- Launch all the trainings with
scripts/launch.py --mode training
(remove the trained models in/log
if you downloaded them for the demo) OR call./log/download_pretrained_models.sh
. - Launch all the inferences in `scripts/launch.py --mode inference
This populates a global Pandas dataframe with the results called results.csv
.
-
Make the tables :
python scripts/make_papertables_from_results.py
and check them in thehtml/
folder. -
Make the curves : Run the notebook : The values are hard-coded from the results of my own experiments.
-
Make the figures : check the generated mesh in the
figures/
folder
- The Meta network predict a Fully Connected Layer (Matrix
input_size,ouput_size
and Bias) instead of a Adaptive Normalization (Vectorinput_size
and Bias) - The Generalized ablation study suggests that on could get rid of the chamfer loss enterily for slightly better performances.