PaddleX/docs/support_list/model_list_gcu.en.md

---

comments: true

PaddleX Model List (Enflame GCU)

PaddleX incorporates multiple pipelines, each containing several modules, and each module encompasses various models. You can select the appropriate models based on the benchmark data below. If you prioritize model accuracy, choose models with higher accuracy. If you prioritize model size, select models with smaller storage requirements.

Image Classification Module

Model Name	Top-1 Accuracy (%)	Model Size (M)	Model Download Link
ResNet50	76.96	90.8 M	Inference Model/Trained Model

Note: The above accuracy metrics refer to Top-1 Accuracy on the ImageNet-1k validation set.

Object Detection Module

Model Name	mAP (%)	Model Size (M)	Model Download Link
PP-YOLOE_plus-L	52.8	185.3 M	Inference Model/Trained Model
PP-YOLOE_plus-M	49.7	83.2 M	Inference Model/Trained Model
PP-YOLOE_plus-S	43.6	28.3 M	Inference Model/Trained Model
PP-YOLOE_plus-X	54.7	349.4 M	Inference Model/Trained Model
RT-DETR-H	56.3	435.8 M	Inference Model/Trained Model
RT-DETR-L	53.0	113.7 M	Inference Model/Trained Model
RT-DETR-R18	46.5	70.7 M	Inference Model/Trained Model
RT-DETR-R50	53.1	149.1 M	Inference Model/Trained Model
RT-DETR-X	54.8	232.9 M	Inference Model/Trained Model

Note: The above accuracy metrics are for COCO2017 validation set mAP(0.5:0.95).

Text Detection Module

Model Name	Detection Hmean (%)	Model Size (M)	Model Download Link
PP-OCRv4_mobile_det	77.79	4.2 M	Inference Model/Trained Model
PP-OCRv4_server_det	82.69	100.1 M	Inference Model/Trained Model

Note: The above accuracy metrics are evaluated on PaddleOCR's self-built Chinese dataset, covering street scenes, web images, documents, and handwritten scenarios, with 500 images for detection.

Text Recognition Module

Model Name	Recognition Avg Accuracy (%)	Model Size (M)	Model Download Link
PP-OCRv4_mobile_rec	78.20	10.6 M	Inference Model/Trained Model
PP-OCRv4_server_rec	79.20	71.2 M	Inference Model/Trained Model

Note: The above accuracy metrics are evaluated on PaddleOCR's self-built Chinese dataset, covering street scenes, web images, documents, and handwritten scenarios, with 11,000 images for text recognition.