Skip to content

Latest commit

 

History

History
131 lines (76 loc) · 4.77 KB

training.md

File metadata and controls

131 lines (76 loc) · 4.77 KB

Training

Autoencoder [default]
python train.py --shapenet13 --dir_name log/atlasnet_autoencoder_25_squares --nb_primitives 25 --template_type "SQUARE"
Single-View Reconstruction [default]
python train.py --shapenet13 --dir_name log/atlasnet_singleview_25_squares_tmp --nb_primitives 25 --template_type SQUARE --SVR --reload_decoder_path log/atlasnet_autoencoder_25_squares --train_only_encoder

🖐️ Monitor your training on http://localhost:8890/

🖐️ See report on a completed training at http://localhost:8891/{DIR_NAME}

visdom

Options : --flag, default, help

Training parameters
  • --batch_size, 32, Training batch size
  • --batch_size_test, 32, Testing batch size
  • --workers, 0, Number of data loading workers
  • --nepoch, 150, Number of training epochs
  • --start_epoch, 0, Start directly from epoch [X]
  • --random_seed, False, Fix random seed
  • --lrate, 0.001, Learning rate
  • --lr_decay_1, 120, Learning rate decay 1
  • --lr_decay_2, 140, Learning rate decay 2
  • --lr_decay_3, 145, Learning rate decay 3
  • --multi_gpu, [0], Replacing CUDA_VISIBLE_DEVICES=X. i.e to use gpu 1 and 2 --multi_gpu 0 1
  • --loop_per_epoch, 1, Number of data loop per epoch
  • --no_learning, False, No backdrop + network in eval mode
  • --train_only_encoder, False, Only train the encoder, freeze decoder
  • --run_single_eval, False, Flag to evaluate a trained network reloaded with --reload_model_path
  • --demo, False, Run forward pass on an input and save mesh output. Input provided by --demo_input_path, network reloaded with --reload_model_path
Data

  • --SVR, False, Single_view Reconstruction mode. Affects dataloader and network architecture

  • --shapenet13, False, Load the 13 usual shapenet categories

  • --class_choice, ["airplane"], Choose classes to train on

  • --random_rotation, False, Random rotation

  • --normalization, "UnitBall" , Choices are 'UnitBall', 'BoundingBox', or 'Identity'

  • --sample, True, Sample the input pointclouds

  • --number_points, 2500, Number of point sampled on the object

  • --number_points_eval, 2500, Number of points sampled on the object during evaluation

  • --data_augmentation_axis_rotation, False, Axial rotation

  • --data_augmentation_random_flips, False, Random flips

  • --random_translation, False, Random translation

  • --anisotropic_scaling, False, Anisotropic scaling

Paths
  • --demo_input_path, "./doc/pictures/plane_input_demo.png", Input test path, requires --demo
  • --reload_model_path, "./training/trained_models/atlasnet_AE_25_patches.pth", Path to pre-trained model
  • --reload_decoder_path, "./training/trained_models/atlasnet_AE_25_patches.pth", Path to pre-trained model
  • --env, "Atlasnet", Visdom environment
  • --visdom_port, 8890, Visdom port
  • --http_port, 8891, Http port
  • --dir_name, "", Name of the log folder.
  • --id, 0, Training id
Network

  • --nb_primitives, 1, Number of primitives in atlasnet decoder

  • --template_type, "SQUARE", Choices are "SPHERE", or "SQUARE"

  • --num_layers, 2, Number of hidden MLP Layer in atlasnet decoder

  • --hidden_neurons, 512, Number of neurons in each hidden layer of atlasnet decoder

  • --bottleneck_size, 1024, Bottleneck_size for the auto encoder

  • --activation, 'relu', Choices are "relu", "sigmoid", "softplus", "logsigmoid", "softsign", or "tanh"

  • --remove_all_batchNorms, False, Replace all batchnorms by identity

Loss
  • --no_metro, False, Skip metro distance evaluation.

Quantitative Results

Method Chamfer (*1) Fscore (*2) Metro (*3) Total Train time (min)
Autoencoder 25 Squares 1.35 82.3% 6.82 731
Autoencoder 1 Sphere 1.35 83.3% 6.94 548
SingleView 25 Squares 3.78 63.1% 8.94 1422
SingleView 1 Sphere 3.76 64.4% 9.01 1297
  • (*1) x1000. Computed between 2500 ground truth points and 2500 reconstructed points.
  • (*2) The threshold is 0.001
  • (*3) x100. Metro is ran on unormalized point clouds (which explains a difference with the paper's numbers)

Paper reproduction

To reproduce main results from the paper : python ./training/launcher.py

In case you need the results of ICP on PointSetGen output :