Commit 4f9ed092 authored by williamzhangNU's avatar williamzhangNU
Browse files

add text-based

parent 358ab1b1

vagen/examples/release_experiments_text/grpo_mask_loss.sh

0 → 100644
+81 −0
Original line number Diff line number Diff line 
set -x





export VLLM_ATTENTION_BACKEND=XFORMERS

export PYTHONHASHSEED=0



python -m vagen.env.sokoban.create_dataset \

    --data_dir data/sokoban-text-6-step \

    --max_action_length 6 \

    --dim_room 6 6 \

    --num_boxes 1 \

    --max_steps 100 \

    --search_depth 30 \

    --start_seed 0 \

    --train_ratio 0.8 \

    --max_action_per_step 3 \

    --max_action_penalty -0.1 \

    --format_reward 0.5 \

    --n_candidate 20000



# max_trajectory_length = max_prompt_length + max_response_length



python3 -m vagen.trainer.main_ppo \

    algorithm.adv_estimator=grpo \

    algorithm.high_level_gamma=0.95 \

    data.train_files=data/sokoban-text-6-step/train.parquet \

    data.val_files=data/sokoban-text-6-step/test.parquet \

    data.train_batch_size=16 \

    data.max_prompt_length=768 \

    data.max_response_length=128 \

    data.max_trajectory_length=1152 \

    data.image_key=images \

    data.truncation=left \

    actor_rollout_ref.model.path=Qwen/Qwen2.5-0.5B-Instruct \

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

    actor_rollout_ref.model.use_remove_padding=False \

    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=1 \

    actor_rollout_ref.rollout.name=vllm \

    actor_rollout_ref.rollout.gpu_memory_utilization=0.4 \

    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=False \

    critic.model.path=Qwen/Qwen2.5-0.5B-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-text' \

    trainer.experiment_name='grpo_mask_loss' \

    trainer.n_gpus_per_node=1 \

    trainer.nnodes=1 \

    trainer.save_freq=100 \

    trainer.test_freq=20 \

    trainer.total_epochs=15 \

    rollout_manager.max_turns=3 \

    rollout_manager.window_size=5 \

    rollout_manager.use_multi_turn_reward=False \

    rollout_manager.use_loss_mask=True \

    trainer.val_before_train=True \

    trainer.val_generations_to_log_to_wandb=8 \

    rollout_manager.n_trajectory=8 \

    2>&1 | tee grpo_mask_loss.log

vagen/examples/release_experiments_text/mask_gae_mask_loss.sh

0 → 100644
+82 −0
Original line number Diff line number Diff line 
set -x





export VLLM_ATTENTION_BACKEND=XFORMERS

export PYTHONHASHSEED=0



python -m vagen.env.sokoban.create_dataset \

    --data_dir data/sokoban-text-6-step \

    --max_action_length 6 \

    --dim_room 6 6 \

    --num_boxes 1 \

    --max_steps 100 \

    --search_depth 30 \

    --start_seed 0 \

    --train_ratio 0.8 \

    --max_action_per_step 3 \

    --max_action_penalty -0.1 \

    --format_reward 0.5 \

    --n_candidate 20000



# 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/sokoban-text-6-step/train.parquet \

    data.val_files=data/sokoban-text-6-step/test.parquet \

    data.train_batch_size=128 \

    data.max_prompt_length=768 \

    data.max_response_length=128 \

    data.max_trajectory_length=1152 \

    data.image_key=images \

    data.truncation=left \

    actor_rollout_ref.model.path=Qwen/Qwen2.5-0.5B-Instruct \

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

    actor_rollout_ref.model.use_remove_padding=False \

    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=1 \

    actor_rollout_ref.rollout.name=vllm \

    actor_rollout_ref.rollout.gpu_memory_utilization=0.4 \

    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=False \

    critic.model.path=Qwen/Qwen2.5-0.5B-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-text' \

    trainer.experiment_name='mask_gae_mask_loss' \

    trainer.n_gpus_per_node=1 \

    trainer.nnodes=1 \

    trainer.save_freq=100 \

    trainer.test_freq=20 \

    trainer.total_epochs=15 \

    rollout_manager.max_turns=3 \

    rollout_manager.window_size=5 \

    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 mask_gae_mask_loss.log

vagen/examples/release_experiments_text/mask_gae_mask_loss_turnwise_reward_bi_level.sh

0 → 100644
+82 −0
Original line number Diff line number Diff line 
set -x





export VLLM_ATTENTION_BACKEND=XFORMERS

export PYTHONHASHSEED=0



python -m vagen.env.sokoban.create_dataset \

    --data_dir data/sokoban-text-6-step \

    --max_action_length 6 \

    --dim_room 6 6 \

    --num_boxes 1 \

    --max_steps 100 \

    --search_depth 30 \

    --start_seed 0 \

    --train_ratio 0.8 \

    --max_action_per_step 3 \

    --max_action_penalty -0.1 \

    --format_reward 0.5 \

    --n_candidate 20000



# max_trajectory_length = max_prompt_length + max_response_length



python3 -m vagen.trainer.main_ppo \

    algorithm.adv_estimator=bi_level_gae \

    algorithm.high_level_gamma=0.95 \

    data.train_files=data/sokoban-text-6-step/train.parquet \

    data.val_files=data/sokoban-text-6-step/test.parquet \

    data.train_batch_size=128 \

    data.max_prompt_length=768 \

    data.max_response_length=128 \

    data.max_trajectory_length=1152 \

    data.image_key=images \

    data.truncation=left \

    actor_rollout_ref.model.path=Qwen/Qwen2.5-0.5B-Instruct \

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

    actor_rollout_ref.model.use_remove_padding=False \

    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=1 \

    actor_rollout_ref.rollout.name=vllm \

    actor_rollout_ref.rollout.gpu_memory_utilization=0.4 \

    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=False \

    critic.model.path=Qwen/Qwen2.5-0.5B-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-text' \

    trainer.experiment_name='mask_gae_mask_loss_turnwise_reward_bi_level' \

    trainer.n_gpus_per_node=1 \

    trainer.nnodes=1 \

    trainer.save_freq=100 \

    trainer.test_freq=20 \

    trainer.total_epochs=15 \

    rollout_manager.max_turns=3 \

    rollout_manager.window_size=5 \

    rollout_manager.use_multi_turn_reward=True \

    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 mask_gae_mask_loss_turnwise_reward_bi_level.log

vagen/trainer/ppo/ray_trainer.py

+1 −1
Original line number Diff line number Diff line
@@ -181,7 +181,7 @@ def compute_advantage(data: DataProto, adv_estimator, gamma=1.0, lam=1.0, num_re 
            loss_mask = data.batch['loss_mask'][:, -response_length:]

        else:

            loss_mask=data.batch['attention_mask'][:, -response_length:]

        advantages, returns = core_algos.compute_BI_LEVEL_GAE_advantage_return(token_level_rewards=data.batch['token_level_rewards'],

        advantages, returns = core_algos.compute_bi_level_gae_advantage_return(token_level_rewards=data.batch['token_level_rewards'],

                                                                        values=values,

                                                                        loss_mask=loss_mask,

                                                                        gamma=gamma,

CRA Git | Maintained and supported by SUSTech CRA and CCSE