minor bug fixing added validation config (eed12d53) · Commits · 张泽凯 / VAGEN

vagen/examples/sokoban/debug_qwen2_5_vl.sh

+3 −3

Original line number	Diff line number	Diff line
		@@ -10,7 +10,7 @@ python3 -m vagen.trainer.main_ppo \
		data.val_files=data/sokoban/test.parquet \
		data.train_batch_size=16 \
		data.max_prompt_length=2048 \
		data.max_response_length=512 \
		data.max_response_length=256 \
		data.max_trajectory_length=3072 \
		data.image_key=images \
		actor_rollout_ref.model.path=Qwen/Qwen2.5-VL-3B-Instruct \
		@@ -41,8 +41,8 @@ python3 -m vagen.trainer.main_ppo \
		trainer.experiment_name='qwen2_5_vl_3b_function_rm' \
		trainer.n_gpus_per_node=4 \
		trainer.nnodes=1 \
		trainer.save_freq=-1 \
		trainer.test_freq=-1 \
		trainer.save_freq=50 \
		trainer.test_freq=2 \
		trainer.total_epochs=15 \
		+max_turns=2 \
		2>&1 \| tee debug.log

+2 −2

Original line number	Diff line number	Diff line
		@@ -20,7 +20,7 @@ from vagen.env.base import EnvConfig,IMAGE_PLACEHOLDER
		@dataclass
		class QwenVLRolloutConifg:
		window_size: int = 5
		max_trajectories_length: int = 3072
		max_trajectory_length: int = 3072
		max_turns: int = 5
		n_gpu_per_node: int = 1 # used for multigpu batch balancing
		sptk_for_loss_mask: List[str] = field(default_factory=lambda: ['<\|box_start\|>', '<\|box_end\|>'])
		@@ -420,7 +420,7 @@ class QwenVLRolloutManger():

		input_ids_response, attention_mask_response = verl_F.tokenize_and_postprocess_data(prompt=response_with_chat_template,
		tokenizer=self.tokenizer,
		max_length=self.config.max_trajectories_length,
		max_length=self.config.max_trajectory_length,
		pad_token_id=self.tokenizer.pad_token_id,
		left_pad=False,
		truncation=self.truncation)

+2 −0

Original line number	Diff line number	Diff line
		@@ -184,3 +184,5 @@ trainer:
		remove_previous_ckpt_in_save: False
		del_local_ckpt_after_load: False
		default_local_dir: checkpoints/${trainer.project_name}/${trainer.experiment_name}
		val_before_train: True
		val_only: False
		No newline at end of file

+8 −8

Original line number	Diff line number	Diff line
		@@ -707,7 +707,7 @@ class RayPPOTrainer(object):

		if self.test_rollout_config==None:
		self.test_rollout_config = QwenVLRolloutConifg(
		max_trajectories_length=self.config.data.max_trajectories_length,
		max_trajectory_length=self.config.data.max_trajectory_length,
		max_turns=self.config.max_turns,
		n_gpu_per_node=self.config.trainer.n_gpus_per_node,
		)
		@@ -956,19 +956,19 @@ class RayPPOTrainer(object):
		# perform validation before training
		# currently, we only support validation using the reward_function.
		# TODO implement validation
		# if self.val_reward_fn is not None and self.config.trainer.get('val_before_train', True):
		# val_metrics = self._validate()
		# pprint(f'Initial validation metrics: {val_metrics}')
		# logger.log(data=val_metrics, step=self.global_steps)
		# if self.config.trainer.get('val_only', False):
		# return
		if self.val_reward_fn is not None and self.config.trainer.get('val_before_train', True):
		val_metrics = self._validate()
		pprint(f'Initial validation metrics: {val_metrics}')
		logger.log(data=val_metrics, step=self.global_steps)
		if self.config.trainer.get('val_only', False):
		return

		# we start from step 1
		self.global_steps += 1


		rollout_config = QwenVLRolloutConifg(
		max_trajectories_length=self.config.data.max_trajectories_length,
		max_trajectory_length=self.config.data.max_trajectory_length,
		max_turns=self.config.max_turns,
		n_gpu_per_node=self.config.trainer.n_gpus_per_node,
		)