianstenbit changed the title ~~[WIP] Example custom c++ op~~ Example custom c++ op Oct 7, 2022

bhack reviewed Oct 7, 2022

View reviewed changes

ianstenbit changed the title ~~Example custom c++ op~~ Example custom c++ op (CPU) Oct 7, 2022

Move loss to internal, update README and CONTRIBUTING.md

4d62a16

tanzhenyu reviewed Oct 18, 2022

View reviewed changes

bhack suggested changes Oct 18, 2022

View reviewed changes

Review comments

982d3ba

ianstenbit requested a review from tanzhenyu October 18, 2022 18:15

ianstenbit added 3 commits October 18, 2022 14:22

Make IoU a metric, not a loss. Also update Cpp implementation to comp…

eb585bc

…ute only major axis of pairwise IoU matrix

typo

c4a7fa6

Make test case more demonstrative

f7f6b73

ianstenbit requested a review from LukeWood October 18, 2022 20:30

Switch to a single binary, remove binary from git

b0b4263

bhack reviewed Oct 19, 2022

View reviewed changes

bhack mentioned this pull request Oct 19, 2022

IOU3d and IOU3d pairwise tensorflow/tensorflow#58166

Open

ianstenbit changed the title ~~Add IoU3DLoss with a custom c++ op (CPU)~~ Add IoU3D metric with a custom c++ op (CPU) Oct 19, 2022

tanzhenyu approved these changes Oct 20, 2022

View reviewed changes

keras_cv/metrics/iou_3d.py Outdated Show resolved Hide resolved

Move IoU3D from metrics to ops

131872a

Switch to full matrix computation of iou

2b138d9

Formatting

8a54df4

ianstenbit changed the title ~~Add IoU3D metric with a custom c++ op (CPU)~~ Add IoU3D as a custom c++ op (CPU) Oct 21, 2022

LukeWood approved these changes Oct 21, 2022

View reviewed changes

ianstenbit merged commit 3d6467e into keras-team:master Oct 21, 2022

ianstenbit deleted the bin branch October 21, 2022 18:05

bhack mentioned this pull request Nov 15, 2022

Reorganize [ops] and [custom_ops] #1010

Closed

seanpmorgan mentioned this pull request Feb 10, 2023

TensorFlow Addons Wind Down tensorflow/addons#2807

Open

		@@ -0,0 +1,23 @@
		# Copyright 2022 The KerasCV Authors. All Rights Reserved.

Add IoU3D as a custom c++ op (CPU) #890

Add IoU3D as a custom c++ op (CPU) #890

Conversation

ianstenbit commented Oct 7, 2022

ianstenbit commented Oct 7, 2022

ianstenbit commented Oct 7, 2022

ianstenbit commented Oct 7, 2022

ianstenbit commented Oct 7, 2022

ianstenbit commented Oct 7, 2022

ianstenbit commented Oct 7, 2022

bhack commented Oct 7, 2022 • edited Loading

Choose a reason for hiding this comment

bhack commented Oct 7, 2022 • edited Loading

ianstenbit commented Oct 7, 2022

bhack commented Oct 7, 2022 • edited Loading

ianstenbit commented Oct 7, 2022

ianstenbit commented Oct 7, 2022

tanzhenyu left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

bhack left a comment

Choose a reason for hiding this comment

ianstenbit commented Oct 18, 2022

ianstenbit commented Oct 18, 2022

bhack Oct 19, 2022 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

tanzhenyu left a comment

Choose a reason for hiding this comment

tanzhenyu commented Oct 20, 2022

ianstenbit commented Oct 21, 2022 • edited Loading

ianstenbit commented Oct 21, 2022

ianstenbit commented Oct 21, 2022

ianstenbit commented Oct 21, 2022

bhack commented Oct 7, 2022 •

edited

Loading

bhack commented Oct 7, 2022 •

edited

Loading

bhack commented Oct 7, 2022 •

edited

Loading

bhack Oct 19, 2022 •

edited

Loading

ianstenbit commented Oct 21, 2022 •

edited

Loading