simet.transforms.inception

simet.transforms.inception

InceptionTransform

Bases: Transform

Preprocessing pipeline for Inception-v3 feature extraction.

Produces a torchvision.transforms.Compose that: 1) resizes the shorter side to 299 px, 2) center-crops to 299×299, 3) converts to a tensor, 4) normalizes each channel to the range ~[-1, 1] via mean=0.5, std=0.5.

Notes

• The normalization here uses mean=[0.5, 0.5, 0.5] and std=[0.5, 0.5, 0.5] (mapping 0..1 → −1..1). If you are using torchvision’s pretrained Inception-v3 weights (Inception_V3_Weights.IMAGENET1K_V1), the typical normalization is ImageNet mean/std ([0.485, 0.456, 0.406] / [0.229, 0.224, 0.225]). Ensure your transform matches the expectations of the model you use for feature extraction to avoid feature drift.
• The center crop after an exact 299 resize is partly redundant but ensures consistent 299×299 framing for non-square inputs.

Example

tfm = InceptionTransform().get_transform() img_t = tfm(PIL_image) # shape [3, 299, 299], roughly in [-1, 1]

get_transform

get_transform()

Return the composed preprocessing pipeline for Inception-v3.

Returns:

Name	Type	Description
Compose	Compose	Resize→CenterCrop→ToTensor→Normalize(mean=0.5, std=0.5).

Source code in simet/transforms/inception.py

@override
def get_transform(self) -> Compose:
    """Return the composed preprocessing pipeline for Inception-v3.

    Returns:
        Compose: Resize→CenterCrop→ToTensor→Normalize(mean=0.5, std=0.5).
    """
    return Compose(
        [
            transforms.Resize(299),  # InceptionV3 expects 299x299
            transforms.CenterCrop(299),  # Ensures centered 299×299 crop
            transforms.ToTensor(),  # PIL→Tensor in [0, 1]
            transforms.Normalize(
                mean=[0.5, 0.5, 0.5],
                std=[0.5, 0.5, 0.5],
            ),  # maps to roughly [-1, 1]
        ]
    )