Scalar Type Error

[natedailey7]

Hello,

I'm trying to classify objects in aerial imagery using Raster Vision, and I'm following the training example in the Raster Vision tutorials. I'm hitting an error when I train the model. My training and validation images/labels are all 326x326 pixels (starting out with small images just to figure out this library). My code is as follows:

class_config = ClassConfig(
    names=['background', '0'],
    colors=['lightgray', 'darkred'],
    null_class='background')

0 is the label for objects I want to identify. The rest I want to classify as background.

train_image_uri = "path_to_geotiff"
train_label_uri = "path_to_geojson"
val_image_uri   = "path_to_geotiff"
val_label_uri   = "path_to_geojson"

train_ds = SemanticSegmentationSlidingWindowGeoDataset.from_uris(
    class_config=class_config,
    image_uri=train_image_uri,
    label_vector_uri=train_label_uri,
    label_vector_default_class_id=class_config.get_class_id('0'),
    size=163,
    stride=163,
)

val_ds = SemanticSegmentationSlidingWindowGeoDataset.from_uris(
    class_config=class_config,
    image_uri=val_image_uri,
    label_vector_uri=val_label_uri,
    label_vector_default_class_id=class_config.get_class_id('0'),
    size=163,
    stride=163,
)

For train_ds and val_ds, these window settings (size and stride) yield 4 windows that fit perfectly in a square pattern for the image. That is, print(len(train_ds)) outputs 4 (and the same for val_ds).

model = torch.hub.load(
    'AdeelH/pytorch-fpn:0.3',
    'make_fpn_resnet',
    name='resnet18',
    fpn_type='panoptic',
    num_classes=len(class_config),
    fpn_channels=128,
    in_channels=3,
    out_size=(163, 163),
    pretrained=True)

data_cfg = SemanticSegmentationGeoDataConfig(
    class_names=class_config.names,
    class_colors=class_config.colors,
    num_workers=0,
)

solver_cfg = SolverConfig(
    batch_sz=8,
    lr=3e-2,
    class_loss_weights=[1., 10.]
)

learner_cfg = SemanticSegmentationLearnerConfig(data=data_cfg, solver=solver_cfg)

learner = SemanticSegmentationLearner(
    cfg=learner_cfg,
    output_dir='./train-test/',
    model=model,
    train_ds=train_ds,
    valid_ds=val_ds,
)

Everything runs fine up to this point. Next, I train the model and get this error:

learner.train(epochs=3)

RuntimeError: expected scalar type Long but found Int

I'm not sure how I can debug this error since it seems like it's happening deeper in Pytorch. Maybe one of my parameters are of the wrong type? Wondering if anyone has any advice.

(Also, this is an offshoot of a previous question I had, which @AdeelH answered 🙏)

[AdeelH]

Please share the full stack trace.

[natedailey7]

Yes, that code gave me the following output:

torch.Size([163, 163]) torch.uint8

[AdeelH]

Okay, that is definitely a bug. I think the conversion to Long is not getting triggered because there's no transform.

As a workaround, you can try the following:

train_ds = SemanticSegmentationSlidingWindowGeoDataset.from_uris(
    ...
    transform=A.Resize(163, 163)
)

This won't actually transform the image since it's already 163x163, but it should trigger this line.

[natedailey7]

Hmmm... I added the line:

transform=A.Resize(163, 163)

to my SemanticSegmentationSlidingWindowGeoDatasets (training and validation), but I'm still getting the same error: RuntimeError: expected scalar type Long but found Int. Maybe there is a way I can manually convert to Long.

In case it's helpful information, I get the same error when I try creating my training and validation datasets with random windows.

[AdeelH]

Okay, interesting. I will need to take a closer look at it then. In the meantime, something like the following might serve as a quick-and-dirty patch:

class SemanticSegmentationSlidingWindowGeoDatasetCustom(SemanticSegmentationSlidingWindowGeoDataset):
	def __getitem__(self, key):
		x, y = super().__getitem__(key)
		y = y.long()
		return x, y

And then use SemanticSegmentationSlidingWindowGeoDatasetCustom in place of SemanticSegmentationSlidingWindowGeoDataset.

The y returned by the dataset should now have the dtype torch.int64.

[natedailey7]

That fix did the trick! I can't thank you enough for digging into this.

[AdeelH]

Should be fixed on master now (via #1954). The fix will be included in the next patch release (v0.21.3).

[natedailey7]

That's awesome! Thank you! :)

[cpuah]

Hi there. Need help. Found this thread. Have encountered the same error when I execute learner.train(epochs=3) going through the train a model tutorial scripts (here: https://docs.rastervision.io/en/latest/usage/tutorials/train.html).

I tried updating RasterVision to 0.21.3 by using "pip install rastervision==0.21.3"

but am still getting this error regarding scalar long.

I am quite a novice and am running python 3.10 from vscode.

any advise would be appreciated.

many thanks
c