Skip to content

Latest commit

 

History

History
49 lines (38 loc) · 3.44 KB

training.md

File metadata and controls

49 lines (38 loc) · 3.44 KB

Input

  1. Image
  2. Image RoI
  3. Use common mask
  4. Use spatial encoder
  5. Upsample mask

Train Time

  1. Train for 8 epochs, batch_size=24, LR=1E-4
  2. Optimize binary and unary predictions together after 4 epochs

Output

  1. Unaries (Translation, Scale, Rotation)
  2. Binaries (Relative Translation, Relative Scale, Relative Direction)

Train SunCG

Box3D base model

nice -n 20 python -m relative3d.experiments.suncg.box3d --plot_scalars --save_epoch_freq=4 --batch_size=24 --name=box3d_base_spatial_mask_common_upsample --use_context --pred_voxels=False --classify_rot --shape_loss_wt=10 --n_data_workers=8 --num_epochs=8 --suncg_dir /nvme-scratch/nileshk/suncg/ --pred_relative=True --visdom --display_port=8094 --display_id=1 --rel_opt=True --display_freq=100 --display_visuals  --auto_rel_opt=5  --use_spatial_map=True --use_mask_in_common=True  --upsample_mask=True

DWR model

Finetune predictions on detections

python -m relative3d.experiments.suncg.dwr --plot_scalars --save_epoch_freq=1 --batch_size=24 --name=dwr_base_spatial_mask_common_upsample --use_context --pred_voxels=False --classify_rot --box3d_ft --box3d_pretrain_name=box3d_base_spatial_mask_common_upsample --shape_loss_wt=10 --n_data_workers=8 --num_epochs=1 --suncg_dir /nvme-scratch/nileshk/suncg/ --pred_relative=True --visdom --display_port=8094 --display_id=1 --rel_opt=True --display_freq=100 --display_visuals  --use_spatial_map=True --use_mask_in_common=True  --upsample_mask=True

DWR FT

Fineture shape decoder

python -m relative3d.experiments.suncg.dwr --name=dwr_base_spatial_mask_common_upsample_ft --classify_rot --shape_dec_ft --use_context --plot_scalars --display_visuals --save_epoch_freq=1 --display_freq=1000 --display_id=202 --shape_loss_wt=2 --label_loss_wt=10 --batch_size=24 --num_epochs=1 --ft_pretrain_epoch=1 --ft_pretrain_name=dwr_base_spatial_mask_common_upsample --split_size=1.0 --display_port=8094 --suncg_dir=/nvme-scratch/nileshk/suncg/ --n_data_workers=4 --visdom=True --use_spatial_map=True --use_mask_in_common=True --upsample_mask=True --pred_relative=True --rel_opt=True

Train on NYU

Box3D base model

We are going to fine tune the model trained on SunCG

python -m relative3d.experiments.nyu.box3d --plot_scalars --save_epoch_freq=4 --batch_size=24 --name=nyu_box3d_base_spatial_mask_common_upsample --use_context --pred_voxels=False --classify_rot --shape_loss_wt=10 --n_data_workers=8 --num_epochs=16 --nyu_dir /nfs.yoda/imisra/nileshk/nyud2/ --pred_relative=True --visdom --display_port=8094 --display_id=1 --rel_opt=True --display_freq=100 --display_visuals   --use_spatial_map=True --use_mask_in_common=True  --upsample_mask=True --ft_pretrain_name=box3d_base_spatial_mask_common_upsample --ft_pretrain_epoch=8

DWR base model

We are going to fine tune the model trained on SunCG for detection. We do not fine tune the shape decoder on NYUv2 as the dataset has very few CAD models.

python -m relative3d.experiments.nyu.dwr --plot_scalars --save_epoch_freq=1 --batch_size=8 --name=nyu_dwr_base_spatial_mask_common_upsample --use_context --pred_voxels=False --classify_rot --ft_pretrain_name=dwr_base_spatial_mask_common_upsample --ft_pretrain_epoch=1 --shape_loss_wt=10 --n_data_workers=0 --num_epochs=16 --nyu_dir /nfs.yoda/imisra/nileshk/nyud2/ --pred_relative=True --visdom --display_port=8094 --display_id=1 --rel_opt=True --display_freq=100 --display_visuals  --use_spatial_map=True --use_mask_in_common=True  --upsample_mask=True