metadata

license: afl-3.0
tags:
  - feature_extraction
  - image
  - perceptual_metric
datasets:
  - Jorgvt/TID2008
  - TID2013
metrics:
  - pearsonr
model-index:
  - name: PerceptNet
    results:
      - task:
          type: feature_extraction
          name: Perceptual Distance
        dataset:
          type: image
          name: tid2013
        metrics:
          - type: pearsonr
            value: 0.93
            name: PearsonR (MOS)

PerceptNet

PercepNet model trained on TID2008 and validated on TID2013, obtaining 0.97 and 0.93 Pearson Correlation respectively.

Link to the run: https://wandb.ai/jorgvt/PerceptNet/runs/28m2cnzj?workspace=user-jorgvt

Usage

There are two alternatives to use the model: install our development repo and load the pretrained weights manually, and load the model using from_pretrained_keras:

Loading weights manually

As of now to use the model you have to install the PerceptNet repo to get access to the PerceptNet class where you will load the weights available here like this:

from perceptnet.networks import PerceptNet
from tensorflow.keras.utils import get_file

weights_path = get_file(fname='perceptnet_rgb.h5',
                        origin='https://huggingface.co/Jorgvt/PerceptNet/resolve/main/tf_model.h5')
model = PerceptNet(kernel_initializer='ones', gdn_kernel_size=1, learnable_undersampling=False)
model.build(input_shape=(None, 384, 512, 3))
model.load_weights(weights_path)

PerceptNet requires wandb to be installed. It's something we're looking into.

Directly from the Hub

As every other Keras model in the Hub, it can be loaded as follows:

from huggingface_hub import from_pretrained_keras
model = from_pretrained_keras("Jorgvt/PerceptNet", compile=False)

Keep in mind that the model uses grouped convolutions and, at least in Colab, Unimplemented Errors may arise when using it in CPU.