TensorFlow Lite Examples Snap

This is a snapped version of the TensorFlow Lite examples for the Raspberry Pi which are available at https://www.tensorflow.org/lite/examples.

Supported platforms

This snap has been tested using a USB webcam on the following platforms.

Build the snap

Clone this repo, then inside the cloned directory execute:

snapcraft -v

Install the snap

Currently this snap is not available from the store, so you need to build it yourself. After building the snap according to the previous section, you should have a .snap file in your working directory.

Install it by running:

sudo snap install --dangerous ./tf-lite-examples_*+snap_amd64.snap

Connect interfaces

Most of the examples use the system camera. Please connect the camera interface to give the snap access to the camera:

sudo snap connect tf-lite-examples:camera

If you want to use a Coral USB Accelerator, also connect the raw-usb plug:

sudo snap connect tf-lite-examples:raw-usb

Edge TPU

TensorFlow Lite supports Coral Edge TPUs. The snap has been tested with the Coral USB Accelerator. See supported platforms.

The correct udev rules need to be installed. For the USB Accelerator you can add the rules with this command:

sudo sh -c "cat >>/etc/udev/rules.d/60-libedgetpu1-std.rules" <<-EOF
    SUBSYSTEM=="usb",ATTRS{idVendor}=="1a6e",ATTRS{idProduct}=="089a",GROUP="plugdev"
    SUBSYSTEM=="usb",ATTRS{idVendor}=="18d1",ATTRS{idProduct}=="9302",GROUP="plugdev"
EOF

And then trigger a udev rule update:

sudo udevadm trigger

On the Raspberry Pi 5 the raw-usb interface has a bug that does not provide the required permissions to the snap to access raw-usb devices. A workaround is to install the snap in developer mode, which removes all confinement from the snap and disables automatic upgrades. This is insecure and not recommended, but currently the only way to use a Coral USB Accelerator with this snap on a Raspberry Pi 5. To install the snap in developer mode, do:

sudo snap install --dangerous ./tf-lite-examples_*+snap_amd64.snap --devmode

Running the examples

Image classification

On the CPU:

tf-lite-examples.image-classification

With a Coral USB Accelerator:

tf-lite-examples.image-classification --enableEdgeTPU --model efficientnet_lite0_edgetpu.tflite

Object detection

On the CPU:

tf-lite-examples.object-detection

With a Coral USB Accelerator:

tf-lite-examples.object-detection --enableEdgeTPU --model efficientdet_lite0_edgetpu.tflite

Pose estimation

tf-lite-examples.pose-estimation

The alternative models for pose estimation is also available. They can be selected by providing a command line argument with the model name:

$ tf-lite-examples.pose-estimation --model posenet
$ tf-lite-examples.pose-estimation --model movenet_lightning
$ tf-lite-examples.pose-estimation --model movenet_thunder
$ tf-lite-examples.pose-estimation --model movenet_multipose

Pose classification is possible by specifying the classifier and label file names:

$ tf-lite-examples.pose-estimation --classifier classifier --label_file tflite/pose_estimation/labels.txt

Segmentation

On the CPU:

tf-lite-examples.image-segmentation

With a Coral USB Accelerator:

tf-lite-examples.image-segmentation --enableEdgeTPU --model deeplabv3_edgetpu.tflite

⚠️ Audio classification

Audio classification does not currently work due to a dependency issue between snap and libportaudio2. See issue #6.

Video classification

tf-lite-examples.video-classification --label tflite/video_classification/kinetics600_label_map.txt

Note: because the running directory is different than the one where the video classification example looks for the label file, one needs to specify the path to it as an argument. See issue #7.