ViT: Patchifying an Image
Intermediate
The problem, from first principles
A ViT treats an image as a sequence of patch tokens. Each patch is flattened into a vector before a linear projection, so the layout of the patches has to be set up carefully.
From theory to code
Implement patchify(img, p), which splits the image into flattened non-overlapping patches.
Constraints
- Patches are listed row by row.
Hints
Hint 1
Reshape to a grid of blocks, move the grid axes to the front, then flatten each block.
Keywords
research-paperscomputer-visiontransformervision
Python
Run your code to see results
No results yet
Run your code to check it against the examples, or submit to run every test.