How to use the TorchInferencer - using the correct image size

[haimat]

I have trained and exported my model like so:

task_type = TaskType.CLASSIFICATION
root_folder = "/data/scratch/mkw-anomalib"
input_size = (256, 256)

datamodule = Folder(
    name="MKW",
    root=os.path.join(root_folder, "images"),
    normal_dir="normal",
    abnormal_dir="abnormal",
    image_size=input_size,
    task=task_type,
    train_batch_size=32,
)
datamodule.prepare_data()
datamodule.setup()

# Create the model
model = Fastflow()
engine = Engine(
    max_epochs=100,
    task=task_type,
)
engine.fit(datamodule=datamodule, model=model)
engine.export(model=model, export_type=ExportType.TORCH, export_root=os.path.join(root_folder, "export"))

Then I want to run inference on the GPU like so:

inferencer = TorchInferencer(path="/data/scratch/mkw-anomalib/export/weights/torch/model.pt", device="gpu")
image = read_image("/data/clients/mkw/images/normal/good00000195.png")
input_img = image.astype(np.float32) / 1.0
image_transposed = np.transpose(input_img, (2, 0, 1))
torch_image = torch.from_numpy(image_transposed)
result = inferencer.predict(torch_image)
print(result)

First, is this the correct way to prepare an image for inferencing?
And: Do I have to resize the image to (256, 256) as well before I call predict(), or is this done automatically by the model?

[samet-akcay]

This section of the documentation might be exactly what you are looking for
https://anomalib.readthedocs.io/en/latest/markdown/guides/how_to/data/transforms.html#transforms-during-inference

[haimat]

@samet-akcay Great, thanks a lot.
So, beside resizing, if I have an image in OpenCV (numpy) format, what steps do I need so that I can pass it to predict()?
Do I still need these following steps for such cv2 images as well?

input_img = image.astype(np.float32) / 1.0
image_transposed = np.transpose(input_img, (2, 0, 1))
torch_image = torch.from_numpy(image_transposed)