Merge branch 'main' of github.com:nv-tlabs/LION into main

script/train_vae_all.sh

@@ -0,0 +1,48 @@
+if [ -z "$1" ]
+    then
+    echo "Require NGPU input; "
+    exit
+fi
+DATA=" ddpm.input_dim 3 data.cates all "
+NGPU=$1 # 
+num_node=1
+BS=32 
+total_bs=$(( $NGPU * $BS ))
+if (( $total_bs > 128 )); then 
+    echo "[WARNING] total batch_size larger than 128 may lead to unstable training, please reduce the size"
+    exit
+fi
+
+ENT="python train_dist.py --num_process_per_node $NGPU "
+kl=0.5  
+lr=1e-3
+latent=1
+skip_weight=0.01 
+sigma_offset=6.0
+loss='l1_sum'
+
+$ENT ddpm.num_steps 1 ddpm.ema 0 \
+    trainer.opt.vae_lr_warmup_epochs 0 \
+    latent_pts.ada_mlp_init_scale 0.1 \
+    sde.kl_const_coeff_vada 1e-7 \
+    trainer.anneal_kl 1 sde.kl_max_coeff_vada $kl \
+    sde.kl_anneal_portion_vada 0.5 \
+    shapelatent.log_sigma_offset $sigma_offset latent_pts.skip_weight $skip_weight \
+    trainer.opt.beta2 0.99 \
+    data.num_workers 4 \
+    ddpm.loss_weight_emd 1.0 \
+    trainer.epochs 8000 data.random_subsample 1 \
+    viz.viz_freq -400 viz.log_freq -1 viz.val_freq 200 \
+    data.batch_size $BS viz.save_freq 2000 \
+    trainer.type 'trainers.hvae_trainer' \
+    model_config default shapelatent.model 'models.vae_adain' \
+    shapelatent.decoder_type 'models.latent_points_ada.LatentPointDecPVC' \
+    shapelatent.encoder_type 'models.latent_points_ada.PointTransPVC' \
+    latent_pts.style_encoder 'models.shapelatent_modules.PointNetPlusEncoder' \
+    shapelatent.prior_type normal \
+    shapelatent.latent_dim $latent trainer.opt.lr $lr \
+    shapelatent.kl_weight ${kl} \
+    shapelatent.decoder_num_points 2048 \
+    data.tr_max_sample_points 2048 data.te_max_sample_points 2048 \
+    ddpm.loss_type $loss cmt "lion" \
+    $DATA viz.viz_order [2,0,1] data.recenter_per_shape False data.normalize_global False data.normalize_shape_box True #True