DLAD Exercise 2

Multitask Semantic Segmentation and Monocular Depth prediction for Autonomous Driving

This is a template project for the practical assignment #2 in Deep Learning for Autonomous Driving course: https://www.trace.ethz.ch/teaching/DLAD

Please refer to the assignment PDF for instructions regarding the assignment scope, and training with Amazon Sagemaker in the cloud.

---

<INSERT A BRIEF VERSION OF YOUR REPORT HERE>

---

Read below only if you wish to use the code in a standalone environment.

Installation and Datasets

Create and activate a new virtual environment, and install python dependencies into it:

python -m venv ~/venv_dlad_ex2
source ~/venv_dlad_ex2/bin/activate
pip install .

The dataset link will be made available in the assignment PDF handout.

Quick Start Training

Only two command line arguments are required to begin training immediately:

• log_dir: path to directory which will be filled with experiment artifacts: logs, tensorboard files, model checkpoints, etc.;
• dataset_root: path to MiniScapes dataset

To train a model use the following command line template:

CUDA_VISIBLE_DEVICES=<gpu_ids> python -m mtl.scripts.train <arguments>    

Example:

CUDA_VISIBLE_DEVICES=0 python -m mtl.scripts.train 
    --log_dir /home/ubuntu/out_dlad_ex2 
    --dataset_root /home/ubuntu/miniscapes/

Visualization with TensorBoard

Each experiment writes TensorBoard logs into the path, specified by log_dir command line argument.

Run TensorBoard with the following command in a separate shell:

tensorboard --logdir <PATH> --port <PORT>

PATH can either be a single experiment log folder, or a parent folder with multiple nested experiments.

PORT value should be chosen such that it is not occupied by other services, e.g. 10000. To access tensorboard, navigate with the browser to machine URL with selected port, e.g. http://localhost:10000 (if running locally), or http://URL:10000 (if running remotely).

When running tensorboard server on a remote machine, which does not expose the selected port online, any port forwarding technique can help:

• SSH. To forward a remote port to localhost via ssh, just create a separate ssh session by running ssh -L <PORT>:localhost:<PORT> <URL>, then you can access localhost:port URL locally while that session is up.
• NGROK. Register at www.ngrok.com (free), then follow instructions - this way TensorBoard can be accessed even by a smartphone.

Command Line Keys

Full description of all command line keys can be found in config.py or by running
python -m mtl.scripts.train --help