All models
This page lists all our trained models that are compiled for the Coral Edge TPU™.
For more information about each model type, including code examples and training scripts, refer to the model-specific pages that are linked on the Models page.
To train a custom model, using transfer learning or by building and training your own model, see our documentation about TensorFlow models on the Edge TPU.
Image classification (pre-trained) link
Model name | Detections/Dataset | Input size | Depth mul. | TF ver. | Latency 1 | Accuracy | Model size | Downloads |
---|---|---|---|---|---|---|---|---|
EfficientNet-EdgeTpu (L)* |
1,000 objects |
300x300x3 | N/A | 1 | 21.3 ms | |
12.8 MB | |
EfficientNet-EdgeTpu (M)* |
1,000 objects |
240x240x3 | N/A | 1 | 7.3 ms | |
8.7 MB | |
EfficientNet-EdgeTpu (S)* |
1,000 objects |
224x224x3 | N/A | 1 | 5.0 ms | |
6.8 MB | |
Inception V1 |
1,000 objects |
224x224x3 | N/A | 1 | 3.4 ms | |
7.0 MB | |
Inception V3 |
1,000 objects |
224x224x3 | N/A | 1 | 13.4 ms | |
12.0 MB | |
Inception V3 |
1,000 objects |
299x299x3 | N/A | 1 | 42.8 ms | |
23.9 MB | |
Inception V4 |
1,000 objects |
299x299x3 | N/A | 1 | 84.7 ms | |
42.9 MB | |
MobileNet V1 |
1,000 objects |
128x128x3 | 0.25 | 1 | 0.9 ms | |
0.7 MB | |
MobileNet V1 |
1,000 objects |
160x160x3 | 0.5 | 1 | 1.4 ms | |
1.6 MB | |
MobileNet V1 |
1,000 objects |
192x192x3 | 0.75 | 1 | 1.8 ms | |
2.8 MB | |
MobileNet V1 |
1,000 objects |
224x224x3 | 1.0 | 1 | 2.8 ms | |
4.4 MB | |
MobileNet V2 |
900+ birds |
224x224x3 | 1.0 | 1 | 2.6 ms | N/A | 4.1 MB | |
MobileNet V2 |
1000+ insects |
224x224x3 | 1.0 | 1 | 2.7 ms | N/A | 4.1 MB | |
MobileNet V2 |
2000+ plants |
224x224x3 | 1.0 | 1 | 2.6 ms | N/A | 5.5 MB | |
MobileNet V2 |
1,000 objects |
224x224x3 | 1.0 | 1 | 2.9 ms | |
4.0 MB | |
MobileNet V1 |
1,000 objects |
224x224x3 | 1.0 | 2 | 2.8 ms | |
4.5 MB | |
MobileNet V2 |
1,000 objects |
224x224x3 | 1.0 | 2 | 3.0 ms | |
4.1 MB | |
MobileNet V3 |
1,000 objects |
224x224x3 | 1.0 | 2 | 3.0 ms | |
4.9 MB | |
ResNet-50 |
1,000 objects |
224x224x3 | N/A | 2 | 42.2 ms | |
25.0 MB | |
Popular Products V1 |
100,000 popular |
224x224x3 | N/A | 1 | 7.0 ms | N/A | 9.8 MB |
1 Latency is the time to perform one inference, as measured with a Coral USB Accelerator on a desktop CPU. Latency varies between systems and is primarily intended for comparison between models. For more comparisons, see the Performance Benchmarks.
* Beware that the EfficientNet family of models have unique input quantization values (scale and zero-point) that you must use when preprocessing your input. For example preprocessing code, see the classify_image.py or classify_image.cc examples.
Image classification (on-device training) link
Model name | Training style | Base dataset | Input size | TF ver. | Model size | Downloads |
---|---|---|---|---|---|---|
EfficientNet-EdgeTpu (L) |
Backpropagation | 1,000 objects |
300x300x3 | 1 | 11.7 MB | |
EfficientNet-EdgeTpu (M) |
Backpropagation | 1,000 objects |
240x240x3 | 1 | 7.6 MB | |
EfficientNet-EdgeTpu (S) |
Backpropagation | 1,000 objects |
224x224x3 | 1 | 5.7 MB | |
MobileNet V1 |
Backpropagation | 1,000 objects |
224x224x3 | 1 | 3.5 MB | |
MobileNet V1 |
Weight imprinting | 1,000 objects |
224x224x3 | 1 | 5.3 MB |
Object detection link
Model name | Detections/Dataset | Input size | TF ver. | Latency 1 | mAP 2 | Model size | Downloads |
---|---|---|---|---|---|---|---|
SSD MobileNet V1 |
90 objects |
300x300x3 | 1 | 6.5 ms | 21.5% | 7.0 MB | |
SSD/FPN MobileNet V1 |
90 objects |
640x640x3 | 2 | 229.4 ms | 31.1% | 37.7 MB | |
SSD MobileNet V2 |
90 objects |
300x300x3 | 1 | 7.3 ms | 25.6% | 6.6 MB | |
SSD MobileNet V2 |
90 objects |
300x300x3 | 2 | 7.6 ms | 22.4% | 6.7 MB | |
SSD MobileNet V2 |
Faces |
320x320x3 | 1 | 5.2 ms | N/A | 6.7 MB | |
SSDLite MobileDet |
90 objects |
320x320x3 | 1 | 9.1 ms | 32.9% | 5.1 MB | |
EfficientDet-Lite0 |
90 objects |
320x320x3 | 2 | 37.4 ms | 30.4% | 5.7 MB | |
EfficientDet-Lite1 |
90 objects |
384x384x3 | 2 | 56.3 ms | 34.3% | 7.6 MB | |
EfficientDet-Lite2 |
90 objects |
448x448x3 | 2 | 104.6 ms | 36.0% | 10.2 MB | |
EfficientDet-Lite3 |
90 objects |
512x512x3 | 2 | 107.6 ms | 39.4% | 14.4 MB | |
EfficientDet-Lite3x* |
90 objects |
640x640x3 | 2 | 197.0 ms | 43.9% | 20.6 MB |
1 Latency is the time to perform one inference, as measured with a Coral USB Accelerator on a desktop CPU. Latency varies between systems and is primarily intended for comparison between models. For more comparisons, see the Performance Benchmarks.
2 mAP is the "mean average precision," as specified by the COCO evaluation metrics. Our evaluation uses a subset of the COCO17 dataset.
* EfficientDet-Lite3x is not compatible with Edge TPUs over USB; it can be used only with PCIe-based devices. Our benchmarks for EfficientDet-Lite3x come from a desktop system paired with the Asus AI Accelerator—your results will vary.
Semantic segmentation link
Model name | Detections/Dataset | Input size | Depth mul. | Output stride | TF ver. | Latency1 | Model size | Downloads |
---|---|---|---|---|---|---|---|---|
U-Net MobileNet v2 |
37 pets |
128x128x3 | N/A | N/A | 1 | 2.7 ms | 7.2 MB | |
U-Net MobileNet v2 |
37 pets |
256x256x3 | N/A | N/A | 1 | 29.0 ms | 7.3 MB | |
MobileNet v2 DeepLab v3 |
20 objects |
513x513x3 | 0.5 | N/A | 1 | 36.8 ms | 1.1 MB | |
MobileNet v2 DeepLab v3 |
20 objects |
513x513x3 | 1.0 | N/A | 1 | 43.0 ms | 2.9 MB | |
EdgeTPU-DeepLab-slim |
28 objects |
513x513x3 | 0.75 | N/A | 1 | 65.9 ms | 3.1 MB | |
MobileNet v1 BodyPix |
24 body parts |
352x480x3 | 0.75 | 16 | 1 | 6.9 ms | 1.6 MB | |
MobileNet v1 BodyPix |
24 body parts |
512x512x3 | 0.75 | 16 | 1 | 10.7 ms | 1.7 MB | |
MobileNet v1 BodyPix |
24 body parts |
480x640x3 | 0.75 | 16 | 1 | 12.3 ms | 1.8 MB | |
MobileNet v1 BodyPix |
24 body parts |
576x768x3 | 0.75 | 16 | 1 | 17.7 ms | 1.8 MB | |
MobileNet v1 BodyPix |
24 body parts |
768x1024x3 | 0.75 | 16 | 1 | 30.8 ms | 2.0 MB | |
MobileNet v1 BodyPix |
24 body parts |
720x1280x3 | 0.75 | 16 | 1 | 38.8 ms | 2.3 MB | |
ResNet-50 BodyPix |
24 body parts |
288x416x3 | N/A | 16 | 1 | 46.9 ms | 24.5 MB | |
ResNet-50 BodyPix |
24 body parts |
480x640x3 | N/A | 16 | 1 | 384.0 ms | 26.6 MB | |
ResNet-50 BodyPix |
24 body parts |
496x768x3 | N/A | 32 | 1 | 87.0 ms | 26.9 MB | |
ResNet-50 BodyPix |
24 body parts |
624x864x3 | N/A | 32 | 1 | 153.5 ms | 28.5 MB | |
ResNet-50 BodyPix |
24 body parts |
672x928x3 | N/A | 16 | 1 | 737.2 ms | 35.3 MB | |
ResNet-50 BodyPix |
24 body parts |
736x960x3 | N/A | 32 | 1 | N/A | 38.6 MB |
1 Latency is the time to perform one inference, as measured with a Coral USB Accelerator on a desktop CPU. Latency varies between systems and is primarily intended for comparison between models. For more comparisons, see the Performance Benchmarks.
If you want to process portrait-orientation images, download BodyPix models for portrait input.
Pose estimation link
Model name | Detections/Dataset | Input size | Output stride | TF ver. | Latency1 | Model size | Downloads |
---|---|---|---|---|---|---|---|
PoseNet MobileNet V1 |
17 body points |
353x481x3 | 16 | 1 | 5.8 ms | 1.5 MB | |
PoseNet MobileNet V1 |
17 body points |
481x641x3 | 16 | 1 | 10.3 ms | 1.7 MB | |
PoseNet MobileNet V1 |
17 body points |
721x1281x3 | 16 | 1 | 32.4 ms | 2.5 MB | |
MoveNet.SinglePose.Lightning |
17 body points |
192x192x3 | 4 | 2 | 7.1 ms | 3.1 MB | |
MoveNet.SinglePose.Thunder |
17 body points |
256x256x3 | 4 | 2 | 13.8 ms | 7.5 MB | |
PoseNet ResNet-50 |
17 body points |
288x416x3 | 16 | 1 | N/A | 24.4 MB | |
PoseNet ResNet-50 |
17 body points |
480x640x3 | 16 | 1 | N/A | 26.4 MB | |
PoseNet ResNet-50 |
17 body points |
496x768x3 | 32 | 1 | N/A | 26.8 MB | |
PoseNet ResNet-50 |
17 body points |
624x864x3 | 32 | 1 | N/A | 28.4 MB | |
PoseNet ResNet-50 |
17 body points |
672x928x3 | 16 | 1 | N/A | 35.0 MB | |
PoseNet ResNet-50 |
17 body points |
736x960x3 | 32 | 1 | N/A | 38.5 MB |
1 Latency is the time to perform one inference, as measured with a Coral USB Accelerator on a desktop CPU. Latency varies between systems and is primarily intended for comparison between models. For more comparisons, see the Performance Benchmarks.
Speech recognition link
Model name | Detections/Dataset | Input size | Model size | Downloads |
---|---|---|---|---|
Keyword Spotter v0.8 |
140+ phrases |
198x32x1 | 578 KB |