1
0
mirror of https://github.com/immich-app/immich.git synced 2024-12-23 02:06:15 +02:00
immich/machine-learning/app/models/transforms.py
Mert 87a0ba3db3
feat(ml): export clip models to ONNX and host models on Hugging Face (#4700)
* export clip models

* export to hf

refactored export code

* export mclip, general refactoring

cleanup

* updated conda deps

* do transforms with pillow and numpy, add tokenization config to export, general refactoring

* moved conda dockerfile, re-added poetry

* minor fixes

* updated link

* updated tests

* removed `requirements.txt` from workflow

* fixed mimalloc path

* removed torchvision

* cleaner np typing

* review suggestions

* update default model name

* update test
2023-10-31 05:02:04 -05:00

36 lines
1.1 KiB
Python

import numpy as np
from PIL import Image
from app.schemas import ndarray_f32
_PIL_RESAMPLING_METHODS = {resampling.name.lower(): resampling for resampling in Image.Resampling}
def resize(img: Image.Image, size: int) -> Image.Image:
if img.width < img.height:
return img.resize((size, int((img.height / img.width) * size)), resample=Image.BICUBIC)
else:
return img.resize((int((img.width / img.height) * size), size), resample=Image.BICUBIC)
# https://stackoverflow.com/a/60883103
def crop(img: Image.Image, size: int) -> Image.Image:
left = int((img.size[0] / 2) - (size / 2))
upper = int((img.size[1] / 2) - (size / 2))
right = left + size
lower = upper + size
return img.crop((left, upper, right, lower))
def to_numpy(img: Image.Image) -> ndarray_f32:
return np.asarray(img.convert("RGB")).astype(np.float32) / 255.0
def normalize(img: ndarray_f32, mean: float | ndarray_f32, std: float | ndarray_f32) -> ndarray_f32:
return (img - mean) / std
def get_pil_resampling(resample: str) -> Image.Resampling:
return _PIL_RESAMPLING_METHODS[resample.lower()]