Skip to main content
Version: v1.7

Additional Models

Models verified on Metis that are not yet in the Model Zoo. They fall into two groups: those with no YAML configuration, which you deploy by adapting an existing template, and those that already have a YAML configuration but are still being verified.


Models without a YAML configuration

These have been verified on Metis but have no dedicated YAML configuration. Deploy them by adapting an existing template.

Image Classification

These classification models have been compiled and accuracy-verified on Metis. To use one, copy the mobilenetv4_small-imagenet.yaml template and update the timm_model_args.name field and preprocessing configuration to match your target model.

ModelAccuracy drop vs FP32
dla34.in1k0.59
dla60.in1k0.55
dla60_res2net.in1k0.15
dla102.in1k0.03
dla169.in1k0.27
efficientnet_es.ra_in1k0.02
efficientnet_es_pruned.in1k0.13
efficientnet_lite0.ra_in1k0.22
dla46_c.in1k1.54
fbnetc_100.rmsp_in1k0.24
gernet_m.idstcv_in1k0.05
gernet_s.idstcv_in1k0.18
mnasnet_100.rmsp_in1k0.28
mobilenetv2_050.lamb_in1k0.92
mobilenetv2_120d.ra_in1k0.44
mobilenetv2_140.ra_in1k0.89
res2net50_14w_8s.in1k0.17
res2net50_26w_4s.in1k0.17
res2net50_26w_6s.in1k0.06
res2net50_48w_2s.in1k0.09
res2net50d.in1k0.00
res2net101_26w_4s.in1k0.19
res2net101d.in1k0.08
resnet10t.c3_in1k1.61
resnet14t.c3_in1k0.85
resnet50c.gluon_in1k0.03
resnet50s.gluon_in1k0.19
resnet101c.gluon_in1k0.08
resnet101d.gluon_in1k0.10
resnet101s.gluon_in1k0.18
resnet152d.gluon_in1k0.15
selecsls42b.in1k0.25
selecsls60.in1k0.05
selecsls60b.in1k0.20
spnasnet_100.rmsp_in1k0.25
tf_efficientnet_es.in1k0.26
tf_efficientnet_lite0.in1k0.33
tf_mobilenetv3_large_minimal_100.in1k1.68
wide_resnet101_2.tv2_in1k0.26

Accuracy drop is measured as FP32 top-1 accuracy minus quantized (int8 on AIPU) top-1 accuracy.


Models with a YAML configuration

These models have a YAML configuration but have not yet been fully verified for speed and accuracy. They move to the Model Zoo once they are.

Image Classification

ModelONNXRepoResolutionDatasetRef FP32 Top1Model license
MobileNetV3-large🔗🔗224x224ImageNet-1K74.05BSD-3-Clause
MobileNetV3-small🔗🔗224x224ImageNet-1K67.67BSD-3-Clause

Object Detection

ModelONNXRepoResolutionDatasetRef FP32 mAPModel license
YOLOv4🔗416x416COCO201725.00GPL-3.0
YOLOv4-CSP-Leaky🔗640x640COCO201729.57GPL-3.0

Keypoint Detection

ModelONNXRepoResolutionDatasetRef FP32 mAPModel license
YOLO26x-pose🔗🔗640x640COCO201772.75AGPL-3.0

Semantic Segmentation

ModelONNXRepoResolutionDatasetRef FP32 mIoUModel license
YOLO26n-sem🔗🔗1024x1024Cityscapes71.42AGPL-3.0
YOLO26s-sem🔗🔗1024x1024Cityscapes76.50AGPL-3.0
YOLO26m-sem🔗🔗1024x1024Cityscapes79.23AGPL-3.0
YOLO26l-sem🔗🔗1024x1024Cityscapes79.73AGPL-3.0
YOLO26x-sem🔗🔗1024x1024Cityscapes80.76AGPL-3.0

Image Enhancement Super Resolution

ModelONNXRepoResolutionDatasetRef FP32 PSNRModel license
Real-ESRGAN-x4plus🔗🔗128x128SuperResolutionCustomSet128x12824.77BSD-3-Clause


See also