Training with a set of ungeoreferenced image files

[MathiasBaumgartinger]

Hello!

I am using a dataset that consists of individual ungeoreferenced image files for training (.bmp in my specific case)? As suggested in the docs:

If > 1 URI is provided, a VRT will be created to mosaic together the individual images.

Consequently, I get Warning 1: gdalbuildvrt does not support ungeoreferenced image. Skipping /path/to/image.bmp and the subsequent error rasterio.errors.RasterioIOError: /tmp/tmp7z6zermj/tmphkqhto1f/tmprocf63op/index.vrt: No such file or directory.

Is there any possibility to circumvent this problem for training? I am not using a plain pytorch model because the predictions shall be performed on large tiffs again.

[AdeelH]

• What version/branch/docker-image of Raster Vision are you using?
• Are these .bmp files small chips/patches or large scenes?
• Raster Vision can do predictions on large TIFFs just fine. What is your exact use case?

[AdeelH]

This causes an error when calling the function learner = SemanticSegmentationLearner.from_model_bundle(bundle_uri, training=False).

Can you share the stack trace for this error?

[MathiasBaumgartinger]

Traceback (most recent call last):
  File "/opt/src/code/rv_predict.py", line 21, in <module>
    learner = SemanticSegmentationLearner.from_model_bundle(bundle_uri, training=False)
  File "/opt/src/rastervision_pytorch_learner/rastervision/pytorch_learner/learner.py", line 998, in from_model_bundle
    cfg.model.pretrained = False
AttributeError: 'NoneType' object has no attribute 'pretrained'

Edit: maybe i should add I am running inside a quay.io/azavea/raster-vision:pytorch-latest-container

As said earlier, the pipeline-config lacks a model field, which I think causes the error:

{
  "root_uri": null,
  "rv_config": {},
  "plugin_versions": {
    "rastervision.pytorch_backend": 1,
    "rastervision.pytorch_learner": 4,
    "rastervision.core": 7,
    "rastervision.gdal_vsi": 0,
    "rastervision.aws_batch": 0,
    "rastervision.aws_s3": 0,
    "rastervision.pipeline": 0
  },
  "type_hint": "learner_pipeline",
  "learner": {
    "model": null,
    ...

[AdeelH]

Thank you. Yeah, that's a bug. The cfg.model.pretrained = False thing should only happen if cfg.model is not None.

Looks like you were able to work around this.

This works in theory, but the resulting predictions seem so far off (I tested multiple different learning settings and epochs) that i fear the Resnet is not configured properly.

You are right. If you pass in a model, RV does not make any changes to it:
https://github.com/azavea/raster-vision/blob/master/rastervision_pytorch_learner/rastervision/pytorch_learner/learner.py#L341-L359

There should ideally be an error or a warning if a user passes in both a model and a model_weights_path.

The solution is to load the weights in yourself before passing it to the SemanticSegmentationLearner. You can do that via

model.load_state_dict(torch.load(model_weights_path))

LMK if that works.

[AdeelH]

@MathiasBaumgartinger

We just released v0.20. We have added a new tutorial for working with non-GeoTIFF datastets. We have also added a new section to the training tutorial about how to use model-bundles; notably, the bug you ran into has been fixed and also passing weights as model_weights_path (as you were doing before) should work now.

[MathiasBaumgartinger]

Much appreciated, will look into the new version =]

Your workaround with loading the state dict did indeed work! Thanks for your support.