GitLab

Summary:
Pull Request resolved: https://github.com/facebookresearch/d2go/pull/409

`assert_close` is preferred over `assert_allclose`: https://github.com/pytorch/pytorch/issues/61844

The `assert_allclose` was removed yesterday in https://github.com/pytorch/pytorch/pull/87974, causing test to fail, eg. https://github.com/facebookresearch/d2go/actions/runs/3389194553/jobs/5632021291

Reviewed By: sstsai-adl

Differential Revision: D41000306

fbshipit-source-id: 7bd1cb9d5edf0a4609a909e2283df411bcabdf13

Summary:
Pull Request resolved: https://github.com/facebookresearch/d2go/pull/408

Update the DPE training script in D2 (https://github.com/facebookresearch/d2go/commit/87374efb134e539090e0b5c476809dc35bf6aedb)Go to support Turing DPE QAT.

Reviewed By: newstzpz

Differential Revision: D40612406

fbshipit-source-id: 9379e4be248045b995293c5a522bab05e0b13c6e

Summary:
Pull Request resolved: https://github.com/facebookresearch/d2go/pull/407

As is, observers can be periodically updated even if observer is disabled. Add a check on whether observer is disabled before updating observers periodically.

Reviewed By: tglik

Differential Revision: D40734564

fbshipit-source-id: fd3e50e95c95ad39bd993f71e02dc7acc79744cc

Summary:
Pull Request resolved: https://github.com/facebookresearch/d2go/pull/403

`cfg.SOLVER.AMP.ENABLED` enabled mixed precision, but this only works for V100 GPUs.
For A100s, the equivalent is to enable TF32.

Reviewed By: tglik

Differential Revision: D40675242

fbshipit-source-id: 5cc3d12cd3d7ec76665e0907ecc87fc5f64d73f0

Summary:
Pull Request resolved: https://github.com/facebookresearch/d2go/pull/404

`get_default_cfg` is now class method since stack of D37294926 (https://github.com/facebookresearch/d2go/commit/b077a2c13845d4ef8481979d64345368864fe5ff), this diff updates call sites using biggrep to replace "Runner().get_default_cfg" with "Runner.get_default_cfg"

Reviewed By: itomatik

Differential Revision: D40707898

fbshipit-source-id: 2b56545769d930d34dad8814d5bfeba4c54224fd

Summary:
Pull Request resolved: https://github.com/facebookresearch/d2go/pull/392

1. Moved scale adjustment to a separate function and expose the option to disable it
2. Add option to keep the original image instead of creating a square image

Reviewed By: wat3rBro

Differential Revision: D40403705

fbshipit-source-id: 6c35a9a1fe3ef868e5f0b2204874fd028776e26a

Summary:
Pull Request resolved: https://github.com/facebookresearch/d2go/pull/401

as followup on D40001329 (https://github.com/facebookresearch/d2go/commit/69bf820c64cd0ffb6a84f465199c9134814cf58e).
The export is running main func without launching distributed workers, so it need to set the shared context explicitly.

Reviewed By: wat3rBro

Differential Revision: D40708631

fbshipit-source-id: 7689a45dff383ba2cce01d33d3be95d612269fbe

Summary:
Pull Request resolved: https://github.com/facebookresearch/d2go/pull/402

print out the optimizer for easier debug

Reviewed By: newstzpz

Differential Revision: D40701959

fbshipit-source-id: 7b610e8f5771409632ae056cb9d34138b331adbc

Summary:
Pull Request resolved: https://github.com/facebookresearch/d2go/pull/399

Freezing the model before running quantization causes an issue with loading a saved checkpoint bc fusing does not support FrozenBatchNorm2d (which means that the checkpoint could have a fused weight conv.bn.weight whereas the model would have an unfused weight bn.weight). The longer term solution is to add FrozenBatchNorm2d to the fusing support but there are some subtle issues there that will take some time to fix:
* need to move FrozenBatchNorm2d out of D2 (https://github.com/facebookresearch/d2go/commit/87374efb134e539090e0b5c476809dc35bf6aedb) and into mobile_cv lib
* current fuser has options to add new bn ops (e.g., FrozenBatchNorm2d) which we use with ops like SyncBN but this currently is only tested with inference so we need to write some additional checks on training

The swap will make freezing compatible with QAT and should still work with standard models. One subtle potential issue is that the current BN swap assumes that BN is a leaf node. If a user runs QAT without fusing BN, the BN will no longer be the leaf node as it will obtain an activation_post_process module in order to record the output. The result is that BN will not be frozen in this specific instance. This should not occur as BN is usually fused. A small adjustment to the BN swap would just be to swap the BN regardless of whether it is a leaf node (but we have to check whether activation_post_process module is retained). Another long term consideration is moving both freezing and quant to modeling hooks so the user can decide the order.

Reviewed By: wat3rBro

Differential Revision: D40496052

fbshipit-source-id: 0d7e467b833821f7952cd2fce459ae1f76e1fa3b

Summary:
Pull Request resolved: https://github.com/facebookresearch/d2go/pull/400

Before the diff, the bootstrap can't handle the following gramma:

https://www.internalfb.com/code/fbsource/[010f09214704]/fbcode/mobile-vision/d2go/tests/modeling/test_modeling_distillation.py?lines=231-236

The fix is recursively applying the truncating trick.

Reviewed By: itomatik

Differential Revision: D40701375

fbshipit-source-id: 946b6be47aa4b879e2e247b879a0d8b9ac13822b

Summary: Pull Request resolved: https://github.com/facebookresearch/d2go/pull/389

Reviewed By: itomatik

Differential Revision: D39631903

fbshipit-source-id: 1668a8b06260d02b40208b3dda3cbade0a12bc16

Summary:
Pull Request resolved: https://github.com/facebookresearch/d2go/pull/393

Add gaussian blur augmentation.

Reviewed By: tglik

Differential Revision: D40404772

fbshipit-source-id: d04774cc8aa9dff00f2b85e9c7feb1b8709edc9e

Summary:
Pull Request resolved: https://github.com/facebookresearch/d2go/pull/391

1. Add unit test for affine augmentation
2. Fix off by one affine scaling (Note: this changes augmentation behavior)

Reviewed By: tglik

Differential Revision: D40374538

fbshipit-source-id: ea037195b9a7dc1b4e254bf35216a8dac610bf29

Summary:
Pull Request resolved: https://github.com/facebookresearch/d2go/pull/395

Having a module as instance variable prevents multiprocessing from spawning a children using Pickle and creates a deadlock, making Motion Blur incompatible with multiprocessing

Reviewed By: wat3rBro

Differential Revision:
D40411481

LaMa Project: L1082110

fbshipit-source-id: b3e44b438367100044b948b5ea616c5c6fd41d3d

Summary:
X-link: https://github.com/facebookresearch/mobile-vision/pull/116

Pull Request resolved: https://github.com/facebookresearch/d2go/pull/398

D2 (https://github.com/facebookresearch/d2go/commit/87374efb134e539090e0b5c476809dc35bf6aedb)Go doesn't have per node initialization api, but only per worker initialization that happens per subprocess.
Some projects (like IOBT) need to way to do shared initialization before spawning all the workers in subprocess and pass this initialized shared context to the workers.
This diff adds API to create a shared context object before launching workers and then use this shared context by the runners inside the workers after launch.

Reviewed By: wat3rBro

Differential Revision: D40001329

fbshipit-source-id: 231a4e7e4da7b5db50849176c58b104c4565306a

Summary: Pull Request resolved: https://github.com/facebookresearch/d2go/pull/394

Reviewed By: wrlife

Differential Revision: D40533013

fbshipit-source-id: c4c0b08b8afb0c5c622a945bd2ef9c3e682f3039

Summary:
Pull Request resolved: https://github.com/facebookresearch/d2go/pull/384

Random resize augmentation. Randomly pick a scale from the shape_list and resize to that scale.

  D2 (https://github.com/facebookresearch/d2go/commit/87374efb134e539090e0b5c476809dc35bf6aedb)GO_DATA:
    AUG_OPS:
      TRAIN: [
        RandomResizeOp::{"shape_list": [[224, 224], [256, 256], [320, 320]]}
      ]

Reviewed By: XiaoliangDai

Differential Revision: D40230332

fbshipit-source-id: 60a48f85240aef673033d48db4662899dc90bef4

Summary:
Pull Request resolved: https://github.com/facebookresearch/d2go/pull/386

Pull Request resolved: https://github.com/facebookresearch/d2go/pull/382

Pull Request resolved: https://github.com/facebookresearch/d2go/pull/364

As title,
when `QUANTIZATION.BACKEND` is set to `turing`, call `odai.transforms` APIs instead of OSS quantization.

Also updated `image_classification` as an example to get Turing quantized model via `custom_prepare_fx/custom_convert_fx`

Reviewed By: wat3rBro

Differential Revision: D40282390

fbshipit-source-id: 7d6509969cfe8537153e1d59f21967eeb7801fd1

Summary:
Pull Request resolved: https://github.com/facebookresearch/d2go/pull/383

Our optimizer tests have become flaky due to often timing out:
* https://www.internalfb.com/intern/test/281475048520501?ref_report_id=0
* https://www.internalfb.com/intern/test/281475048520502?ref_report_id=0

{F778290241}{F778290240}

This diff splits tests that run multiple optimizer through training.
Also reduced number of iterations and number of datapoints for evaluation. At the moment we aren't really verifying end result value, so I assume this reduction shouldn't matter.

Reviewed By: tglik

Differential Revision: D40124949

fbshipit-source-id: 5d8f309106dd5f1829f291784d36768dab2e9eca

Summary:
X-link: https://github.com/facebookresearch/mobile-vision/pull/110

Pull Request resolved: https://github.com/facebookresearch/d2go/pull/380

- remove `alias`
- only annotate different implementation with `fb_overwrite`
- fix lint

Reviewed By: itomatik

Differential Revision: D39981383

fbshipit-source-id: 9739b7026510b3f1a2e69fe1de5b3f721759a209

Summary:
Pull Request resolved: https://github.com/facebookresearch/d2go/pull/381

Introduce extra parameter SOLVER.AMP.PRECISION which can be sued to control the mixed precision training when lightning  backend is used.

Previous value `precision: "mixed"` was worng and the training failed (See screenshot below)
{F777576618}

I had to make AMP.PRECISION as string and make sure that it can work with two values: "float16" and "bfloat16". Before feeding it to the Trainer we convert "float16" string to integer value 16. Such a workaround was unavoidable because D2 (https://github.com/facebookresearch/d2go/commit/87374efb134e539090e0b5c476809dc35bf6aedb)Go's config value cannot be of int and str at the same time.

Reviewed By: wat3rBro

Differential Revision: D40035367

fbshipit-source-id: ed4f615ab29a2258164cbe179a9adba11559d804

Summary:
Pull Request resolved: https://github.com/facebookresearch/d2go/pull/378

Some hooks need access to cfg to be initialized correctly. Pass cfg down the hook registration method.

Reviewed By: ertrue, miqueljubert

Differential Revision: D39303862

fbshipit-source-id: 931c356c7045f95fc0af5b20c7782ea4d1aff138

Summary:
Pull Request resolved: https://github.com/facebookresearch/d2go/pull/376

Make sure the return result is consistent for ResultCache
* When gather = True, it will always return a list.

Reviewed By: itomatik

Differential Revision: D39874274

fbshipit-source-id: ce042261432cdce5544d10b2d4b88f5e2d0d1b68

Summary:
Pull Request resolved: https://github.com/facebookresearch/d2go/pull/379

This diff ~~~prototypes~~~ implements replacing the `custom_convert_fx` API with a callback.

Reviewed By: LiamZhuuu

Differential Revision: D39859228

fbshipit-source-id: 34719d1758c4afa7e47930c12d3443813d3f4546

Summary:
Pull Request resolved: https://github.com/facebookresearch/d2go/pull/377

Only automatically rescale the lr for sgd optimizers.
* Seems that only sgd needs scaling the lr, so we change to not to scale lr automatically by default. This will work better for newly added optimizers (like adam).

Reviewed By: itomatik, lg-zhang

Differential Revision: D39899434

fbshipit-source-id: d6eebc5b07d4489b401c1fc3cea00f5a060fe19d

Summary:
Pull Request resolved: https://github.com/facebookresearch/d2go/pull/375

`patch_d2_meta_arch` had already been removed in D37246483 (https://github.com/facebookresearch/d2go/commit/d1518ff66205089115df56010e8eafcf94efb08d), need to also update the `beginner.ipynb`.

Reviewed By: itomatik

Differential Revision: D39861148

fbshipit-source-id: bac80efbff1f99a023d604a66c3667bc94f8c6f4

Summary:
Pull Request resolved: https://github.com/facebookresearch/d2go/pull/374

AMP trained with mixed precision is implemented for the Native d2go Runner, but not for Lightning Tasks.

Now we pass params SOLVER.AMP* and SOLVER.CLIP_GRADIENTS* to the lightning Trainer as well.

Reviewed By: wat3rBro

Differential Revision: D39798007

fbshipit-source-id: e48560a91d37c21c56d953eed141876d8c759329

Summary:
Pull Request resolved: https://github.com/facebookresearch/d2go/pull/371

In a previous iteration of this diff, we were specifying the teacher model in the same config as the student model, something like:
```
# config.py
MODEL:
  FBNET_V2:
  ...
DISTILLATION:
  TEACHER:
    MODEL:
      FBNET_V2:
      ...
      WEIGHTS: /path/to/teacher/weights
...
```

This leads to some oddities in the code, like we have to have a default config that adds all the required keys in the distillation teacher model.

In this diff, we just let the user supply a teacher config (and optionally runner_name and overwrite opts) and use the supplied runner to build the model:
```
# new_config.py
MODEL:
  FBNET_V2:
...
DISTILLATION:
  TEACHER:
    CONFIG_FNAME: /path/to/teacher/config
    RUNNER_NAME:
...
```

This should make it very easy to specify the teacher as the user could potentially just reuse the trained_config generated in d2go.

Reviewed By: newstzpz

Differential Revision: D37640041

fbshipit-source-id: 088a636c96f98279c9a04e32d1674f703451aec3

Summary:
X-link: https://github.com/facebookresearch/mobile-vision/pull/108

Pull Request resolved: https://github.com/facebookresearch/d2go/pull/373

Asserting that methods replaced by fb_overwritable are also annotated with corresponding decorator.

Reviewed By: wat3rBro

Differential Revision: D39674347

fbshipit-source-id: a4cf007419760aa07d929ab1cf819ba54b11b9da

Summary:
This diff is reverting D38436675 (https://github.com/facebookresearch/d2go/commit/d98f74aa467a6576c9905accbf4a2b6279599f9c)
D38436675 (https://github.com/facebookresearch/d2go/commit/d98f74aa467a6576c9905accbf4a2b6279599f9c) has been identified to be causing the following test or build failures:
Tests affected:
- https://www.internalfb.com/intern/test/844425001950025/
- https://www.internalfb.com/intern/test/844425001950027/

Here's the Multisect link:
https://www.internalfb.com/intern/testinfra/multisect/1259258
Here are the tasks that are relevant to this breakage:
T120995919: 51 tests started failing for oncall d2go in the last 2 weeks
We're generating a revert to back out the changes in this diff, please note the backout may land if someone accepts it.

Reviewed By: wat3rBro

Differential Revision: D39594147

fbshipit-source-id: 56c489bb9feea2d60a2a5f0e89941ed7c0f3f675

Summary: Pull Request resolved: https://github.com/facebookresearch/d2go/pull/372

Reviewed By: itomatik

Differential Revision: D39628884

fbshipit-source-id: bb1d5d77eeb965dff675c17a8fbc36e4da4e25cd

Summary:
Pull Request resolved: https://github.com/facebookresearch/d2go/pull/369

Pull Request resolved: https://github.com/facebookresearch/d2go/pull/368

Adding image subsampling in evaluation to avoid saving to TB sequential frames for video-like data.

Reviewed By: wat3rBro

Differential Revision: D39233843

fbshipit-source-id: a49bbfe1b9114c11565ed04db1f2f186675ea9ed

Summary:
Pull Request resolved: https://github.com/facebookresearch/d2go/pull/364

As title,
when `QUANTIZATION.BACKEND` is set to `turing`, call `odai.transforms` APIs instead of OSS quantization.

Also updated `image_classification` as an example to get Turing quantized model via `custom_prepare_fx/custom_convert_fx`

Reviewed By: wat3rBro

Differential Revision: D38436675

fbshipit-source-id: e4f0e02290512bce8b18c2369a67ed9b0f116825

Summary:
Pull Request resolved: https://github.com/facebookresearch/d2go/pull/367

EZ

Reviewed By: xiecong

Differential Revision: D39407416

fbshipit-source-id: d0e6fa09ff926780e98c210bfce955e6b8eec7f6

Summary:
Pull Request resolved: https://github.com/facebookresearch/d2go/pull/366

Users often time just use the default value without knowing they need to change  this config, and getting low accuracy after PTQ. Therefore increase the default value.

Reviewed By: miqueljubert

Differential Revision: D39331680

fbshipit-source-id: 6b05773c3c4cd48e6298d341d77a0e373d5439f6

Summary:
Pull Request resolved: https://github.com/facebookresearch/d2go/pull/355

switch to use inference_on_dataset_with_checkpointing in default runner.

Reviewed By: HarounH

Differential Revision: D37215292

fbshipit-source-id: c006784ce0b31700bcbb1f79c303fd791f1561ff

Summary:
Pull Request resolved: https://github.com/facebookresearch/d2go/pull/354

Allow skipping inference when running evaluation.
* `inference_on_dataset_with_checkpointing` works similar to `inference_on_dataset` in d2 but allows skipping the inference step if the evaluator has cached the results.
* If the evaluator has a function `could_skip_process` and returns True, inference will be skipped and only `evaluator. reset()` and `evaluator.evaluate()` are called.

Reviewed By: wat3rBro

Differential Revision: D37213004

fbshipit-source-id: d12cc480589ff04fd8dbb42b22633ab34bc4bf63

Summary:
Pull Request resolved: https://github.com/facebookresearch/d2go/pull/362

X-link: https://github.com/facebookresearch/detectron2/pull/4491

Recently landed D35518556 (https://github.com/facebookresearch/d2go/commit/1ffc801bf5a1d4fe926b815ba93f21632f0980f9) / Github: 36a65a0907d90ed591479b2ebaa8b61cfa0b4ef0 throws an exception with older versions of PyTorch, due to a missing library for import. This has been reported by multiple members of the PyTorch community at https://github.com/facebookresearch/detectron2/commit/36a65a0907d90ed591479b2ebaa8b61cfa0b4ef0

This change uses `try/except` to check for libraries and set flags on presence/absence to later guard code that would use them.

Reviewed By: wat3rBro

Differential Revision: D38879134

fbshipit-source-id: 72f5a7a8d350eb82be87567f006368bf207f5a74

Summary:
Pull Request resolved: https://github.com/facebookresearch/d2go/pull/358

Avoid calling scheduler.step() after the last training iteration is done

Reviewed By: wat3rBro

Differential Revision: D38605135

fbshipit-source-id: 87a55309bf6d1f7e598b567cc2372b00b8885c7c

Enable torch tracing by changing assertions in d2go forwards to allow for torch.fx.proxy.Proxy type.

Summary:
X-link: https://github.com/facebookresearch/detectron2/pull/4227

Pull Request resolved: https://github.com/facebookresearch/d2go/pull/241

Torch FX tracing propagates a type of `torch.fx.proxy.Proxy` through the graph.

Existing type assertions in the d2go code base trigger during torch FX tracing, causing tracing to fail.

This adds a check for FX tracing in progress and  adds a helper function `assert_fx_safe()`, that can be used in place of a standard assertion. This function only applies the assertion if one is not tracing, allowing d2go assertion tests to be compatible with FX tracing.

Reviewed By: wat3rBro

Differential Revision: D35518556

fbshipit-source-id: a9b5d3d580518ca74948544973ae89f8b9de3282