TensorFlow Docker containers with Improvements and fixes for training Objekt Detection models.
Can also be used to train models for classification of entire images.
Prebuilt containers can be retrieved from here:
https://hub.docker.com/r/jacobpeddk/tensorflow-improved/
Information abut the TensorFlow containers, which these containers is based on, can be seen at the following location:
https://hub.docker.com/r/tensorflow/tensorflow/
https://github.com/tensorflow/tensorflow/blob/master/tensorflow/tools/docker/README.md
I'd recommend mapping a directory into the containers, so you can have the training data and results saved on the host machine.
This and related documents will expect the host folder with the data to be mapped to: /root/sharedfolder
For launching the container that uses the CPU for training, use the following command:
docker run --rm -it -p 8888:8888 -p 6006:6006 -v <host path>:/root/sharedfolder:Z tensorflow-improved:latest
Remember to replace <host path> with the actual path on your host machine!
Example:
docker run --rm -it -p 8888:8888 -p 6006:6006 -v /home/jacob/andet/training/docker-training-shared:/root/sharedfolder:Z tensorflow-improved:latest
For launching the container that uses the GPU for training, use the following command:
nvidia-docker run --rm -it -p 8888:8888 -p 6006:6006 -v <host path>:/root/sharedfolder:Z tensorflow-improved:latest-gpu
For an Optimus laptop that's using Bumblebee, you can use above command after optirun. Example:
optirun nvidia-docker run --rm -it -p 8888:8888 -p 6006:6006 -v <host path>:/root/sharedfolder:Z tensorflow-improved:latest-gpu
Remember to replace <host path> with the actual path on your host machine!
Example:
nvidia-docker run --rm -it -p 8888:8888 -p 6006:6006 -v /home/jacob/andet/training/docker-training-shared:/root/sharedfolder:Z tensorflow-improved:latest-gpu
Be aware! GPU training is a lot faster than CPU, but it requires an Nvidia GPU and Nvidia Cuda to be installed and working on the host machine.
It also requires the host to have NVIDIA Container Runtime for Docker installed.
You can read more about it here: https://github.com/NVIDIA/nvidia-docker
You can test if it's working by running the following test container with the command below. If it doesn't throw an error, you're golden.
nvidia-docker run --rm nvidia/cuda nvidia-smi
Port 8888 is used for jupyter notebook, which is a fairly cool web based teaching tool that's part of the TensorFlow containers. Port 6006is used for TensorFlow Tensorboard websitet, where one can evaluate the training process.
docker build -t tensorflow-improved -f Dockerfile .
docker build -t tensorflow-improved-gpu -f Dockerfile.gpu .