Hello, I'm looking for some help with predictions using the Predictor class. I wasn't able to find any examples in the documentation or GitHub.
I'm using the PyTorch SpaceNet Vegas Buildings predict package from the model zoo along with the sample image.
I define a new instance of the Predictor class and then predict on an image:
predict = rv.Predictor("predict_package.zip","tmp").predict("1929.tif")
The return is: rastervision.data.label.semantic_segmentation_labels.SemanticSegmentationLabels
How do I access an array of labels from my prediction?
Looking at the SemanticSegmentatationLabels class I thought it was something like this:
windows = predict.get_windows()label_array = predict.get_label_array(windows[0])
But I'm getting this error:
ActivationError: RasterSource must be activated before use
Can anyone please let me know if I'm on the right track or going about this the wrong way. Thanks!
@g2giovanni There isn't any direct way to make prediction using a model file. A workaround (as you observed) is to modify an existing predict package which contains necessary metadata. The files in the zip file have to be at the root of the zip file, and I think that might be causing the problem. Another potentially deeper issue is that the model file needs to have just the weights
state_dict, not the whole pickled model. (See https://pytorch.org/tutorials/beginner/saving_loading_models.html). And if you supply the weights, they need to correspond to a model architecture that Raster Vision knows about (ie. the ones used by the PyTorch backends).
Thank you @lewfish . My first problem actually is I want to make a prediction using a custom model architecture, not included in Pytorch backends.
I'm planning to develop a new backend plugin in order to load a custom Pytorch models and a simple CLI python module to create a mock bundle_config.json file
predict CLI command I think things should work. So I would follow this as a template for what to do: https://github.com/azavea/raster-vision/blob/master/rastervision/cli/main.py#L273-L281 Mainly the thing to follow here is passing an output_uri to predict so that it writes the output to disk. That may be inefficient for your use case, but I think it should work.
rastervision2 package) which is under development now and is in a minimally usable state. So I'm reluctant to try to debug the old version of it. Unfortunately I don't have the corresponding rv2 predict package to share at the moment.
3 replies
1) Does the pytorch object detection pipeline normalize the image channels to what the pretrained ImageNet models were trained with? https://pytorch.org/hub/pytorch_vision_resnet/
2) Are there any plans for integration with MLFlow?
3) Is work still being done on the rastervision project, or should things be moved to rastervision2? Are you still accepting pull requests for rastervision1. I've made a few tiny quality of life tweaks that I think might be useful (e.g., using the environment's TORCH_HOME if defined vs setting it to /opt/data/torch-cache, making the example docker images not run as the root user, flushing the tb_writer after each epoch, etc.)
We will be releasing the next version of Raster Vision, which is a major refactoring sometime before June 1. The code is currently being developed in the rastervision2 package and is in a usable, but poorly documented state. The major changes include:
- a simpler configuration system that uses a single
Configclass based on Pydantic in place of the existingConfig,ConfigBuilder, and Protobufs, and that does validation, serialization, and deserialization automatically using a declarative representation. - generalizing and extracting out the functionality for running configurable pipelines in the cloud. What used to be
Experimentsare now just instances ofPipeline, and thepipelinepackage can be used independent of the rest of RV. - elimination of the Tensorflow backends, and abstraction of the PyTorch/torchvision backend code, so that all tasks are implemented in a uniform way and it's easier to customize aspects of the training loop.
Before releasing the big refactor of RV, we will have a final release of the existing RV codebase, which will be version 0.11. If there are any small changes you want to get into that final release, please make a PR by next Friday May 5. After that release is made, the old version of RV will be removed from the master branch, the code in rastervision2 will be moved to rastervision, and that will be released in version 0.12.
train_label_source = rv.LabelSourceConfig.builder(rv.OBJECT_DETECTION) \
AttributeError: 'ObjectDetectionLabelSourceConfigBuilder' object has no attribute 'with_raster_source'
anyone have any solution ?
2 replies
3 replies
Hello,
for a semantic segmentation task, When I add a with_vector_output to the labelStore builder to output labels as polygons, It works for the evaluation scenes after training is completed. The output in the predict directory will contain predicted labels in both tif and geojson formats. But when I use the bundle to predict new scenes, it only outputs labels in raster format (.tif).
I tried both versions 0.10 and 0.11 of rastervision but both produce the same results.
Here is the labelStore code that is added to the scenes using .with_label_store:
vector_output = {'mode': 'polygons', 'class_id': 1, 'denoise': 3}
label_store = rv.LabelStoreConfig.builder(rv.SEMANTIC_SEGMENTATION_RASTER) \
.with_vector_output([vector_output]).build()Here is the scene,groundTruthLabelSource and predictionLabelStore sections of the bundle_config.json:
`
"scene": {
"id": "train_austin1",
"rasterSource": {
"sourceType": "RASTERIO_SOURCE",
"channelOrder": [
0,
1,
2
],
"rasterioSource": {
"uris": [
"BUNDLE"
],
"xShiftMeters": 0.0,
"yShiftMeters": 0.0
}
},
"groundTruthLabelSource": {
"sourceType": "SEMANTIC_SEGMENTATION_RASTER",
"semanticSegmentationLabelSource": {
"source": {
"sourceType": "RASTERIZED_SOURCE",
"rasterizedSource": {
"vectorSource": {
"geojson": {
"uri": "/opt/src/code/inria/train/small_test_vector/austin1.geojson"
},
"sourceType": "GEOJSON_SOURCE",
"classIdToFilter": {
"1": [
"==",
"DN",
255.0
]
}
},
"rasterizerOptions": {
"backgroundClassId": 2,
"allTouched": false
}
}
}
}
},
"predictionLabelStore": {
"storeType": "SEMANTIC_SEGMENTATION_RASTER",
"semanticSegmentationRasterStore": {
"rgb": false,
"vectorOutput": [
{
"denoise": 2,
"uri": "",
"mode": "polygons",
"classId": 1,
"buildingOptions": {
"minAspectRatio": 1.6180000305175781,
"elementWidthFactor": 0.5,
"elementThickness": 0.0010000000474974513
}
}
]
}
}
Anyone knows what am I missing?
1 reply
raster-vision v0.10 using aws batch and now there are errors due the reason of breaking changes of the latest docker image.
I tried to update it to 0.12 and now I willing to execute it in aws batch, I miss few things
in
0.10I spin up an image like thatdocker run --rm -it -p 6006:6006 -p 8888:8888 -p 8000:8000 \ -e AWS_PROFILE=default -v /home/xristos/.aws:/root/.aws:ro \ -v /home/xristos/.rastervision:/root/.rastervision \ -v `pwd`/notebooks:/opt/notebooks \ -v ${RV_CODE_DIR}:/opt/src/code \ -v ${RV_EXP_DIR}:/opt/data \ quay.io/azavea/raster-vision:pytorch-0.10Now how can I pass the aws profile in the image? the same way?
I really miss the
-nargs where checks everything is OK and in place
I found the spacenet dataset for RIO here:
https://spacenet.ai/spacenet-buildings-dataset-v1/
aws s3 ls s3://spacenet-dataset/spacenet/SN1_buildings/
I was able to download it, but it seems to be in a different format from what the spacenet_rio_data_prep.ipymb was created for.
7 replies
However, when I upload the template.yml file in CloudFormation, this error pops-up:
The following resource types are not supported for resource import: AWS::ECR::Repository,AWS::Batch::JobDefinition,AWS::EC2::LaunchTemplate,AWS::Batch::JobDefinition,AWS::Batch::JobDefinition,AWS::Batch::JobDefinition,AWS::EC2::LaunchTemplate,AWS::Batch::ComputeEnvironment,AWS::Batch::ComputeEnvironment,AWS::Batch::JobQueue,AWS::Batch::JobQueueThat's quite a number of unsupported resources... am I missing something here?
2 replies
hi, i'm new to all of this. I am trying to run the local example with the command "rastervision predict https://s3.amazonaws.com/azavea-research-public-data/raster-vision/examples/model-zoo-0.12/spacenet-vegas-buildings-ss/ model-bundle.zip https://s3.amazonaws.com/azavea-research-public-data/raster-vision/examples/model-zoo-0.12/spacenet-vegas-buildings-ss/1929.tif prediction.tif "
It gives me the following error:
File "/opt/src/rastervision_aws_s3/rastervision/aws_s3/s3_file_system.py", line 182, in write_bytes
raise NotWritableError ('Could not write {}'. format (uri)) from e
rastervision.pipeline.file_system.file_system.NotWritableError: Could not write s3: //raster-vision-lf-dev/examples/spacenet-vegas-buildings-ss/output_6_27c/predict/1332-0-polygons.json
root @ 89c196e3f178: / opt / src # "
Does anyone know how to fix it?
Predictor(model_bundle, tmp_dir, update_stats, channel_order) I keep getting the following error: rastervision.pipeline.config.ConfigError: The plugin_versions field contains an unrecognized plugin name: rastervision.core. Not sure why I'm getting an error with the Predictor class when the CLI works fine.
2 replies
6 replies
1 reply
1 reply