segmentation

TensorFlow implementation of ENet (https://arxiv.org/pdf/1606.02147.pdf) based on the official Torch implementation (https://github.com/e-lab/ENet-training) and the Keras implementation by PavlosMelissinos (https://github.com/PavlosMelissinos/enet-keras), trained on the Cityscapes dataset (https://www.cityscapes-dataset.com/).

• Youtube video of results (https://youtu.be/HbPhvct5kvs):

• 

• The results in the video can obviously be improved, but because of limited computing resources (personally funded Azure VM) I did not perform any further hyperparameter tuning.

You might get the error "No gradient defined for operation 'MaxPoolWithArgmax_1' (op type: MaxPoolWithArgmax)". To fix this, I had to add the following code to the file /usr/local/lib/python2.7/dist-packages/tensorflow/python/ops/nn_grad.py:

@ops.RegisterGradient("MaxPoolWithArgmax")  
def _MaxPoolGradWithArgmax(op, grad, unused_argmax_grad):  
  return gen_nn_ops._max_pool_grad_with_argmax(op.inputs[0], grad, op.outputs[1], op.get_attr("ksize"), op.get_attr("strides"), padding=op.get_attr("padding"))  

Documentation:

preprocess_data.py:

• ASSUMES: that all Cityscapes training (validation) image directories have been placed in data_dir/cityscapes/leftImg8bit/train (data_dir/cityscapes/leftImg8bit/val) and that all corresponding ground truth directories have been placed in data_dir/cityscapes/gtFine/train (data_dir/cityscapes/gtFine/val).
• DOES: script for performing all necessary preprocessing of images and labels.

  ---

model.py:

• ASSUMES: that preprocess_data.py has already been run.
• DOES: contains the ENet_model class.

  ---

utilities.py:

• ASSUMES: -
• DOES: contains a number of functions used in different parts of the project.

  ---

train.py:

• ASSUMES: that preprocess_data.py has already been run.
• DOES: script for training the model.

  ---

run_on_sequence.py:

• ASSUMES: that preprocess_data.py has already been run.
• DOES: runs a model checkpoint (set in line 56) on all frames in a Cityscapes demo sequence directory (set in line 30) and creates a video of the result.

Training details:

• In the paper the authors suggest that you first pretrain the encoder to categorize downsampled regions of the input images, I did however train the entire network from scratch.

• Batch size: 4.

• For all other hyperparameters I used the same values as in the paper.

• Training loss:

• 

• Validation loss:

• 

• The results in the video above was obtained with the model at epoch 23, for which a checkpoint is included in segmentation/training_logs/best_model in the repo.

Training on Microsoft Azure:

To train the model, I used an NC6 virtual machine on Microsoft Azure. Below I have listed what I needed to do in order to get started, and some things I found useful. For reference, my username was 'fregu856':

• Download Cityscapes.

• Install docker-ce:

• • $ curl -fsSL https://download.docker.com/linux/ubuntu/gpg | sudo apt-key add -
• • $ sudo add-apt-repository "deb [arch=amd64] https://download.docker.com/linux/ubuntu $(lsb_release -cs) stable"
• • $ sudo apt-get update
• • $ sudo apt-get install -y docker-ce

• Install CUDA drivers (see "Install CUDA drivers for NC VMs" in https://docs.microsoft.com/en-us/azure/virtual-machines/linux/n-series-driver-setup):

• • $ CUDA_REPO_PKG=cuda-repo-ubuntu1604_8.0.61-1_amd64.deb
• • $ wget -O /tmp/${CUDA_REPO_PKG} http://developer.download.nvidia.com/compute/cuda/repos/ubuntu1604/x86_64/${CUDA_REPO_PKG}
• • $ sudo dpkg -i /tmp/${CUDA_REPO_PKG}
• • $ rm -f /tmp/${CUDA_REPO_PKG}
• • $ sudo apt-get update
• • $ sudo apt-get install cuda-drivers
• • Reboot the VM

• Install nvidia-docker:

• • $ wget -P /tmp https://github.com/NVIDIA/nvidia-docker/releases/download/v1.0.1/nvidia-docker_1.0.1-1_amd64.deb
• • $ sudo dpkg -i /tmp/nvidia-docker.deb && rm /tmp/nvidia-docker.deb
• • $ sudo nvidia-docker run --rm nvidia/cuda nvidia-smi

• Download the latest TensorFlow docker image with GPU support (tensorflow 1.3):

• • $ sudo docker pull tensorflow/tensorflow:latest-gpu

• Create start_docker_image.sh containing:

  
  #!/bin/bash

DEFAULT VALUES

GPUIDS="0" NAME="fregu856_GPU"

NV_GPU="$GPUIDS" nvidia-docker run -it --rm \ -p 5584:5584 \ --name "$NAME""$GPUIDS" \ -v /home/fregu856:/root/ \ tensorflow/tensorflow:latest-gpu bash

- /root/ will now be mapped to /home/fregu856 (i.e., $ cd -- takes you to the regular home folder). 

- To start the image:
- - $ sudo sh start_docker_image.sh 
- To commit changes to the image:
- - Open a new terminal window.
- - $ sudo docker commit fregu856_GPU0 tensorflow/tensorflow:latest-gpu
- To stop the image when it’s running:
- - $ sudo docker stop fregu856_GPU0
- To exit the image without killing running code:
- - Ctrl-P + Q
- To get back into a running image:
- - $ sudo docker attach fregu856_GPU0
- To open more than one terminal window at the same time:
- - $ sudo docker exec -it fregu856_GPU0 bash

- To install the needed software inside the docker image:
- - $ apt-get update
- - $ apt-get install nano
- - $ apt-get install sudo
- - $ apt-get install wget
- - $ sudo apt-get install libopencv-dev python-opencv
- - Commit changes to the image (otherwise, the installed packages will be removed at exit!)