Raster Vision 0.20.2#
Update example plugin
__init__.pyfiles to include
Add error handling for empty
Only set default stride if stride value is missing in
Minor doc and type-hint fixes and refactoring for OD (#1675, #1676)
Raster Vision 0.20.1#
Do not install
rastervision_gdal_vsiby default (#1622)
Do not set
Fix docker build errors (#1629)
Raster Vision 0.20#
This release brings major improvements to Raster Vision’s usability as well as its usefulness.
Whereas previously Raster Vision was a framework where users could configure a pipeline and then let it run, it is now also a library from which users can pick individual components and use them to build new things.
We have also significantly improved the documentation. Most notably, it now contains detailed tutorial notebooks as well a full API reference. The documentation for the Raster Vision pipeline, which used to make up most of the documentation in previous versions, is now located in the The Raster Vision Pipeline section.
In terms of features, some highlights are:
Support for multiband imagery, introduced in v0.13 for semantic segmentation, is now also available for chip classification and object detection. (#1345)
Improved data fusion: the
MultiRasterSourcecan now combine
RasterSourceswith varying extents and resolutions. (#1308)
You can now discard edges of predicted chips in semantic segmentation in order to reduce boundary artifacts (#1486). This can be used in addition to the previously introduced ability to average overlapping regions in adjacent chips.
Progress-bars will now be shown for all downloads and uploads as well as other time-consuming operations that take longer than 5 seconds.
Improved caching of downloads: Raster Vision can now cache downloads. Also a bug that caused Raster Vision to download the same image multiple times has been fixed, resulting in significant speedups.
This release breaks backward-compatibility with previous versions.
Extend multiband support to all tasks (#1345)
Add support for external models for object detection (#1337)
MultiRasterSourceto read from sub raster sources with non-identical extents and resolutions (#1308)
Allow discarding edges of predicted chips in semantic segmentation (#1486)
Add numpy-like array indexing and slicing to
RandomWindowGeoDatasetmore efficient when sampling chips from scenes with sparse AOIs (#1225)
Add support for Albumentations’ lambda transforms (#1368)
Provide grouping mechanism for scenes and use it in the
Update STAC-reading functionality to make it compatible with STAC v1.0.* (#1243)
Add progress bars for downloads and uploads (#1343)
Allow caching downloads (#1450)
Learnerand related configs to be more flexible and easier to use in a notebook (#1413)
Refactor to make it easier to programmatically make predictions on new scenes (#1434)
Evaluatoreasier to use independently (#1438)
GeoDataset.from_uris()for convenient initialization of
Labels.save()convenience method (#1486)
Factor out dataset visualization into a
GeoDataFrame-based spatial joins in
RGBClassTransformerby an order of magnitude (#1485)
rastervision_pipelineentry point to ensure commands from other plugins are available (#1250)
Fix incorrect F1 scores when aggregating evals for scenes in the eval stage (#1386)
Fix bug in semantic segmentation prediction output paths (#1354)
Do not zero out null class pixels when creating semantic segmentation training chips (#1556)
Fix a bug in
DataConfigvalidation and refactor
Make the semantic segmentation integration test more deterministic (#1261)
Migrate from Travis to GitHub Actions (#1218)
Update cloudformation template to allow use of on-demand GPU instances (#1482)
docker/runautomatically find a free port for Jupyter server if the default port is already taken (#1558)
Set tutorial-notebooks path as the default jupyter path in
Raster Vision 0.13#
This release presents a major jump in Raster Vision’s power and flexibility. The most significant changes are:
Support arbitrary models and loss functions (#985, #992)#
Raster Vision is no longer restricted to using the built in models and loss functions. It is now possible to import models and loss functions from a GitHub repo or a URI or a zip file as long as they interface correctly with RV’s learner code. This means that you can now easily swap models in your existing training pipelines, allowing you to take advantage of the latest models or to make customizations that help with your specific task; all with minimal changes.
This is made possible by PyTorch’s
Currently not supported for Object Detection.
Support for multiband images (even with Transfer Learning) (#972)#
It is now possible to train on imagery with more than 3 channels. Raster Vision automatically modifies the model to be able to accept more than 3 channels. If using pretrained models, the pre-learned weights are retained.
The model modification cannot be performed automatically when using an external model. But as long as the external model supports multiband inputs, it will work correctly with RV.
Currently only supported for Semantic Segmentation.
Support for reading directly from raster sources during training without chipping (#1046)#
It is no longer necessary to go through a
chip stage to produce a training dataset. You can instead provide the
DatasetConfig directly to the PyTorch backend and RV will sample training chips on the fly during training. All the examples now use this as the default. Check them out to see how to use this feature.
Support for arbitrary Albumentations transforms (#1001)#
It is now possible to supply an arbitrarily complicated Albumentations transform for data augmentation. In the
DataConfig subclasses, you can specify a
base_transform that is applied every time (i.e. in training, validation, and prediction), an
aug_transform that is only applied during training, and a
PlotOptions) to ensure that sample images are plotted correctly (e.g. use
plot_transform to rescale a normalized image to 0-1).
Allow streaming reads from Rasterio sources (#1020)#
It is now possible to stream chips from a remote
RasterioSource without first downloading the entire file. To enable, set
allow_streaming=True in the
Analyze stage no longer necessary when using non-uint8 rasters (#972)#
It is no longer necessary to go through an
analyze stage to be able to convert non-
uint8 rasters to
uint8 chips. Chips can now be stored as
numpy arrays, and will be normalized to
float during training/prediction based on their specific data type. See
spacenet_vegas.py for example usage.
Currently only supported for Semantic Segmentation.
Add support for multiband images #972
Add support for vector output to predict command #980
Add support for weighted loss for classification and semantic segmentation #977
Add multi raster source #978
Add support for fetching and saving external model definitions #985
Add support for external loss definitions #992
Upgrade to pyproj 2.6 #1000
Add support for arbitrary albumentations transforms #1001
Minor tweaks to regression learner #1013
Add ability to specify number of PyTorch reader processes #1008
Make img_sz specifiable #1012
Add ignore_last_class capability to segmentation #1017
Add filtering capability to segmentation sliding window chip generation #1018
Add raster transformer to remove NaNs from float rasters, add raster transformers to cast to arbitrary numpy types #1016
Add plot options for regression #1023
Add ability to use fewer channels w/ pretrained models #1026
Remove 4GB file size limit from VSI file system, allow streaming reads #1020
Add reclassification transformer for segmentation label rasters #1024
Allow filtering out chips based on proportion of NODATA pixels #1025
Allow ignore_last_class to take either a boolean or the literal ‘force’; in the latter case validation of that argument is skipped so that it can be used with external loss functions #1027
Add ability to crop raster source extent #1030
Accept immediate geometries in SceneConfig #1033
Only perform normalization on unsigned integer types #1028
Make group_uris specifiable and add group_train_sz_rel #1035
Make number of training and dataloader previews independent of batch size #1038
Allow continuing training from a model bundle #1022
Allow reading directly from raster source during training without chipping #1046
Remove external commands (obsoleted by external architectures and loss functions) #1047
Allow saving SS predictions as probabilities #1057
Update CUDA version from 10.1 to 10.2 #1115
Add integration tests for the nochip functionality #1116
Update examples to make use of the nochip functionality by default #1116
Raster Vision 0.12#
This release presents a major refactoring of Raster Vision intended to simplify the codebase, and make it more flexible and customizable.
To learn about how to upgrade existing experiment configurations, perhaps the best approach is to read the source code of the Examples to get a feel for the new syntax. Unfortunately, existing predict packages will not be usable with this release, and upgrading and re-running the experiments will be necessary. For more advanced users who have written plugins or custom commands, the internals have changed substantially, and we recommend reading Architecture and Customization.
Since the changes in this release are sweeping, it is difficult to enumerate a list of all changes and associated PRs. Therefore, this change log describes the changes at a high level, along with some justifications and pointers to further documentation.
Simplified Configuration Schema#
We are still using a modular, programmatic approach to configuration, but have switched to using a
Config base class which uses the Pydantic library. This allows us to define configuration schemas in a declarative fashion, and let the underlying library handle serialization, deserialization, and validation. In addition, this has allowed us to DRY up the configuration code, eliminate the use of Protobufs, and represent configuration from plugins in the same fashion as built-in functionality. To see the difference, compare the configuration code for
ChipClassificationLabelSource in 0.11 (label_source.proto and chip_classification_label_source_config.py), and in 0.12 (chip_classification_label_source_config.py).
Abstracted out Pipelines#
Raster Vision includes functionality for running computational pipelines in local and remote environments, but previously, this functionality was tightly coupled with the “domain logic” of machine learning on geospatial data in the
Experiment abstraction. This made it more difficult to add and modify commands, as well as use this functionality in other projects. In this release, we factored out the experiment running code into a separate rastervision.pipeline package, which can be used for defining, configuring, customizing, and running arbitrary computational pipelines.
Reorganization into Plugins#
The rest of Raster Vision is now written as a set of optional plugins that have
Pipelines which implement the “domain logic” of machine learning on geospatial data. Implementing everything as optional (
pip installable) plugins makes it easier to install subsets of Raster Vision functionality, eliminates separate code paths for built-in and plugin functionality, and provides (de facto) examples of how to write plugins. See Codebase Overview for more details.
More Flexible PyTorch Backends#
The 0.10 release added PyTorch backends for chip classification, semantic segmentation, and object detection. In this release, we abstracted out the common code for training models into a flexible
Learner base class with subclasses for each of the computer vision tasks. This code is in the
rastervision.pytorch_learner plugin, and is used by the
rastervision.pytorch_backend. By decoupling
Learners, it is now easier to write arbitrary
Pipelines and new
Backends that reuse the core model training code, which can be customized by overriding methods such as
build_model. See Customizing Raster Vision.
Removed Tensorflow Backends#
The Tensorflow backends and associated Docker images have been removed. It is too difficult to maintain backends for multiple deep learning frameworks, and PyTorch has worked well for us. Of course, it’s still possible to write
Backend plugins using any framework.
For simplicity, we moved the contents of the raster-vision-examples and raster-vision-aws repos into the main repo. See Examples and Setup AWS Batch using CloudFormation.
To help people bootstrap new projects using RV, we added Bootstrap new projects with a template.
All the PyTorch backends now offer data augmentation using albumentations.
We removed the ability to automatically skip running commands that already have output, “tree workflows”, and “default providers”. We also unified the
Taskclasses into a single
Pipelineclass which is subclassed for different computer vision (or other) tasks. These features and concepts had little utility in our experience, and presented stumbling blocks to outside contributors and plugin writers.
Although it’s still possible to add new
VectorSourcesand other classes for reading data, our philosophy going forward is to prefer writing pre-processing scripts to get data into the format that Raster Vision can already consume. The
VectorTileVectorSourcewas removed since it violates this new philosophy.
We previously attempted to make predictions for semantic segmentation work in a streaming fashion (to avoid running out of RAM), but the implementation was buggy and complex. So we reverted to holding all predictions for a scene in RAM, and now assume that scenes are roughly < 20,000 x 20,000 pixels. This works better anyway from a parallelization standponit.
We switched to writing chips to disk incrementally during the
CHIPcommand using a
SampleWriterclass to avoid running out of RAM.
The term “predict package” has been replaced with “model bundle”, since it rolls off the tongue better, and
BUNDLEis the name of the command that produces it.
Class ids are now indexed starting at 0 instead of 1, which seems more intuitive. The “null class”, used for marking pixels in semantic segmentation that have not been labeled, used to be 0, and is now equal to
aws_batchrunner was renamed
batchdue to a naming conflict, and the names of the configuration variables for Batch changed. See Running on AWS Batch.
The next big features we plan on developing are:
the ability to read and write data in STAC format using the label extension. This will facilitate integration with other tools such as GroundWork.
Raster Vision 0.11#
Raster Vision 0.10#
Raster Vision 0.10.0#
Notes on switching to PyTorch-based backends#
The current backends based on Tensorflow have several problems:
They depend on third party libraries (Deeplab, TF Object Detection API) that are complex, not well suited to being used as dependencies within a larger project, and are each written in a different style. This makes the code for each backend very different from one other, and unnecessarily complex. This increases the maintenance burden, makes it difficult to customize, and makes it more difficult to implement a consistent set of functionality between the backends.
Tensorflow, in the maintainer’s opinion, is more difficult to write and debug than PyTorch (although this is starting to improve).
The third party libraries assume that training images are stored as PNG or JPG files. This limits our ability to handle more than three bands and more that 8-bits per channel. We have recently completed some research on how to train models on > 3 bands, and we plan on adding this functionality to Raster Vision.
Therefore, we are in the process of sunsetting the Tensorflow backends (which will probably be removed) and have implemented replacement PyTorch-based backends. The main things to be aware of in upgrading to this version of Raster Vision are as follows:
Instead of there being CPU and GPU Docker images (based on Tensorflow), there are now tf-cpu, tf-gpu, and pytorch (which works on both CPU and GPU) images. Using
./docker/build --pytorchwill only build the TF or PyTorch images, respectively.
Using the TF backends requires being in the TF container, and similar for PyTorch. There are now
--pytorch-gpuoptions for the
./docker/runcommand. The default setting is to use the PyTorch image in the standard (CPU) Docker runtime.
The raster-vision-aws CloudFormation setup creates Batch resources for TF-CPU, TF-GPU, and PyTorch. It also now uses default AMIs provided by AWS, simplifying the setup process.
To easily switch between running TF and PyTorch jobs on Batch, we recommend creating two separate Raster Vision profiles with the Batch resources for each of them.
The way to use the
ConfigBuildersfor the new backends can be seen in the examples repo and the Backend reference
Add confusion matrix as metric for semantic segmentation #788
Add predict_chip_size as option for semantic segmentation #786
Handle “ignore” class for semantic segmentation #783
Add stochastic gradient descent (“SGD”) as an optimizer option for chip classification #792
Add option to determine if all touched pixels should be rasterized for rasterized RasterSource #803
Script to generate GeoTIFF from ZXY tile server #811
Remove QGIS plugin #818
Add PyTorch backends and add PyTorch Docker image #821 and #823.
Raster Vision 0.9#
Raster Vision 0.9.0#
Add requester_pays RV config option #762
Unify Docker scripts #743
Switch default branch to master #726
Merge GeoTiffSource and ImageSource into RasterioSource #723
Simplify/clarify/test/validate RasterSource #721
Simplify and generalize geom processing #711
Predict zero for nodata pixels on semantic segmentation #701
Add support for evaluating vector output with AOIs #698
Conserve disk space when dealing with raster files #692
Optimize StatsAnalyzer #690
Include per-scene eval metrics #641
Make and save predictions and do eval chip-by-chip #635
Decrease semseg memory usage #630
Add support for vector tiles in .mbtiles files #601
Add support for getting labels from zxy vector tiles #532
Add ability to shift raster images by given numbers of meters. #573
Add ability to generate GeoJSON segmentation predictions. #575
Add ability to run the DeepLab eval script. #653
Submit CPU-only stages to a CPU queue on Aws. #668
Parallelize CHIP and PREDICT commands #671
update_for_commandto split out the IO reporting into
Add Multi-GPU Support to DeepLab Backend #590
Handle multiple AOI URIs #617
train_restart_dirDefault Value #626
`maketo manage local execution #664
Optimize vector tile processing #676
Fix Deeplab resume bug: update path in checkpoint file #756
Allow Spaces in
Fix error when using predict packages with AOIs #674
Correct checkpoint name #624
Allow using default stride for semseg sliding window #745
Fix filter_by_aoi for ObjectDetectionLabels #746
Load null channel_order correctly #733
Handle Rasterio crs that doesn’t contain EPSG #725
Fixed issue with saving semseg predictions for non-georeferenced imagery #708
Fixed issue with handling width > height in semseg eval #627
Fixed issue with experiment configs not setting key names correctly #576
Fixed issue with Raster Sources that have channel order #576