These can be used on the fly with minimal or no changes to test deploy visual detection models to the Clarifai platform. See the required files section for each model below and deployment instruction.
YOLOF Requirements to run tests locally:
Download checkpoint and save it in yolof/config/
:
$ wget -P yolof/config https://download.openmmlab.com/mmdetection/v2.0/yolof/yolof_r50_c5_8x8_1x_coco/yolof_r50_c5_8x8_1x_coco_20210425_024427-8e864411.pth
Install dependecies to test locally
$ pip install -r yolof/requirements.txt
Torch serve model format faster-rcnn_torchserve
To utilize a Torch serve model (.mar file) created by running torch-model-archiver – essentially a zip file containing the model checkpoint, Python code, and other components – within this module, follow these steps:
- Unzip the .mar file to obtain your checkpoint.
- Implement your postprocess method in inference.py.
For example: Faster-RCNN example, suppose you already have .mar file following the torch serve example
unzip it to ./faster-rcnn_torchserve/model_store/hub/checkpoints
as the Torch cache is configured to use this folder in torch serve inference.py.
$ unzip faster_rcnn.mar -d ./faster-rcnn_torchserve/model_store/hub/checkpoints/
# in model_store/hub/checkpoints you will have
model_store/hub/checkpoints/
├── MAR-INF
│ └── MANIFEST.json
├── model.py
└── fasterrcnn_resnet50_fpn_coco-258fb6c6.pth
Install dependecies to test locally
$ pip install -r faster-rcnn_torchserve/requirements.txt
YOLOX Requirements to run tests locally:
Download checkpoint and save it in yolox/configs/yolox/
, e.g download x
type of model:
$ wget -P yolox/configs/yolox/ https://download.openmmlab.com/mmdetection/v2.0/yolox/yolox_x_8x8_300e_coco/yolox_x_8x8_300e_coco_20211126_140254-1ef88d67.pth
Note: If you want to use a different model type or checkpoint, remember to update the
checkpoint
andconfig_path
in theinference.py
file accordingly.
Install dependecies to test locally
$ pip install -r yolox/requirements.txt
Steps to deploy one of above examples after downloading weights and testing to the Clarifai platform.
Note: set
--no-test
flag forbuild
andupload
command to disable testing
- Build
$ clarifai build model <path/to/folder> # either `faster-rcnn_torchserve` or `yolof` or `yolox`
upload *.clarifai
file to storage to obtain direct download url
- Upload
$ clarifai upload model <path/to/folder> --url <your_url>