Final project for Deep Learning (02456)

Comparison and implementation (from scratch) of YOLOv1 and U-Net.

Dataset: Airbus Ship Detection Challenge

Paper

See paper.pdf

Poster

Useful tips

We're training on the DTU HPC LSF 10 Cluster. All used scripts are available under jobscripts/.

How to submit a job

From the root of the repo run, e.g.

bsub < ./jobscripts/unet/jobscript_train.sh

Then you can use bstat to check the job status.

How to see real-time output

You can use tail to see the job log in real time, e.g.:

tail gpu_23113941.out -f

tqdm output interpretation

Epoch 3, lr 0.0001:  18%|█▊        | 7376/40432 [02:56<13:06, 42.02it/s, loss=0.01581]

• 02:56: This indicates the elapsed time since the progress bar started. In this case, no significant time has elapsed since the start of the epoch.
• 13:06: This shows the estimated remaining time to complete the current epoch or task at the current processing speed.
• 42.02it/s: This represents the processing speed, indicating that your model is processing 42.02 iterations (samples) per second.

Evaluate on interactive GPU

To just evaluate the model you can use the interactive nodes.

Type:

# Enter an interactive node (alt: sxm2sh, voltash)
a100sh                         

cd ~/deep-learning-project
clear && ./jobscripts/unet/jobscript_evaluate.sh

Synopsis