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.
| Model | Accuracy drop vs FP32 |
|---|
dla34.in1k | 0.59 |
dla60.in1k | 0.55 |
dla60_res2net.in1k | 0.15 |
dla102.in1k | 0.03 |
dla169.in1k | 0.27 |
efficientnet_es.ra_in1k | 0.02 |
efficientnet_es_pruned.in1k | 0.13 |
efficientnet_lite0.ra_in1k | 0.22 |
dla46_c.in1k | 1.54 |
fbnetc_100.rmsp_in1k | 0.24 |
gernet_m.idstcv_in1k | 0.05 |
gernet_s.idstcv_in1k | 0.18 |
mnasnet_100.rmsp_in1k | 0.28 |
mobilenetv2_050.lamb_in1k | 0.92 |
mobilenetv2_120d.ra_in1k | 0.44 |
mobilenetv2_140.ra_in1k | 0.89 |
res2net50_14w_8s.in1k | 0.17 |
res2net50_26w_4s.in1k | 0.17 |
res2net50_26w_6s.in1k | 0.06 |
res2net50_48w_2s.in1k | 0.09 |
res2net50d.in1k | 0.00 |
res2net101_26w_4s.in1k | 0.19 |
res2net101d.in1k | 0.08 |
resnet10t.c3_in1k | 1.61 |
resnet14t.c3_in1k | 0.85 |
resnet50c.gluon_in1k | 0.03 |
resnet50s.gluon_in1k | 0.19 |
resnet101c.gluon_in1k | 0.08 |
resnet101d.gluon_in1k | 0.10 |
resnet101s.gluon_in1k | 0.18 |
resnet152d.gluon_in1k | 0.15 |
selecsls42b.in1k | 0.25 |
selecsls60.in1k | 0.05 |
selecsls60b.in1k | 0.20 |
spnasnet_100.rmsp_in1k | 0.25 |
tf_efficientnet_es.in1k | 0.26 |
tf_efficientnet_lite0.in1k | 0.33 |
tf_mobilenetv3_large_minimal_100.in1k | 1.68 |
wide_resnet101_2.tv2_in1k | 0.26 |
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
Object Detection
| Model | ONNX | Repo | Resolution | Dataset | Ref FP32 mAP | Model license |
|---|
| YOLOv4 | | 🔗 | 416x416 | COCO2017 | 25.00 | GPL-3.0 |
| YOLOv4-CSP-Leaky | | 🔗 | 640x640 | COCO2017 | 29.57 | GPL-3.0 |
Keypoint Detection
| Model | ONNX | Repo | Resolution | Dataset | Ref FP32 mAP | Model license |
|---|
| YOLO26x-pose | 🔗 | 🔗 | 640x640 | COCO2017 | 72.75 | AGPL-3.0 |
Semantic Segmentation
Image Enhancement Super Resolution
| Model | ONNX | Repo | Resolution | Dataset | Ref FP32 PSNR | Model license |
|---|
| Real-ESRGAN-x4plus | 🔗 | 🔗 | 128x128 | SuperResolutionCustomSet128x128 | 24.77 | BSD-3-Clause |
See also