FIACCEL in PyTorch

Introduction

This is a PyTorch implementation of the FIACCEL

Running the code

To facilitate training we use pytorch_lightning and hydra to manage model configurations.

To create an environment with all the requirements for running this code, we suggest first verifying that the PyTorch version in the fiaccel.yml file is appropriate for your OS and server and selecting an appropriate one if it is not. Then create a new conda environment by running:

conda env create -f fiaccel.yml
conda activate fiaccel

You can train a model with model_train.py but remember to modify the the training configuration (config/train_config_*.yaml) and include paths to an appropriate training dataset.

Training data

fiaccel is trained on a proprietary dataset with one million internet video clips, each comprising 3 frames.

• We use the publcicly available Vimeo-90k dataset (Xue et al., 2019), which is commonly used dataset for video frame interpolation

FIACCEL: Implementation details

TOP

## TOP_SUB

## Ecore

Dcore

PE

Experimental_settings

>