profile
viewpoint
Timo Furrer timofurrer __switzerland__ https://tuxtimo.me Enthusiastic Hacker who loves Free Open Source Software, Linux, Python, Machine Learning and :penguin:. Commits are my own.

gabrielfalcao/sure 653

idiomatic assertion toolkit with human-friendly failure messages, inspired on RSpec Expectations and should.js

Roche/lxdhub 98

Display, search and copy LXD images using a web interface.

click-contrib/click-man 72

Automate generation of man pages for python click applications :star:

randombenj/docat 70

Host your docs. Simple. Versioned. Fancy.

click-contrib/click-didyoumean 51

Enable git-like did-you-mean feature in click :sparkles:

mogria/msnake 35

a command line based snake game written in C using ncurses

click-contrib/click-aliases 20

enable aliases for click

randombenj/35c3-ctf 9

Documentation for the 35c3 junior capture the flag event

randombenj/yelims 8

:notebook::smiley:

BrunnerLivio/agile-ipsum 5

Generate an Agile Lorem Ipsum placeholder text

startedjoshiemoore/snakeware

started time in a day

fork timofurrer/web-abstract-template

Template für das Web-Abstract (Bachelorarbeit HSLU)

fork in 3 days

startedvincentdoerig/latex-css

started time in 8 days

delete branch timofurrer/ramlient

delete branch : pyup-update-flake8-3.7.5-to-3.8.1

delete time in 8 days

push eventtimofurrer/ramlient

pyup-bot

commit sha 9ea84c8f421e26926e64e4b31041758225895bd6

Update flake8 from 3.7.5 to 3.8.2

view details

push time in 8 days

create barnchtimofurrer/ramlient

branch : pyup-update-flake8-3.7.5-to-3.8.2

created branch time in 8 days

push eventradish-bdd/radish

Timo Furrer

commit sha d63fa17955ef655ca05be044205dfe86013314eb

remove stale action for now

view details

push time in 11 days

push eventtimofurrer/detectron2

Marc Szafraniec

commit sha 79fe87f2f4fb647107a9a4e81859ec675155d243

Minor print correction in anchor_generator.py Summary: Minor print correction in anchor_generator.py. It printed `{num_features}` instead of the actual value Reviewed By: patricklabatut Differential Revision: D21202639 fbshipit-source-id: 61324a483a85f6c1d1ddcd6d2cb251a40908d6e6

view details

Yuxin Wu

commit sha a5381dce40a85d075c712e2ac500260343883f0e

fix the use of some contiguous call Summary: similar fix was sent to torchvision as well Reviewed By: rbgirshick Differential Revision: D21188203 fbshipit-source-id: fd1ba680ad448fe15b42d6d7b9630a3ff7122203

view details

Yuxin Wu

commit sha fe46290e0e3af626fcc1b77075fc87c9172be1b1

9x SyncBN/GN configs Reviewed By: rbgirshick, jcjohnson Differential Revision: D21189755 fbshipit-source-id: 3e6a454be5cb5fdf64c4c4a40be5aa0e39f69297

view details

Timo Furrer

commit sha 7c377f2662ab43c17f21ecb8159cf6942d9e6a16

Implement RandomApply which applies a transformation with a given probability Summary: As discussed in https://github.com/facebookresearch/detectron2/issues/1251 this change implements a `RandomApply` transformation generator which can be used to only apply a given transformation with a certain probability. `RandomApply` supports wrapping of `T.Transform` and `T.TransformGen`. Pull Request resolved: https://github.com/facebookresearch/detectron2/pull/1254 Reviewed By: rbgirshick Differential Revision: D21199280 Pulled By: ppwwyyxx fbshipit-source-id: 271e740da59a588c4a81d4c5098ec3e7be6ec9b2

view details

tkhe

commit sha e6e2a716d95b505d4027ff165c9bba2a08e290f9

update setup.py Summary: Detectron2 requires PyTorch >= 1.4 now. Pull Request resolved: https://github.com/facebookresearch/detectron2/pull/1290 Reviewed By: rbgirshick Differential Revision: D21225795 Pulled By: ppwwyyxx fbshipit-source-id: 4402bb36004d3e6cabdd859da9c1c14ef840fab4

view details

Claudio Verdú Ruiz

commit sha a68aa1556cc37076264f1178c1b73ff053c52efc

Fix script docstring Summary: Thanks for your contribution! If you're sending a large PR (e.g., >50 lines), please open an issue first about the feature / bug, and indicate how you want to contribute. See more at https://detectron2.readthedocs.io/notes/contributing.html#pull-requests about how we handle PRs. Before submitting a PR, please run `dev/linter.sh` to lint the code. Pull Request resolved: https://github.com/facebookresearch/detectron2/pull/1291 Reviewed By: rbgirshick Differential Revision: D21226830 Pulled By: ppwwyyxx fbshipit-source-id: 756545a70a1f24170d8ff121638ec2d9948cafc6

view details

David Biertimpel

commit sha 6bb5a42af0c97c37fc735e130f72bdf51a2b9f02

Adding histograms to tensorboard Summary: Hey, as this is a small PR and I already implemented this feature for myself I did not open an issue. Description: I changed EventStorage and TensorboardXWriter such that histograms (for monitoring weights or gradients) can be saved to tensorboard. To make this implementation as efficient as possible I calculate the histogram statistics for the add_histogram_raw function of the PyTorch SummaryWriter (instead of saving weight matrices). In general, I tried to match the information flow to that of the put_image function. In my personal implementation I extended the EventStorage and TensorboardXWriter, but I thought that this might be a feature that Detectron could offer out of the box. A possible example usage could be a Hook: ``` class HistogramHook(HookBase): def after_step(self): storage = self.trainer.storage hist_period = self.trainer.cfg.HIST_PERIOD if hist_period > 0 and storage.iter % hist_period == 0: self.add_histogram(self.trainer.model, storage) def add_histogram(self, model, storage): for i, (tag, value) in enumerate(model.named_parameters()): # E.g. add histograms for parameters of the Mask R-CNN mask head if tag.split(".")[1] == "mask_head": storage.put_histogram(tag, value.data.cpu().numpy()) ``` Feedback is very much appreciated! :) Best, David Pull Request resolved: https://github.com/facebookresearch/detectron2/pull/1265 Reviewed By: rbgirshick Differential Revision: D21223968 Pulled By: ppwwyyxx fbshipit-source-id: 92a7b3d04fd1b45b4fd35fb9a814e529ba40ab9d

view details

Yuxin Wu

commit sha e6b1fd3313666d3286a95c4a6f2e42f60c10b75c

access to image/histogram in EventStorage is private Summary: Make them private given how they are current being used. Reviewed By: rbgirshick Differential Revision: D21224649 fbshipit-source-id: 8a9012475f31491bfd3f511cd3b9caaf9951c024

view details

Vasil Khalidov

commit sha f4761e20fea11750d447250eaed941ebc5626fc0

correct evaluation for external masks Summary: Evaluation with external masks is required for DP Chimps dataset, which does not have DensePose specific segmentation annotations. However, `minival_100` dataset contains some pure mask annotations that are taken into consideration by this logic. This fixes the logic to suit both use cases. Reviewed By: ppwwyyxx Differential Revision: D21229297 fbshipit-source-id: a01b8d6578acd48c726294128fb643a0b585e979

view details

Vasil Khalidov

commit sha 6cb9defcfa4bbfa2a906392372695f1ad72d2bbb

add categories metadata, add dataset name to each instance Summary: 1. Adds category metadata to keep the original mapping `category_id` -> `category_name`; this is required to correctly merge datasets with objects from different but compatible category sets; 2. Adds `dataset_name` entry for every annotation instance; this would allow one to combine annotation instances from datasets with different metadata and use the correct metadata in transforms, evaluations etc. Reviewed By: patricklabatut Differential Revision: D21227014 fbshipit-source-id: ff77ab9d11a262e10221d585bd86e7b9c1c1eeaa

view details

Vasil Khalidov

commit sha 92084c8cab10d015b39778e833199656ce35a5a7

support combining datasets with different but compatible categories Summary: 1. Adds possibility to combine datasets with different but compatible object category sets (i.e. different categories must have different IDs). This allows one to combine COCO instance segmentation datasets with COCO densepose datasets, for instance; 2. Segregate keep instance predicates (NB: ideally these should be available from the corresponding modules) Reviewed By: patricklabatut Differential Revision: D21201505 fbshipit-source-id: 8cd22eabdf0dc8fcff7a0155007da6551205d235

view details

Vasil Khalidov

commit sha 89ec4f9f65c9aa985f7b299041520d38829509eb

enable category agnostic training Summary: 1. Adds per-dataset category whitelists to control which objects are used for training; 2. Adds per-dataset category maps to consider objects from different categories as different representations of some general category; this is used for category-agnostic training; Reviewed By: patricklabatut Differential Revision: D21201936 fbshipit-source-id: 16d628b8c3bf635d5ff66667bab3677ef64a5f87

view details

Vasil Khalidov

commit sha 58cc9bfdfef71ae94b3eeae68f2a77cbef2c1623

add convenient Instance typedef Summary: Introduces a convenient typedef (`Instance = Dict[str, Any]`) for annotation instance specification (`Dict[str, Any]`). Reviewed By: patricklabatut Differential Revision: D21220935 fbshipit-source-id: 84c801a94020ae076e9c1a5ecf813fe9e8a76cdb

view details

Vasil Khalidov

commit sha 9a07526dec6c04d3cc987066361053730e3b9f99

add data and configs for class agnostic detection Summary: 1. Add config for class agnostic detection training 2. Adds required dataset definitions Reviewed By: patricklabatut Differential Revision: D21203261 fbshipit-source-id: 52209297ceb08a9768fd412473a95ae5b61c8ad0

view details

Yuxin Wu

commit sha 403f1d39af74bb1d35b5fd7efd11cc57eb5a409a

fix tensormask Summary: 1. self.device is not available 2. similar fix of contiguous Reviewed By: rbgirshick Differential Revision: D21257233 fbshipit-source-id: 82a3f40c541a7e7426a9c6a74a408fe9b678d296

view details

siyuanfeng-tri

commit sha 6640501c1a8e2b67519ffbab4e350841263be003

Fix RLE segmentation to json for coco annotation. Summary: I got `TypeError: Object of type 'uint32' is not JSON serializable` for `area`, and `TypeError: Object of type 'bytes' is not JSON serializable` for `segmentation["counts"]`during json dump. Relates to https://github.com/facebookresearch/detectron2/issues/200 Pull Request resolved: https://github.com/facebookresearch/detectron2/pull/1275 Reviewed By: rbgirshick Differential Revision: D21207121 Pulled By: ppwwyyxx fbshipit-source-id: 1c0b554ff9f980827c9134d6cb2b18a5cf9c78cc

view details

Yuxin Wu

commit sha 0d3face8cbb1946aeaf9de6c4beebba3134a2acc

updates in docker&docs Summary: Pull Request resolved: https://github.com/fairinternal/detectron2/pull/402 Reviewed By: rbgirshick Differential Revision: D21245938 Pulled By: ppwwyyxx fbshipit-source-id: 3c0972929431a5fa70870c39ae18cf38fa0fa6ad

view details

Yuxin Wu

commit sha 9e17a17bdd0a11ff0a439d6f6de9e449c4046df0

9x scratch model zoo Reviewed By: rbgirshick, jcjohnson Differential Revision: D21252160 fbshipit-source-id: eac6eab3c1700ca8546dbedb9482ca6cf3f3d41f

view details

Yuxin Wu

commit sha 35bb340684265527c5e47a1b616d379986ec7155

don't duplicate anchors for each image Summary: They are identical and should not be duplicated Reviewed By: rbgirshick Differential Revision: D21244959 fbshipit-source-id: 140f5d70b519f4738850692520e4f5b9b495e857

view details

Patrick Labatut

commit sha 4915330cac34570fc2ea4fa2c8b4088de448a973

Add unit tests for loading evolution configs Summary: Add unit tests for loading evolution configs. Changed logic to find configs to validate (consider only regular files and skip `.yaml` files with `Base-` prefix) Reviewed By: vkhalidov Differential Revision: D21276063 fbshipit-source-id: 57b1439013581bfbe40c334a7840fd2350bf72c4

view details

push time in 11 days

push eventtimofurrer/maya

Timo Furrer

commit sha b12a8dad11aec99ef62b063b5631dce0f528bcb4

Remove the stale bot for now

view details

push time in 12 days

IssuesEvent
IssuesEvent
IssuesEvent
IssuesEvent
IssuesEvent
IssuesEvent
IssuesEvent
IssuesEvent
IssuesEvent

push eventtimofurrer/maya

Jacob Floyd

commit sha d0d3b0136eff38268504c938e70b7cfca5be7ccf

Add timezone aware snap modifiers This allows for modifications in the given timezone. ie, snapping to the beginning of the day with @d uses midnight in the given timezone instead of in UTC. As usual, the maya object returned is back in UTC, only the modifications happen in the requested timezone.

view details

Timo Furrer

commit sha 0a7061c1258650f8b2ad4479f3cdecbc56fcf7de

Merge pull request #186 from cognifloyd/snap_tz Add timezone aware snap modifiers

view details

push time in 12 days

PR merged timofurrer/maya

Add timezone aware snap modifiers

This allows for modifications in the given timezone. ie, snapping to the beginning of the day with @d uses midnight in the given timezone instead of in UTC. As usual, the maya object returned is back in UTC, only the modifications happen in the requested timezone.

This is the same as #185. But there was no response in that PR and the Github Actions bot closed it. Since I can't reopen the PR (as it is still valid, and will still merge cleanly on master), I'm opening a new PR.

+36 -0

1 comment

3 changed files

cognifloyd

pr closed time in 12 days

pull request commenttimofurrer/maya

Add timezone aware snap modifiers

Looks good! Thanks :tada:

I'm gonna to disable the bot .. it's causing more harm then good.

Cheers :tada:

cognifloyd

comment created time in 12 days

issue commentradish-bdd/radish

Import `each_scenario` fails

Or you could also be brave and use the pre release of radish - which I'll try to release at the end of this summer. I'll have time to keep working on this in June.

dschiller

comment created time in 12 days

issue commentradish-bdd/radish

Import `each_scenario` fails

The latest docs are for the current master which is only released as pre version. So please use the stable version on read the docs: https://radish.readthedocs.io/en/stable/tutorial.html

dschiller

comment created time in 12 days

issue closedradish-bdd/radish

Import `each_scenario` fails

I am trying to execute this Example: https://radish.readthedocs.io/en/latest/usage.html#implementing-hooks

from radish import each_scenario

fails with

 cannot import name 'each_scenario' from 'radish' (c:\program files\python37\lib\site-packages\radish\__init__.py)
radish==0.1.10
radish-bdd==0.13.1

closed time in 12 days

dschiller

push eventtimofurrer/awesome-asyncio

Manuel Miranda

commit sha 552a0d4537bbdaf300acdae3f170bbba97f9b99d

Add aiocache

view details

Timo Furrer

commit sha ecb15e3e50ef92bd68ae62033b971c8453692f08

Merge pull request #70 from argaen/master Add aiocache

view details

push time in 14 days

pull request commenttimofurrer/awesome-asyncio

add core-service to misc section

Thanks for the suggestion! :tada:

It still seams to be a very fresh and "unknown" package - let's start a vote here!

flu2020

comment created time in 14 days

startedjaparic/rust-cross

started time in 14 days

startedtr1pzz/Realtime_PyAudio_FFT

started time in 18 days

startedyhirose/cpp-httplib

started time in 18 days

delete branch timofurrer/ramlient

delete branch : pyup-update-flake8-3.7.5-to-3.7.9

delete time in 19 days

push eventtimofurrer/ramlient

pyup-bot

commit sha 92f5d1ce4c7716242f6112dc0220433d13ff486f

Update flake8 from 3.7.5 to 3.8.1

view details

push time in 19 days

create barnchtimofurrer/ramlient

branch : pyup-update-flake8-3.7.5-to-3.8.1

created branch time in 19 days

startedalfg/opendrinks

started time in 19 days

delete branch timofurrer/ramlient

delete branch : pyup-update-pytest-4.4.1-to-5.4.1

delete time in 22 days

push eventtimofurrer/ramlient

pyup-bot

commit sha 10fe13968a8be1bc3fcb7c4980fb94680b3d2fdb

Update pytest from 4.4.1 to 5.4.2

view details

push time in 22 days

create barnchtimofurrer/ramlient

branch : pyup-update-pytest-4.4.1-to-5.4.2

created branch time in 22 days

issue commentpypa/setuptools_scm

Versions according to the Debian Policy Manual

I totally agree that this might be the topic for another separate tool and let setuptools_scm do what it currently does.

Feel free to close this issue for now

I think once there are efforts made towards an integration into setuptools_scm - a new issue will come up any ways :)

timofurrer

comment created time in a month

startedMachine-Learning-Tokyo/Interactive_Tools

started time in a month

startedpatrickbucher/web-abstract-template

started time in a month

pull request commentfacebookresearch/detectron2

Implement `crop` option in RotationTransform

Is there something I can do to move this PR forward?

timofurrer

comment created time in a month

startedpdoc3/pdoc

started time in a month

startedfangfufu/Linux-Fake-Background-Webcam

started time in a month

issue commentpypa/pypi-support

Transfer of s3monkey to the primary maintainer

I got push permissions some time ago from Kenneth. I can confirm that @kobayashi is the current maintainer of the project (Kenneth transferred the ownership).

It is okay for me to handover sole ownership of the project to @kobayashi - you can even remove myself as a maintainer on PyPI.

Cheers and good look with the project! :tada:

kobayashi

comment created time in a month

issue commentfacebookresearch/detectron2

Visualize transformed annotations

The issue was in the transformation of the tensor to an actual image, where I mixed up the width and height.

timofurrer

comment created time in a month

push eventtimofurrer/detectron2

Vasil Khalidov

commit sha 5525cf500ddb5010fedf38fabc97471986f21a13

add CocoDatasetInfo data class for COCO datasets Summary: Summary of changes: 1. Extract all COCO-specific logic for builtin datasets into `coco.py` 2. Switch to Manifold-based image storage for COCO-2014-train 3. Change required Python version (>= 3.7 for dataclasses) These changes are useful for adding other types of builtin datasets with different data layout or metadata. Reviewed By: patricklabatut Differential Revision: D21132895 fbshipit-source-id: b12161eb8c8fc5695c65ed859b98b06bfec5fa27

view details

Yuxin Wu

commit sha 4bd92bb9b993312c7598a59fb7c203ad24f3e4a5

add model analysis tool Summary: Support flop/activation counting for 100 inputs. Reviewed By: rbgirshick Differential Revision: D20988652 fbshipit-source-id: 394bd3050fd4edc0d726bdb431edd40211f66810

view details

Yuxin Wu

commit sha 273e3b9746473a7469550bf793bbfb4c0a47c7d5

misc small changes Summary: on testing, typing Reviewed By: rbgirshick Differential Revision: D21158243 fbshipit-source-id: 84237575f3a65816b6988fa8994a8106853497e7

view details

Yuxin Wu

commit sha 5d967edf0c36b5bd3ab46ead629eb73dae4b6e05

implement inference code in box_predictor Summary: keep some methods in FastRCNNOutputs since they are still used Reviewed By: rbgirshick Differential Revision: D21073561 fbshipit-source-id: dd55f485eb576f775453227775cde27332d1c189

view details

Marc Szafraniec

commit sha dc37777d02737aa4e993d7f20516356f22d2f77a

Slight change in densepose.vis.densepose.py to accomodate refactoring Summary: Slight change in densepose.vis.densepose.py to accomodate refactoring Reviewed By: vkhalidov Differential Revision: D21156357 fbshipit-source-id: 25bd9c7752ff5e3f4a7ca732622e59e7c9c0c3ea

view details

Marc Szafraniec

commit sha 46e5460f6c1f1d8052838b0dee6d7b53564055a6

Remove overhead for rotation transform when angle = 0 Summary: Remove overhead for rotation transform when angle = 0 Reviewed By: patricklabatut Differential Revision: D21175810 fbshipit-source-id: 8242908d25ff5f9e760877d3a160f42d5c41fb80

view details

Vasil Khalidov

commit sha 4caa60602ca0a684aae51853abcd920715954ffe

add DP Chimps dataset Summary: Add DensePose Chimps dataset to the list of builtin datasets: 1. Add dataset specs to `builtin` and `builtin_fb` datasets 2. Remove default `"thing_classes": ["person"]` metadata: it should be filled directly from the dataset; (NB: later diffs will add object class metadata merging logic to be able to join category-compatible datasets) 3. Add json annotation path conversion to DensePose evaluator 4. Corrected evaluation to work for arrays of valid distances of length 1 Reviewed By: patricklabatut Differential Revision: D21134576 fbshipit-source-id: 1384f754f3849e445b46ca0a3fb2d038cb27317f

view details

Vasil Khalidov

commit sha 01483f342d311a2e4a74c16ed767c1318a480401

fix references to densepose structures Summary: Structures were moved to `densepose.data` module in D21129619. This corrects the `query_db` tool accordingly. The other modification has already been made in D21156357. Reviewed By: patricklabatut Differential Revision: D21175367 fbshipit-source-id: 1e244f7613358043102955b419ae3ce5e978101a

view details

Vasil Khalidov

commit sha 8a2b8709cced7cc48b1d9f14b78bea3f9f6ea751

add support for non-local entry paths in query_db Summary: Add support for entries with non-local image file paths to `query_db` tool Reviewed By: patricklabatut Differential Revision: D21176271 fbshipit-source-id: f67f77df1f68f3bab99412b4ff45b80c6ae58a20

view details

Yuxin Wu

commit sha 0c045f269d69ecbf8a4c978ee9a9f61a46861b91

fix some tests Summary: Fix some interesting numerical issues. Use latest torch release in docker. Reviewed By: rbgirshick Differential Revision: D21175202 fbshipit-source-id: f47be6ae647139db46079d79329bf5a8402ac8a0

view details

Timo Furrer

commit sha ea541618eac2b286fbe819d98852ff71627f7ded

Implement RandomApply which applies a transformation with a given probability

view details

push time in a month

push eventtimofurrer/detectron2

Vasil Khalidov

commit sha 5525cf500ddb5010fedf38fabc97471986f21a13

add CocoDatasetInfo data class for COCO datasets Summary: Summary of changes: 1. Extract all COCO-specific logic for builtin datasets into `coco.py` 2. Switch to Manifold-based image storage for COCO-2014-train 3. Change required Python version (>= 3.7 for dataclasses) These changes are useful for adding other types of builtin datasets with different data layout or metadata. Reviewed By: patricklabatut Differential Revision: D21132895 fbshipit-source-id: b12161eb8c8fc5695c65ed859b98b06bfec5fa27

view details

Yuxin Wu

commit sha 4bd92bb9b993312c7598a59fb7c203ad24f3e4a5

add model analysis tool Summary: Support flop/activation counting for 100 inputs. Reviewed By: rbgirshick Differential Revision: D20988652 fbshipit-source-id: 394bd3050fd4edc0d726bdb431edd40211f66810

view details

Yuxin Wu

commit sha 273e3b9746473a7469550bf793bbfb4c0a47c7d5

misc small changes Summary: on testing, typing Reviewed By: rbgirshick Differential Revision: D21158243 fbshipit-source-id: 84237575f3a65816b6988fa8994a8106853497e7

view details

Yuxin Wu

commit sha 5d967edf0c36b5bd3ab46ead629eb73dae4b6e05

implement inference code in box_predictor Summary: keep some methods in FastRCNNOutputs since they are still used Reviewed By: rbgirshick Differential Revision: D21073561 fbshipit-source-id: dd55f485eb576f775453227775cde27332d1c189

view details

Marc Szafraniec

commit sha dc37777d02737aa4e993d7f20516356f22d2f77a

Slight change in densepose.vis.densepose.py to accomodate refactoring Summary: Slight change in densepose.vis.densepose.py to accomodate refactoring Reviewed By: vkhalidov Differential Revision: D21156357 fbshipit-source-id: 25bd9c7752ff5e3f4a7ca732622e59e7c9c0c3ea

view details

Marc Szafraniec

commit sha 46e5460f6c1f1d8052838b0dee6d7b53564055a6

Remove overhead for rotation transform when angle = 0 Summary: Remove overhead for rotation transform when angle = 0 Reviewed By: patricklabatut Differential Revision: D21175810 fbshipit-source-id: 8242908d25ff5f9e760877d3a160f42d5c41fb80

view details

Vasil Khalidov

commit sha 4caa60602ca0a684aae51853abcd920715954ffe

add DP Chimps dataset Summary: Add DensePose Chimps dataset to the list of builtin datasets: 1. Add dataset specs to `builtin` and `builtin_fb` datasets 2. Remove default `"thing_classes": ["person"]` metadata: it should be filled directly from the dataset; (NB: later diffs will add object class metadata merging logic to be able to join category-compatible datasets) 3. Add json annotation path conversion to DensePose evaluator 4. Corrected evaluation to work for arrays of valid distances of length 1 Reviewed By: patricklabatut Differential Revision: D21134576 fbshipit-source-id: 1384f754f3849e445b46ca0a3fb2d038cb27317f

view details

Vasil Khalidov

commit sha 01483f342d311a2e4a74c16ed767c1318a480401

fix references to densepose structures Summary: Structures were moved to `densepose.data` module in D21129619. This corrects the `query_db` tool accordingly. The other modification has already been made in D21156357. Reviewed By: patricklabatut Differential Revision: D21175367 fbshipit-source-id: 1e244f7613358043102955b419ae3ce5e978101a

view details

Vasil Khalidov

commit sha 8a2b8709cced7cc48b1d9f14b78bea3f9f6ea751

add support for non-local entry paths in query_db Summary: Add support for entries with non-local image file paths to `query_db` tool Reviewed By: patricklabatut Differential Revision: D21176271 fbshipit-source-id: f67f77df1f68f3bab99412b4ff45b80c6ae58a20

view details

Yuxin Wu

commit sha 0c045f269d69ecbf8a4c978ee9a9f61a46861b91

fix some tests Summary: Fix some interesting numerical issues. Use latest torch release in docker. Reviewed By: rbgirshick Differential Revision: D21175202 fbshipit-source-id: f47be6ae647139db46079d79329bf5a8402ac8a0

view details

push time in a month

push eventtimofurrer/detectron2

Vasil Khalidov

commit sha 5525cf500ddb5010fedf38fabc97471986f21a13

add CocoDatasetInfo data class for COCO datasets Summary: Summary of changes: 1. Extract all COCO-specific logic for builtin datasets into `coco.py` 2. Switch to Manifold-based image storage for COCO-2014-train 3. Change required Python version (>= 3.7 for dataclasses) These changes are useful for adding other types of builtin datasets with different data layout or metadata. Reviewed By: patricklabatut Differential Revision: D21132895 fbshipit-source-id: b12161eb8c8fc5695c65ed859b98b06bfec5fa27

view details

Yuxin Wu

commit sha 4bd92bb9b993312c7598a59fb7c203ad24f3e4a5

add model analysis tool Summary: Support flop/activation counting for 100 inputs. Reviewed By: rbgirshick Differential Revision: D20988652 fbshipit-source-id: 394bd3050fd4edc0d726bdb431edd40211f66810

view details

Yuxin Wu

commit sha 273e3b9746473a7469550bf793bbfb4c0a47c7d5

misc small changes Summary: on testing, typing Reviewed By: rbgirshick Differential Revision: D21158243 fbshipit-source-id: 84237575f3a65816b6988fa8994a8106853497e7

view details

Yuxin Wu

commit sha 5d967edf0c36b5bd3ab46ead629eb73dae4b6e05

implement inference code in box_predictor Summary: keep some methods in FastRCNNOutputs since they are still used Reviewed By: rbgirshick Differential Revision: D21073561 fbshipit-source-id: dd55f485eb576f775453227775cde27332d1c189

view details

Marc Szafraniec

commit sha dc37777d02737aa4e993d7f20516356f22d2f77a

Slight change in densepose.vis.densepose.py to accomodate refactoring Summary: Slight change in densepose.vis.densepose.py to accomodate refactoring Reviewed By: vkhalidov Differential Revision: D21156357 fbshipit-source-id: 25bd9c7752ff5e3f4a7ca732622e59e7c9c0c3ea

view details

Marc Szafraniec

commit sha 46e5460f6c1f1d8052838b0dee6d7b53564055a6

Remove overhead for rotation transform when angle = 0 Summary: Remove overhead for rotation transform when angle = 0 Reviewed By: patricklabatut Differential Revision: D21175810 fbshipit-source-id: 8242908d25ff5f9e760877d3a160f42d5c41fb80

view details

Vasil Khalidov

commit sha 4caa60602ca0a684aae51853abcd920715954ffe

add DP Chimps dataset Summary: Add DensePose Chimps dataset to the list of builtin datasets: 1. Add dataset specs to `builtin` and `builtin_fb` datasets 2. Remove default `"thing_classes": ["person"]` metadata: it should be filled directly from the dataset; (NB: later diffs will add object class metadata merging logic to be able to join category-compatible datasets) 3. Add json annotation path conversion to DensePose evaluator 4. Corrected evaluation to work for arrays of valid distances of length 1 Reviewed By: patricklabatut Differential Revision: D21134576 fbshipit-source-id: 1384f754f3849e445b46ca0a3fb2d038cb27317f

view details

Vasil Khalidov

commit sha 01483f342d311a2e4a74c16ed767c1318a480401

fix references to densepose structures Summary: Structures were moved to `densepose.data` module in D21129619. This corrects the `query_db` tool accordingly. The other modification has already been made in D21156357. Reviewed By: patricklabatut Differential Revision: D21175367 fbshipit-source-id: 1e244f7613358043102955b419ae3ce5e978101a

view details

Vasil Khalidov

commit sha 8a2b8709cced7cc48b1d9f14b78bea3f9f6ea751

add support for non-local entry paths in query_db Summary: Add support for entries with non-local image file paths to `query_db` tool Reviewed By: patricklabatut Differential Revision: D21176271 fbshipit-source-id: f67f77df1f68f3bab99412b4ff45b80c6ae58a20

view details

Yuxin Wu

commit sha 0c045f269d69ecbf8a4c978ee9a9f61a46861b91

fix some tests Summary: Fix some interesting numerical issues. Use latest torch release in docker. Reviewed By: rbgirshick Differential Revision: D21175202 fbshipit-source-id: f47be6ae647139db46079d79329bf5a8402ac8a0

view details

Timo Furrer

commit sha 3d3e66416b40da3048814c0654c98f499a3e1944

Merge branch 'master' into feature/random-rotation-crop

view details

push time in a month

pull request commentfacebookresearch/detectron2

Implement RandomApply which applies a transformation with a given probability

Tests are failing because of linting issues present in master.

timofurrer

comment created time in a month

push eventtimofurrer/detectron2

Vasil Khalidov

commit sha 45e202943cd96f96d8b45fc35efd402e2d1fbfd2

create densepose.data module and follow D2 structure Summary: Created `densepose.data` package and aligned its structure to follow D2: - builtin datasets are defined in files `builtin.py` and `builtin_fb.py` in `densepose/data/datasets/` - structures reside in `densepose/data/` Changes all references accordingly Reviewed By: patricklabatut Differential Revision: D21129619 fbshipit-source-id: 5d328e2c7a340cf957e63ff12d07fd49aecdd1bb

view details

Aleksey Nozdryn-Plotnicki

commit sha e924f6ce39cc5b5c9379e2671c3122cdb156dd6c

Removed superfluous code from box_iou_rotated_cpu_kernel Summary: I think these unnecessary lines do nothing. Pull Request resolved: https://github.com/facebookresearch/detectron2/pull/1243 Reviewed By: rbgirshick Differential Revision: D21101523 Pulled By: ppwwyyxx fbshipit-source-id: 4339771f861aba6d202951d81ef407df18004efb

view details

Yuxin Wu

commit sha c6c56323e2d71177ca93ea76bf191c0644b37076

update to sphinx3.0 Summary: Latest version of these tools no longer support "url_resolver", so an alternative is implemented to resolve github urls. Pull Request resolved: https://github.com/fairinternal/detectron2/pull/399 Reviewed By: rbgirshick Differential Revision: D21025270 Pulled By: ppwwyyxx fbshipit-source-id: 8df9c8d59718e360b5a521dd6322d34b40c8fb60

view details

Yuxin Wu

commit sha 4643ec27815787e88df60d349b4f5e00fe545441

use configurable in anchor generator Reviewed By: rbgirshick Differential Revision: D21111568 fbshipit-source-id: a5040b9c1912ad68fd95afd6c29fdfb347b5bbb9

view details

Yuxin Wu

commit sha a06299df47dd855f3ebd0f4a95d822423b788273

let checkpointer._load_model return incompatible keys just like pytorch Reviewed By: rbgirshick Differential Revision: D21107752 fbshipit-source-id: c1942c4d17ad2fbe46bda8a7bb6f1b14f4b20028

view details

Yuxin Wu

commit sha 9ced00a1305a9a0d4da4cc7fc998cc429af24c14

Enable model.to(device) Summary: Each meta_arch's pixel_mean & pixel_std is not tracked as module's buffers. As a result, `model.cpu()` or `model.to(device)` will not change the device of pixel_mean/pixel_std, and will not work properly. This PR declares pixel_mean&pixel_std as buffers to enable `model.to(device)`. The missing key warning was silenced explicitly. Pull Request resolved: https://github.com/fairinternal/detectron2/pull/361 Reviewed By: rbgirshick Differential Revision: D19254365 Pulled By: ppwwyyxx fbshipit-source-id: 1e8251290a5a5d69e60576bc8713e6c727f69d3b

view details

Tongliang Liao

commit sha 1e3bd77c4b977e01108931bbeb9b65d2c40059a0

Make padding shape tracable. Summary: Previously `max_size` will be hardcoded in traced ONNX model since some dataflow/computation is outside PyTorch. The entire chunk between `[Sub, Div]`(whitening) and `Pad` shows how this shape inferencing works, with `size_divisibility==32` for this backbone. `ceil()` is replaced with `[Add, Div, Mul]` in the middle, i.e. `(n + (m - 1)) // m * m` with constant folding for `m-1`. <img width="479" alt="image" src="https://user-images.githubusercontent.com/5203025/79173373-e9d48780-7e29-11ea-880b-dc8612e027e2.png"> Pull Request resolved: https://github.com/facebookresearch/detectron2/pull/1208 Reviewed By: rbgirshick Differential Revision: D21076153 Pulled By: ppwwyyxx fbshipit-source-id: c6d37114dd5d9cb0ec65f561fba7961ce26d74ec

view details

Yuxin Wu

commit sha bb9f5d8e613358519c9865609ab3fe7b6571f2ba

organize unittests to directories Reviewed By: rbgirshick Differential Revision: D21104831 fbshipit-source-id: 67dac8e40abb4ce3320e3878ee160fee975d7bb1

view details

Timo Furrer

commit sha 8b39590f400e4bd15b590c3a8560ced8d5d02c0c

Implement `crop` option in RotationTransform The `crop` option in the `RotationTransform` and `RandomRotation` transformations can be used to crop the rotated image to the largest possible axis-aligned rectangle with a maximal area within the rotated image. It's disabled by default to ensure backwards compatibility.

view details

push time in a month

push eventtimofurrer/detectron2

Vasil Khalidov

commit sha 45e202943cd96f96d8b45fc35efd402e2d1fbfd2

create densepose.data module and follow D2 structure Summary: Created `densepose.data` package and aligned its structure to follow D2: - builtin datasets are defined in files `builtin.py` and `builtin_fb.py` in `densepose/data/datasets/` - structures reside in `densepose/data/` Changes all references accordingly Reviewed By: patricklabatut Differential Revision: D21129619 fbshipit-source-id: 5d328e2c7a340cf957e63ff12d07fd49aecdd1bb

view details

Aleksey Nozdryn-Plotnicki

commit sha e924f6ce39cc5b5c9379e2671c3122cdb156dd6c

Removed superfluous code from box_iou_rotated_cpu_kernel Summary: I think these unnecessary lines do nothing. Pull Request resolved: https://github.com/facebookresearch/detectron2/pull/1243 Reviewed By: rbgirshick Differential Revision: D21101523 Pulled By: ppwwyyxx fbshipit-source-id: 4339771f861aba6d202951d81ef407df18004efb

view details

Yuxin Wu

commit sha c6c56323e2d71177ca93ea76bf191c0644b37076

update to sphinx3.0 Summary: Latest version of these tools no longer support "url_resolver", so an alternative is implemented to resolve github urls. Pull Request resolved: https://github.com/fairinternal/detectron2/pull/399 Reviewed By: rbgirshick Differential Revision: D21025270 Pulled By: ppwwyyxx fbshipit-source-id: 8df9c8d59718e360b5a521dd6322d34b40c8fb60

view details

Yuxin Wu

commit sha 4643ec27815787e88df60d349b4f5e00fe545441

use configurable in anchor generator Reviewed By: rbgirshick Differential Revision: D21111568 fbshipit-source-id: a5040b9c1912ad68fd95afd6c29fdfb347b5bbb9

view details

Yuxin Wu

commit sha a06299df47dd855f3ebd0f4a95d822423b788273

let checkpointer._load_model return incompatible keys just like pytorch Reviewed By: rbgirshick Differential Revision: D21107752 fbshipit-source-id: c1942c4d17ad2fbe46bda8a7bb6f1b14f4b20028

view details

Yuxin Wu

commit sha 9ced00a1305a9a0d4da4cc7fc998cc429af24c14

Enable model.to(device) Summary: Each meta_arch's pixel_mean & pixel_std is not tracked as module's buffers. As a result, `model.cpu()` or `model.to(device)` will not change the device of pixel_mean/pixel_std, and will not work properly. This PR declares pixel_mean&pixel_std as buffers to enable `model.to(device)`. The missing key warning was silenced explicitly. Pull Request resolved: https://github.com/fairinternal/detectron2/pull/361 Reviewed By: rbgirshick Differential Revision: D19254365 Pulled By: ppwwyyxx fbshipit-source-id: 1e8251290a5a5d69e60576bc8713e6c727f69d3b

view details

Tongliang Liao

commit sha 1e3bd77c4b977e01108931bbeb9b65d2c40059a0

Make padding shape tracable. Summary: Previously `max_size` will be hardcoded in traced ONNX model since some dataflow/computation is outside PyTorch. The entire chunk between `[Sub, Div]`(whitening) and `Pad` shows how this shape inferencing works, with `size_divisibility==32` for this backbone. `ceil()` is replaced with `[Add, Div, Mul]` in the middle, i.e. `(n + (m - 1)) // m * m` with constant folding for `m-1`. <img width="479" alt="image" src="https://user-images.githubusercontent.com/5203025/79173373-e9d48780-7e29-11ea-880b-dc8612e027e2.png"> Pull Request resolved: https://github.com/facebookresearch/detectron2/pull/1208 Reviewed By: rbgirshick Differential Revision: D21076153 Pulled By: ppwwyyxx fbshipit-source-id: c6d37114dd5d9cb0ec65f561fba7961ce26d74ec

view details

Yuxin Wu

commit sha bb9f5d8e613358519c9865609ab3fe7b6571f2ba

organize unittests to directories Reviewed By: rbgirshick Differential Revision: D21104831 fbshipit-source-id: 67dac8e40abb4ce3320e3878ee160fee975d7bb1

view details

Timo Furrer

commit sha 3a8627d39e800566edfe215695a3dd768f920034

Implement RandomApply which applies a transformation with a given probability

view details

push time in a month

push eventtimofurrer/detectron2

Vasil Khalidov

commit sha 45e202943cd96f96d8b45fc35efd402e2d1fbfd2

create densepose.data module and follow D2 structure Summary: Created `densepose.data` package and aligned its structure to follow D2: - builtin datasets are defined in files `builtin.py` and `builtin_fb.py` in `densepose/data/datasets/` - structures reside in `densepose/data/` Changes all references accordingly Reviewed By: patricklabatut Differential Revision: D21129619 fbshipit-source-id: 5d328e2c7a340cf957e63ff12d07fd49aecdd1bb

view details

Aleksey Nozdryn-Plotnicki

commit sha e924f6ce39cc5b5c9379e2671c3122cdb156dd6c

Removed superfluous code from box_iou_rotated_cpu_kernel Summary: I think these unnecessary lines do nothing. Pull Request resolved: https://github.com/facebookresearch/detectron2/pull/1243 Reviewed By: rbgirshick Differential Revision: D21101523 Pulled By: ppwwyyxx fbshipit-source-id: 4339771f861aba6d202951d81ef407df18004efb

view details

Yuxin Wu

commit sha c6c56323e2d71177ca93ea76bf191c0644b37076

update to sphinx3.0 Summary: Latest version of these tools no longer support "url_resolver", so an alternative is implemented to resolve github urls. Pull Request resolved: https://github.com/fairinternal/detectron2/pull/399 Reviewed By: rbgirshick Differential Revision: D21025270 Pulled By: ppwwyyxx fbshipit-source-id: 8df9c8d59718e360b5a521dd6322d34b40c8fb60

view details

Yuxin Wu

commit sha 4643ec27815787e88df60d349b4f5e00fe545441

use configurable in anchor generator Reviewed By: rbgirshick Differential Revision: D21111568 fbshipit-source-id: a5040b9c1912ad68fd95afd6c29fdfb347b5bbb9

view details

Yuxin Wu

commit sha a06299df47dd855f3ebd0f4a95d822423b788273

let checkpointer._load_model return incompatible keys just like pytorch Reviewed By: rbgirshick Differential Revision: D21107752 fbshipit-source-id: c1942c4d17ad2fbe46bda8a7bb6f1b14f4b20028

view details

Yuxin Wu

commit sha 9ced00a1305a9a0d4da4cc7fc998cc429af24c14

Enable model.to(device) Summary: Each meta_arch's pixel_mean & pixel_std is not tracked as module's buffers. As a result, `model.cpu()` or `model.to(device)` will not change the device of pixel_mean/pixel_std, and will not work properly. This PR declares pixel_mean&pixel_std as buffers to enable `model.to(device)`. The missing key warning was silenced explicitly. Pull Request resolved: https://github.com/fairinternal/detectron2/pull/361 Reviewed By: rbgirshick Differential Revision: D19254365 Pulled By: ppwwyyxx fbshipit-source-id: 1e8251290a5a5d69e60576bc8713e6c727f69d3b

view details

Tongliang Liao

commit sha 1e3bd77c4b977e01108931bbeb9b65d2c40059a0

Make padding shape tracable. Summary: Previously `max_size` will be hardcoded in traced ONNX model since some dataflow/computation is outside PyTorch. The entire chunk between `[Sub, Div]`(whitening) and `Pad` shows how this shape inferencing works, with `size_divisibility==32` for this backbone. `ceil()` is replaced with `[Add, Div, Mul]` in the middle, i.e. `(n + (m - 1)) // m * m` with constant folding for `m-1`. <img width="479" alt="image" src="https://user-images.githubusercontent.com/5203025/79173373-e9d48780-7e29-11ea-880b-dc8612e027e2.png"> Pull Request resolved: https://github.com/facebookresearch/detectron2/pull/1208 Reviewed By: rbgirshick Differential Revision: D21076153 Pulled By: ppwwyyxx fbshipit-source-id: c6d37114dd5d9cb0ec65f561fba7961ce26d74ec

view details

Yuxin Wu

commit sha bb9f5d8e613358519c9865609ab3fe7b6571f2ba

organize unittests to directories Reviewed By: rbgirshick Differential Revision: D21104831 fbshipit-source-id: 67dac8e40abb4ce3320e3878ee160fee975d7bb1

view details

push time in a month

push eventtimofurrer/detectron2

Timo Furrer

commit sha 54ac1d4318e5eaf81ca7851999d15112533af2a1

Implement RandomApply which applies a transformation with a given probability

view details

push time in a month

Pull request review commentfacebookresearch/detectron2

Implement RandomApply which applies a transformation with a given probability

 def __repr__(self):     __str__ = __repr__  +class RandomApply(TransformGen):+    """+    Randomly apply the wrapper transformation with a given probability.+    """++    def __init__(self, transform, prob=0.5):+        """+        Args:+            prob (float): probability between 0.0 and 1.0 that

done

timofurrer

comment created time in a month

push eventtimofurrer/detectron2

Timo Furrer

commit sha 48e00f47b18f089b813e0b8c18bf8b610fc4dcf1

Implement RandomApply which applies a transformation with a given probability

view details

push time in a month

push eventtimofurrer/detectron2

Timo Furrer

commit sha bb11a70822156bc5c267d6453a9f0189b9da6d06

Implement `crop` option in RotationTransform The `crop` option in the `RotationTransform` and `RandomRotation` transformations can be used to crop the rotated image to the largest possible axis-aligned rectangle with a maximal area within the rotated image. It's disabled by default to ensure backwards compatibility.

view details

push time in a month

issue commenttimofurrer/colorful

colorful dicts in system service app

Converting p to str explicitly helps, right?

guydepauw

comment created time in a month

push eventtimofurrer/detectron2

Timo Furrer

commit sha b3342bbb6d6267051484baddf4cd0d0a59ec9326

Implement `crop` option in RotationTransform The `crop` option in the `RotationTransform` and `RandomRotation` transformations can be used to crop the rotated image to the largest possible axis-aligned rectangle with a maximal area within the rotated image. It's disabled by default to ensure backwards compatibility.

view details

push time in a month

PR opened facebookresearch/detectron2

Implement `crop` option in RotationTransform

The crop option in the RotationTransform and RandomRotation transformations can be used to crop the rotated image to the largest possible axis-aligned rectangle with a maximal area within the rotated image.

It's disabled by default to ensure backwards compatibility.

The following image depicts the crop option:

image

+86 -6

0 comment

3 changed files

pr created time in a month

create barnchtimofurrer/detectron2

branch : feature/random-rotation-crop

created branch time in a month

issue commenttimofurrer/colorful

colorful dicts in system service app

Can you please share a little bit more about how you use it? Like the code involved here?

guydepauw

comment created time in a month

issue commentfacebookresearch/detectron2

Visualize transformed annotations

I'm not saying it's a detectron2 issue - maybe my usage is just wrong.

I've modified the Tutorial notebook slightly and added the code in question to illustrate the problem. The notebook is here: https://drive.google.com/file/d/1aPRkNFnAEiYLg7HPXRlyLOVQ5trzhiqF/view?usp=sharing

And a screenshot of the results for convenience:

image

timofurrer

comment created time in a month

push eventtimofurrer/detectron2

Timo Furrer

commit sha 40772b141b0baf5c86ee5a3885cdc34116f567cc

Implement RandomApply which applies a transformation with a given probability

view details

push time in a month

PR opened facebookresearch/detectron2

Implement RandomApply which applies a transformation with a given probability

As discussed in #1251 this change implements a RandomApply transformation generator which can be used to only apply a given transformation with a certain probability.

RandomApply supports wrapping of T.Transform and T.TransformGen.

+69 -0

0 comment

2 changed files

pr created time in a month

create barnchtimofurrer/detectron2

branch : feature/random-apply-transform

created branch time in a month

issue openedfacebookresearch/detectron2

Visualize transformed annotations

I'm having troubles visualizing the transformed annotations from sampled images using the DefaultTrainer.build_train_loader(cfg) loader. By using this trainer / builder I'm also using the default crop, resize and random flip transformers.

I've tried to use the Visualizers overlay_instances method, like this:

train_loader = DefaultTrainer.build_train_loader(cfg)

sample_image_batches = 2
f = plt.figure(figsize=(20, 10 * sample_image_batches))
idx = 1
for sample_image_batch_idx, train_image_batch in enumerate(train_loader):

    for idx, train_image in enumerate(train_image_batch):
        image = train_image["image"].numpy().transpose(2, 1, 0)

        # visualize ground truth
        gt_visualizer = Visualizer(
            image[:, :, ::-1], metadata=dataset_train_metadata, scale=1
        )
        gt_image = gt_visualizer.overlay_instances(
            boxes=train_image["instances"].gt_boxes, 
            masks=train_image["instances"].gt_masks, 
            labels=train_image["instances"].gt_classes
        )

        f.add_subplot(sample_image_batches, 2, (2 * sample_image_batch_idx) + idx + 1)
        plt.title(f"Image {train_image['file_name']}")
        plt.imshow(gt_image.get_image()[:, :, ::-1])

        break

plt.show()

I don't understand why the annotations are not in the correct place. As far as I can tell those should be transformed here: https://github.com/facebookresearch/detectron2/blob/d1c7a4a63b4e24b02564cbef0bd18d28f757bc0d/detectron2/data/dataset_mapper.py#L127

created time in a month

PR closed facebookresearch/detectron2

Add `prob` argument for RandomRotations CLA Signed

The prob argument for RandomRotation (similar to RandomFlip) allows to control the probability the rotation angle will be applied.

This allows for the original image to be sampled with a certain probability.

The prob argument is appended to the argument list for backwards compatibility.

+58 -2

3 comments

2 changed files

timofurrer

pr closed time in a month

pull request commentfacebookresearch/detectron2

Add `prob` argument for RandomRotations

Sounds good!

I'll create another PR!

timofurrer

comment created time in a month

pull request commentfacebookresearch/detectron2

Add `prob` argument for RandomRotations

I think that the test failures are unrelated to my change, can you confirm?

timofurrer

comment created time in a month

startedbernedom/bertrand

started time in a month

push eventtimofurrer/detectron2

Timo Furrer

commit sha 3ca4aa1cb537b51c277e2b682544129bab90ad55

Add `prob` argument for RandomRotations The `prob` argument for `RandomRotation` (similar to `RandomFlip`) allows to control the probability the rotation angle will be applied. This allows for the original image to be sampled with a certain probability.

view details

push time in a month

push eventtimofurrer/detectron2

Timo Furrer

commit sha fb47bfd76c220c7c69a134484f91ff06a4ae844f

Add `prob` argument for RandomRotations The `prob` argument for `RandomRotation` (similar to `RandomFlip`) allows to control the probability the rotation angle will be applied. This allows for the original image to be sampled with a certain probability.

view details

push time in a month

push eventtimofurrer/detectron2

Timo Furrer

commit sha bf0587263f049f8158afef9346eb359e31d59b3d

Add `prob` argument for RandomRotations The `prob` argument for `RandomRotation` (similar to `RandomFlip`) allows to control the probability the rotation angle will be applied. This allows for the original image to be sampled with a certain probability.

view details

push time in a month

startedsmddzcy/cv

started time in a month

push eventtimofurrer/detectron2

Timo Furrer

commit sha aee9edc7f567213d32e985924651f3214661fe60

Add `prob` argument for RandomRotations The `prob` argument for `RandomRotation` (similar to `RandomFlip`) allows to control the probability the rotation angle will be applied. This allows for the original image to be sampled with a certain probability.

view details

push time in a month

PR opened facebookresearch/detectron2

Add `prob` argument for RandomRotations

The prob argument for RandomRotation (similar to RandomFlip) allows to control the probability the rotation angle will be applied.

This allows for the original image to be sampled with a certain probability.

Thanks for your contribution!

If you're sending a large PR (e.g., >50 lines), please open an issue first about the feature / bug, and indicate how you want to contribute. See more at https://detectron2.readthedocs.io/notes/contributing.html#pull-requests about how we handle PRs.

Before submitting a PR, please run dev/linter.sh to lint the code.

+65 -15

0 comment

2 changed files

pr created time in a month

create barnchtimofurrer/detectron2

branch : feature/rotation-prob

created branch time in a month

push eventtimofurrer/detectron2

Yuxin Wu

commit sha 5f7bbed5bfacdf4e2597788c2b4c77fd6388a65e

use more kw-only arguments for configurable Reviewed By: rbgirshick Differential Revision: D21010578 fbshipit-source-id: 79d7e480ab730a79df14a3b8d6252e470a78afbe

view details

Aleksey Nozdryn-Plotnicki

commit sha 2c346e0b6ed95a6966564d588da8746f66adab38

filter_empty_instances with box threshold Summary: Minor change to enable passing a threshold through filter_empty_instances. Due to floating point issues I had boxes with width 1e-14 that passed the 0 threshold but blew up later. Also fixed the type hinting to float. These are floats. Pull Request resolved: https://github.com/facebookresearch/detectron2/pull/1198 Reviewed By: rbgirshick Differential Revision: D21030901 Pulled By: ppwwyyxx fbshipit-source-id: 5eedb9c4b801bcb85e5b040cd915628f0ec279e6

view details

Jing Huang

commit sha 24a273e20d60b21528e807ce969eb9d58a9260e4

Fix rotated box conversion that leads to incorrect IoU Summary: Fix https://github.com/facebookresearch/detectron2/issues/1207 When converting the (x_center, y_center, width, height, angle) to 4 points representation for rotated box, we need to take into account the fact that the right-handed image coordinate system is yOx, where y is top->down and x is left->right. The conversion should thus be calculated as follows: P_i(rotated) - P_c = R * (P_i(horizontal) - P_c), where R is the standard rotation matrix in right-handed system, P_i(rotated) = (y_i, x_i), i = 0, 1, 2, 3 are the 4 coordinates after rotation, P_i(horizontal), i = 0, 1, 2, 3 are the 4 coordinates before rotation. and we also make use of the fact that P_0 + P_2 = P_c * 2, P_1 + P_3 = P_c * 2 to compute P_2 and P_3 from P_0, P_1 and P_c substituting with the coordinates, we have: [y_0 - y_c, x_0 - x_c]^T = [[cos A, -sin A], [sin A, cos A]]^T * [h/2, w/2]^T [y_1 - y_c, x_1 - x_c]^T = [[cos A, -sin A], [sin A, cos A]]^T * [-h/2, w/2]^T which leads to the new formula we're using: x_0 = x_c + (sin A / 2) * h + (cos A / 2) * w y_0 = y_c + (cos A / 2) * h - (sin A / 2) * w x_1 = x_c - (sin A / 2) * h + (cos A / 2) * w y_1 = y_c - (cos A / 2) * h - (sin A / 2) * w Reviewed By: ppwwyyxx Differential Revision: D21036508 fbshipit-source-id: f33baa0b3a1353be0486495d89b10de12b06d711

view details

Yuxin Wu

commit sha e8e6dbf276d202902c2cd625ebf76936fc211e04

improve resnet.py Summary: 1. ResNetBlockBase->CNNBlockBase, because the concept seems general enough for any "blocks". BasicStem any other non-resnet blocks can benefit from the same abstraction as well. 2. Move `freeze()` into `ResNet`. Note that this can make it work with explicit arg, like `backbone=ResNet(...).freeze(2)`. But this is not enough for it to support type-dict style config. 3. update some docs Reviewed By: rbgirshick Differential Revision: D21014145 fbshipit-source-id: ba9079eda824e23a122648bc84cd6963c681e24a

view details

Yuxin Wu

commit sha f26697961573d39ea1a332cba851be74768d5d73

update docs; TridentNet fix; Summary: TridentNet was using old ROIHeads. fix https://github.com/facebookresearch/detectron2/issues/1220 fix https://github.com/facebookresearch/detectron2/issues/1214 Pull Request resolved: https://github.com/fairinternal/detectron2/pull/400 Reviewed By: rbgirshick, lyttonhao Differential Revision: D21045748 Pulled By: ppwwyyxx fbshipit-source-id: 562f1d5c2b81a2a72ec06b293cf5f1201d1acbb9

view details

Yuxin Wu

commit sha 02478d2e332d4ff501be136794b2919029ebb6be

use official SyncBatchNorm Reviewed By: rbgirshick Differential Revision: D21050713 fbshipit-source-id: 81c95876e20cbc79b061f8d16108579484d878f3

view details

Yuxin Wu

commit sha 83c96d95858b38c3b25491da338cd418b9bd642c

make unittest "test_export" public Reviewed By: rbgirshick Differential Revision: D21075118 fbshipit-source-id: a921c682796819daa8435347eca998da5516f824

view details

Yanghan Wang

commit sha e6b95f32a5d5543706b2868d5d780766737b9e1a

use PathManager and PIL to load sem_seg_file Reviewed By: ppwwyyxx Differential Revision: D21090513 fbshipit-source-id: 41ddc682deaa5966372fbe0746a13310e5dc1e53

view details

Yuxin Wu

commit sha 91159007074984aaa4559abf7d2abcf3db4802d2

remove some redundancy in tests Summary: Pull Request resolved: https://github.com/fairinternal/detectron2/pull/401 Reviewed By: rbgirshick Differential Revision: D21073942 Pulled By: ppwwyyxx fbshipit-source-id: 91d2f3aca54749a56b85cef0d9ca849f7f8ab7ab

view details

Yuxin Wu

commit sha d613277956599a34053775741f796defd2885752

make Boxes.cat a classmethod Summary: As suggested in D14660715. This should work for subclasses including RotatedBoxes and Caffe2Boxes This will also support empty list `Boxes.cat([])` Reviewed By: rbgirshick Differential Revision: D21021837 fbshipit-source-id: 675237a21c28aeb9cddfcb7219bba8174b5b8c40

view details

Yanghan Wang

commit sha d1c7a4a63b4e24b02564cbef0bd18d28f757bc0d

also return per-class IoU and ACC from SemSegEvaluator Summary: Update the SemSegEvaluator to also return per-class metrics. - The format follows detection's `"AP-class_name"` - It seems per-class results won't show in the copypaste, perhaps it won't make the log messy. - Change zero-initialization to nan, which represents non-valid. - Assume `stuff_classes` always exists and is valid. Reviewed By: ppwwyyxx Differential Revision: D21074976 fbshipit-source-id: 3458828f02ffd1b6a2afb8f4eb7e1d13d0f7d886

view details

push time in a month

startedwillmcgugan/rich

started time in a month

started8051Enthusiast/regex2fat

started time in a month

delete branch timofurrer/ramlient

delete branch : pyup-update-coverage-4.5.1-to-5.0.4

delete time in 2 months

push eventtimofurrer/ramlient

pyup-bot

commit sha 2ce21690764b31f33c51ef25ce5fbc22604e2312

Update coverage from 4.5.1 to 5.1

view details

push time in 2 months

create barnchtimofurrer/ramlient

branch : pyup-update-coverage-4.5.1-to-5.1

created branch time in 2 months

startedkarpathy/micrograd

started time in 2 months

issue openedfacebookresearch/detectron2

Visualize FPN and RPN outputs

I'm working with the Mask R-CNN, ResNet50 with FPN configuration.

I'd like to "visualize" the outputs from the FPN and RPN Heads, much like it's being shown in the following two blog posts:

Another good example is the matterport Mask-RCNN library:

I think an easy way to access those outputs from detectron2 would help to better understand the models we are training .

What do you think? Do you have any example where you are already doing this? If not, can you briefly elaborate on how one could achieve this (even if it's no easy accessible from detectron2 yet, but just to make it work) ?

created time in 2 months

issue commentjsbroks/coco-annotator

Exporting return empty images and annotations

@Aaronzinhoo Yes, that's correct - but that isn't too much of a problem .. The COCO data format documentation gives you pretty good start for that ;)

Aaronzinhoo

comment created time in 2 months

push eventtimofurrer/detectron2

Yuxin Wu

commit sha 738b253da6e073c6493bf0d84aca60567bc762fc

remove model_init.pbtxt Summary: it's not used, and sometimes this line fails if the weights file are too large. Reviewed By: dhuber23b Differential Revision: D20798810 fbshipit-source-id: 1f403c7031240dec8c923e951b7c818fd009599b

view details

Yuxin Wu

commit sha 95a8d6cf0e19a72a8e2ac14aff59a51bafce2ec2

fix cascade rcnn inference Summary: Should add cascade r-cnn to quick tests Pull Request resolved: https://github.com/fairinternal/detectron2/pull/397 Reviewed By: rbgirshick Differential Revision: D20796977 Pulled By: ppwwyyxx fbshipit-source-id: 0673817f9009b77ee5484e7b5163b8d2ad5f7fb6

view details

Yuxin Wu

commit sha 5a18ece9ec421e9466ee1d9f83a7483640cffa46

add Cascade R-CNN to tests Summary: n/a Reviewed By: rbgirshick Differential Revision: D20802974 fbshipit-source-id: eb89812484943257875f3a31f47d198bebd29cb9

view details

Timo Furrer

commit sha ffff8acc35ea88ad1cb1806ab0f00b4c1c5dbfd9

Clarify `DatasetEvaluator` and `DatasetEvaluators` API Summary: This change clarifies the usage of the `DatasetEvaluator` and `DatasetEvaluators` by being more precise and complete in the doc strings of the API. Pull Request resolved: https://github.com/facebookresearch/detectron2/pull/1133 Reviewed By: rbgirshick Differential Revision: D20803403 Pulled By: ppwwyyxx fbshipit-source-id: 9453efb6be3aceeeec32c3d42efcb0ba2ae81f01

view details

Yuxin Wu

commit sha 41ab438621d9f8a214096f45ee1c26881f020fe1

use configurable for box head Reviewed By: rbgirshick Differential Revision: D20808474 fbshipit-source-id: 7790a22f20c7c119f99a7d4a29246e52959c7f5c

view details

Yuxin Wu

commit sha 474c6b9ebc1a2475c5ffa2fe34c4a84ae4cacc63

fix deprecated use of torch.nonzero Summary: as_tuple is a required argument. trigger warnings otherwise. Reviewed By: rbgirshick Differential Revision: D20832106 fbshipit-source-id: 5a7440527d3c4f3366c0aeaa6483fc5cfc1b771e

view details

Yuxin Wu

commit sha 6e150b3adb3be1f18efd0bf940206d99c95dbbd8

use configurable for FastRCNNOutputLayers Reviewed By: rbgirshick Differential Revision: D20713231 fbshipit-source-id: f558a501e66ab5c5397b9376cef81f357f619c89

view details

Yuxin Wu

commit sha c5fe3c314fc2d8ffd594afaf0c63b2b499fbc461

support configurable in mask head Reviewed By: rbgirshick Differential Revision: D20866132 fbshipit-source-id: c54fdbe207eba86a87fc92866a16d83e79e59c0e

view details

Yuxin Wu

commit sha 483f775a0c4900e0ac8a563a99b44d738c756be0

remove the assertion on canonical level Summary: close https://github.com/facebookresearch/detectron2/issues/1160 It seems legit to me to have canonical level outside the range of min/max level and get clipped later. For example if the dataset is full of small objects, it might make sense to use a very small max level. Reviewed By: rbgirshick Differential Revision: D20866217 fbshipit-source-id: 4b28deec8d2893edd7fe256444f82a79ee77bf64

view details

c1aris

commit sha b8f4eebeb06827d49e31738c0997f919dac26aba

Support empty predictions when visualizing results Summary: This PR fixes a potential error when no predictions found in `tools/visualize_json_results.py`. Pull Request resolved: https://github.com/facebookresearch/detectron2/pull/1166 Reviewed By: rbgirshick Differential Revision: D20888849 Pulled By: ppwwyyxx fbshipit-source-id: a53d67a7f3ee63070defedb6ca3812d0ae5d6cc9

view details

Yuxin Wu

commit sha cd9ac61861e83856ed8854c98ebaf383b77950ae

update path for cityscapes semantic seg Reviewed By: rbgirshick Differential Revision: D20898635 fbshipit-source-id: 4ad2c4487040bb3e7429b1cac6f4692d57bc9dfe

view details

Yuxin Wu

commit sha 576cf9a5da8592f9aa6379e9f4b5592042d46488

make sure cityscapes's gt_dir is local Summary: otherwise, reading all the json file will be very slow Reviewed By: rbgirshick Differential Revision: D20906336 fbshipit-source-id: f21652a1532ff1bcbab5074c432317ca02354cc7

view details

Tarang Shah

commit sha 93693826c66d1200983d91ad89ebe3bc6cbfc393

Updated typos and wording for custom datasets page Summary: An small update to the documentation. Fixed some typos and sentences in the Custom Datasets page. Pull Request resolved: https://github.com/facebookresearch/detectron2/pull/1177 Reviewed By: rbgirshick Differential Revision: D20924485 Pulled By: ppwwyyxx fbshipit-source-id: 6e693de9021d77af3f4dc73af0f4e1c87ffe6d7e

view details

Marc Szafraniec

commit sha 63ea9001838bfa1078c60a878db9caf6c4246adb

Adding rotations to transform.py Summary: Adding rotations to transform.py Reviewed By: ppwwyyxx Differential Revision: D20488988 fbshipit-source-id: 39cb9ebf9aaa79b284b044d726dce9a55a8f7561

view details

Yuxin Wu

commit sha 7cbc2f4bead75b5d68cf5353cf4a34b1c7fc800a

support configurable in keypoint head Reviewed By: rbgirshick Differential Revision: D20944994 fbshipit-source-id: 15c67e8939957d87f8ba38118224d75b18bf468c

view details

Yuxin Wu

commit sha 8affda03a30d576e789aac49879b502757f7ab1e

use conv_dims/fc_dims as arguments to Box/Mask head. Summary: More general and simpler than "num_conv+conv_dim". Reviewed By: rbgirshick Differential Revision: D20956553 fbshipit-source-id: c380a037288cef9fd22f9650b94950eee48e7241

view details

Marc Szafraniec

commit sha 2c00dc055a2a6de033fad17b9375d9ceae39cfe2

Add Rotation Transform to transform_gen.py Summary: Adding rotations to transform_gen.py Reviewed By: ppwwyyxx Differential Revision: D20623641 fbshipit-source-id: 53b4c751e6c97b91120006bf41e620362e7d0d6e

view details

Yuxin Wu

commit sha 2ca36e3cbfb2c84c18502221564b629f3877e8be

update docs; assert all @configurable are experimental Summary: Pull Request resolved: https://github.com/fairinternal/detectron2/pull/398 Reviewed By: rbgirshick Differential Revision: D20980850 Pulled By: ppwwyyxx fbshipit-source-id: c2032312fa1c58c5fd3fc68e8f492de64edcc98c

view details

push time in 2 months

issue closedtimofurrer/w1thermsensor

Resolution = Precision ?

In order to change sensor resolution, the documentation refers to sensor.set_resolution() (Python case) w1thermsensor --resolution (CLI case)

But it does not work : the rights syntax seems to be sensor.set_precision() (Python case) w1thermsensor --precision (CLI case)

Side note : root is required for such resolution change, I think this may be written in the doc also.

closed time in 2 months

dusscloud

issue commenttimofurrer/w1thermsensor

Resolution = Precision ?

But it does not work : the rights syntax seems to be

Depending on the version of w1thermsensor that's true.

The precision phrasing was changed to resolution in: https://github.com/timofurrer/w1thermsensor/commit/3e91314dc892e8177011405669f0f3ef254a9146. Because resolution is actually the correct term here.

You can use 1.3.0 from PyPI to have to change. (might be the case that an older version is deployed in rasbian apt sources)

Side note : root is required for such resolution change, I think this may be written in the doc also.

Yes, good catch! I've added this in: https://github.com/timofurrer/w1thermsensor/commit/519010e47e3e7441e05c221e1c6e1cc2f2a83252

dusscloud

comment created time in 2 months

push eventtimofurrer/w1thermsensor

Timo Furrer

commit sha 519010e47e3e7441e05c221e1c6e1cc2f2a83252

Add hint about resolution change root privileges

view details

push time in 2 months

startedmitsuhiko/proximity-tracing

started time in 2 months

push eventtimofurrer/w1thermsensor

KlausHuber

commit sha afc8a52e802f52b58af9239c71ec8623f5336ae1

disconnected sensors still appear under /sys for a while but report w… (#76) * disconnected sensors still appear under /sys for a while but report wrong values * add unit tests for partially disconnected sensor which only report zero values * add empty line to make tests happy Co-authored-by: Klaus Huber <git@klaus-huber.com>

view details

push time in 2 months

PR merged timofurrer/w1thermsensor

disconnected sensors still appear under /sys for a while but report w…

…rong values If sensors are disconnected they remain in the system for a while until they are detected as disconnected. If they are read in this time frame, they report wrong values of zeros only.

+21 -2

3 comments

3 changed files

KlausHuber

pr closed time in 2 months

pull request commenttimofurrer/w1thermsensor

disconnected sensors still appear under /sys for a while but report w…

Awesome!

Thanks for the contribution! :tada:

KlausHuber

comment created time in 2 months

startedPaddlePaddle/VisualDL

started time in 2 months

issue commentfacebookresearch/detectron2

RandomRotate for training augmentation

@ppwwyyxx Thanks for implementing this. When will this be available from https://dl.fbaipublicfiles.com/detectron2/wheels/cu100/index.html ?

timofurrer

comment created time in 2 months

push eventtimofurrer/pycococreator

Timo Furrer

commit sha feb8d5ce842bb09e60b3ea7bd2731681d9bc5880

Add function to create multiple annotations from the same image. Closes #29 The functionality is implemented in a new function `create_annotation_infos` with a similar API as the already existing `create_annotation_info`. I think it makes sense to have a dedicated function for this use-case. An example is at `samples/multi-annotations`.

view details

push time in 2 months

push eventtimofurrer/pycococreator

Timo Furrer

commit sha b0e03f15c8cca790da64ed8988aaa5715ae04c43

Support already labelled images

view details

push time in 2 months

push eventtimofurrer/pycococreator

Timo Furrer

commit sha 623f12017c676399488a42b2bf3a727ad02972a8

Support already labelled images

view details

push time in 2 months

push eventtimofurrer/pycococreator

Timo Furrer

commit sha 8e3d7b4e866b9fa39384fc76e3499e41f1d8489d

Support already labelled images

view details

push time in 2 months

more