TensorFlow models on the Edge TPU
In order for the Edge TPU to provide highspeed neural network performance with a lowpower cost, the Edge TPU supports a specific set of neural network operations and architectures. This page describes what types of models are compatible with the Edge TPU and how you can create them, either by compiling your own TensorFlow model or retraining an existing model with transferlearning.
If you're looking for information about running your model, read the Edge TPU inferencing overview.
Compatibility overview
The Edge TPU is capable of executing deep feedforward neural networks such as convolutional neural networks (CNN). It supports only TensorFlow Lite models that are fully 8bit quantized and then compiled specifically for the Edge TPU.
If you're not familiar with TensorFlow Lite, it's a lightweight version of TensorFlow designed for mobile and embedded devices. It achieves lowlatency inference in a small binary size—both the TensorFlow Lite models and interpreter kernels are much smaller. TensorFlow Lite models can be made even smaller and more efficient through quantization, which converts 32bit parameter data into 8bit representations (which is required by the Edge TPU).
You cannot train a model directly with TensorFlow Lite; instead you must convert your model from a
TensorFlow file (such as a .pb
file) to a TensorFlow Lite file (a .tflite
file), using the
TensorFlow Lite converter.
Figure 1 illustrates the basic process to create a model that's compatible with the
Edge TPU. Most of the workflow uses standard TensorFlow tools. Once you have a TensorFlow Lite
model, you then use our Edge TPU compiler to create a .tflite
file
that's compatible with the Edge TPU.
However, you don't need to follow this whole process to create a good model for the Edge TPU. Instead, you can leverage existing TensorFlow models that are compatible with the Edge TPU by retraining them with your own dataset. For example, MobileNet is a popular image classification/detection model architecture that's compatible with the Edge TPU. We've created several versions of this model that you can use as a starting point to create your own model that recognizes different objects. To get started, see the next section about how to retrain an existing model with transfer learning.
If you have designed—or plan to design—your own model architecture, then you should read the section below about model requirements.
Transfer learning
Instead of building your own model and then training it from scratch, you can retrain an existing model that's already compatible with the Edge TPU, using a technique called transfer learning (sometimes also called "fine tuning").
Training a neural network from scratch (when it has no computed weights or bias) can take daysworth of computing time and requires a vast amount of training data. But transfer learning allows you to start with a model that's already trained for a related task and then perform further training to teach the model new classifications using a smaller training dataset. You can do this by retraining the whole model (adjusting the weights across the whole network), but you can also achieve very accurate results by simply removing the final layer that performs classification, and training a new layer on top that recognize your new classes.
Using this process, with sufficient training data and some adjustments to the hyperparameters, you can create a highly accurate TensorFlow model in a single sitting. Once you're happy with the model's performance, simply convert it to TensorFlow Lite and then compile it for the Edge TPU. And because the model architecture doesn't change during transfer learning, you know it will fully compile for the Edge TPU (assuming you start with a compatible model).
If you're already familiar with transfer learning, check out our Edge TPUcompatible models that you can use as a starting point to create your own model. Just click to download "All model files" to get the TensorFlow model and pretrained checkpoints you need to begin transfer learning.
If you're new to this technique and want to quickly see some results, try the following tutorials that simplify the process to retrain a MobileNet model with new classes:
Transfer learning ondevice
If you're using an image classification model, you can also perform accelerated transfer learning on the Edge TPU. Our Edge TPU Python API offers two different techniques for ondevice transfer learning:
 Weight imprinting on the last layer
(
ImprintingEngine
)  Backpropagation on the last layer
(
SoftmaxRegression
)
In both cases, you must provide a model that's specially designed to allow training on the last layer. The required model structure is different for each API, but the result is basically the same: the last fullyconnected layer where classification occurs is separated from the base of the graph. Then only the base of the graph is compiled for the Edge TPU, which leaves the weights in the last layer accessible for training. More detail about the model architecture is available in the corresponding documents below. For now, let's compare how retraining works for each technique:

Weight imprinting takes the output (the embedding vectors) from the base model, adjusts the activation vectors with L2normalization, and uses those values to compute new weights in the final layer—it averages the new vectors with those already in the last layer's weights. This allows for effective training of new classes with very few sample images.

Backpropagation is an abbreviated version of traditional backpropagation. Instead of backpropagating new weights to all layers in the graph, it updates only the fullyconnected layer at the end of the graph with new weights. This is the more traditional training strategy that generally achieves higher accuracy, but it requires more images and multiple training iterations.
When choosing between these training techniques, you might consider the following factors:

Training sample size: Weight imprinting is more effective if you have a relatively small set of training samples: anywhere from 1 to 200 sample images for each class (as few as 5 can be effective and the API sets a maximum of 200). If you have more samples available for training, you'll likely achieve higher accuracy by using them all with backpropagation.

Training sample variance: Backpropagation is more effective if your dataset includes large intraclass variance. That is, if the images within a given class show the subject in significantly different ways, such as in angle or size, then backpropagation probably works better. But if your application operates in an environment where such variance is low, and your training samples thus also have little intraclass variance, then weight imprinting can work very well.

Adding new classes: Only weight imprinting allows you to add new classes to the model after you've begun training. If you're using backpropagation, adding a new class after you've begun training requires that you restart training for all classes. Additionally, weight imprinting allows you to retain the classes from the pretrained model (those trained before converting the model for the Edge TPU); whereas backpropagation requires all classes to be learned ondevice.

Model compatibility: Backpropagation is compatible with more model architectures "out of the box"; you can convert existing, pretrained MobileNet and Inception models into embedding extractors that are compatible with ondevice backpropagation. To use weight imprinting, you must use a model with some very specific layers and then train it in a particular manner before using it for ondevice training (currently, we offer a version of MobileNet v1 with the proper modifications).
In both cases, the vast majority of the training process is accelerated by the Edge TPU. And when performing inferences with the retrained model, the Edge TPU accelerates everything except the final classification layer, which runs on the CPU. But because this last layer accounts for only a small portion of the model, running this last layer on the CPU should not significantly affect your inference speed.
To learn more about each technique and try some sample code, see the following pages:
 Retrain a classification model ondevice with weight imprinting
 Retrain a classification model ondevice with backpropagation
Model requirements
If you want to build your own TensorFlow model that takes full advantage of the Edge TPU at runtime, it must meet the following requirements:
 Tensor parameters are quantized (8bit fixedpoint numbers).
 Tensor sizes are constant at compiletime (no dynamic sizes).
 Model parameters (such as bias tensors) are constant at compiletime.
 Tensors are either 1, 2, or 3dimensional. If a tensor has more than 3 dimensions, then only the 3 innermost dimensions may have a size greater than 1.
 The model uses only the operations supported by the Edge TPU (see table 1 below).
Supported operations
When building your own model architecture, be aware that only the operations in the following table are supported by the Edge TPU. If your architecture uses operations not listed here, then only a portion of the model will execute on the Edge TPU, as described in the section below about compiling.
Operation name  Known limitations 

Logistic  
Relu  
Relu6  
ReluN1To1  
Tanh  
Add  
Maximum  
Minimum  
Mul  
Sub  
AveragePool2d  No fused activation function. 
Concatenation  No fused activation function. If any input is a compiletime constant tensor, there must be only 2 inputs, and this constant tensor must be all zeroes (effectively, a zeropadding op). 
Conv2d  
DepthwiseConv2d  Dilated conv kernels are not supported. 
FullyConnected  Only default format supported for fullyconnected weights. Output tensor is onedimensional. 
L2Normalization  
MaxPool2d  No fused activation function. 
Mean  Supports reduction along x and/or ydimensions only. 
Pad  Supports padding along x and/or ydimensions only. 
Reshape  
ResizeBilinear  Input/output is a 3dimensional tensor. Depending on input/output size, this operation may not be mapped to the Edge TPU to avoid loss in precision. 
ResizeNearestNeighbor  Input/output is a 3dimensional tensor. Depending on input/output size, this operation may not be mapped to the Edge TPU to avoid loss in precision. 
Slice  
Softmax  Supports only 1D input tensor with a max of 16,000 elements. 
SpaceToDepth  
Split  
Squeeze  Supported only when input tensor dimensions that have leading 1s (that is, no relayout needed). For example input tensor with [y][x][z] = 1,1,10 or 1,5,10 is ok. But [y][x][z] = 5,1,10 is not supported. 
StridedSlice  Supported only when all strides are equal to 1 (that is, effectively a Stride op), and with ellipsisaxismask == 0, and newaxismax == 0. 
Quantization
Quantizing your model means converting all the 32bit floatingpoint numbers (such as weights and activation outputs) to the nearest 8bit fixedpoint numbers. This makes the model smaller and faster. And although these 8bit representations can be less precise, the inference accuracy of the neural network is not significantly affected.
For compatibility with the Edge TPU, you must use either quantizationaware training (recommended) or full integer posttraining quantization.
Quantizationaware training uses "fake" quantization nodes in the neural network graph to simulate the effect of 8bit values during training. Thus, this technique requires modification to the network before initial training. This generally results in a higher accuracy model (compared to posttraining quantization) because it makes the model more tolerant of lower precision values, due to the fact that the 8bit weights are learned through training rather than being converted later. It's also currently compatible with more operations than posttraining quantization.
Full integer posttraining quantization doesn't require any modifications to the network, so you can use this technique to convert a previouslytrained network into a quantized model. However, this conversion process requires that you supply a representative dataset. That is, you need a dataset that's formatted the same as the original training dataset (uses the same data range) and is of a similar style (it does not need to contain all the same classes, though you may use previous training/evaluation data). This representative dataset allows the quantization process to measure the dynamic range of activations and inputs, which is critical to finding an accurate 8bit representation of each weight and activation value.
However, not all TensorFlow Lite operations are currently implemented with an integeronly specification (they cannot be quantized using posttraining quantization). When converting a model, the TensorFlow Lite converter leaves those operations in their float format. You can change this behavior to instead throw an error (as documented in the posttraining quantization guide). Of course, the model won't be compatible with the Edge TPU Compiler either way. So until TensorFlow Lite adds integeronly specifications for all your model's operations, you can either try quantizationaware training or use one of the model architectures provided here, which are all compatible with posttraining quantization.
For more details about how quantization works, read the TensorFlow Lite 8bit quantization spec.
Quantizing Keras models
Currently, the Keras API does not officially support quantizationaware training. So if you're using Keras to build your model, then you must use posttraining quantization.
However, we have encountered inaccurate results after quantizing some Keras models. So far, we have verified that the following Keras models do successfully convert to TensorFlow Lite using posttraining quantization:
There are several models we have not tested yet, so we expect this list will grow as we verify more models. But currently, we cannot guarantee support for any models not listed above.
Compiling
After you train and convert your model to TensorFlow Lite (with quantization), the final step is to compile it with the Edge TPU Compiler.
If your model does not meet all the requirements listed at the top of this section, it can still compile, but only a portion of the model will execute on the Edge TPU. At the first point in the model graph where an unsupported operation occurs, the compiler partitions the graph into two parts. The first part of the graph that contains only supported operations is compiled into a custom operation that executes on the Edge TPU, and everything else executes on the CPU, as illustrated in figure 2.
If you inspect your compiled model (with a tool such as
visualize.py
),
you'll see that it's still a TensorFlow Lite model except it now has a custom operation at the
beginning of the graph. This custom operation is the only part of your model that is actually
compiled—it contains all the operations that run on the Edge TPU. The rest of the graph (beginning
with the first unsupported operation) remains the same and runs on the CPU.
If part of your model executes on the CPU, you should expect a significantly degraded inference speed compared to a model that executes entirely on the Edge TPU. We cannot predict how much slower your model will perform in this situation, so you should experiment with different architectures and strive to create a model that is 100% compatible with the Edge TPU. That is, your compiled model should contain only the Edge TPU custom operation.
Is this content helpful?