v0.6.0.dev0

🚧 Dev release. 🚧

KerasCV is adding experimental support for Jax and PyTorch backends on top of the Keras Core library. Read the anouncement, and browse the full library documentation, including how to specify the backend when running your code.

Support for both Jax and PyTorch is still experimental, expect some rough edges and please give us feedback!

Summary

This release should be equivalent to 0.5.1 with the addition of multi-backend support.
The following API components do not yet support Jax + PyTorch:

• keras_cv.models.StableDiffusion[V2]
• keras_cv.models.MultiHeadCenterPillar
• keras_cv.training.ContrastiveTrainer
• All models under keras_cv.models.legacy

Note that there are two ways you can run on top of Tensorflow.
If you run your scripts/colab without any changes, KerasCV will use tf.keras for all layer and modeling implementations. This should be a no-op from previous releases of the library.

If you run your scripts/colab with KERAS_BACKEND={jax, torch, tensorflow}, you will be trying the new Keras Core library, using the specified backend. This is a great way to test out the future of the library!

Full details on runtime specification is available along with the Keras Core documentation.

What's Changed

• Add support for PyTorch and Jax to KerasCV by @tirthasheshpatel in #1925
• Update version to dev release by @ianstenbit in #1927

New Contributors

• @tirthasheshpatel made their first contribution in #1925

Full Changelog: v0.5.1...v0.6.0.dev0

KerasCV 0.5.1

What's Changed

• Add epsilon for NaN prevention + test case to RetinaNet by @LukeWood in #1684
• Fixing typos by @ID6109 in #1778
• Remove randomly zoomed crop by @LukeWood in #1769
• Add a numeric preset test for our MobileNet backbone by @ianstenbit in #1786
• Refactor correct_pad_downsample as a util by @ianstenbit in #1787
• Remove an unused parameter in the CenterNet label encoder by @ianstenbit in #1788
• bug fix for csp_darknet presets by @haifeng-jin in #1790
• YOLOV8Detector fpn_depth defaults to 2 by @haifeng-jin in #1791
• Rename MaybeApply to RandomApply by @avinair in #1792
• Fix MobileNetBackbone.feature_pyramid_inputs by @barrypitman in #1785
• Mark RetinaNetSmokeTest "extra_large" by @jbischof in #1797
• Fix param order for YOLOV8 backbone by @ianstenbit in #1796
• Fix BoxCOCOMetrics by @james77777778 in #1799
• Support the updated 3D data format in most 3D preprocessing KPLs by @ianstenbit in #1789
• Mark contrastive test suit as "extra_large" by @jbischof in #1802
• Mark saving & fitting tests "large" and test models/ on CI by @jbischof in #1804
• Run large tests on GCB by @jbischof in #1803
• Test legacy on GCB by @jbischof in #1805
• Fix init.py behavior for legacy folder by @jbischof in #1810
• Fix DeepLabV3 tests by @ianstenbit in #1817
• Use result instead of evaluate in WaymoEvaluationCallback by @ianstenbit in #1821
• Add saving tests for object detection by @jbischof in #1815
• Fix error thrown by YOLOV8Detector.summary() by @ianstenbit in #1823
• Exclude the backbone from Task model summaries by @ianstenbit in #1824
• Mark the RetinaNetSmoke test as large, since #1725 is fixed by @ianstenbit in #1825
• Support Segmentation Masks in Mixup Layer by @cosmo3769 in #1584
• Move requirements.txt to the top-level of the repo by @ianstenbit in #1827
• Clean up references to LukeWood@ and tanzheny@ by @ianstenbit in #1847
• Change pyramid_level_inputs to have string keys by @avinair in #1826
• Throw a better error when ragged boxes are passed to YOLOv8 by @ianstenbit in #1844
• Fix broken StableDiffusion license link by @ianstenbit in #1843
• Fix AugMix/RandomShear issue by @ianstenbit in #1846
• Support updating the NMS decoder in YOLOv8 by @ianstenbit in #1841
• Refactor MobileNetBackbone for readability by @ID6109 in #1838
• Handle new import path for private Keras modules in 2.13 by @ianstenbit in #1850
• Temporarily disable CI on Windows by @ianstenbit in #1855
• Refactor ResNetBackbone for readability by @ID6109 in #1839
• CIoU Loss by @IMvision12 in #1845
• Remove custom_object_scope from mobilenet_v3_backbone.py by @ID6109 in #1857
• backbone compatibility for objection detection models [RetinaNet] by @IMvision12 in #1852
• Segmentation map in Resizing layer by @cosmo3769 in #1834
• Update YOLOv8 Loss by @IMvision12 in #1858
• Remove Windows CI by @ianstenbit in #1868
• Retinanet Tests by @IMvision12 in #1863
• Refactor ResNetV2Backbone for readability by @ID6109 in #1842
• Refactor ResNetBackbone for readability part2 by @ID6109 in #1856
• Migrate DenseNet to subclass Backbone by @ID6109 in #1759
• Fix RetinaNet anchor generator guard clause by @ianstenbit in #1867
• Add: Visualization API for segmentation maps by @ariG23498 in #1866
• Rationalize pyramid_level_inputs tests by @jbischof in #1873
• [Vizualization] Swap np.interp with utils.transform_value_range by @ariG23498 in #1881
• Add ragged support to YOLOV8LabelEncoder by @james77777778 in #1870
• Create presets_without_weights classproperty by @jbischof in #1872
• Fix off-by-one issue with RetinaNet anchor generation by @ianstenbit in #1882
• [Contrastive Trainer] Check for duplicate projectors by @ariG23498 in #1888
• Add ragged groundtruth support to BoxCOCOMetrics by @james77777778 in #1864
• Support for tf.data.Dataset added in plot_image_gallery, removed hard dependency on rows and cols params by @suvadityamuk in #1837
• Mark TF saving tests as "extra_large" by @jbischof in #1898
• Backbone Compatibility Test for YOLOv8 by @IMvision12 in #1890
• Fixing retinanet.py training example by @ID6109 in #1889
• Assert metric reset works correctly in RetinaNet by @ianstenbit in #1899
• Add "feature_pyramid" as parameter inside RetinaNet init by @gianlucasama in #1885
• Add YOLOV8 backbone presets to the global backbone registry by @ianstenbit in #1900
• Create a separate license structure for Waymo-contributed 3D augmentations by @ianstenbit in #1598
• Bump version to 0.5.1 by @ianstenbit in #1915

New Contributors

• @avinair made their first contribution in #1792
• @ariG23498 made their first contribution in #1866
• @suvadityamuk made their first contribution in #1837
• @gianlucasama made their first contribution in #1885

Full Changelog: v0.5.0...v0.5.1

KerasCV 0.5.0

With KerasCV 0.5.0, we're releasing some shiny new functionality, as well a new modeling API to bring us into alignment with our sister package, KerasNLP.

Check out our newly updated API documentation on KerasIO
(note that the new documentation will not be live until approximately 1 day after this release)

Our exciting new offerings in 0.5.0 include:

• The new from_preset API including pretrained presets for your favorite backbones
• A new elegant Backbone API for easy model composition and transfer learning
• YOLOV8 for object detection
• Vectorized augmentation layers to bring optimal performance to data preprocessing

We've also partnered with Waymo to offer a 3D object detection API for data preprocessing and modeling in KerasCV.
In 0.5.0 we're releasing an initial version of this API, with more to come! Look out for an upcoming guide on 3D object detection in KerasCV in partnership with Waymo.

As always, please let us know if you encounter any issues with the release by opening an Issue here on GitHub.

Thanks!
The KerasCV Team

v0.4.2

What's Changed

• Fix demos with dict format by @LukeWood in #1259
• Fix inference for jittered resize by @LukeWood in #1260
• API polish on NMS layer by @LukeWood in #1261
• Refactor preserve_rel to a more readable set of methods by @ianstenbit in #1264
• Update resizing demo to match other demos and use new bbox format by @ianstenbit in #1266
• add utility function to mask invalid detections by @LukeWood in #1265
• [Cleanup] Use flag-specified values in RetinaNet example script by @ianstenbit in #1271
• [YOLOX] Step 1/? : add GIoU loss and update IoU loss by @quantumalaviya in #1006
• Wrong install documentation string on readme by @MrinalTyagi in #1284
• Fix callback import error by @LukeWood in #1279
• [Cleanup] remove print statements in tests by @ianstenbit in #1287
• Update the internal UNet API to return a Keras model by @ianstenbit in #1286
• Make 3d namespacing consistent by @ianstenbit in #1288
• Fix the pascal VOC loader to use the dictionary format by @LukeWood in #1289
• Make include_rescaling and include_top required positionals by @LukeWood in #1291
• Fixed a small error in the RetinaNet training script by @quantumalaviya in #1295
• Test compilation of cv_layers.MultiClassNonMaxSuppression by @bhack in #1270
• Added missing name to DarknetConvBlock variables. by @sebastian-sz in #1298
• fix example training script for RetinaNet by @LukeWood in #1302
• Remove gt_ prefix from OD models compute_loss by @ianstenbit in #1304
• Another small fix to RetinaNet training example by @quantumalaviya in #1306
• [3D Object Detection] Rename GlobalRandomFlipY -> GlobalRandomFlip by @ianstenbit in #1303
• [Custom ops] Update docs + build without custom ops by default by @ianstenbit in #1312
• Implement RepeatedAugmentation as a KerasCV API by @LukeWood in #1293
• [1/3~] Begin cleaning up ops namespace by @LukeWood in #1299
• Add RandomTranslation to KerasCV by @LukeWood in #1314
• Fix layers/init.py by @ianstenbit in #1316
• Add RandomTranslation to serialization test by @LukeWood in #1315
• Update multi_class_non_max_suppression.py by @LukeWood in #1318
• Update RetinaNet + RCNN to use 'classification' instead of cls by @LukeWood in #1313
• Add ConvMixer Models by @IMvision12 in #1281
• Remove extra models from CI tests for ViT by @ianstenbit in #1322
• [api-polish] Renaming api's by @arnavrneo in #1267
• Refine readme.md by @LukeWood in #1292
• Update README.md by @LukeWood in #1328
• Fix typo in FasterRCNN compile error message by @ianstenbit in #1324
• fix AnchorGenerator output shape for object detection models by @tomrtk in #1305
• Switch to core Keras GroupNormalization by @fchollet in #1333
• Reformat with black 23.1 to fix CI by @ianstenbit in #1337
• Remove unused target_gather layer. This lives in utils by @ianstenbit in #1341
• Add pre-trained weights for ConvMixer_512_16 by @ianstenbit in #1325
• Add gs:// path loading to weights by @LukeWood in #1340
• Support pure-dictionary input format for OD models by @ianstenbit in #1348
• Removed the redundant call by @soma2000-lang in #1352
• add serialization test to losses by @LukeWood in #1345
• Fixes weight loading for gs:// by @LukeWood in #1355
• remove pad_batch_to_shape by @LukeWood in #1346
• v0.4.2 by @LukeWood in #1354

New Contributors

• @MrinalTyagi made their first contribution in #1284
• @arnavrneo made their first contribution in #1267
• @tomrtk made their first contribution in #1305
• @soma2000-lang made their first contribution in #1352

Full Changelog: v0.4.1...v0.4.2

v0.4.1

What's Changed

• Use keras application backbone for better result. by @tanzhenyu in #1220
• Demo files clean up by @divyashreepathihalli in #1214
• Add augment bounding box support for RandomCrop by @divyashreepathihalli in #1207
• revert RetinaNet to use KPL. by @tanzhenyu in #1222
• add CenterPoint. by @tanzhenyu in #1224
• Speed up the Waymo Evaluation callback by removing manual batching by @ianstenbit in #1239
• Minor updates to 3D preprocessing layers by @lengzq in #1238
• Added CSPDarkNet variants and as_backbone as a util by @quantumalaviya in #1232
• Add BinaryCenternetCrossentropy. by @tanzhenyu in #1241
• missing vgg16 in init by @IMvision12 in #1243
• Update bounding box format to a dictionary. by @LukeWood in #1121
• Add: support for negative prompts by @miguelCalado in #1225
• Add new transform functions for converting to vehicle frame and padding. by @Guowang in #1240
• Fix installation, currently import fails with an import error by @LukeWood in #1254
• Add pretrained weights for CSPDarkNetTiny by @ianstenbit in #1251
• 0.4.1 by @LukeWood in #1255

New Contributors

• @miguelCalado made their first contribution in #1225

Full Changelog: v0.4...v0.4.1

v0.4

What's Changed

• v0.3.5 by @LukeWood in #1123
• [Object Detection] Update backbone structure for FasterRCNN by @ianstenbit in #1122
• Fix a bug in gourp_points_by_bounding_boxes data augmentation by @lengzq in #1126
• Add get_config to all 3d data augmentations by @lengzq in #1127
• Skip waymo tests in devcontainer by @ianstenbit in #1125
• Fix bugs in frustum-based data augmentations by @lengzq in #1135
• Adding ViT Models by @DavidLandup0 in #1014
• [Preprocessing] Migrate RandomContrast to KerasCV by @ianstenbit in #1133
• Fix GPU testing by @chenmoneygithub in #1142
• Migrate RandomCrop to Keras CV by @divyashreepathihalli in #1141
• ViT Name Refactor by @DavidLandup0 in #1143
• Format weight conversion script by @LukeWood in #1149
• Migrate RandomZoom by @divyashreepathihalli in #1136
• Fix typo in iou_3d docs by @ianstenbit in #1147
• [RetinaNet Refactor] Step 1 Fix bug and Reuse box delta encoding and decoding logic. by @tanzhenyu in #1152
• (Step 4/5) Add KPLs to RetinaNet training script by @LukeWood in #1148
• [Preprocessing] Migrate RandomBrightness to KerasCV by @ianstenbit in #1138
• fix forward for RetinaNet. by @tanzhenyu in #1155
• [Backbones] Update basic_training to support weight decay and ema by @ianstenbit in #1124
• [Refactor RetinaNet] Step 2 Use label dict in label encoder and base model. by @tanzhenyu in #1156
• pin bazel version by @tanzhenyu in #1165
• [Refactor RetinaNet] Step 3, Batch implementation of label encoding. by @tanzhenyu in #1163
• Migrate VGG16 to keras_cv.models by @IMvision12 in #1164
• Separating MBConvBlock, FusedMBConvBlock and Refactoring EfficientNetV2 by @DavidLandup0 in #1146
• [Refactor RetinaNet] Step 4 Encode labels in model forward pass. by @tanzhenyu in #1173
• Switch to new EfficientNet weights (post-PR #1146) by @ianstenbit in #1151
• [Refactor RetinaNet] Step 5 remove the inheritance from ObjectDetectionBaseModel. by @tanzhenyu in #1175
• Add black jupyter extension and fix skip tests without custom ops by @bhack in #1150
• Gate Waymo-dependent tests with a flag by @ianstenbit in #1177
• Add an internal UNet implementation by @ianstenbit in #1167
• Add missing serialization params for two layers by @ianstenbit in #1178
• [Refactor RetinaNet] Step 6 Unify FasterRCNN to decode during predict. by @tanzhenyu in #1182
• Fix bug in GlobalRandomDroppingPoints by @lengzq in #1160
• Reduce unit test dataset size for OD models by @ianstenbit in #1185
• [Refactor RetinaNet] Step 8 Unify FasterRCNN NMS decode to return dict. by @tanzhenyu in #1187
• [Refactor RetinaNet] Step 7, unify nms decoder outputs for both models. by @tanzhenyu in #1194
• add voxelization layer. by @tanzhenyu in #1186
• [Codebase Consistency] MLPMixer Refactoring by @DavidLandup0 in #1024
• Improve FlipY by @lengzq in #1195
• Add support for Stable Diffusion 2.1 by @Jobayer-000 in #1130
• [Refactor RetinaNet] Step 9 unify NMSDecoder. by @tanzhenyu in #1197
• [0.4 Release SemSeg] Refactor Deeplab to be a functional subclassing model by @DavidLandup0 in #1119
• Basic Semantic Segmentation Training Script by @DavidLandup0 in #1183
• [3D Object Detection] Waymo Evaluation Callback by @ianstenbit in #1129
• Upgrade to the last TF stable version by @bhack in #1170
• Fix gpu tests after pinning bazel version in #1188 by @ianstenbit in #1199
• Make GPU CI tests run custom op tests by @ianstenbit in #1200
• Update release process to build wheels with custom ops by @ianstenbit in #1188
• Adding ViT weights by @DavidLandup0 in #1145
• Add ABI warning to custom ops loader by @ianstenbit in #1201
• [Refactor RetinaNet] Step 10, Unify call and loss reduction for Retin… by @tanzhenyu in #1204
• add center_net_label_encoder by @tanzhenyu in #1202
• [Object Detection] Remove backbone_weights from RetinaNet by @ianstenbit in #1203
• Add pretrained weights for ResNet50 by @ianstenbit in #1211
• Add top-5 accuracy for imagenet training script by @ianstenbit in #1212
• [Refactor FasterRCNN] remove unknown batch size restriction to be ready for Functional Subc… by @tanzhenyu in #1208
• [RetinaNet training] init optimizer inside strategy scope by @ianstenbit in #1213
• Migrate Rescaling to Keras_CV by @divyashreepathihalli in #1209
• Fix weights typo for ResNet50 by @ianstenbit in #1215
• [Object Detection] Use functional subclass paradigm for RetinaNet by @ianstenbit in #1205
• Revert "[Object Detection] Use functional subclass paradigm for RetinaNet" by @ianstenbit in #1216
• RCNN model API finalization. by @tanzhenyu in #1217
• RetinaNet API Finalization by @ianstenbit in #1218
• Update version to 0.4 by @ianstenbit in #1219

New Contributors

• @chenmoneygithub made their first contribution in #1142
• @IMvision12 made their first contribution in #1164
• @Jobayer-000 made their first contribution in #1130

Full Changelog: v0.3.5...v0.4

v0.3.5

What's Changed

• [Object Detection] Update backbone structure for FasterRCNN by @ianstenbit in #1122
• Fix a bug in gourp_points_by_bounding_boxes data augmentation by @lengzq in #1126
• Add get_config to all 3d data augmentations by @lengzq in #1127
• Skip waymo tests in devcontainer by @ianstenbit in #1125

Full Changelog: v0.3.5-rc0...v0.3.5-rc3

v0.3.5

v0.3.5-rc2

Skip waymo tests in devcontainer (#1125)

v0.3.5

What's Changed

• Update unit test and golden numbers in KPL by @qlzh727 in #864
• Support no image size in PascalVOC loader by @LukeWood in #862
• Make separate benchmark methods for single and multi-gpu benchmarks by @ianstenbit in #867
• revert keypoint refactoring. by @tanzhenyu in #869
• Add deeplab_v3 model with keras cv components only by @qlzh727 in #856
• Update the instructions to update the cloudbuild cluster by @LukeWood in #832
• Address review comments from #874 by @qlzh727 in #875
• Support mixed-precision training in basic_training script by @ianstenbit in #878
• Add FasterRCNN model. by @tanzhenyu in #870
• Fix the unit test for segmentation head serialization. by @qlzh727 in #882
• Add Atrous Spatial Pyramid Pooling as a feature decoder layer by @qlzh727 in #884
• Add contributing docs about new model weights by @ianstenbit in #889
• Add devcontainer CI by @bhack in #893
• Improve FasterRCNN model quality. by @tanzhenyu in #898
• Minor performance improvements to basic ImageNet training by @ianstenbit in #902
• Add pre-trained weights for EfficientNetV2B0 by @ianstenbit in #906
• Filter pull request directory by @bhack in #899
• Make RandomFlip perform horizontal flipping only by default by @ianstenbit in #903
• Add distributed training, and further improve training quality. by @tanzhenyu in #914
• Add some pre-trained weight info to README by @ianstenbit in #913
• Adds the VAEEncoder to StableDiffusion by @LukeWood in #873
• Allow compute_iou to take boxes of different shape. by @tanzhenyu in #919
• Add pretrained EfficientNetV2b1 weights by @ianstenbit in #923
• update readme and roadmap. by @tanzhenyu in #934
• Fix typo in name of devcontainer workflow by @ianstenbit in #940
• Add model fit for FasterRCNN. by @tanzhenyu in #925
• Update MobileNet docs to point to the correct paper by @ianstenbit in #948
• Add pascal voc 2012 dataset with segmentation data. by @qlzh727 in #943
• Add RandomCropAndResize to ImageNet classification training script by @ianstenbit in #935
• Add better weights for ResNet50V2 by @ianstenbit in #952
• Fixed efficientnet v1 correct pad. by @sebastian-sz in #927
• Update the RetinaNet example by @LukeWood in #942
• Add tf.function(jit_compile=True) for the PNG decode function. by @qlzh727 in #953
• Add IoU3D as a custom c++ op (CPU) by @ianstenbit in #890
• bug fix in image_encoder.py by @LukeWood in #955
• Pre-trained weights for EfficientNetV2B2 by @ianstenbit in #956
• Training script improvement proposal by @DavidLandup0 in #908
• Devcontainer GPU support #945 by @jacoverster in #946
• Fix build target for custom ops by @ianstenbit in #959
• Step 1/3. add dummy semantic segmentation training script. by @tanzhenyu in #957
• Step 2/3: add Sbd data by @tanzhenyu in #961
• Fix casting issue with basic training learning rate schedule by @ianstenbit in #965
• Add waymo dataset loader by @ianstenbit in #966
• Add weights for EfficientNetV2S by @ianstenbit in #971
• citing reference. by @tanzhenyu in #972
• Fix reproducability issue of seeded StableDiffusion runs by @ianstenbit in #962
• Added mixed_precision support for preprocessing layers. by @sebastian-sz in #967
• Save best weights only for basic_training by @ianstenbit in #977
• 630/add efficientnet lite by @sebastian-sz in #921
• Shuffle dataset every epoch for basic_training by @ianstenbit in #983
• Track encoded boxes in the RetinaNet in a Keras metric by @LukeWood in #964
• Step 3/3: Update DeeplabV3 model by @tanzhenyu in #975
• Add README about Waymo Open Dataset dependency by @ianstenbit in #985
• Lazy-load components of StableDiffusion and make them visible to users by @ianstenbit in #989
• Add text-conditioned in-painting functionality to Keras Stable Diffusion by @bobqywei in #973
• Add support for add_tokens() to StableDiffusion in order to support TextualInversion by @LukeWood in #831
• Add bounding box support to resizing by @LukeWood in #861
• Implement bounding boxes for RandomCropAndResize by @AmedeoBiolatti in #994
• Update README.md with disclaimer by @LukeWood in #993
• Avoid retracing with python lazy_loading by @bhack in #997
• Restructure StableDiffusion namespace and folder hierarchy by @ianstenbit in #1002
• fix serializationt test. by @tanzhenyu in #1004
• SimCLR Training Script by @ianstenbit in #717
• Add seed=seed, param_name to factor parsing for layers.Mosaic() by @LukeWood in #881
• Adding ViT Components/Layers by @DavidLandup0 in #929
• Add base augmentation layer for 3D preception. by @lengzq in #986
• [Object Detection] Update retinanet predict loop to enable TPU support by @ianstenbit in #1017
• [Object Detection] Introduce a callback-based coco metric evaluator by @ianstenbit in #1000
• fixed transformer encoder dims per head by @DavidLandup0 in #1013
• Revert "[Object Detection] Update retinanet predict loop to enable TPU support" by @ianstenbit in #1022
• ResNet and ResNetV2 - Typo fix in docstrings by @DavidLandup0 in #1023
• Adding RegNets to keras_cv by @AdityaKane2001 in #739
• Add global_z_rotation data augmentation and wrap_angle_rad helper function. by @lengzq in #1012
• [Backbones] Add pre-trained weights for DarkNet53 by @ianstenbit in #1029
• [Object Detection] RetinaNet: custom predict_function by @ianstenbit in #1021
• [Object Detection] Speed up the PyCOCO callback by only calling model.predict once by @ianstenbit in #1026
• Add mixed precision for deeplab. by @tanzhenyu in #1032
• Update group_normalization.py by @charbull in #1035
• [Backbone Training] added 50/50 between cutmix and mixup, changed default hold steps by @DavidLandup0 in #980
• IoULoss documentation typo fix by @DavidLandup0 in #1038
• Add four global data augmentations for 3D by @lengzq in #1028
• Add pre-trained weights for CSPDarkNet by @ianstenbit in #1036
• [Draft] Polish the RetinaNet by @LukeWood in #984
• fix for dynamic shape. by @tanzhenyu in #1042
• fix fasterrcnn train dataset. by @tanzhenyu in #1046
• [Stable Diffusion] Noise scheduler for fine-tuning and textual inversion by @ianstenbit in #1045
• polish roadmap and readme. by @tanzhenyu in #1009
• Introduced a new internal util to preserve relative format by @Quantumal...

v0.3.4

What's Changed

• Introduce a new function to convert an application model to backbone model. by @qlzh727 in #820
• Add mosaic data augmentation by @quantumalaviya in #799
• Support ragged images in converter by @bhack in #848
• Readability improvement to SD by @LukeWood in #847
• Fix label encoder when all -1s are provided by @LukeWood in #846
• Small performance optimization for StableDiffusion by @ianstenbit in #851
• Respect box_head in constructor by @LukeWood in #857
• Update TF stable release by @bhack in #837
• Add martins comment by @LukeWood in #859
• Add encode_text and generate_image to StableDiffusion API by @ianstenbit in #853
• Add licensing notice and information about model safety to model card by @LukeWood in #863

Full Changelog: v0.3.3...v0.3.4