Car edges-to-photo pix2pix in PyTorch

This reop is forked in this repe. Our team's goal was replace dataset to car-dataset.

We use car-dataset from stanford.

First, we remove image's background by Mask-Rcnn. Mask-Rcnn help us gereate mask only for the biggest car in image and crop image by it's bounding box.

Data preprocessing

Tarin

cd [path to pytorch-car-pix2pix]

python train.py --dataroot /home/u2546764/pytorch-car-pix2pix/datasets/car_data --name car_pix2pix --model pix2pix --direction BtoA --batch_size 2000 --gpu_ids 0,1,2,3,4,5,6,7 --num_threads 16

Test

python test.py --dataroot /home/u2546764/pytorch-car-pix2pix/datasets/car_data --direction BtoA --model pix2pix --name car_pix2pix

Demo

Facing problem

1. data preprocessing resizing
2. data i/o speed method: Load all data to memory result: reduce about 20% training time

Hardware

32 core cpu
Telsa V100 *8
480G memory