2_inference.md

II. Inference

Pre-trained Checkpoint

We provide our ShineOn model checkpoint on Google Drive.

Command Line

Reconstruction Task

Reconstruction tests how well model can synthesize the person re-wearing their ORIGINAL clothes.

1. Warp Module

   The warp module takes the product images

   python test.py \
   --name reconstruction_warp \
   --model warp \
   --workers 4 \
   --batch 4 \
   --dataset vvt \
   --datamode test \
   --checkpoint experiments/path/to/WARP/checkpoint.ckpt

2. Try-on Module

   python test.py \
   --name reconstruction_try_on \
   --model unet \
   --workers 4 \
   --batch 4 \
   --dataset vvt \
   --datamode test \
   --checkpoint experiments/path/to/UNET/checkpoint.ckpt \
   --warp_cloth_dir \
   test_results/reconstruction/checkpoint.ckpt/test/VVTDataset/warp-cloth

Try-on Task

Try-on tests how well the model can synthesize the person wearing a NEW article of clothing.

Use the flag --tryon_list to choose your CSV file that specifies cloth-person try on pairs.

The CSV file should be formatted with two columns (no headers):

path/to/cloth/product_image.png, VIDEO_ID

where path/to/cloth/product_image.png is a path to the cloth image, and VIDEO_ID is the name of a video folder containing the frames for one person.

1. Warp Module

   The warp-cloth folder must be generated for every tryon_file.csv you have.

   python test.py \
   --name warp_try_on \
   --model warp \
   --workers 4 \
   --batch 4 \
   --dataset vvt \
   --datamode test \
   --checkpoint experiments/path/to/WARP/checkpoint.ckpt \
   --tryon_list path/to/tryon_file.csv

2. Try-on Module

   python test.py \
   --name complete_try_on \
   --model unet \
   --workers 4 \
   --batch 4 \
   --dataset vvt \
   --datamode test \
   --checkpoint experiments/path/to/UNET/checkpoint.ckpt \
   --tryon_list path/to/tryon_file.csv \
   --warp_cloth_dir \
   test_results/tryon/checkpoint.ckpt/test/VVTDataset/warp-cloth