princetonlips / vitruvion Goto Github PK

Hi,

I am trying to reimplement the sketch autocomplete results in Table 2. It seems we can only do the process_single_sketch one sample each iteration, which caused the autocompletion super slow. Is there any way we can do that paralleled?

Best

Hi, thanks for your interesting and impressive work!

I'm curious about what can be the possible reasons that I'm getting a low-score for re-training image-to-primitive model. All the hyperparameters were the same with the original model and I adapted your slurm script to run the model on a similar setup (4 v100 gpus). Final accuracy for the image to primitive model was around 0.47.

I used the following slurm script to retrain the model.

#!/bin/bash -l

#SBATCH -p gpu
#SBATCH -N 1
#SBATCH --gpus=4
#SBATCH --constraint=volta32
#SBATCH --job-name train_image_to_primitive
#SBATCH --time=24:00:00
source activate vitruvion

python -m img2cad.train_image_to_primitive +cluster=akaradeniz +compute=4xv100 batch_size=2048

These are the graphs from the training logs:

Thank you in advance!

Thank you for opening your code.
How do I enter a sketch picture I draw by hand into your network and get visual network predictions?