Sampling training data

[jbbarre]

hello,
I'm starting a project with RV. As a very first step, I've followed the sampling training data instructions, but got an error.

I load this data:

pwd = "/media/jbbarre/data/1_PROJETS/FireAccess/fireacess/data/interim/"
image_uri = pwd+"chalabrais_0.5_shaded_TRAIN_filled.tif"
label_uri = pwd+"chalabrais_train_groundtruth.geojson"

and then using the same code as in the example, to visualize data :

from rastervision.core.data import ClassConfig
from rastervision.pytorch_learner import (
    SemanticSegmentationSlidingWindowGeoDataset, SemanticSegmentationVisualizer)

import albumentations as A

class_config = ClassConfig(
    names=['background', 'dfci'], 
    colors=['lightgray', 'darkred'],
    null_class='background')

ds = SemanticSegmentationSlidingWindowGeoDataset.from_uris(
    class_config=class_config,
    image_uri=image_uri,
    label_vector_uri=label_uri,
    label_vector_default_class_id=class_config.get_class_id('dfci'),
    image_raster_source_kw=dict(allow_streaming=True),
    size=200,
    stride=200,
    transform=A.Resize(256, 256)
)

I've got these errors:

def map_to_pixel_coords(geojson: dict,
    234                         crs_transformer: 'CRSTransformer') -> dict:
    235     """Convert a GeoJSON dict from map to pixel coordinates."""
    236     return map_geoms(
--> 237         lambda g, **kw: crs_transformer.map_to_pixel(g),
    238         geojson,
    239         progressbar_kw=dict(desc='Transforming to pixel coords'))

.......

File ~/.local/lib/python3.10/site-packages/rasterio/transform.py:294, in rowcol(transform, xs, ys, zs, op, precision, **rpc_options)
    292 transformer_cls = get_transformer(transform, **rpc_options)
    293 with transformer_cls() as transformer:
--> 294     return transformer.rowcol(xs, ys, zs=zs, op=op)

File ~/.local/lib/python3.10/site-packages/rasterio/transform.py:392, in TransformerBase.rowcol(self, xs, ys, zs, op, precision)
    390         return (op(new_rows[0]), op(new_cols[0]))
    391     else:
--> 392         return ([op(r) for r in new_rows], [op(c) for c in new_cols])
    393 except TypeError:
    394     raise TransformError("Invalid inputs")

File ~/.local/lib/python3.10/site-packages/rasterio/transform.py:392, in <listcomp>(.0)
    390         return (op(new_rows[0]), op(new_cols[0]))
    391     else:
--> 392         return ([op(r) for r in new_rows], [op(c) for c in new_cols])
    393 except TypeError:
    394     raise TransformError("Invalid inputs")

ValueError: cannot convert float NaN to integer

->> The problem is related to the GEOJSON file. My question is: Do we need a specific setup for the Geojson ? I created it by exporting the vector layer in QGIS in GEOJSON format. Its CRS is ESPG 2154. I've created the geojson file a new time to ensured I didn't miss something but with no success.

chalabrais_train_groundtruth.zip

I don't where to start to fix this problem. Any help whould be appreciated. thanks

[AdeelH]

RV expects the GeoJSON to be EPSG:4326. Can you try again with the vector data exported to that CRS?

Edit 2024-04-11: this is no longer true in v0.30. RV now supports GeoJSONs with arbitrary CRS's.

[jbbarre]

Thanks Adeel.
I've converted the geotiff and the GeoJSON into 4326 and everything is back to normal.
Further on in the notebook, another message appears, but it's for tomorow.