Skip to main content

Post-training Quantization for Edge TPU

In Tensorflow website, there is quite a bit of explanation for post-training quantization but there is not much on transfer learning. The sample shown on Coral website is using Tensorflow 1.x and requires to execute the transfer learning inside a docker.

In this blog post, I am going to demonstrate on how to perform post-training quantization using Tensorflow 2.0 for Mobilenet V1 and V2. All the steps can be performed on Colab notebook (thus making use of free GPU from Google, Thank you Google!!!). The steps are almost the same for both versions except at the base model I have changed the model. The tflite model is then converted to Edge TPU tflite model which can be used for realtime inferencing.

For both the models, I am using the flower dataset to perform the transfer learning. Readers can use this as a base for another class of classification. In the future blog post, I may try more advanced models such as Inception, Resnet etc. A lot depends on the Edge TPU compiler because the compiler must be able to compile the layers to be supported by Edge TPU.

Some observations:
The final TPU tflite model is smaller for Mobilenet V2. For V1 is about 3.6MB and V2 is about 2MB
The training accuracy is higher for V2 as compared to V1. For 20 epochs, for V1 is about 93% and V2 is about 97%

Looks like V2 is definitely an improvement over V1.

Mobilenet V1


Mobilenet V2

Comments

Popular posts from this blog

DIY Sonoff RF Bridge

Tasmota recently supported the RF bridge by iTead which allowed for RF 433 remote devices to be controlled/controlling using Sonoff products. e.g A handphone can be used to control a remote RF device or an RF remote control can use to control a WIFI enabled device.

By using Tasmota, I can control the device using MQTT. This idea fits exactly to my smart home plan. Here is the schematic:



I cheated a bit by having ESP01 mounted on a breakout board. This breakout board converts all the external pins to 5v which simplified the wiring and I do not need to use a voltage level shifter. Moreover, the Arduino 3.3v power does not have enough current to power the ESP reliably and I have to power it via the 5V from Arduino Uno. Next, I used an Arduino Uno prototype shield with a tiny breadboard to hold all the components together. For the RF 433 transmitter/receiver, I am using the development kits when I purchased the all-in-one Arduino development kits. An antenna is added for more sensitivity.



Using ESP-Link transparent bridge (ESP-01 and Arduino Pro Mini)

Recently stumbled across an interesting open source project ESP-Link. Its main purpose is to network-enable a non-network microcontroller (MCU) such as Arduino Uno, Pro mini or Nano using ESP8266. The author termed it as "Transparent Bridge". The ESP and MCU  communicate via the serial link and there is a companion Arduino library EL-Client for the MCU to connect up the network using MQTT, REST, TCP and UDP.
Setup I have put together an ESP-01 and an Arduino Pro Mini for this experiment. I have chosen a 3.3 version Pro mini so that I do not need to do any voltage level shifting between the I/O pins. In order to have a stable voltage source, the ESP8266 is powered by Pro Mini and the Pro Mini "RAW" pin is connected to a 5v USB power source. The RAW pin can take voltage up to 12V. The reset pin of Pro Mini is connected to GPIO 0 of ESP-01. This enables the ESP-01 to reset the Pro Mini.




I have linked up an APDS 9960 sensor to it and periodically send the luminosity a…