archive-gh-me/robosat
1
0
Fork 0
mirror of https://github.com/mapbox/robosat.git synced 2025-12-08 20:14:01 +00:00
Code Issues Projects Releases Wiki Activity

Commit Graph

  • cbb1c73328 Fixed slowdown in graph building and merging by pre-initializing pyproj projections master venkatesh10 2020-08-25 17:36:47 +05:30
  • b593a6b129 Adds development status note daniel-j-h 2020-05-08 09:46:33 +02:00
  • e4598404d7 updated pip requirements to be more strict Devin A. Conley 2020-01-30 05:12:45 -06:00
  • b8cfa36cf2 add rs entrypoint to pip Devin A. Conley 2020-01-30 05:06:56 -06:00
  • e7bbc79911 add package version Devin A. Conley 2020-01-28 18:46:26 -06:00
  • fb0ed891ab add setup.py Devin A. Conley 2020-01-27 19:13:22 -06:00
  • 3ae9c7e1d5 corrects spelling of 'PREPARATION' Michael Norelli 2019-11-11 13:32:50 -08:00
  • 990db527a1 Compiles Pillow-SIMD with libwebp support, closes #193 Daniel J. Hofmann 2019-11-10 17:46:09 +01:00
  • 1f2d271835 Fixed deprecated pyproj syntax venkatesh10 2019-10-31 14:41:47 +05:30
  • 97f83fb0e9 Fixes contour extraction, resolves #179 Daniel J. Hofmann 2019-09-14 15:41:57 +02:00
  • e844099038 Compiles latest libjpeg-turbo release for AVX2 Daniel J. Hofmann 2019-09-14 14:34:22 +02:00
  • a50367a142 Bumps pip and pip-tools to latest stable release Daniel J. Hofmann 2019-09-14 14:33:38 +02:00
  • 9847877a9e Provides convenient Makefile for development Daniel J. Hofmann 2019-08-14 23:17:12 +02:00
  • 910a7fd0ff Uses libjpeg-turbo and pillow-simd with AVX2 support Daniel J. Hofmann 2019-08-14 23:16:49 +02:00
  • 4a3c1237bb Uses EfficientNetB0 as segmentation model encoder backbone efficientnet Daniel J. Hofmann 2019-07-28 18:00:14 +02:00
  • 1d0cf506cd Updates readme with recent diary posts about robosat Daniel J. Hofmann 2019-06-08 18:07:02 +02:00
  • a855a8a009 Triggers travis to docker build and push, see #153 v1.2.0 Daniel J. Hofmann 2019-05-31 18:42:25 +02:00
  • 8783c312d2 Adds feature pyramid attention (FPA) module, resolves #167 fpa Daniel J. Hofmann 2019-05-30 23:35:59 +02:00
  • 27833dae80 Removes adam's l2 regularization until fixed in pytorch#4429 Daniel J. Hofmann 2019-05-30 19:51:37 +02:00
  • e31b206b4c Fixes travis trusty packages Daniel J. Hofmann 2019-05-30 16:00:24 +02:00
  • ff7eb208de Upates readme with instructions on how to use the docker images Daniel J. Hofmann 2019-05-23 21:41:12 +02:00
  • a5ea73c2e7 Pins dependencies checking hashes with pip-tools Daniel J. Hofmann 2019-05-30 11:13:54 +02:00
  • aa57aaaa13 Installs the torchvision wheel on travis Daniel J. Hofmann 2019-05-29 19:38:08 +02:00
  • a9caea75a9 Updates pip on travis Daniel J. Hofmann 2019-05-29 19:37:52 +02:00
  • 54e20dc379 Updates dependency management in docker images Daniel J. Hofmann 2019-05-28 22:47:12 +02:00
  • 20850edbce Replaces deprecated upsample with interpolate in model Daniel J. Hofmann 2019-05-27 20:56:14 +02:00
  • 0c1bdc9dfa Make sure there are tiles in the dataset upfront, resolves #141 Daniel J. Hofmann 2019-05-27 20:21:22 +02:00
  • c8fe2d35ac Uses NaN in metrics in case of errors Daniel J. Hofmann 2019-05-27 20:34:19 +02:00
  • e3310bed0e Fixes potential div-by-zero in miou metric Daniel J. Hofmann 2019-05-27 20:33:59 +02:00
  • e69fd80b2e fixed typo Erik 2019-05-23 22:34:17 +02:00
  • bde13538ab Splits docker build into cache-able layers for faster iteration Daniel J. Hofmann 2019-05-23 01:22:00 +02:00
  • 5c0ada68bf
         Merge pull request #151 from daniel-j-h/prefetch-model jqtrde 2019-05-22 12:21:49 -04:00
  • ee2a35c4c1 Bakes pretrained model into docker image Daniel J. Hofmann 2019-05-22 00:52:01 +02:00
  • c4b70a93fa
         Merge pull request #144 from daniel-j-h/update-deps jqtrde 2019-05-19 15:08:36 -04:00
  • c825b57075
         Merge pull request #148 from daniel-j-h/batch-extract jqtrde 2019-05-19 15:06:58 -04:00
  • 916e10663d
         Merge pull request #149 from daniel-j-h/entrypoint jqtrde 2019-05-19 15:06:09 -04:00
  • 61f9f95f11 Uses rs command as default docker entrypoint Daniel J. Hofmann 2019-05-18 09:21:46 +02:00
  • c9cb7c3999 Implements batched feature extraction, resolves #120 Daniel J. Hofmann 2019-05-17 14:38:31 +02:00
  • bf81550540 Uses cuda 10, pytorch 1.1 Daniel J. Hofmann 2019-03-30 20:21:25 +01:00
  • 9f79d27b10 configurable input channels issue/56 lily-chai 2018-11-09 13:37:02 -08:00
  • a8e0e3d676 Removes ununsed multi-class PSPNet model Daniel J. Hofmann 2018-10-24 16:40:47 +03:00
  • 8e99e5b776 Allows for batch-rasterization in rs rasterize, closes #25 Daniel J. Hofmann 2018-10-23 19:35:10 +03:00
  • 5277aeb539 Bumps requirements for latest PyTorch 0.4.1 Daniel J. Hofmann 2018-08-21 09:33:41 +02:00
  • 62840febb8 Implements topology-aware loss function; resolves #133 issue/133 Daniel J. Hofmann 2018-10-23 16:54:40 +03:00
  • 8876768ffd Assert image resolution has to be divisible by 32 for resnet in fpn issue/60 Bhargav Kowshik 2018-08-10 10:02:55 +05:30
  • 7d8bed505e Implements Feature Pyramid Network (FPN), closes #60 Daniel J. Hofmann 2018-07-10 18:36:04 +02:00
  • b7f6ebf79d Disables shared-memory workers by default, see #126 Daniel J. Hofmann 2018-10-09 13:19:01 +02:00
  • dcea15d97b Division by zero handling in metrics. Cf https://github.com/mapbox/robosat/issues/124 ocourtin 2018-10-08 19:18:00 +02:00
  • 036e2aef33 Requires a sub-command from users, closes #102 Daniel J. Hofmann 2018-10-02 15:24:41 +02:00
  • 0f4185ba9b Filters buildings based on location tags, closes #115 Daniel J. Hofmann 2018-10-02 15:32:09 +02:00
  • b2e8d71ad2 Upon choosen loss, weight dataset values could be mandatory. Olivier Courtin 2018-09-26 18:24:21 +02:00
  • f551f91545 Fix wrong hyperparameter log label. Update related tabs ocourtin 2018-09-24 00:09:39 +02:00
  • 862b6ea8bd A bit more generic Log handling, allowing using also stderr as optionnal output messages ocourtin 2018-09-24 00:01:59 +02:00
  • db74b08209 black session ocourtin 2018-09-23 23:49:11 +02:00
  • 5248d9266e Put log stuff in a dedicated module. First step to something more generic ocourtin 2018-09-21 14:42:39 +02:00
  • 9c058a80fc Black session ! ocourtin 2018-09-21 14:06:53 +02:00
  • 248c299f59 Default loss to Lovasz rather than CrossEntropy. Don't communicate on focal at this point ocourtin 2018-09-21 14:05:11 +02:00
  • fead5e372f Update model structure change in export too. Related to resume train feature, cf ec17c32 ocourtin 2018-09-21 09:28:58 +02:00
  • 25deb1ede4 Update config file comment related to Lovasz loss ocourtin 2018-09-20 18:00:16 +02:00
  • b192e237dd Add Lovasz Loss ocourtin 2018-09-20 17:51:33 +02:00
  • ef4fc72da0 Create a dedicate Log Class. Add Hyperparameters logs. Remove useless pixel accuracy metrics, and add MCC metric. Small cleanup changes ocourtin 2018-09-19 21:05:25 +02:00
  • 334952fad0 As loss become too small, it leads to numerical instability. Fix this, to be both more reliable, and more efficient, throught max(m_iou, x_entropy). Compute loss per image rather than per batch before mean(). ocourtin 2018-09-19 21:01:01 +02:00
  • 6925e883e7 Remove size_average and ignore_index options. ocourtin 2018-09-14 23:44:03 +02:00
  • 95de2648ed add IoU foreground and background computation. Use foreground one userland ocourtin 2018-09-14 22:57:23 +02:00
  • c738a06994 Use both mIoU and CrossEntropy in the same loss. If not, with mIoU alone, only contours are learned for surfacic features. ocourtin 2018-09-14 22:53:05 +02:00
  • e3ffb61a11 Add log text handling. ocourtin 2018-09-14 15:03:54 +02:00
  • 94491cc174 Add resume optionnal train parameter: Allow to retrieve optimizer setting from a previous train. Modify/enhance structure saved in .pth files to do so. Lead to brake compatibility with previous .pth files generated ocourtin 2018-09-14 14:24:04 +02:00
  • 3fec9ed05b remove useless import and blank lines ocourtin 2018-09-14 05:27:38 +02:00
  • 9b3c7084e2 Add mIoU loss. Add [opt][loss] config parameter, to select which loss to use ocourtin 2018-09-14 00:09:33 +02:00
  • 4a486c33d7 mIoU metric code refactor. Add Pixel accuracy metric. ocourtin 2018-09-13 17:00:58 +02:00
  • cf8f9a8093 Remove useless requires_grad = False. Thanks to Daniel-J-H to point this one. ocourtin 2018-09-11 18:34:44 +02:00
  • 5d273e0093 rewrite it slightly to be still readable by human after black formating. ocourtin 2018-09-08 10:44:09 +02:00
  • 5983577b17 format stuff. black no option ocourtin 2018-09-07 16:34:03 +02:00
  • e2c849fc70 Switch confusion matrix computation from CPU to GPU. Add workers support in train. Perf in mind. ocourtin 2018-09-07 15:39:09 +02:00
  • a6b87997da Add checkpoint optionnal argument in train tool. Allow to retrain or fine tune an existing model Olivier Courtin 2018-08-12 18:58:46 +02:00
  • 7a85d9d384 Generalizes post-processing handlers across zoom levels, closes #51 issue/51 Daniel J. Hofmann 2018-07-20 12:16:15 +02:00
  • 68679933ad Fixes flake8 issues in road handler Daniel J. Hofmann 2018-09-26 18:43:31 +02:00
  • af1edfa63d Formats road handler with black Daniel J. Hofmann 2018-09-26 14:54:16 +02:00
  • 05be4a7911 Guards against osm road data issues, resolves #114 Daniel J. Hofmann 2018-09-26 14:38:07 +02:00
  • a9ad0d348d Decorate code DragonEmperorG 2018-09-12 18:40:54 +08:00
  • b29cc55ece Make the code robust and format it DragonEmperorG 2018-09-12 11:07:33 +08:00
  • 0bf0ef09c0 Create highway_attributes dict and add EARTH_MEAN_RADIUS const DragonEmperorG 2018-09-08 10:29:18 +08:00
  • d6f88d9158 Add automatically generate a road training set based on OpenStreetMap DragonEmperorG 2018-09-07 19:13:02 +08:00
  • be9d6a6bdc Formats code with black v1.1.0 Daniel J. Hofmann 2018-09-12 11:05:04 +03:00
  • d073f4a797 Remove useless requires_grad = False. Thanks to Daniel-J-H to point this one. ocourtin 2018-09-11 18:34:44 +02:00
  • b8f5532354 rewrite it slightly to be still readable by human after black formating. ocourtin 2018-09-08 10:44:09 +02:00
  • 2dafd48bf1 format stuff. black no option ocourtin 2018-09-07 16:34:03 +02:00
  • 5e191c4d45 fix missing num_workers handling ocourtin 2018-09-07 16:31:10 +02:00
  • 284e879aab remove scikit learning dependancies ocourtin 2018-09-07 16:19:05 +02:00
  • 9a8d91022c Switch confusion matrix computation from CPU to GPU. Add workers support in train. Perf in mind. ocourtin 2018-09-07 15:39:09 +02:00
  • ea6de6e936 Update merge.py Alexander Usoltsev 2018-09-10 11:25:09 +02:00
  • 078afcca2c Address issue #103 Alexander Usoltsev 2018-09-07 16:04:42 +02:00
  • 868a5aa296 compatable no_grad decorator for pytorch 0.4.0 Jesse Hinrichsen 2018-09-04 14:47:02 +02:00
  • 7b7b1d18a3 no_grad() as decorator on validate() instead of in main() Jesse Hinrichsen 2018-09-04 14:02:43 +02:00
  • a73c52cd2a added torch.no_grad() (available since pytorch 0.4.0) to avoid 'out of memory'-issue during validation Jesse Hinrichsen 2018-08-29 10:00:54 +02:00
  • d43d69feb7 Add checkpoint optionnal argument in train tool. Allow to retrain or fine tune an existing model Olivier Courtin 2018-08-12 18:58:46 +02:00
  • 495edfba66
         Merge pull request #104 from mapbox/assert-resolution Bhargav Kowshik 2018-08-10 09:59:39 +05:30
  • d65190bb13 Assert image resolution has to be divisible by 32 for resnet Bhargav Kowshik 2018-08-09 22:31:37 +05:30
  • d33ae08001
         Merge pull request #101 from mapbox/assert-messages Bhargav Kowshik 2018-08-09 13:05:44 +05:30
  • 729b212e1f Use images and label instead of inputs and target Bhargav Kowshik 2018-08-08 20:43:00 +05:30