1# Build TensorFlow Lite models 2 3This page provides guidance for building 4your TensorFlow models with the intention of converting to the TensorFlow 5Lite model format. The machine learning (ML) models you use with TensorFlow 6Lite are originally 7built and trained using TensorFlow core libraries and tools. Once you've built 8a model with TensorFlow core, you can convert it to a smaller, more 9efficient ML model format called a TensorFlow Lite model. 10 11* If you have a model to convert already, see the 12 [Convert models overview](../convert/) 13 page for guidance on converting your model. 14 15* If you want to modify an existing model instead of starting from scratch, 16 see the [Modify models overview](../modify/model_maker) for guidance. 17 18 19## Building your model 20 21If you are building a custom model for your specific use case, 22you should start with developing and training a TensorFlow model or extending 23an existing one. 24 25### Model design constraints 26 27Before you start your model development process, you should be aware of the 28constraints for TensorFlow Lite models and build your model with these 29constraints in mind: 30 31* **Limited compute capabilities** - Compared to fully equipped servers with 32 multiple CPUs, high memory capacity, and specialized processors such as GPUs 33 and TPUs, mobile and edge devices are much more limited. While they are 34 growing in compute power and specialized hardware compatibility, the models 35 and data you can effectively process with them are still comparably limited. 36* **Size of models** - The overall complexity of a model, including data 37 pre-processing logic and the number of layers in the model, increases the 38 in-memory size of a model. A large model may run unacceptably slow or simply 39 may not fit in the available memory of a mobile or edge device. 40* **Size of data** - The size of input data that can be effectively processed 41 with a machine learning model is limited on a mobile or edge device. Models 42 that use large data libraries such as language libraries, image libraries, or 43 video clip libraries may not fit on these devices, and may require 44 off-device storage and access solutions. 45* **Supported TensorFlow operations** - TensorFlow Lite runtime environments 46 support a subset of machine learning model operations compared to 47 regular TensorFlow models. As you develop a model for use with TensorFlow 48 Lite, you should track the compatibility of your model against the 49 capabilities of TensorFlow Lite runtime environments. 50 51For more information building effective, compatible, high performance models 52for TensorFlow Lite, see 53[Performance best practices](../../performance/best_practices). 54 55### Model development 56 57To build a TensorFlow Lite model, you first need to build a model using the 58TensorFlow core libraries. TensorFlow core libraries are the lower-level 59libraries that provide APIs to build, train and deploy ML models. 60 61 62 63TensorFlow provides two paths for doing this. You can develop 64your own custom model code or you can start with a model implementation 65available in the TensorFlow 66[Model Garden](https://www.tensorflow.org/tfmodels). 67 68#### Model Garden 69 70The TensorFlow Model Garden provides implementations of many state-of-the-art 71machine learning (ML) models for vision and natural language processing (NLP). 72You'll also find workflow tools to let you quickly configure and run those 73models on standard datasets. The machine learning models in the 74Model Garden include full code so you can 75test, train, or re-train them using your own datasets. 76 77Whether you are looking to benchmark performance for a 78well-known model, verify the results of recently released research, or extend 79existing models, the Model Garden can help you drive your ML goals. 80 81#### Custom models 82 83If your use case is outside of those supported by the models in Model Garden, 84you can use a high level library like 85[Keras](https://www.tensorflow.org/guide/keras/sequential_model) to 86develop your custom training code. To learn the fundamentals of TensorFlow, see 87the [TensorFlow guide](https://www.tensorflow.org/guide/basics). To get started 88with examples, see the 89[TensorFlow tutorials overview](https://www.tensorflow.org/tutorials) which 90contain pointers to beginning to expert level tutorials. 91 92### Model evaluation 93 94Once you've developed your model, you should evaluate its performance and test 95it on end-user devices. 96TensorFlow provides a few ways to do this. 97 98* [TensorBoard](https://www.tensorflow.org/tensorboard/tensorboard_profiling_keras) 99 is a tool for providing the measurements and visualizations needed during 100 the machine learning workflow. It enables tracking experiment metrics like 101 loss and accuracy, visualizing the model graph, projecting embeddings to a 102 lower dimensional space, and much more. 103* [Benchmarking tools](https://www.tensorflow.org/lite/performance/measurement) 104 are available for each supported platform such as the Android benchmark app 105 and the iOS benchmark app. Use these tools to measure and calculate statistics 106 for important performance metrics. 107 108### Model optimization 109 110With the [constraints](#model_constraints) on resources specific to 111TensorFlow Lite models, model optimization can help to ensure your 112model performance well and uses less compute resources. Machine learning model 113performance is usually a balance between size and speed of inference vs 114accuracy. TensorFlow Lite currently supports optimization via quantization, 115pruning and clustering. See the 116[model optimization] (https://www.tensorflow.org/lite/performance/model_optimization) 117topic for more details on these techniques. TensorFlow also provides a 118[Model optimization toolkit](https://www.tensorflow.org/model_optimization/guide) 119which provides an API that implements these techniques. 120 121## Next steps 122 123* To start building your custom model, see the 124 [quick start for beginners](https://www.tensorflow.org/tutorials/quickstart/beginner) 125 tutorial in TensorFlow core documentation. 126* To convert your custom TensorFlow model, see the 127 [Convert models overview](../convert). 128* See the 129 [operator compatibility](../../guide/ops_compatibility) guide to determine 130 if your model is compatible with TensorFlow Lite or if you'll need to take 131 additional steps to make it compatible. 132* See the 133 [performance best practices guide](https://www.tensorflow.org/lite/performance/best_practices) 134 for guidance on making your TensorFlow Lite models efficient and performant. 135* See the [performance metrics guide](../../performance/measurement) to learn 136 how to measure the performance of your model using benchmarking tools. 137 138 139
