https://github.com/nobuotsukamoto/edge_tpu_mnist
Create Edge TPU Model (MNIST fully_connected_feed example)
https://github.com/nobuotsukamoto/edge_tpu_mnist
Last synced: 5 months ago
JSON representation
Create Edge TPU Model (MNIST fully_connected_feed example)
- Host: GitHub
- URL: https://github.com/nobuotsukamoto/edge_tpu_mnist
- Owner: NobuoTsukamoto
- License: other
- Created: 2019-05-30T12:46:37.000Z (over 6 years ago)
- Default Branch: master
- Last Pushed: 2019-06-03T12:36:54.000Z (over 6 years ago)
- Last Synced: 2025-05-07T23:42:24.762Z (5 months ago)
- Language: Jupyter Notebook
- Size: 206 KB
- Stars: 8
- Watchers: 1
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Awesome Lists containing this project
README
# Create Own Edge TPU Model.
## About
The Edge TPU has a great sample.
However, the samples are only image classification(slim) and object detection(object detection api), and there is no way to build and learn your own model.
This is a sample for converting your own model from learning to Edge TPU Model.
The model uses TensorFlow's MNIST Network sample (fully_connected_feed).
## Environment
- Coral USB Accelerator(May be Devboard)
## Installation
- Tensorflow (Learning only)
- Edge TPU Python library [(Get started with the USB Accelerator)](https://coral.withgoogle.com/tutorials/accelerator/)
## How to Train
### 1.Create Model and Quantization-aware training
It is a sample of Jupyter notebook.
This sample is based on the Tensorflow tutorial [(MNIST fully_connected_feed.py)](https://github.com/tensorflow/tensorflow/blob/master/tensorflow/examples/tutorials/mnist/fully_connected_feed.py).
Open train/MNIST_fully_connected_feed.ipynb and Run All.
When learning is complete, a model is generated in the following directory:
```
train
+ logs
| + GraphDef
+ checkpoint
+ model.ckpt-xxxx.data-00000-of-00001
+ model.ckpt-xxxx.index
+ model.ckpt-xxxx.meta
+ mnist_fully_connected_feed_graph.pb
```
model.ckpt is a ckecpoint file.
The number of learning steps is recorded in "xxxx".
The default is 4999 maximum.
mnist_fully_connected_feed_graph.pb is a GrarhDef file.
The GraphDef directory contains the structure of the model.
You can check with tensorboard.
(Move dir and ```tensorboard --logdir=./```)
"Quantization-aware training" is required to generate Edge TPU Model.
See [Quantization-aware training](https://github.com/tensorflow/tensorflow/tree/master/tensorflow/contrib/quantize) for details.
### 2.Freeze Graph
Converts checkpoint variables into Const ops in a standalone GraphDef file.
Use the freeze_graph command.
The freeze_mnist_fully_connected_feed_graph.pb file is generated in the logs directory.
```
$ freeze_graph \
--input_graph=./logs/mnist_fully_connected_feed_graph.pb \
--input_checkpoint=./logs/model.ckpt-4999 \
--input_binary=true \
--output_graph=./logs/freeze_mnist_fully_connected_feed_graph.pb \
--output_node_names=softmax_tensor
```
### 3.Convert TF-Lite Model
Generate a TF-Lite model using the tflite_convert command.
The output_tflite_graph.tflite file is generated in the logs directory.
```
$ tflite_convert \
--output_file=./logs/output_tflite_graph.tflite \
--graph_def_file=./logs/freeze_mnist_fully_connected_feed_graph.pb \
--inference_type=QUANTIZED_UINT8 \
--input_arrays=input_tensor \
--output_arrays=softmax_tensor \
--mean_values=0 \
--std_dev_values=255 \
--input_shapes="1,784"
```
### 4.Compile Edge TPU Model
Compile the TensorFlow Lite model for compatibility with the Edge TPU.
See [Edge TPU online compiler](https://coral.withgoogle.com/web-compiler) for details.
# How to Run a model on the Edge TPU
First, Copy the Edge TPU Model to the classify directory.
The directory structure is as follows.
```
classify
+ classify_mnist.py
+ mnist_data.py
+ xxxx_edgetpu.tflite
```
Note: "xxxx_edgetpu.tflite" is Edge TPU Model. This repository contains "mnist_tflite_edgetpu.tflite"
Then, classify MNIST data:
```
$ python3 classify_mnist.py --model=./mnist_tflite_edgetpu.tflite
```
The result of classification with images is displayed on the console as follows.
"Inference latency:" and "Output tensor:" are return values of the RunInference method of eddypu.basic.basic_engine.BasicEngine.
(See details [Python API](https://coral.withgoogle.com/docs/reference/edgetpu.basic.basic_engine/#edgetpu.basic.basic_engine.BasicEngine.RunInference))
"Inference" is the most accurate value of "output tensor" (correct and predicted numbers).
"Label" is the correct label.
Note: Full command-line options:
```
$ python3 classify_mnist.py -h
```
## How to Run a model on the TF-Lite
If you want to use the TF-Lite API, take a look at the notebook sample.
Classify MNIST using TF-Lite Model.
For more information, please refer [my blogs](https://nextremer-nbo.blogspot.com/2019/06/edge-tpu.html).
(comming soon. but sorry japanese only.)