Skip to content

Lane segmentation model trained with tensorflow implementation MobileNetV2 based U-Net

Notifications You must be signed in to change notification settings

xadrianzetx/mobileunet-tensorflow

Repository files navigation

MobileUnet

Lane segmentation model trained with tensorflow implementation of MobileNetV2 based U-Net

Setup

Training environment is contenerized, however you still need some stuff installed on host machine. This includes:

  • Docker
  • NVIDIA driver package
  • nvidia-docker

This will allow to take full advantage of host GPUs from docker container.

If you are running Ubuntu 16.04 then use envsetup.sh which will take care of setting up your host with all necessities.

sudo dos2unix envsetup.sh && \ 
sudo chmod -x envsetup.sh && \
sudo bash envsetup.sh

Training

First, build docker image with

docker build . --tag mobileunet:v1

then you can start training with

docker run -d --gpus all -v $(pwd)/data:/app/data \
-e --mode="train" \
-e --data-train=${DATA_DIR} \
-e --epochs=${EPOCHS} \
-e --model-name=${MODEL_NAME} mobileunet:v1

Container has volume mapped to $(pwd)/data so make sure such path exists on your host. This directory should also contain training data. Baseline for this model was trained on CU Lane Dataset. Tensorboard files and json logs will be available at $(pwd)/data/logs, model checkpoints and .h5 are saved to $(pwd)/data/models.

Testing on CPU

Run python test.py -m image -f ${IMG_PATH} -s ${SAVEPOINT} to test the model. You can pass -m image to run model on .jpg or -m video to run it on video feed. FlatBuffer model (.tflite) supported by passing --flatbuff

Testing on TPU

In order to test model on Coral USB Accelerator, prepare local env and make sure your model has been compiled, then run python tpu_test.py -m ${MODEL_PATH} -v ${VIDEO_PATH}

References

About

Lane segmentation model trained with tensorflow implementation MobileNetV2 based U-Net

Topics

Resources

Stars

Watchers

Forks