Visual Parser (ViP)

This is the official implementation of the paper Visual Parser: Representing Part-whole Hierarchies with Transformers.

We also released the codebase for ViP object detection & instance segmentation.

Key Features & TLDR

1. PyTorch Implementation of the ViP network. Check it out at models/vip.py

2. A fast and neat implementation of the relative positional encoding proposed in HaloNet, BOTNet and AANet.

3. A transformer-friendly FLOPS & Param counter that supports FLOPS calculation for einsum and matmul operations.

Prerequisite

Please refer to get_started.md.

Results and Models

All models listed below are evaluated with input size 224x224

To load the pretrained checkpoint, e.g. ViP-Tiny, simply run:

# first download the checkpoint and name it as vip_t_dict.pth
from models.vip import vip_tiny
model = vip_tiny(pretrained="vip_t_dict.pth")

Evaluation

To evaluate a pre-trained ViP on ImageNet val, run:

python3 main.py <data-root> --model <model-name> -b <batch-size> --eval_checkpoint <path-to-checkpoint>

Training from scratch

To train a ViP on ImageNet from scratch, run:

bash ./distributed_train.sh <job-name> <config-path> <num-gpus>

For example, to train ViP with 8 GPU on a single node, run:

ViP-Tiny:

bash ./distributed_train.sh vip-t-001 configs/vip_t_bs1024.yaml 8

ViP-Small:

bash ./distributed_train.sh vip-s-001 configs/vip_s_bs1024.yaml 8

ViP-Medium:

bash ./distributed_train.sh vip-m-001 configs/vip_m_bs1024.yaml 8

ViP-Base:

bash ./distributed_train.sh vip-b-001 configs/vip_b_bs1024.yaml 8

Profiling the model

To measure the throughput, run:

python3 test_throughput.py <model-name>

For example, if you want to get the test speed of Vip-Tiny on your device, run:

python3 test_throughput.py vip-tiny

To measure the FLOPS and number of parameters, run:

python3 test_flops.py <model-name>

Citing ViP

@article{vip,
  title={Visual Parser: Representing Part-whole Hierarchies with Transformers},
  author={Sun, Shuyang and Yue, Xiaoyu, Bai, Song and Torr, Philip},
  journal={arXiv preprint arXiv:2107.05790},
  year={2021}
}

Contact

If you have any questions, don't hesitate to contact Shuyang (Kevin) Sun. You can easily reach him by sending an email to [email protected].