Commit ff7f23cf authored by jameskrw's avatar jameskrw
Browse files

minor

parent ebce6a88

vagen/env/README.md

+6 −4
Original line number Diff line number Diff line
@@ -8,11 +8,13 @@ python vagen/env/server.py --debug 


# For headless servers, additional setup is required:

# Install required packages

apt-get install -y pciutils

apt-get install -y xorg xserver-xorg-core xserver-xorg-video-dummy



#apt-get install -y pciutils

#apt-get install -y xorg xserver-xorg-core xserver-xorg-video-dummy

# Start X server in a tmux window

python vagen/env/navigation/startx.py 1

#python vagen/env/navigation/startx.py 1



sudo apt-get -y install libvulkan1

sudo apt install vulkan-tools

```



### Manipulation

vagen/examples/navigation_aico/env_config.yaml

0 → 100644
+8 −0
Original line number Diff line number Diff line 
env1:

    env_name: navigation  

    env_config:

        render_mode: "vision"

    train_size: 10000  

    test_size: 16

vagen/examples/navigation_aico/run.sh

0 → 100755
+76 −0
Original line number Diff line number Diff line 
set -x





export VLLM_ATTENTION_BACKEND=XFORMERS

export PYTHONHASHSEED=0



SCRIPT_DIR="$( cd "$( dirname "${BASH_SOURCE[0]}" )" && pwd )"



python -m vagen.env.create_dataset \

    --yaml_path "$SCRIPT_DIR/env_config.yaml" \

    --train_path "data/navigation-vision-debug/train.parquet" \

    --test_path "data/navigation-vision-debug/test.parquet" \

    --force_gen



# max_trajectory_length = max_prompt_length + max_response_length



python3 -m vagen.trainer.main_ppo \

    algorithm.adv_estimator=masked_gae \

    algorithm.high_level_gamma=0.95 \

    data.train_files=data/navigation-vision-debug/train.parquet \

    data.val_files=data/navigation-vision-debug/test.parquet \

    data.train_batch_size=32 \

    data.max_prompt_length=1024 \

    data.max_response_length=128 \

    data.max_trajectory_length=2400 \

    data.image_key=images \

    data.truncation=left \

    actor_rollout_ref.model.path=Qwen/Qwen2.5-VL-3B-Instruct \

    actor_rollout_ref.actor.optim.lr=1e-6 \

    actor_rollout_ref.model.use_remove_padding=True \

    actor_rollout_ref.actor.ppo_mini_batch_size=32 \

    actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=1 \

    actor_rollout_ref.actor.use_kl_loss=False \

    actor_rollout_ref.actor.kl_loss_coef=0.001 \

    actor_rollout_ref.actor.kl_loss_type=mse \

    actor_rollout_ref.model.enable_gradient_checkpointing=True \

    actor_rollout_ref.actor.fsdp_config.param_offload=False \

    actor_rollout_ref.actor.fsdp_config.optimizer_offload=False \

    actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=1 \

    actor_rollout_ref.rollout.tensor_model_parallel_size=2 \

    actor_rollout_ref.rollout.name=vllm \

    actor_rollout_ref.rollout.gpu_memory_utilization=0.2 \

    actor_rollout_ref.rollout.enable_chunked_prefill=False \

    actor_rollout_ref.rollout.enforce_eager=False \

    actor_rollout_ref.rollout.free_cache_engine=False \

    actor_rollout_ref.rollout.n=1 \

    actor_rollout_ref.ref.log_prob_micro_batch_size_per_gpu=1 \

    actor_rollout_ref.ref.fsdp_config.param_offload=True \

    actor_rollout_ref.rollout.top_p=0.95 \

    actor_rollout_ref.rollout.temperature=0.7 \

    critic.optim.lr=1e-5 \

    critic.model.use_remove_padding=True \

    critic.model.path=Qwen/Qwen2.5-VL-3B-Instruct \

    critic.model.enable_gradient_checkpointing=True \

    critic.ppo_micro_batch_size_per_gpu=1 \

    critic.model.fsdp_config.param_offload=False \

    critic.model.fsdp_config.optimizer_offload=False \

    algorithm.kl_ctrl.kl_coef=0.001 \

    trainer.critic_warmup=0 \

    trainer.logger=['console','wandb'] \

    trainer.project_name='vagen_new' \

    trainer.experiment_name='aico_navigation_vision_service' \

    trainer.n_gpus_per_node=4 \

    trainer.nnodes=1 \

    trainer.save_freq=70 \

    trainer.test_freq=20 \

    trainer.total_training_steps=200 \

    rollout_manager.max_turns=3 \

    rollout_manager.window_size=3 \

    rollout_manager.use_multi_turn_reward=False \

    rollout_manager.use_loss_mask=True \

    rollout_manager.use_gae_mask=True \

    trainer.val_before_train=True \

    trainer.val_generations_to_log_to_wandb=4 \

    rollout_manager.n_trajectory=1 \

    2>&1 | tee aico_navigation_vision.log
 
 No newline at end of file

vagen/examples/primitive_skill_aico/env_config.yaml

0 → 100644
+28 −0
Original line number Diff line number Diff line 
env1:

    env_name: primitive_skill 

    env_config:

        render_mode: text

        env_id: "AlignTwoCube" # AlignTwoCube,PlaceTwoCube,PutAppleInDrawer,StackThreeCube

    train_size: 10000  

    test_size: 8

env2:

    env_name: primitive_skill 

    env_config:

        render_mode: text

        env_id: "PlaceTwoCube" # AlignTwoCube,PlaceTwoCube,PutAppleInDrawer,StackThreeCube

    train_size: 10000  

    test_size: 8

env3:

    env_name: primitive_skill 

    env_config:

        render_mode: text

        env_id: "PutAppleInDrawer" # AlignTwoCube,PlaceTwoCube,PutAppleInDrawer,StackThreeCube

    train_size: 10000  

    test_size: 8

env4:

    env_name: primitive_skill 

    env_config:

        render_mode: text

        env_id: "StackThreeCube" # AlignTwoCube,PlaceTwoCube,PutAppleInDrawer,StackThreeCube

    train_size: 10000  

    test_size: 8
 
 No newline at end of file

vagen/examples/primitive_skill_aico/run.sh

0 → 100755
+76 −0
Original line number Diff line number Diff line 
set -x





export VLLM_ATTENTION_BACKEND=XFORMERS

export PYTHONHASHSEED=0



SCRIPT_DIR="$( cd "$( dirname "${BASH_SOURCE[0]}" )" && pwd )"



python -m vagen.env.create_dataset \

    --yaml_path "$SCRIPT_DIR/env_config.yaml" \

    --train_path "data/primitive_skill-vision-debug/train.parquet" \

    --test_path "data/primitive_skill-vision-debug/test.parquet" \

    --force_gen



# max_trajectory_length = max_prompt_length + max_response_length



python3 -m vagen.trainer.main_ppo \

    algorithm.adv_estimator=masked_gae \

    algorithm.high_level_gamma=0.95 \

    data.train_files=data/primitive_skill-vision-debug/train.parquet \

    data.val_files=data/primitive_skill-vision-debug/test.parquet \

    data.train_batch_size=64 \

    data.max_prompt_length=1024 \

    data.max_response_length=128 \

    data.max_trajectory_length=2400 \

    data.image_key=images \

    data.truncation=left \

    actor_rollout_ref.model.path=Qwen/Qwen2.5-VL-3B-Instruct \

    actor_rollout_ref.actor.optim.lr=1e-6 \

    actor_rollout_ref.model.use_remove_padding=True \

    actor_rollout_ref.actor.ppo_mini_batch_size=32 \

    actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=1 \

    actor_rollout_ref.actor.use_kl_loss=False \

    actor_rollout_ref.actor.kl_loss_coef=0.001 \

    actor_rollout_ref.actor.kl_loss_type=mse \

    actor_rollout_ref.model.enable_gradient_checkpointing=True \

    actor_rollout_ref.actor.fsdp_config.param_offload=False \

    actor_rollout_ref.actor.fsdp_config.optimizer_offload=False \

    actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=1 \

    actor_rollout_ref.rollout.tensor_model_parallel_size=2 \

    actor_rollout_ref.rollout.name=vllm \

    actor_rollout_ref.rollout.gpu_memory_utilization=0.2 \

    actor_rollout_ref.rollout.enable_chunked_prefill=False \

    actor_rollout_ref.rollout.enforce_eager=False \

    actor_rollout_ref.rollout.free_cache_engine=False \

    actor_rollout_ref.rollout.n=1 \

    actor_rollout_ref.ref.log_prob_micro_batch_size_per_gpu=1 \

    actor_rollout_ref.ref.fsdp_config.param_offload=True \

    actor_rollout_ref.rollout.top_p=0.95 \

    actor_rollout_ref.rollout.temperature=0.7 \

    critic.optim.lr=1e-5 \

    critic.model.use_remove_padding=True \

    critic.model.path=Qwen/Qwen2.5-VL-3B-Instruct \

    critic.model.enable_gradient_checkpointing=True \

    critic.ppo_micro_batch_size_per_gpu=1 \

    critic.model.fsdp_config.param_offload=False \

    critic.model.fsdp_config.optimizer_offload=False \

    algorithm.kl_ctrl.kl_coef=0.001 \

    trainer.critic_warmup=0 \

    trainer.logger=['console','wandb'] \

    trainer.project_name='vagen_new' \

    trainer.experiment_name='aico_primitive_skill_vision_service' \

    trainer.n_gpus_per_node=4 \

    trainer.nnodes=1 \

    trainer.save_freq=70 \

    trainer.test_freq=20 \

    trainer.total_training_steps=200 \

    rollout_manager.max_turns=3 \

    rollout_manager.window_size=3 \

    rollout_manager.use_multi_turn_reward=False \

    rollout_manager.use_loss_mask=True \

    rollout_manager.use_gae_mask=True \

    trainer.val_before_train=True \

    trainer.val_generations_to_log_to_wandb=8 \

    rollout_manager.n_trajectory=1 \

    2>&1 | tee aico_primitive_skill_vision.log
 
 No newline at end of file

CRA Git | Maintained and supported by SUSTech CRA and CCSE