<figcaption><i>An illustration of NAS to find Edge TPU optimized models, Each column represents a stage in the natural network, with dots indicating different options, and each color representing a different type of building block. A path from inputs (e.g., an image) to outputs (e.g., per-pixel label predictions) through the matrix represents a candidate neural network. In each iteration of the search, a neural network is formed using the blocks chosen at every stage, and the search algorithm aims to find neural networks that jointly minimize TPU latency and/or energy and maximize accuracy.
</i></figcaption>
</figure>
This repository contains machine learning models optimized for Edge TPU in
This repository contains machine learning models optimized for the Edge TPU in
<figcaption><i>Inverted bottleneck block (IBN) variants: (a) Conventional with depthwise, (b) Fused-IBN, (c)GC-IBN with group convolutions in the expansion phase</i></figcaption>
</figure>
In this work we utilize Group Convolution (GC) as part of the fused expansion in
constructing IBNs (Figure 1). GC based IBN becomes a versatile block that opens
up a large design space between conventional depthwise IBNs and fused
full-convolution IBNs which can be controlled by the group size parameter.
Figure 2 demonstrates the search space enabled by GC-based IBNs that allows a
flexible tradeoff between latency and number of trainable parameters. GC-based
IBNs allow increasing the number of trainable parameters gradually without
requiring the latency cost of full-convolution based IBNs. Moreover, they can
also be faster than conventional IBNs with depthwise convolutions while
<figcaption><i>Comparison of Imagenet top-1 accuracy and Pixel 6 Edge TPU latency of MobileNetEdgeTPUV2 models with other on-device classification models</i></figcaption>
</figure>
#### On-device benchmarking of classification models
Results on on-device benchmarking of various int8 quantized image classification
<figcaption><i>Comparison of Imagenet top-1 accuracy and Pixel 6 latency of MobileNetEdgeTPUV2 models with other on-device classification models</i></figcaption>
</figure>
## Semantic segmentation task
### Using classification models as backbone
We also present segmentation models based on MobileNetEdgeTPUV2 backbone and
DeepLab v3 plus decoder and head (first used
[here](https://arxiv.org/pdf/1802.02611.pdf)). These models optimized for the
next generation Edge TPU accelerators featured in Pixel 6 phones and improve the
latency-accuracy pareto-frontier compared to the their predecessor based on
MobileNetV2 and DeepLabV3+.
The segmentation model is built using the pretrained MobileNetEdgeTPUV2 as a
feature encoder and ASPP decoder in conjunction with a Deeplab V3 Plus head.
Separable convolutions used to reduce the size of the model.
<figcaption><i>Performance of AutosegEdgeTPU and MobileNetEdgeTPUV2+DeeplabV3+ models on the 32-class ADE20K semantic segmentation task.</i></figcaption>
</figure>
Model Name (Checkpoint) | Backbone | Segmentation Head| #Parameters (million)| ADE20K 32-class mIOU| Pixel 6 EdgeTPU latency (ms)| Tflite |
