Detector Recipes: Supported Model Architectures

The following table presents the Detector Recipes that we currently support. The modular nature of LEIP recipes makes it easy to evaluate different architectures for speed and accuracy on your target hardware to find the model that best meets your application requirements.

Model	Default Size	recipe_name	architecture_name
YOLOv5 Large	640	`yolov5`	`yolov5l`
YOLOv5 Medium	640	`yolov5`	`yolov5m`
YOLOv5 Small	640	`yolov5`	`yolov5s`
Mobilenet V1 SSD	300	`ssd`	`mb1-ssd`
Mobilenet V2 Lite SSD	300	`ssd`	`mb2-ssd-lite`
EfficientDet D0	512	`efficientdet`	`efficientdet_d0`
TF EfficientDet D0 **	512	`efficientdet`	`tf_efficientdet_d0`
EfficientDet Q0	512	`efficientdet`	`efficientdet_q0`
TF EfficientDet Lite3	512	`efficientdet`	`tf_efficientdet_lite3`
EfficientDet D1	640	`efficientdet`	`efficientdet_d1`
TF EfficientDet D1 **	640	`efficientdet`	`tf_efficientdet_d1`
EfficientDet Q1	640	`efficientdet`	`efficientdet_q1`
TF EfficientDet Lite4	640	`efficientdet`	`tf_efficientdet_lite4`
EfficientDet V2 DT	768	`efficientdet`	`efficientdetv2_dt`
EfficientDet Q2	768	`efficientdet`	`efficientdet_q2`
TF EfficientDet D2 **	768	`efficientdet`	`efficientdet_q1`
ResDet 50	640	`efficientdet`	`resdet50`

** Known issue: The tf_effcientdet_d# recipes will currently not calibrate for Int8 optimization on Nvidia Ampere devices. If you want to run these models on Ampere devices, you will need to create the .activations files on a pre-Ampere device.

Please contact Latent AI if you require larger to smaller models.