Joshua Lochner

> @xenova Is bicubic/linear not a config option? If it is not, I am afraid we need to patch indeed. Unfortunately not - this seems to be hard-coded in the...

Got it working 🥳 https://github.com/xenova/transformers.js/pull/473

https://github.com/huggingface/transformers/pull/26522 Is now merged into transformers!

Done ✅ (merged main)

Converted to draft and will mark as ready when I've added back the unit tests. @fxmarty once again, it requires a tiny model - do you have a script to...

@fxmarty done ✅ (I just merged main)

If it helps debug, it does work for input size = 4: ```js import FFT from 'https://cdn.jsdelivr.net/npm/[email protected]/+esm' const input = new Float32Array(4); input.fill(1); const f = new FFT(input.length); const out...

Further investigation shows this only occurs for `realTransform`: ```js import FFT from 'https://cdn.jsdelivr.net/npm/[email protected]/+esm' const input = new Float32Array([1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1,...

Also, if you replace this block of code: https://github.com/indutny/fft.js/blob/4a18cf88fcdbd4ad5acca6eaea06a0b462047835/lib/fft.js#L339-L439 with this block of code: https://github.com/indutny/fft.js/blob/4a18cf88fcdbd4ad5acca6eaea06a0b462047835/lib/fft.js#L151-L222 The output is as expected, so it seems to be an error in the optimization.

Oh wow, that was it... thanks so much @daniele-roncaglioni!