Large-scale Product Recognition (CVPR 2021 AliProducts Challenge)

This repository contains source codes of SYSU-Youtu for CVPR 2021 AliProducts Challenge: Large-scale Product Recognition

You can find details from our technical reports Solution for Large-scale Long-tailed Recognition with Noisy Labels.

Our solution obtains 6.4365% mean class error rate in the leaderboard with our ensemble model, which ranks the Second.

Requirements

• Pytorch 1.6+
• Cuda 10

Prerequisite

Using code in tfrecord_tool to convert images from jpg to TF Records and txt.

Run

For transformers

cd Deit

cp -r /path/to/TFR-aliproduct_train /dev/shm/
cp -r /path/to/TFR-aliproduct_val /dev/shm/
cp -r /path/to/*.txt /dev/shm/

python3 -m torch.distributed.launch --nproc_per_node=8 --use_env main.py --model deit_small_patch16_224 --version _v9 --train --lr 5e-5 --epochs 100 2>&1 | tee ./logs/std.log

TFRecord

V4

V9