HelloPain / tf-object-detection-DSVM

The codebase including TensorFlow Object Detection API and shell script for configuring the enviromnent on Azure Data Science Virtual Machine (Ubuntu) to run the training and evaluation out of the box

Environment configuration

First run commands below and this will install CUDA 9.0, cudnn 7.0 and tensorflow-gpu 1.5. These are required for TensorFlow Object Detection API of the version included in this repository. The details about TensorFlow Object Detection API is here

cd envsetup
chmod +x ./*.sh
./prepare.sh

After running the setup script above, virtual enviroment, tensorflow-py3ve is created. From next time, you need to run below to activate the virtual environment first.

source ~/tensorflow-py3ve/bin/activate

Run training

python -m object_detection.train --logtostderr --pipeline_config_path=YOUR_CONFIG_FILE --train_dir=TRAIN_RESULT_OUTPUT_DIRECTORY

You need to configure your pipeline file for training.

Configuring an object detection pipeline

You can download pre-trained object detection models from here.

Tensorflow detection model zoo

Run evaluation

python -m object_detection.eval  --logtostderr  --pipeline_config_path=YOUR_CONFIG_FILE --checkpoint_dir=TRAIN_RESULT_OUTPUT_DIRECTORY --eval_dir=EVAL_RESULT_OUTPUT_DIRECTORY

You have to run this evaluation concurrently with training script. This script periodically checks the update in checkpoint_dir and runs evaluation on them.

TensorBoard setting

tensorboard --logdir YOUR_RESULT_OUTPUT_DIRECTORY
ssh -L 16006:127.0.0.1:6006 YOUR_DSVM_IP_ADDRESS

Browse to 127.0.0.1:16006 and you can see the training/eval results