#821 Feature/sg 735 deci yolo qs

Merged

Ghost merged 1 commits into Deci-AI:feature/SG-736_deci_yolo_rf100 from deci-ai:feature/SG-735_deci_yolo_qs

3 changed files with 68 additions and 0 deletions

DeciYolo Quickstart

We developed a new deep learning architecture that competes with YOLOv8 using their efficient AutoNAC algorithm. We incorporated quantization-aware RepVGG blocks into the model architecture to ensure compatibility with Post-Training Quantization, making it more flexible and usable for different hardware configurations.

In this tutorial, we will go over the basic functionalities of DeciYolo very briefly.

Instantiate a DeciYolo Model

from super_gradients.training import models
from super_gradients.common.object_names import Models

net = models.get(Models.DECIYOLO_S, pretrained_weights="coco")

Predict

prediction = net.predict("https://www.aljazeera.com/wp-content/uploads/2022/12/2022-12-03T205130Z_851430040_UP1EIC31LXSAZ_RTRMADP_3_SOCCER-WORLDCUP-ARG-AUS-REPORT.jpg?w=770&resize=770%2C436&quality=80")

Export to ONNX

models.convert_to_onnx(model=net, input_shape=(3,640,640), out_path="deci_yolo_s.onnx")

Train on RF100

Follow the setup instructions for RF100:

        - Follow the official instructions to download Roboflow100: https://github.com/roboflow/roboflow-100-benchmark?ref=roboflow-blog
            //!\\ To use this dataset, you must download the "coco" format, NOT the yolov5.

        - Your dataset should look like this:
            rf100
            ├── 4-fold-defect
            │      ├─ train
            │      │    ├─ 000000000001.jpg
            │      │    ├─ ...
            │      │    └─ _annotations.coco.json
            │      ├─ valid
            │      │    └─ ...
            │      └─ test
            │           └─ ...
            ├── abdomen-mri
            │      └─ ...
            └── ...

        - Install CoCo API: https://github.com/pdollar/coco/tree/master/PythonAPI

We will use the roboflow_deciyolo_sconfiguration to train the small variant of our DeciYolo, DeciYolo S.

To launch training on one of the RF100 datasets, we pass it through the dataset_name argument:

python -m super_gradients.train_from_recipe --config-name=roboflow_deciyolo_s  dataset_name=<DATASET_NAME> dataset_params.data_dir=<PATH_TO_RF100_ROOT> ckpt_root_dir=<YOUR_CHECKPOINTS_ROOT_DIRECTORY>

Replace <DATASET_NAME> with any of the RF100 datasets that you wish to train on.

Deci-AI / super-gradients connected to https://github.com/Deci-AI/super-gradients.git