20 KiB
Hosted models
The following is an incomplete list of pre-trained models optimized to work with TensorFlow Lite.
To get started choosing a model, visit Models page with end-to-end examples, or pick a [TensorFlow Lite model from TensorFlow Hub] (https://tfhub.dev/s?deployment-format=lite).
Note: The best model for a given application depends on your requirements. For example, some applications might benefit from higher accuracy, while others require a small model size. You should test your application with a variety of models to find the optimal balance between size, performance, and accuracy.
Image classification
For more information about image classification, see Image classification.
Quantized models
Quantized image classification models offer the smallest model size and fastest performance, at the expense of accuracy. The performance values are measured on Pixel 3 on Android 10.
You can find many quantized models from TensorFlow Hub and get more model information there.
| Model name | Paper and model | Model size | Top-1 accuracy | Top-5 accuracy | CPU, 4 threads | NNAPI |
|---|---|---|---|---|---|---|
| Mobilenet_V1_0.25_128_quant | paper, tflite&pb | 0.5 Mb | 39.5% | 64.4% | 0.8 ms | 2 ms |
| Mobilenet_V1_0.25_160_quant | paper, tflite&pb | 0.5 Mb | 42.8% | 68.1% | 1.3 ms | 2.4 ms |
| Mobilenet_V1_0.25_192_quant | paper, tflite&pb | 0.5 Mb | 45.7% | 70.8% | 1.8 ms | 2.6 ms |
| Mobilenet_V1_0.25_224_quant | paper, tflite&pb | 0.5 Mb | 48.2% | 72.8% | 2.3 ms | 2.9 ms |
| Mobilenet_V1_0.50_128_quant | paper, tflite&pb | 1.4 Mb | 54.9% | 78.1% | 1.7 ms | 2.6 ms |
| Mobilenet_V1_0.50_160_quant | paper, tflite&pb | 1.4 Mb | 57.2% | 80.5% | 2.6 ms | 2.9 ms |
| Mobilenet_V1_0.50_192_quant | paper, tflite&pb | 1.4 Mb | 59.9% | 82.1% | 3.6 ms | 3.3 ms |
| Mobilenet_V1_0.50_224_quant | paper, tflite&pb | 1.4 Mb | 61.2% | 83.2% | 4.7 ms | 3.6 ms |
| Mobilenet_V1_0.75_128_quant | paper, tflite&pb | 2.6 Mb | 55.9% | 79.1% | 3.1 ms | 3.2 ms |
| Mobilenet_V1_0.75_160_quant | paper, tflite&pb | 2.6 Mb | 62.4% | 83.7% | 4.7 ms | 3.8 ms |
| Mobilenet_V1_0.75_192_quant | paper, tflite&pb | 2.6 Mb | 66.1% | 86.2% | 6.4 ms | 4.2 ms |
| Mobilenet_V1_0.75_224_quant | paper, tflite&pb | 2.6 Mb | 66.9% | 86.9% | 8.5 ms | 4.8 ms |
| Mobilenet_V1_1.0_128_quant | paper, tflite&pb | 4.3 Mb | 63.3% | 84.1% | 4.8 ms | 3.8 ms |
| Mobilenet_V1_1.0_160_quant | paper, tflite&pb | 4.3 Mb | 66.9% | 86.7% | 7.3 ms | 4.6 ms |
| Mobilenet_V1_1.0_192_quant | paper, tflite&pb | 4.3 Mb | 69.1% | 88.1% | 9.9 ms | 5.2 ms |
| Mobilenet_V1_1.0_224_quant | paper, tflite&pb | 4.3 Mb | 70.0% | 89.0% | 13 ms | 6.0 ms |
| Mobilenet_V2_1.0_224_quant | paper, tflite&pb | 3.4 Mb | 70.8% | 89.9% | 12 ms | 6.9 ms |
| Inception_V1_quant | paper, tflite&pb | 6.4 Mb | 70.1% | 89.8% | 39 ms | 36 ms |
| Inception_V2_quant | paper, tflite&pb | 11 Mb | 73.5% | 91.4% | 59 ms | 18 ms |
| Inception_V3_quant | paper,tflite&pb | 23 Mb | 77.5% | 93.7% | 148 ms | 74 ms |
| Inception_V4_quant | paper, tflite&pb | 41 Mb | 79.5% | 93.9% | 268 ms | 155 ms |
Note: The model files include both TF Lite FlatBuffer and Tensorflow frozen Graph.
Note: Performance numbers were benchmarked on Pixel-3 (Android 10). Accuracy numbers were computed using the TFLite accuracy tool.
Floating point models
Floating point models offer the best accuracy, at the expense of model size and performance. GPU acceleration requires the use of floating point models. The performance values are measured on Pixel 3 on Android 10.
You can find many image classification models from TensorFlow Hub and get more model information there.
| Model name | Paper and model | Model size | Top-1 accuracy | Top-5 accuracy | CPU, 4 threads | GPU | NNAPI |
|---|---|---|---|---|---|---|---|
| DenseNet | paper, tflite&pb | 43.6 Mb | 64.2% | 85.6% | 195 ms | 60 ms | 1656 ms |
| SqueezeNet | paper, tflite&pb | 5.0 Mb | 49.0% | 72.9% | 36 ms | 9.5 ms | 18.5 ms |
| NASNet mobile | paper, tflite&pb | 21.4 Mb | 73.9% | 91.5% | 56 ms | --- | 102 ms |
| NASNet large | paper, tflite&pb | 355.3 Mb | 82.6% | 96.1% | 1170 ms | --- | 648 ms |
| ResNet_V2_101 | paper, tflite&pb | 178.3 Mb | 76.8% | 93.6% | 526 ms | 92 ms | 1572 ms |
| Inception_V3 | paper, tflite&pb | 95.3 Mb | 77.9% | 93.8% | 249 ms | 56 ms | 148 ms |
| Inception_V4 | paper, tflite&pb | 170.7 Mb | 80.1% | 95.1% | 486 ms | 93 ms | 291 ms |
| Inception_ResNet_V2 | paper, tflite&pb | 121.0 Mb | 77.5% | 94.0% | 422 ms | 100 ms | 201 ms |
| Mobilenet_V1_0.25_128 | paper, tflite&pb | 1.9 Mb | 41.4% | 66.2% | 1.2 ms | 1.6 ms | 3 ms |
| Mobilenet_V1_0.25_160 | paper, tflite&pb | 1.9 Mb | 45.4% | 70.2% | 1.7 ms | 1.7 ms | 3.2 ms |
| Mobilenet_V1_0.25_192 | paper, tflite&pb | 1.9 Mb | 47.1% | 72.0% | 2.4 ms | 1.8 ms | 3.0 ms |
| Mobilenet_V1_0.25_224 | paper, tflite&pb | 1.9 Mb | 49.7% | 74.1% | 3.3 ms | 1.8 ms | 3.6 ms |
| Mobilenet_V1_0.50_128 | paper, tflite&pb | 5.3 Mb | 56.2% | 79.3% | 3.0 ms | 1.7 ms | 3.2 ms |
| Mobilenet_V1_0.50_160 | paper, tflite&pb | 5.3 Mb | 59.0% | 81.8% | 4.4 ms | 2.0 ms | 4.0 ms |
| Mobilenet_V1_0.50_192 | paper, tflite&pb | 5.3 Mb | 61.7% | 83.5% | 6.0 ms | 2.5 ms | 4.8 ms |
| Mobilenet_V1_0.50_224 | paper, tflite&pb | 5.3 Mb | 63.2% | 84.9% | 7.9 ms | 2.8 ms | 6.1 ms |
| Mobilenet_V1_0.75_128 | paper, tflite&pb | 10.3 Mb | 62.0% | 83.8% | 5.5 ms | 2.6 ms | 5.1 ms |
| Mobilenet_V1_0.75_160 | paper, tflite&pb | 10.3 Mb | 65.2% | 85.9% | 8.2 ms | 3.1 ms | 6.3 ms |
| Mobilenet_V1_0.75_192 | paper, tflite&pb | 10.3 Mb | 67.1% | 87.2% | 11.0 ms | 4.5 ms | 7.2 ms |
| Mobilenet_V1_0.75_224 | paper, tflite&pb | 10.3 Mb | 68.3% | 88.1% | 14.6 ms | 4.9 ms | 9.9 ms |
| Mobilenet_V1_1.0_128 | paper, tflite&pb | 16.9 Mb | 65.2% | 85.7% | 9.0 ms | 4.4 ms | 6.3 ms |
| Mobilenet_V1_1.0_160 | paper, tflite&pb | 16.9 Mb | 68.0% | 87.7% | 13.4 ms | 5.0 ms | 8.4 ms |
| Mobilenet_V1_1.0_192 | paper, tflite&pb | 16.9 Mb | 69.9% | 89.1% | 18.1 ms | 6.3 ms | 10.6 ms |
| Mobilenet_V1_1.0_224 | paper, tflite&pb | 16.9 Mb | 71.0% | 89.9% | 24.0 ms | 6.5 ms | 13.8 ms |
| Mobilenet_V2_1.0_224 | paper, tflite&pb | 14.0 Mb | 71.8% | 90.6% | 17.5 ms | 6.2 ms | 11.23 ms |
AutoML mobile models
The following image classification models were created using Cloud AutoML. The performance values are measured on Pixel 3 on Android 10.
You can find these models in TensorFlow Hub and get more model information there.
| Model Name | Paper and model | Model size | Top-1 accuracy | Top-5 accuracy | CPU, 4 threads | GPU | NNAPI |
|---|---|---|---|---|---|---|---|
| MnasNet_0.50_224 | paper, tflite&pb | 8.5 Mb | 68.03% | 87.79% | 9.5 ms | 5.9 ms | 16.6 ms |
| MnasNet_0.75_224 | paper, tflite&pb | 12 Mb | 71.72% | 90.17% | 13.7 ms | 7.1 ms | 16.7 ms |
| MnasNet_1.0_96 | paper, tflite&pb | 17 Mb | 62.33% | 83.98% | 5.6 ms | 5.4 ms | 12.1 ms |
| MnasNet_1.0_128 | paper, tflite&pb | 17 Mb | 67.32% | 87.70% | 7.5 ms | 5.8 ms | 12.9 ms |
| MnasNet_1.0_160 | paper, tflite&pb | 17 Mb | 70.63% | 89.58% | 11.1 ms | 6.7 ms | 14.2 ms |
| MnasNet_1.0_192 | paper, tflite&pb | 17 Mb | 72.56% | 90.76% | 14.5 ms | 7.7 ms | 16.6 ms |
| MnasNet_1.0_224 | paper, tflite&pb | 17 Mb | 74.08% | 91.75% | 19.4 ms | 8.7 ms | 19 ms |
| MnasNet_1.3_224 | paper, tflite&pb | 24 Mb | 75.24% | 92.55% | 27.9 ms | 10.6 ms | 22.0 ms |
Note: Performance numbers were benchmarked on Pixel-3 (Android 10). Accuracy numbers were computed using the TFLite accuracy tool.
Object detection
For more information about object detection, see Object detection.
Please find object detection models from TensorFlow Hub.
Pose estimation
For more information about pose estimation, see Pose estimation.
Please find pose estimation models from TensorFlow Hub.
Image segmentation
For more information about image segmentation, see Segmentation.
Please find image segmentation models from TensorFlow Hub.
Question and Answer
For more information about text classification with Mobile BERT, see Question And Answer.
Please find Mobile BERT model from TensorFlow Hub.
Smart reply
For more information about smart reply, see Smart reply.
Please find Smart Reply model from TensorFlow Hub.