LPIPS for MegEngine

MegEngine port for LPIPS metric.

Get Started

Run pip3 install mge-lpips to install this package. The provided mge_lpips.LPIPS is a normal MegEngine M.Module which can be used as a metric or a loss function.

import numpy as np
import megengine as mge
from mge_lpips import LPIPS

metric_lpips = LPIPS(net="alex")

in1 = np.random.uniform(-1, 1, size=(2, 3, 256, 256)).astype(np.float32)
in2 = np.random.uniform(-1, 1, size=(2, 3, 256, 256)).astype(np.float32)

d = metric_lpips(mge.Tensor(in1), mge.Tensor(in2))

Acknowledgements

The code of this project is modified from the original LPIPS project. The backbone model code is modified from torchvision. The weights are from the origian LPIPS project and torchvision.