1# Hosted models 2 3The following is an incomplete list of pre-trained models optimized to work with 4TensorFlow Lite. 5 6To get started choosing a model, visit <a href="../models">Models</a> page with 7end-to-end examples, or pick a 8[TensorFlow Lite model from TensorFlow Hub](https://tfhub.dev/s?deployment-format=lite). 9 10Note: The best model for a given application depends on your requirements. For 11example, some applications might benefit from higher accuracy, while others 12require a small model size. You should test your application with a variety of 13models to find the optimal balance between size, performance, and accuracy. 14 15## Image classification 16 17For more information about image classification, see 18<a href="../models/image_classification/overview.md">Image classification</a>. 19Explore the TensorFlow Lite Task Library for instructions about 20[how to integrate image classification models](../inference_with_metadata/task_library/image_classifier) 21in just a few lines of code. 22 23### Quantized models 24 25<a href="../performance/post_training_quantization">Quantized</a> image 26classification models offer the smallest model size and fastest performance, at 27the expense of accuracy. The performance values are measured on Pixel 3 on 28Android 10. 29 30You can find many 31[quantized models](https://tfhub.dev/s?deployment-format=lite&module-type=image-classification&q=quantized) 32from TensorFlow Hub and get more model information there. 33 34Model name | Paper and model | Model size | Top-1 accuracy | Top-5 accuracy | CPU, 4 threads | NNAPI 35--------------------------- | :-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------: | ---------: | -------------: | -------------: | -------------: | ----: 36Mobilenet_V1_0.25_128_quant | [paper](https://arxiv.org/pdf/1712.05877.pdf), [tflite&pb](https://storage.googleapis.com/download.tensorflow.org/models/mobilenet_v1_2018_08_02/mobilenet_v1_0.25_128_quant.tgz) | 0.5 Mb | 39.5% | 64.4% | 0.8 ms | 2 ms 37Mobilenet_V1_0.25_160_quant | [paper](https://arxiv.org/pdf/1712.05877.pdf), [tflite&pb](https://storage.googleapis.com/download.tensorflow.org/models/mobilenet_v1_2018_08_02/mobilenet_v1_0.25_160_quant.tgz) | 0.5 Mb | 42.8% | 68.1% | 1.3 ms | 2.4 ms 38Mobilenet_V1_0.25_192_quant | [paper](https://arxiv.org/pdf/1712.05877.pdf), [tflite&pb](https://storage.googleapis.com/download.tensorflow.org/models/mobilenet_v1_2018_08_02/mobilenet_v1_0.25_192_quant.tgz) | 0.5 Mb | 45.7% | 70.8% | 1.8 ms | 2.6 ms 39Mobilenet_V1_0.25_224_quant | [paper](https://arxiv.org/pdf/1712.05877.pdf), [tflite&pb](https://storage.googleapis.com/download.tensorflow.org/models/mobilenet_v1_2018_08_02/mobilenet_v1_0.25_224_quant.tgz) | 0.5 Mb | 48.2% | 72.8% | 2.3 ms | 2.9 ms 40Mobilenet_V1_0.50_128_quant | [paper](https://arxiv.org/pdf/1712.05877.pdf), [tflite&pb](https://storage.googleapis.com/download.tensorflow.org/models/mobilenet_v1_2018_08_02/mobilenet_v1_0.5_128_quant.tgz) | 1.4 Mb | 54.9% | 78.1% | 1.7 ms | 2.6 ms 41Mobilenet_V1_0.50_160_quant | [paper](https://arxiv.org/pdf/1712.05877.pdf), [tflite&pb](https://storage.googleapis.com/download.tensorflow.org/models/mobilenet_v1_2018_08_02/mobilenet_v1_0.5_160_quant.tgz) | 1.4 Mb | 57.2% | 80.5% | 2.6 ms | 2.9 ms 42Mobilenet_V1_0.50_192_quant | [paper](https://arxiv.org/pdf/1712.05877.pdf), [tflite&pb](https://storage.googleapis.com/download.tensorflow.org/models/mobilenet_v1_2018_08_02/mobilenet_v1_0.5_192_quant.tgz) | 1.4 Mb | 59.9% | 82.1% | 3.6 ms | 3.3 ms 43Mobilenet_V1_0.50_224_quant | [paper](https://arxiv.org/pdf/1712.05877.pdf), [tflite&pb](https://storage.googleapis.com/download.tensorflow.org/models/mobilenet_v1_2018_08_02/mobilenet_v1_0.5_224_quant.tgz) | 1.4 Mb | 61.2% | 83.2% | 4.7 ms | 3.6 ms 44Mobilenet_V1_0.75_128_quant | [paper](https://arxiv.org/pdf/1712.05877.pdf), [tflite&pb](https://storage.googleapis.com/download.tensorflow.org/models/mobilenet_v1_2018_08_02/mobilenet_v1_0.75_128_quant.tgz) | 2.6 Mb | 55.9% | 79.1% | 3.1 ms | 3.2 ms 45Mobilenet_V1_0.75_160_quant | [paper](https://arxiv.org/pdf/1712.05877.pdf), [tflite&pb](https://storage.googleapis.com/download.tensorflow.org/models/mobilenet_v1_2018_08_02/mobilenet_v1_0.75_160_quant.tgz) | 2.6 Mb | 62.4% | 83.7% | 4.7 ms | 3.8 ms 46Mobilenet_V1_0.75_192_quant | [paper](https://arxiv.org/pdf/1712.05877.pdf), [tflite&pb](https://storage.googleapis.com/download.tensorflow.org/models/mobilenet_v1_2018_08_02/mobilenet_v1_0.75_192_quant.tgz) | 2.6 Mb | 66.1% | 86.2% | 6.4 ms | 4.2 ms 47Mobilenet_V1_0.75_224_quant | [paper](https://arxiv.org/pdf/1712.05877.pdf), [tflite&pb](https://storage.googleapis.com/download.tensorflow.org/models/mobilenet_v1_2018_08_02/mobilenet_v1_0.75_224_quant.tgz) | 2.6 Mb | 66.9% | 86.9% | 8.5 ms | 4.8 ms 48Mobilenet_V1_1.0_128_quant | [paper](https://arxiv.org/pdf/1712.05877.pdf), [tflite&pb](https://storage.googleapis.com/download.tensorflow.org/models/mobilenet_v1_2018_08_02/mobilenet_v1_1.0_128_quant.tgz) | 4.3 Mb | 63.3% | 84.1% | 4.8 ms | 3.8 ms 49Mobilenet_V1_1.0_160_quant | [paper](https://arxiv.org/pdf/1712.05877.pdf), [tflite&pb](https://storage.googleapis.com/download.tensorflow.org/models/mobilenet_v1_2018_08_02/mobilenet_v1_1.0_160_quant.tgz) | 4.3 Mb | 66.9% | 86.7% | 7.3 ms | 4.6 ms 50Mobilenet_V1_1.0_192_quant | [paper](https://arxiv.org/pdf/1712.05877.pdf), [tflite&pb](https://storage.googleapis.com/download.tensorflow.org/models/mobilenet_v1_2018_08_02/mobilenet_v1_1.0_192_quant.tgz) | 4.3 Mb | 69.1% | 88.1% | 9.9 ms | 5.2 ms 51Mobilenet_V1_1.0_224_quant | [paper](https://arxiv.org/pdf/1712.05877.pdf), [tflite&pb](https://storage.googleapis.com/download.tensorflow.org/models/mobilenet_v1_2018_08_02/mobilenet_v1_1.0_224_quant.tgz) | 4.3 Mb | 70.0% | 89.0% | 13 ms | 6.0 ms 52Mobilenet_V2_1.0_224_quant | [paper](https://arxiv.org/abs/1806.08342), [tflite&pb](https://storage.googleapis.com/download.tensorflow.org/models/tflite_11_05_08/mobilenet_v2_1.0_224_quant.tgz) | 3.4 Mb | 70.8% | 89.9% | 12 ms | 6.9 ms 53Inception_V1_quant | [paper](https://arxiv.org/abs/1409.4842), [tflite&pb](https://storage.googleapis.com/download.tensorflow.org/models/inception_v1_224_quant_20181026.tgz) | 6.4 Mb | 70.1% | 89.8% | 39 ms | 36 ms 54Inception_V2_quant | [paper](https://arxiv.org/abs/1512.00567), [tflite&pb](https://storage.googleapis.com/download.tensorflow.org/models/inception_v2_224_quant_20181026.tgz) | 11 Mb | 73.5% | 91.4% | 59 ms | 18 ms 55Inception_V3_quant | [paper](https://arxiv.org/abs/1806.08342),[tflite&pb](https://storage.googleapis.com/download.tensorflow.org/models/tflite_11_05_08/inception_v3_quant.tgz) | 23 Mb | 77.5% | 93.7% | 148 ms | 74 ms 56Inception_V4_quant | [paper](https://arxiv.org/abs/1602.07261), [tflite&pb](https://storage.googleapis.com/download.tensorflow.org/models/inception_v4_299_quant_20181026.tgz) | 41 Mb | 79.5% | 93.9% | 268 ms | 155 ms 57 58Note: The model files include both TF Lite FlatBuffer and Tensorflow frozen 59Graph. 60 61Note: Performance numbers were benchmarked on Pixel-3 (Android 10). Accuracy 62numbers were computed using the 63[TFLite image classification evaluation tool](https://github.com/tensorflow/tensorflow/tree/master/tensorflow/lite/tools/evaluation/tasks/imagenet_image_classification). 64 65### Floating point models 66 67Floating point models offer the best accuracy, at the expense of model size and 68performance. <a href="../performance/gpu">GPU acceleration</a> requires the use 69of floating point models. The performance values are measured on Pixel 3 on 70Android 10. 71 72You can find many 73[image classification models](https://tfhub.dev/s?deployment-format=lite&module-type=image-classification) 74from TensorFlow Hub and get more model information there. 75 76Model name | Paper and model | Model size | Top-1 accuracy | Top-5 accuracy | CPU, 4 threads | GPU | NNAPI 77--------------------- | :---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------: | ---------: | -------------: | -------------: | -------------: | -----: | ----: 78DenseNet | [paper](https://arxiv.org/abs/1608.06993), [tflite&pb](https://storage.googleapis.com/download.tensorflow.org/models/tflite/model_zoo/upload_20180427/densenet_2018_04_27.tgz) | 43.6 Mb | 64.2% | 85.6% | 195 ms | 60 ms | 1656 ms 79SqueezeNet | [paper](https://arxiv.org/abs/1602.07360), [tflite&pb](https://storage.googleapis.com/download.tensorflow.org/models/tflite/model_zoo/upload_20180427/squeezenet_2018_04_27.tgz) | 5.0 Mb | 49.0% | 72.9% | 36 ms | 9.5 ms | 18.5 ms 80NASNet mobile | [paper](https://arxiv.org/abs/1707.07012), [tflite&pb](https://storage.googleapis.com/download.tensorflow.org/models/tflite/model_zoo/upload_20180427/nasnet_mobile_2018_04_27.tgz) | 21.4 Mb | 73.9% | 91.5% | 56 ms | --- | 102 ms 81NASNet large | [paper](https://arxiv.org/abs/1707.07012), [tflite&pb](https://storage.googleapis.com/download.tensorflow.org/models/tflite/model_zoo/upload_20180427/nasnet_large_2018_04_27.tgz) | 355.3 Mb | 82.6% | 96.1% | 1170 ms | --- | 648 ms 82ResNet_V2_101 | [paper](https://arxiv.org/abs/1603.05027), [tflite&pb](https://storage.googleapis.com/download.tensorflow.org/models/tflite_11_05_08/resnet_v2_101.tgz) | 178.3 Mb | 76.8% | 93.6% | 526 ms | 92 ms | 1572 ms 83Inception_V3 | [paper](http://arxiv.org/abs/1512.00567), [tflite&pb](https://storage.googleapis.com/download.tensorflow.org/models/tflite/model_zoo/upload_20180427/inception_v3_2018_04_27.tgz) | 95.3 Mb | 77.9% | 93.8% | 249 ms | 56 ms | 148 ms 84Inception_V4 | [paper](http://arxiv.org/abs/1602.07261), [tflite&pb](https://storage.googleapis.com/download.tensorflow.org/models/tflite/model_zoo/upload_20180427/inception_v4_2018_04_27.tgz) | 170.7 Mb | 80.1% | 95.1% | 486 ms | 93 ms | 291 ms 85Inception_ResNet_V2 | [paper](https://arxiv.org/abs/1602.07261), [tflite&pb](https://storage.googleapis.com/download.tensorflow.org/models/tflite/model_zoo/upload_20180427/inception_resnet_v2_2018_04_27.tgz) | 121.0 Mb | 77.5% | 94.0% | 422 ms | 100 ms | 201 ms 86Mobilenet_V1_0.25_128 | [paper](https://arxiv.org/pdf/1704.04861.pdf), [tflite&pb](https://storage.googleapis.com/download.tensorflow.org/models/mobilenet_v1_2018_02_22/mobilenet_v1_0.25_128.tgz) | 1.9 Mb | 41.4% | 66.2% | 1.2 ms | 1.6 ms | 3 ms 87Mobilenet_V1_0.25_160 | [paper](https://arxiv.org/pdf/1704.04861.pdf), [tflite&pb](https://storage.googleapis.com/download.tensorflow.org/models/mobilenet_v1_2018_02_22/mobilenet_v1_0.25_160.tgz) | 1.9 Mb | 45.4% | 70.2% | 1.7 ms | 1.7 ms | 3.2 ms 88Mobilenet_V1_0.25_192 | [paper](https://arxiv.org/pdf/1704.04861.pdf), [tflite&pb](https://storage.googleapis.com/download.tensorflow.org/models/mobilenet_v1_2018_02_22/mobilenet_v1_0.25_192.tgz) | 1.9 Mb | 47.1% | 72.0% | 2.4 ms | 1.8 ms | 3.0 ms 89Mobilenet_V1_0.25_224 | [paper](https://arxiv.org/pdf/1704.04861.pdf), [tflite&pb](https://storage.googleapis.com/download.tensorflow.org/models/mobilenet_v1_2018_02_22/mobilenet_v1_0.25_224.tgz) | 1.9 Mb | 49.7% | 74.1% | 3.3 ms | 1.8 ms | 3.6 ms 90Mobilenet_V1_0.50_128 | [paper](https://arxiv.org/pdf/1704.04861.pdf), [tflite&pb](https://storage.googleapis.com/download.tensorflow.org/models/mobilenet_v1_2018_02_22/mobilenet_v1_0.5_128.tgz) | 5.3 Mb | 56.2% | 79.3% | 3.0 ms | 1.7 ms | 3.2 ms 91Mobilenet_V1_0.50_160 | [paper](https://arxiv.org/pdf/1704.04861.pdf), [tflite&pb](https://storage.googleapis.com/download.tensorflow.org/models/mobilenet_v1_2018_02_22/mobilenet_v1_0.5_160.tgz) | 5.3 Mb | 59.0% | 81.8% | 4.4 ms | 2.0 ms | 4.0 ms 92Mobilenet_V1_0.50_192 | [paper](https://arxiv.org/pdf/1704.04861.pdf), [tflite&pb](https://storage.googleapis.com/download.tensorflow.org/models/mobilenet_v1_2018_02_22/mobilenet_v1_0.5_192.tgz) | 5.3 Mb | 61.7% | 83.5% | 6.0 ms | 2.5 ms | 4.8 ms 93Mobilenet_V1_0.50_224 | [paper](https://arxiv.org/pdf/1704.04861.pdf), [tflite&pb](https://storage.googleapis.com/download.tensorflow.org/models/mobilenet_v1_2018_02_22/mobilenet_v1_0.5_224.tgz) | 5.3 Mb | 63.2% | 84.9% | 7.9 ms | 2.8 ms | 6.1 ms 94Mobilenet_V1_0.75_128 | [paper](https://arxiv.org/pdf/1704.04861.pdf), [tflite&pb](https://storage.googleapis.com/download.tensorflow.org/models/mobilenet_v1_2018_02_22/mobilenet_v1_0.75_128.tgz) | 10.3 Mb | 62.0% | 83.8% | 5.5 ms | 2.6 ms | 5.1 ms 95Mobilenet_V1_0.75_160 | [paper](https://arxiv.org/pdf/1704.04861.pdf), [tflite&pb](https://storage.googleapis.com/download.tensorflow.org/models/mobilenet_v1_2018_02_22/mobilenet_v1_0.75_160.tgz) | 10.3 Mb | 65.2% | 85.9% | 8.2 ms | 3.1 ms | 6.3 ms 96Mobilenet_V1_0.75_192 | [paper](https://arxiv.org/pdf/1704.04861.pdf), [tflite&pb](https://storage.googleapis.com/download.tensorflow.org/models/mobilenet_v1_2018_02_22/mobilenet_v1_0.75_192.tgz) | 10.3 Mb | 67.1% | 87.2% | 11.0 ms | 4.5 ms | 7.2 ms 97Mobilenet_V1_0.75_224 | [paper](https://arxiv.org/pdf/1704.04861.pdf), [tflite&pb](https://storage.googleapis.com/download.tensorflow.org/models/mobilenet_v1_2018_02_22/mobilenet_v1_0.75_224.tgz) | 10.3 Mb | 68.3% | 88.1% | 14.6 ms | 4.9 ms | 9.9 ms 98Mobilenet_V1_1.0_128 | [paper](https://arxiv.org/pdf/1704.04861.pdf), [tflite&pb](https://storage.googleapis.com/download.tensorflow.org/models/mobilenet_v1_2018_02_22/mobilenet_v1_1.0_128.tgz) | 16.9 Mb | 65.2% | 85.7% | 9.0 ms | 4.4 ms | 6.3 ms 99Mobilenet_V1_1.0_160 | [paper](https://arxiv.org/pdf/1704.04861.pdf), [tflite&pb](https://storage.googleapis.com/download.tensorflow.org/models/mobilenet_v1_2018_02_22/mobilenet_v1_1.0_160.tgz) | 16.9 Mb | 68.0% | 87.7% | 13.4 ms | 5.0 ms | 8.4 ms 100Mobilenet_V1_1.0_192 | [paper](https://arxiv.org/pdf/1704.04861.pdf), [tflite&pb](https://storage.googleapis.com/download.tensorflow.org/models/mobilenet_v1_2018_02_22/mobilenet_v1_1.0_192.tgz) | 16.9 Mb | 69.9% | 89.1% | 18.1 ms | 6.3 ms | 10.6 ms 101Mobilenet_V1_1.0_224 | [paper](https://arxiv.org/pdf/1704.04861.pdf), [tflite&pb](https://storage.googleapis.com/download.tensorflow.org/models/mobilenet_v1_2018_02_22/mobilenet_v1_1.0_224.tgz) | 16.9 Mb | 71.0% | 89.9% | 24.0 ms | 6.5 ms | 13.8 ms 102Mobilenet_V2_1.0_224 | [paper](https://arxiv.org/pdf/1801.04381.pdf), [tflite&pb](https://storage.googleapis.com/download.tensorflow.org/models/tflite_11_05_08/mobilenet_v2_1.0_224.tgz) | 14.0 Mb | 71.8% | 90.6% | 17.5 ms | 6.2 ms | 11.23 ms 103 104### AutoML mobile models 105 106The following image classification models were created using 107<a href="https://cloud.google.com/automl/">Cloud AutoML</a>. The performance 108values are measured on Pixel 3 on Android 10. 109 110You can find these models in 111[TensorFlow Hub](https://tfhub.dev/s?deployment-format=lite&q=MnasNet) and get 112more model information there. 113 114Model Name | Paper and model | Model size | Top-1 accuracy | Top-5 accuracy | CPU, 4 threads | GPU | NNAPI 115---------------- | :------------------------------------------------------------------------------------------------------------------------------------------------------------: | ---------: | -------------: | -------------: | -------------: | ------: | ----: 116MnasNet_0.50_224 | [paper](https://arxiv.org/abs/1807.11626), [tflite&pb](https://storage.cloud.google.com/download.tensorflow.org/models/tflite/mnasnet_0.5_224_09_07_2018.tgz) | 8.5 Mb | 68.03% | 87.79% | 9.5 ms | 5.9 ms | 16.6 ms 117MnasNet_0.75_224 | [paper](https://arxiv.org/abs/1807.11626), [tflite&pb](https://storage.cloud.google.com/download.tensorflow.org/models/tflite/mnasnet_0.75_224_09_07_2018.tgz) | 12 Mb | 71.72% | 90.17% | 13.7 ms | 7.1 ms | 16.7 ms 118MnasNet_1.0_96 | [paper](https://arxiv.org/abs/1807.11626), [tflite&pb](https://storage.cloud.google.com/download.tensorflow.org/models/tflite/mnasnet_1.0_96_09_07_2018.tgz) | 17 Mb | 62.33% | 83.98% | 5.6 ms | 5.4 ms | 12.1 ms 119MnasNet_1.0_128 | [paper](https://arxiv.org/abs/1807.11626), [tflite&pb](https://storage.cloud.google.com/download.tensorflow.org/models/tflite/mnasnet_1.0_128_09_07_2018.tgz) | 17 Mb | 67.32% | 87.70% | 7.5 ms | 5.8 ms | 12.9 ms 120MnasNet_1.0_160 | [paper](https://arxiv.org/abs/1807.11626), [tflite&pb](https://storage.cloud.google.com/download.tensorflow.org/models/tflite/mnasnet_1.0_160_09_07_2018.tgz) | 17 Mb | 70.63% | 89.58% | 11.1 ms | 6.7 ms | 14.2 ms 121MnasNet_1.0_192 | [paper](https://arxiv.org/abs/1807.11626), [tflite&pb](https://storage.cloud.google.com/download.tensorflow.org/models/tflite/mnasnet_1.0_192_09_07_2018.tgz) | 17 Mb | 72.56% | 90.76% | 14.5 ms | 7.7 ms | 16.6 ms 122MnasNet_1.0_224 | [paper](https://arxiv.org/abs/1807.11626), [tflite&pb](https://storage.cloud.google.com/download.tensorflow.org/models/tflite/mnasnet_1.0_224_09_07_2018.tgz) | 17 Mb | 74.08% | 91.75% | 19.4 ms | 8.7 ms | 19 ms 123MnasNet_1.3_224 | [paper](https://arxiv.org/abs/1807.11626), [tflite&pb](https://storage.cloud.google.com/download.tensorflow.org/models/tflite/mnasnet_1.3_224_09_07_2018.tgz) | 24 Mb | 75.24% | 92.55% | 27.9 ms | 10.6 ms | 22.0 ms 124 125Note: Performance numbers were benchmarked on Pixel-3 (Android 10). Accuracy 126numbers were computed using the 127[TFLite image classification evaluation tool](https://github.com/tensorflow/tensorflow/tree/master/tensorflow/lite/tools/evaluation/tasks/imagenet_image_classification). 128 129## Object detection 130 131For more information about object detection, see 132<a href="../models/object_detection/overview.md">Object detection</a>. Explore 133the TensorFlow Lite Task Library for instructions about 134[how to integrate object detection models](../inference_with_metadata/task_library/object_detector) 135in just a few lines of code. 136 137Please find 138[object detection models](https://tfhub.dev/s?deployment-format=lite&module-type=image-object-detection) 139from TensorFlow Hub. 140 141## Pose estimation 142 143For more information about pose estimation, see 144<a href="../models/pose_estimation/overview.md">Pose estimation</a>. 145 146Please find 147[pose estimation models](https://tfhub.dev/s?deployment-format=lite&module-type=image-pose-detection) 148from TensorFlow Hub. 149 150## Image segmentation 151 152For more information about image segmentation, see 153<a href="../models/segmentation/overview.md">Segmentation</a>. Explore the 154TensorFlow Lite Task Library for instructions about 155[how to integrate image segmentation models](../inference_with_metadata/task_library/image_segmenter) 156in just a few lines of code. 157 158Please find 159[image segmentation models](https://tfhub.dev/s?deployment-format=lite&module-type=image-segmentation) 160from TensorFlow Hub. 161 162## Question and Answer 163 164For more information about question and answer with MobileBERT, see 165<a href="../models/bert_qa/overview.md">Question And Answer</a>. Explore the 166TensorFlow Lite Task Library for instructions about 167[how to integrate question and answer models](../inference_with_metadata/task_library/bert_question_answerer) 168in just a few lines of code. 169 170Please find [Mobile BERT model](https://tfhub.dev/tensorflow/mobilebert/1) from 171TensorFlow Hub. 172 173## Smart reply 174 175For more information about smart reply, see 176<a href="../models/smart_reply/overview.md">Smart reply</a>. 177 178Please find [Smart Reply model](https://tfhub.dev/tensorflow/smartreply/1) from 179TensorFlow Hub. 180