The section's substance is the preprocessing contract, not the
download: the image must be brought to the size the pretrained model
expects (224 x 224 for ConvNeXt) and the pixel intensities
standardized per colour channel using ImageNet's means and standard
deviations. Those constants are NOT printed in the extracted text,
so mean and sd are required arguments rather than baked-in defaults.
Usage
morie_pretprep(image, size, mean, sd, logits = NULL, topk = 1L)
Arguments
- image
height by width by channels array
- size
the square edge the model expects
- mean, sd
per-channel standardization constants
- logits
optional class logits, for the p. 459 argmax read-out
- topk
how many classes to report
Value
list(pixels, size, channels, cropside, pixelmean, pixelmax,
and pred/topk/topprob when logits are supplied)
Details
ponytail: centre crop to a square, then nearest-neighbour subsample.
weights.transforms() uses bilinear resizing; swap the sampler if you
need to match torchvision to the last decimal.
Examples
set.seed(1)
img <- array(runif(8 * 8 * 3), dim = c(8, 8, 3))
out <- morie_pretprep(img, size = 4, mean = c(0.5, 0.5, 0.5),
sd = c(0.25, 0.25, 0.25))
dim(out$image)
#> NULL
