tasks=open_microwave

demos=5

demo_path=data # path to the dataset

batch_size=32

start_seed=0

logdir=logs # path to save logs

training_iterations=20000 # increase training_iterations when number of demos increase

save_freq=$((training_iterations/25))

method=SGR

lr=0.003

color_drop=0.2

use_semantic=[5]

pretrained_model=clip

tag=sgrv1-demos_${demos}-iter_${training_iterations}

model=pointnext-s_cls

test_type=missing5 # eval last 5 checkpoints

test_envs=5

test_episodes=50

i=0 # GPU id

# Training

CUDA_VISIBLE_DEVICES=$i python train.py \

rlbench.tasks=$tasks rlbench.demos=$demos rlbench.demo_path=$demo_path/train replay.batch_size=$batch_size \

framework.start_seed=$start_seed framework.logdir=$logdir framework.save_freq=$save_freq framework.training_iterations=$training_iterations \

method=$method method.lr=$lr method.color_drop=$color_drop method.use_semantic=$use_semantic method.pretrained_model=$pretrained_model \

method.tag=$tag model=$model

# Evaluation

CUDA_VISIBLE_DEVICES=$i python eval.py \

rlbench.tasks=$tasks rlbench.demo_path=$demo_path/test framework.start_seed=$start_seed framework.logdir=$logdir \

framework.eval_type=$test_type framework.eval_envs=$test_envs framework.eval_episodes=$test_episodes \

method.name=$method method.tag=$tag model.name=$model