2) cd to the created directory
3) Created directories code and rv_root using the commands:
export RV_QUICKSTART_CODE_DIR=
pwd/codeexport RV_QUICKSTART_EXP_DIR=
pwd/rv_rootmkdir -p ${RV_QUICKSTART_CODE_DIR} ${RV_QUICKSTART_EXP_DIR}
4) Copied tiny_spacenet.py to /code
5) Then run: docker run --rm -it -p 6006:6006 -v ${RV_QUICKSTART_CODE_DIR}:/opt/src/code -v ${RV_QUICKSTART_EXP_DIR}:/opt/data quay.io/azavea/raster-vision:cpu-latest /bin/bash
7) run the commands:
cd /opt/src/code
rastervision run local -p tiny_spacenet.py -n
I got the following error message, please see the picture attached above.
quay.io/azavea/raster-vision:pytorch-0.10 but you are using quay.io/azavea/raster-vision:cpu-latest.
``` rastervision.registry.RegistryError: No command found for type EXAMPLE ``` rastervision -p example run local -e split_images_sta EXAMPLE -a root_uri /opt/rvspace/sindarasterpro/rastervision/dm/ -a image_uri /opt/rvspace/sindarasterpro/rastervision/data/shptemp/train.tif
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).
register_plugin function? See https://docs.rastervision.io/en/latest/commands.html#custom-aux-commands
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