Swin-Transformer-Serve

Deploy Pre-trained Swin-Transformer classifier trained on ImageNet 1K using TorchServe

Create and activate virtual env

virtualenv env --python=python3
source env/bin/activate

Clone and Install dependencies

git clone https://github.com/kamalkraj/Swin-Transformer-Serve.git
cd Swin-Transformer-Serve
# clone TorchServe
git clone https://github.com/pytorch/serve.git
cd serve
# Refer to readme in serve repo for CUDA enabled local installtion
# The instruction below is for CPU
python ./ts_scripts/install_dependencies.py
pip install torchserve torch-model-archiver

Pretrained Weights and Config

Download weights.

name	pretrain	resolution	acc@1	acc@5	1K model
Swin-T	ImageNet-1K	224x224	81.2	95.5	github/baidu
Swin-S	ImageNet-1K	224x224	83.2	96.2	github/baidu
Swin-B	ImageNet-1K	224x224	83.5	96.5	github/baidu
Swin-B	ImageNet-1K	384x384	84.5	97.0	github/baidu
Swin-B	ImageNet-22K	224x224	85.2	97.5	github/baidu
Swin-B	ImageNet-22K	384x384	86.4	98.0	github/baidu
Swin-L	ImageNet-22K	224x224	86.3	97.9	github/baidu
Swin-L	ImageNet-22K	384x384	87.3	98.2	github/baidu

For demo we will use Tiny model.

wget https://github.com/SwinTransformer/storage/releases/download/v1.0.0/swin_tiny_patch4_window7_224.pth -O weights/swin_tiny_patch4_window7_224.pth

Copy the corresonding model config file from to swin_config.yaml

cp configs/swin_tiny_patch4_window7_224.yaml swin_config.yaml

TorchServe

Create MAR file using torch-model-archiver.

torch-model-archiver --model-name swin -v 1.0 --model-file swin_transformer.py --serialized-file weights/swin_tiny_patch4_window7_224.pth --handler swin_handler.py --extra-files index_to_name.json,swin_config.yaml --requirements-file requirements.txt
# move the swin.mar to model_store folder
mv swin.mar model_store

Start the torchserve using below cmd

in the config.properties we have set the configurations.Refer

torchserve --start --model-store model_store --models swin=swin.mar

test prediction using cURL

curl http://127.0.0.1:8080/predictions/swin -T kitten_small.jpg

{
  "tabby": 0.44951513409614563,
  "tiger_cat": 0.1962115466594696,
  "lynx": 0.16013166308403015,
  "Egyptian_cat": 0.08244507014751434,
  "tiger": 0.015334611758589745
}

As we set in the configuration file only one worker will be created per model. To increase workers and batch inference can done using Management API

Increase worker

curl -v -X PUT "http://localhost:8081/models/swin?min_worker=3"

This will increase minimum workers to 3.

Batch Inference.

Reference

#if the model is already running dergister it first
curl -X DELETE "http://localhost:8081/models/swin/1.0"
# The following command will register a model "swin.mar" and configure TorchServe to use a batch_size of 8 and a max batch delay of 50 milli seconds.
curl -X POST "localhost:8081/models?url=swin.mar&batch_size=8&max_batch_delay=50"

Stop torchserve

torchserve --stop

Model deploy using docker

Update config.properties file with following lines

install_py_dep_per_model=true
default_workers_per_model=1

Execute the below cmds to build a cpu docker image

cd serve/docker
sudo ./build_image.sh

gpu docker image

sudo ./build_image.sh -g -cv cu102

On successful docker build

CPU

cd Swin-Transformer-Serve
sudo docker run --rm -it -p 8080:8080 -p 8081:8081 -p 8082:8082 -p 7070:7070 -p 7071:7071 -v $(pwd)/model_store:/home/model-server/model-store pytorch/torchserve:latest-cpu

GPU

cd Swin-Transformer-Serve
sudo docker run --rm -it --gpus all -p 8080:8080 -p 8081:8081 -p 8082:8082 -p 7070:7070 -p 7071:7071 -v $(pwd)/model_store:/home/model-server/model-store pytorch/torchserve:latest-gpu

Register model

curl -X POST "localhost:8081/models?url=swin.mar&batch_size=1&max_batch_delay=50"

Creating mar file for torchscript mode model

Generate serialized-file using TorchScript

import torch

from swin_handler import get_config
from swin_transformer import SwinTransformer

config = get_config("swin_config.yaml")
model = SwinTransformer(img_size=config.DATA.IMG_SIZE,
                                patch_size=config.MODEL.SWIN.PATCH_SIZE,
                                in_chans=config.MODEL.SWIN.IN_CHANS,
                                num_classes=config.MODEL.NUM_CLASSES,
                                embed_dim=config.MODEL.SWIN.EMBED_DIM,
                                depths=config.MODEL.SWIN.DEPTHS,
                                num_heads=config.MODEL.SWIN.NUM_HEADS,
                                window_size=config.MODEL.SWIN.WINDOW_SIZE,
                                mlp_ratio=config.MODEL.SWIN.MLP_RATIO,
                                qkv_bias=config.MODEL.SWIN.QKV_BIAS,
                                qk_scale=config.MODEL.SWIN.QK_SCALE,
                                drop_rate=config.MODEL.DROP_RATE,
                                drop_path_rate=config.MODEL.DROP_PATH_RATE,
                                ape=config.MODEL.SWIN.APE,
                                patch_norm=config.MODEL.SWIN.PATCH_NORM,
                                use_checkpoint=config.TRAIN.USE_CHECKPOINT)
model.load_state_dict(torch.load("weights/swin_tiny_patch4_window7_224.pth",map_location="cpu")["model"])
model.eval()
example_input = torch.rand(1, 3, config.DATA.IMG_SIZE, config.DATA.IMG_SIZE)
traced_script_module = torch.jit.trace(model, example_input)

traced_script_module.save("swin.pt")

TorchServe

torch-model-archiver --model-name swin -v 1.0 --serialized-file weights/swin_tiny_patch4_window7_224.pth --handler swin_handler.py --extra-files index_to_name.json,swin_config.yaml --requirements-file requirements.txt

Other steps are same for both eager mode and torchscript