conda create -n fashionmatrix python=3.9
conda activate fashionmatrix
cd FashionMatrix
pip install -r requirements.txt
For Detectron2, Segment Anything and GroundingDINO, install them locally:
# Detectron2(DensePose), SAM and GroundingDINO
python -m pip install -e detectron2
pip install git+https://github.com/facebookresearch/detectron2@main#subdirectory=projects/DensePose
python -m pip install -e segment_anything
python -m pip install -e GroundingDINO
mkdir checkpoints
# For most checkpoints, we provide them in HuggingFace, just clone it.
git clone https://huggingface.co/zhengchong/FashionMatrix ./checkpoints
# For ControlNet and BLIP, You need to clone them manually.
cd checkpoints
git clone https://huggingface.co/lllyasviel/control_v11p_sd15_lineart
git clone https://huggingface.co/Salesforce/blip-vqa-capfilt-large
The checkpoints folder should look like this:
checkpoints
├── blip-vqa-capfilt-large
├── control_v11p_sd15_lineart
├── realisticVisionV40_v40VAE
├── densepose
├── Base-DensePose-RCNN-FPN.yaml
├── densepose_rcnn_R_50_FPN_s1x.yaml
├── model_final_162be9.pkl
├── graphonomy
├── inference.pth
├── grounded-sam
├── groundingdino_swint_ogc.pth
├── sam_vit_b_01ec64.pth
├── Annotators
├── sk_model.pth
├── sk_model2.pth
To deploy Vicuna-13B in OpenAI API, follow this instructions.
Firstly, deploy the models on a local server:
# You can change anther host:port, but make sure to change the corresponding address in api.py
CUDA_VISIBLE_DEVICES='YOUR_DEVICE_ID' nohup python -u server.py --host=0.0.0.0 --port=8123 >server.log 2>&1 &
It is worth mentioning that all models (except LLM) can be run on a single consumer-grade GPU with 13G+ VRAM.
Then, run the gradio app:
python app_label.py