Chainercv Versions Save

This release is not a substantial one. This release includes a patch to failing doc build. https://github.com/chainer/chainercv/pull/885

This release only supports Chainer v6 and not Chainer v5. For those users using Chainer v5, please update the version of Chainer or use ChainerCV v0.13.

Features

• add MaskRCNN #781
• DeepLabV3+ #757 (Thank you 69guitar1015)

No Compat Changes

• FPN related implementation changed (e.g. head.py is renamed to bbox_head.py) #781
• When chainercv.global_config.cv_read_image_backend == 'cv2', ChainerCV is changed to raise error when cv2 is not installed. #852
• Point data format has changed from (K, 2) to [(K, 2)]
  • change interface of point/transforms #808
  • Change interface of vis_point #809
• Requires name as key for SlicableDataset #833

New CI

We upgraded our CI to pfnCI (PFN’s in-house CI system). Now GPU tests and example tests run (related)

• pfnCI (pytests) #823
• pfnCI (examples tests) #842
• Enable model cache in pfnCI #882
• pfnCI on CuPy repository #865
• pfnCI on Chainer repository #854
• Include dockerfile for PFN CI and add tests with chainer master #848

Enhancements

• Support Chainer-style initialization in TransformDataset #832
• add arg --dataset in examples demo #877 (Thank you knorth55)
• refactor eval_semantic_segmentation #875 (Thank you knorth55)
• refactor eval_imagenet.py #874 (Thank you knorth55)
• refactor eval codes #871 (Thank you knorth55)
• Support MN mode for evaluators #868 (yuyu2172)
• DetectionVisReport save tight figure #863 (Thank you 23pointsNorth)
• use pure_nccl in examples/ #845 (Thank you knorth55)
• add fcis examples_tests #840 (Thank you knorth55)
• Create .lock directory #839
• Add argument of alpha to MixUpSoftLabelDataset #836 (Thank you crcrpar)
• Change links to override forward #835 (Thank you crcrpar)
• FPN - Add min/max size params for pre-processing. #830 (Thank you 23pointsNorth)
• Support file-like object in image I/O #824
• add left time in progress hook #817 (Thank you knorth55)
• Show the cause of ValueError #814 (Thank you ktns)
• args --batch-size -> --batchsize in examples #811 (Thank you knorth55)
• change assert_is_point, assert_is_point_dataset, CUB datasets #806
• Change doc style of eval_* #804
• Change doc style of vis_* #803
• add sort_by_score in vis_bbox #801 (Thank you knorth55)
• add scale_mask #799
• rename psroi_pooling -> ps_roi_average_pooling and refactor #795 (Thank you knorth55)
• make mask_to_bbox not copy to CPU #792
• add eval_instance_segmentation_multi.py #789 (Thank you knorth55)
• Lock dataset directory before extraction #788
• Unify eval detection #786
• use scales in faster_rcnn_tests #783 (Thank you knorth55)
• simplify coco instance datasets #778
• add sort_by_score option to vis_instance_segmentation #777
• set n_gpu * 0.0005 as fcis lr default #776 (Thank you knorth55)
• Fix a typo #864 (Thank you mitmul)
• Change the condition in mask\_to\_segm\(\) to eliminate invalid bboxes #860 (Thank you mitmul)
• Use with block instead of setting global config #857
• Set allow_pickle=True #853
• fix typo #851 (Thank you okdshin)
• Fix SSD train_multi.py #844
• fix missing link in doc #834
• Fix ImageNet example #831 (Thank you keisukefukuda)
• Fix inference configuration on ADE20K in DeepLab #826 (Thank you 69guitar1015)
• Add link to deeplab page in document #805
• fix resize when 0 length image is given #798
• fix eval_instance_segmentation_coco when pred\_masks is empty #780
• Prevent mutating input of FeaturePredictor.prepare #779
• add test_xception.py #879 (Thank you knorth55)
• update semantic_segmentation/README.md #878 (Thank you knorth55)
• Update readme #872 (Thank you knorth55)
• Update to Chainer 6.0 #867
• use - instead of _ #866
• fix instance_segmentation readme #862 (Thank you knorth55)
• Make Travis CI smaller #855
• Use Ours and Original when reporting performance #846
• update readme #837 (Thank you knorth55)
• Add deeplab pretrained test #827 (Thank you 69guitar1015)
• Fix README #822 (Thank you Hakuyume)
• Fix docstring of DirectoryParsingLabelDataset #820 (Thank you iwiwi)
• Fix typo: Citåyscapes -> Cityscapes #813 (Thank you soskek)
• Use table for the documentation of bbox transforms #807
• add PS RoI Max Pooling #797 (Thank you knorth55)
• add PS RoI max/average align #796 (Thank you knorth55)
• cache conda for travis build #791 (Thank you knorth55)
• Add chainermn support to apply_to_iterator #785
• test FCISResNet101 loss #784 (Thank you knorth55)

Spotlight

• Feature Pyramid Networks #685
• Add ResNet training code #436
• Add training part of PSPNet #634
• Add FCIS coco demo, training and evaluation #724 (Thank you knorth55!)
• Add YOLOv2 tiny #711

New Features

• Add COCOSemanticSegmentationDataset #736
• Add interpolation and fill options to rotate #739
• Add make_shift #726

API change

• change range of labels for ADE20K #743

Enhancement

• Enable cv2/InfiniBand hack #769
• Perform batch size check before converting scales #762 (Thank you 23pointsNorth!)
• Unify evaluation for semantic segmentation #749
• Use PIL and cv2 for rotate #740
• use scales (tuple of floats) in region_proposal_network #729 (Thank you knorth55!)
• Remove version specification of OpenCV #728
• raise error when year=2017 and split in [minival, valminusminival] #725 (Thank you knorth55!)
• use cupy.RawKernel in non_maximum_suppression #723 (Thank you knorth55!)
• Use f.convert('L') to read grayscale in read_image() #695 (Thank you 69guitar1015!)
• Fix PickableSequentialChain.copy() #764 (Thank you ktns!)
• Fix docstring of ProposalTargetCreator #758 (Thank you nai62!)
• return empty bbox when input mask is empty in mask_to_bbox #754 (Thank you knorth55!)
• fix typo in mxnet2npz.py script #752 (Thank you knorth55!)
• Fix resnet50 fb url #751 (Thank you knorth55!)
• Fix VOCSemanticSegmentationDataset bug with cv2 #746
• Stop extracting dataset when ADE20k is extracted already #745
• Return rgb when rgba is loaded #734
• Update README for FPN #771
• doc source link to github source page #768 (Thank you knorth55!)
• Fix rst of YOLO #767
• Add truncated\_index into params of crop_bbox #761 (Thank you ktns!)
• Fix typo in SSD's test_train_multi.sh #760
• remove use_pretrained_class_weights #755
• Refactor PSPNet, support Res50 backbone and ImageNet pretrained weights #748
• Use chainer.global_config in doc #742
• add table in COCOInstanceSegmentationDataset #741 (Thank you knorth55!)
• fix typo in transforms/image/resize_contain.py #733 (Thank you knorth55)
• Fix typo in transforms/image #732 (Thank you takaaki82!)
• Fix typo in links connection #731 (Thank you takaaki82!)
• Fix README #722
• fix typo, 'bounding box' #721 (Thank you disktnk!)
• Fix doc of TupleDataset #720
• Make backends of read_image and resize selectable #713
• Add examples tests #708

This release only supports Chainer v5 and not Chainer v4. For those users using Chainer v4, please update the version of Chainer or use ChainerCV v0.10.

Spotlight

• Add FCIS training link #620 (Thank you knorth55!)
• Implement SE-ResNeXt. #659 (Thank you g-votte!)
• Implement SE-ResNet #655 (Thank you g-votte!)

New Featrues

• add InstanceSegmentationCOCOEvaluator #674 (Thank you knorth55!)
• Add COCO instance segmentation evaluation #671 (Thank you knorth55!)
• Add COCO Instance segmentation dataset #665 (Thank you knorth55!)
• Add DetectionCOCOEvaluator #648
• Add COCOBboxDataset #453
• Add eval_detection_coco #456
• add rotate_bbox #692 (Thank you knorth55!)
• add rotate transforms for image #690 (Thank you knorth55!)

API Change

• The order of the arguments of resize_contain changed #660

Implemented Enhancements

• Remove angle range in rotate_bbox and rotate #699 (Thank you knorth55!)
• fix travis with Chainerv5 #717
• Fix failing CUB tests #716
• Release objects in tests #703
• Use return_indices in SegNet#704
• allreduce -> allreduce_obj #705
• nosetests -> pytest #706
• add run_module #707
• Fix installation of pycocotools #709
• Fix division in eval_cityscapes_multi.py #710
• fix failing flake8 #712
• fix doc badge link #700 (Thank you knorth55!)
• Fix DeprecationWarning #702 (Thank you ktns!)
• make mask_to_bbox support all False mask #701
• Use COCO2017 #694
• refine transforms_tests/bbox_tests #689 (Thank you knorth55!)
• assert flip image ndim == 3 #688 (Thank you knorth55!)
• Refactor FCIS codes #672 (Thank you knorth55!)
• Fix doc and style of SEResNet #656
• Use unittest.skipUnless #641
• Support bool in SliceableDataset #638
• Fix coco instance dataset test #687 (Thank you knorth55!)
• use kwargs for ResBlock in pspnet #686 (Thank you knorth55!)
• fix cocoapi repo in .travis.yml #675 (Thank you knorth55!)
• fix typo in examples/fcis/README.md #658 (Thank you knorth55!)
• weight -> width in README #657
• fix ssd.multibox_loss with comm #654
• Do not use lambdas for add_getter #652
• Fix downloading COCO data #650
• Use http in OnlineProductsDataset #649
• fix eval_detection_coco to work when classes are missing #647
• Fix typo in COCOBBoxDataset #639
• deprecate random_rotate #693
• Fix typo in caffe2npz.py #678
• Add Mxnet model to Chainer convert script for FCIS #664 (Thank you knorth55!)
• Fix doc of eval_detection_coco #645
• Add Ogawa as author #642
• fix typo in faster_rcnn_train_chain #640 (Thank you knorth55!)
• Change assert_is_bbox_dataset to pass with zero-sized bbox #452
• Check that zero-sized bounding box can be generated by generate_random_bbox #451
• Check that assert_is_bbox passes 0 sized bbox #450
• Add a test for OnlineProductsDataset #400

This release only supports ChainerMN v1.3.

Spotlight Feature

We added various algorithms.

• Inference code of FCIS ResNet101 #568 (Thank you knorth55 !)
• Inference code of YOLOv3 #586
• Inference code of YOLOv2 #586
• Inference code of PSPNet #610

We wrote our license term more explicitly (check here).

• Note pretrained models' license #631

We use AWS to host pretrained weights

• Use official hosting #633

API Changes

• Stop using nobias=False for ResNet101 and ResNet152 with the He architecture #621

New Features

• multi GPU evaluation of semantic segmentation #629

New tutorials

• Add fine-tuning tutorial (#601, tutorial)

Implemented enhancements:

• ChainerMN 1.3 #619
• Use loc, obj, conf in YOLO #618
• use --foo-bar style in argparse #617
• args --pretrained_model -> --pretrained-model #616 (Thank you knorth55!)
• fix typo in examples/instance_segmentation/eval_sbd.py #615 (Thank you knorth55!)
• Chainer 5.0.0b1 #614
• run_module missing in test_fcis_resnet101 #611
• add instance segmentation voc evaluator extensions #609 (Thank you knorth55!)
• update fcis variable names #608 (Thank you knorth55!)
• Simplify mask voting #607
• Delete unnecessary assignment in ProposalCreator #604 (Thank you t2kasa!)
• Improve pretrained tests #599
• Delete attr.disk #598
• misc fix for YOLOv3 #593
• add prepare_pretrained_model #591
• update fcis variables names #587 (Thank you knorth55!)
• Use prob more strictly inside SSD and Faster R-CNN #584
• add more badges in README #581 (Thank you knorth55!)
• Add exclude option to style_checker.py #441
• Fix bbox_to_mask to return np.float32 #625 (Thank you knorth55!)
• add prob clipping for mask_voting #606 (Thank you knorth55!)
• Fix hacking version #597
• Let YOLOv3.predict accept input image other than float32 #594 (Thank you ronekko!)
• Add FCIS to README #632
• Add PSPNet to README #630
• Fix pickle issue in train_multi.py in SSD #623
• Add YOLO to README #590
• str --> string #583

This release only supports Chainer v4 and not Chainer v3. For those users using Chainer v3, please update the version of Chainer or use ChainerCV v0.8.

Spotlight Feature

• Slicable datasets (chainercv.chainer_experimental.datasets.slicable, tutorial). All dataset classes in ChainerCV support the functionality.

API Changes

We renamed APIs relating to points.

• Change name to CUBPointDataset and add tests #528
• Change naming conventions in transforms for points #526
• vis_keypoint --> vis_point #525

Some public attributes of datasets have been changed

• Use SliceableDataset #457

We improved and renamed the method apply_to_iteartor

• apply_prediction_to_iterator --> apply_to_iterator #523

We changed the interface of vis_semantic_segmentation to be consistent with the rest of the visualizers.

• Change vis_semantic_segmentation to take image #576

New Features

• Add SliceableDataset #454, (Check the tutorial!)
• SSD train example with multiple GPUs #504 (Check the training script!)
• Add ResNet #321
• Add SiameseDataset #505
• Support MixUpSoftLabelDataset. #497 (thank you fukatani!)
• Add random_size_crop #432

Utilities for Instance Segmentation

• add VOC instance segmentation evaluation #554 (thank you knorth55)
• Add vis instance segmentation func #541 (thank you knorth55)
• Add VOC and SBD instance segmentation dataset #540 (thank you knorth55)
• add mask_iou #552 (thank you knorth55)
• Add PSROIPooling Function #545 (thank you knorth55)

Implemented enhancements:

• fix _check_available in SBDInstanceSegmentationDataset #577
• plot --> plt #574
• Change the author email address #573
• Chainer 4.0.0 #572
• Support keys assign in GetterDataset #571
• remove _chainermn_available #563
• Make dataset dependent part clear for SegNet example #557
• Change InstanceSegmentationDataset interface #555
• Add dilate and bn_kwargs to Bottleneck and ResBlock #549
• Update interface of vis_instance_segmentation and add color option #546
• Do not use vis_image in vis_instance_segmentation #544
• Support None option for activ #529
• Add assert_is_point #524
• list()/dict() -> []/{} #522
• Allow padding to be different in y and x directions for tile_images #512
• fix test_assert_is_point #566
• Fix ade_label_names and ade_label_colors #556
• Fix build scripts not to include binary files and .pyx files #550 (thank you gwtnb)
• Fix ResNets when initialW is specified #548
• Fix the range of offset in random_crop #534 (thank you akitotakeki)
• Fix pretrained weight for ResNet #533
• Dynamically import matplotlib.pyplot #519
• Fix a test failure due to a rounding error #513
• Support list in sliceable datasets #575
• Improve sliceable tutorial #570
• bach -> batch #565
• Support using vis_bbox together with vis_instance_segmentation #560
• update LICENSE #551 (thank you knorth55)
• fix doc #542 (thank you fukatani)
• Add a simple example for SiameseDataset #531
• add tuple check to style_checker #530
• Add assert_is_point_dataset #527
• Fix typos in detection tutorial #515
• Fix example code in a detection tutorial #514 (thank you koki0702)
• Add dilate option and MultiNodeBatchNormalization to Conv2DActiv and Conv2DBNActiv #494

This release only supports Chainer v3 and not Chainer v2. For those users using Chainer v2, please update the version of Chainer or use ChainerCV v0.7.

API Changes

• The default value of option data_dir for CityscapesDataset has changed from None to 'auto'. #448

New Features

• Add ProgressHook #473
• Add CityscapesTestImageDataset #440

New Tutorial

• Add a tutorial on Object Detection #434

Enhancements and Bug Fixes

• Test Chainer 3.2.0 on Travis #499
• Use ProgressHook in segnet #498
• Improve reference of tutorial #479
• Fix doc of download to refer to chainer.dataset.download properly #478
• Improve appearence of y_offset and x_offset in doc #475
• Use 'data' as an attribute for BlobProto for SSD #491
• Retruns -> Returns #483
• Fix flip_keypoint to accept keypoints on the edge #481
• Fix grammatical mistakes in doc #477
• Fix refs to NMS in SSD doc #476
• Fix doc of crop_bbox #474
• Fix variable names of eval_imagenet #472
• Fix datasets docs #464
• fix bbox format in flip_bbox #458 (Thank you @knorth55)
• Add testing.run_module to eval_detection_voc #455
• Fix doc of bbox2loc and loc2bbox #439
• Add attr.disk #493
• Use math notations for bounding box coordinates #485
• Improve reference links #482
• Fix documentation of SegNet #471
• Improve the top page of the documentation #470
• Fix letter case in README #469
• Improve code examples of detection.rst #467
• Fix datasets docs #466
• Simplify README #465
• fix setting parameters of random_distort #461 (thank you @peisuke)
• avoid zero division in bbox2loc #459 (thank you @knorth55)
• Fix doc of Conv2DBNActiv #447
• Chainer 3.0.0 #445

This release is targeted at Chainer v2.

Major Improvements

• We added an evaluation script for ImageNet trained models (link).

API Changes

• CUBLabelDataset and CUBKeypointDataset
  • Add return_prob_map option to CUBDataset and delete return_mask option #443
  • Add return_bb option to CUBDatasets and delete crop_bbox option #399
• VOCDetectionDataset
  • Use bbox* instead of detection* for datasets and label_names (e.g. VOCDetectionDataset -> VOCBboxDataset. The examples for SSD and Faster R-CNN no longer work with the previous versions) #419
• vis_label
  • Change function name: vis_label -> vis_semantic_segmentation #420
  • Show in legend only classes that appear in an image for vis_semantic_segmentation #345
• FasterRCNNVGG16
  • Organization of weights for Faster R-CNN has changed. The weights trained by the previous version of ChainerCV can not be loaded to the new model.#265
• Set train=False in predict #407

New Features

• Add ADE20K dataset #429
• Add tile_images #422
• Add Cityscapes semantic segmentation dataset #392
• Add FeaturePredictor #383
• Add Conv2DBNActiv #390
• Add Conv2DActiv #384
• Add ImageFolderDataset (Renamed to DirectoryParsingLabelDataset later) #271
• Add write_image #377
• Add sequential_feature_extractor (Renamed to PickableSequentialChain later) #342
• Add VGG16 #265

Enhancements and Bug Fixes

• 3.0.0b1 -> 3.0.0rc1 #442
• Fix doc of bbox2loc and loc2bbox #439
• Refactor faster_rcnn.predict #438 (@knorth55 thanks!)
• Change variable names for Cityscapes to be specific to semantic segmentation #437
• Use old conda version to make travis work again #431
• Change SentialFeatureExtractor to PickableSequentialChain #427
• fix a typo in assert_is_semantic_segmentation_dataset #424
• Fix doc of semantic segmentation datasets to not use "task" #423
• Make a directory for chainercv.utils.image #421
• Raise helpful warnings with Faster R-CNN train script #418
• Fix README for VGG #414
• Add interpolation options to scale #412 (@akitotakeki thanks!)
• n_class before pretrained_model #408
• Use LabelDataset naming convention #405
• Fix resize_contain with odd margin #404
• Remove cityscapes_labels #403
• Change default nobias option for Conv2DBNActiv #402
• Change variable names: filenames to paths #398
• Fix a typo #397
• Improve style checker #396
• Add style checker #393
• use write_image in tests #382
• Remove get_device #381 (@naoto0804 thanks!)
• Add citation #378
• Delete ConcatenatedDataset doc #374
• Fix a typo in eval_semantic_segmentation.py #372
• Fix use_gpu in examples #371
• Support Chainer 2.0.1 and 3.0.0a1 #370 (@kkk669 thanks!)
• update to Chainer v3.0.0b1 #369
• Add training result of SSD512 #366
• Stop PixelwiseSoftmaxClassifier from requring n_class #364
• Remove matplotlib backend specification line from example code #355
• Fix Chainer version in Travis #351
• Add Faster R-CNN VOC07+12 experiment #350
• Sequential feature extractor remove_unused #348
• Fix SequentialFeatureExtractor doc to properly compile #347
• Add all_feature_names to SequentialFeatureExtractor #343
• Stop using invoke_before_training option in Faster R-CNN example #340
• Fix ConcatenatedDataset #339
• use init_scope in examples/ssd/train.py #337
• Add numpy installation to docs #336

Annoncements ChainerCV changed CI to test both on Chainer v2 and Chainer v3 (alpha). Since there are users that use development version and stable version of Chainer, ChainerCV will try to support both versions of Chainer as much as possible. The priority is on stable version of Chainer, and ChainerCV will be made sure to always work on the stable version. The development version will be supported while it is possible to support the two versions simultaneously. This means that CI tests will run with the development version and the stable version. In ChainerCV, new features added to Chainer will not be used. If a new feature is necessary, the relevant part can be copied to ChainerCV temporarily. In this case, the code will be deleted once the development version is released.

Major Improvements

• Add training code of SSD #206

API Change

• Add training code of SSD #206
• Change eval_detection_voc_ap to eval_detection_voc #277
• Delete eval_pck #276
• Change eval_semantic_segmentation_iou to eval_semantic_segmentation #275
• Directory structure of extensions has changed #238

New features

• Add SemanticSegmentationEvaluator #238
• Add download_model #260
• Add crop_bbox #291

Implemented enhancements:

• Improve setup.py to work on a machine without numpy or cython #327
• Import cupy as cp #316
• Stop raising a warning when resize is not used #307
• Support (C, 1, 1) for fill argument #306
• Add version to chainercv.__init__.py #301
• Improve caffe2npz in example/ssd #263
• Improve CubDataset #259
• Add cub_label_names #253
• Redirect to doc on stable version from README #302
• Notify users about evaluation code #299
• bottom left, top right -> top left, bottom right #294
• Show dialog before download #287
• Add tests for VOCDetectionDataset #177
• Add assert_is_*_link #285
• Add assertions #278
• Mark TransformDataset as deprecated (#335)
• remove ConcatenatedDataset (#331)
• Travis CI for development version (#330)

Fixed bugs:

• Fix a bug in Faster R-CNN #266
• fix zero division #322
• Fix clip in Faster RCNN's predict #319
• Stop declaring transform inside main #311
• Use autoclass for datasets doc #308
• Fix test_assert_is_detection_link to always raise assert error #304
• Update documentation version #288
• prevent duration = 0.0 and divding by zero #284 (@sitifukujin thanks!)
• Fix a typo in the yml files. #282 (@zori thanks!)
• Fix error message for ten_crop #280
• Fix a bug in scale #279
• Fix pip version in README #270
• ChainerV2 calls trigger outside of snapshot object #272 (@MannyKayy thanks!)

This release fixes a bug found in Faster R-CNN model (#266).

Full list of features added during v0.5 development will be summarized in v0.6 release note.