sokoban grpo text debug finished (367cf380) · Commits · 张泽凯 / VAGEN

vagen/configs/sokoban/debug_1_step.yaml

deleted100644 → 0

+0 −11

Original line number	Diff line number	Diff line
		seed: 42
		train_path: "./data/sokoban-text-1-step/train.parquet"
		test_path: "./data/sokoban-text-1-step/test.parquet"
		"env1": {
		"env_name": "sokoban",
		'env_config': {
		'num_boxes': 1
		},
		"train_size": 1000,
		"test_size": 100,
		}
		No newline at end of file

vagen/env/base_env.py

+2 −2

Original line number	Diff line number	Diff line
		@@ -14,11 +14,11 @@ class BaseEnv(ABC):

		obs: {
		'obs_str': "This is the obs template, you see <image> and <image>, you heard <audio> and <audio>",
		'multi_modal_inputs':{
		'multi_modal_data':{
		'<image>':[list of images],
		'<audio>':[list of audios],
		}
		# num of <image> and <audio> in the obs_str should match len(multi_modal_inputs['<image>']) and len(multi_modal_inputs['<audio>'])
		# num of <image> and <audio> in the obs_str should match len(multi_modal_data['<image>']) and len(multi_modal_data['<audio>'])
		}
		info: {
		"metrics": {

vagen/env/create_dataset.py

+11 −14

Original line number	Diff line number	Diff line
		@@ -5,19 +5,19 @@ import yaml
		import argparse
		from datasets import Dataset, load_dataset
		from vagen.env.utils.env_utils import permanent_seed
		def create_dataset_from_yaml(yaml_file_path: str, force_gen=False):
		def create_dataset_from_yaml(yaml_file_path: str, force_gen=False,seed=42,train_path='./train.parquet',test_path='./test.parquet'):
		"""
		Create dataset from a YAML configuration file.

		Args:
		yaml_file_path (str): Path to the YAML configuration file
		force_gen (bool): Whether to force regeneration of existing datasets
		seed (int): Seed for random number generation
		train_path (str): Path to save the training dataset
		test_path (str): Path to save the testing dataset

		The YAML file should have the following structure:
		```
		seed: 42
		train_path: path/to/train.parquet
		test_path: path/to/test.parquet
		env1:
		env_name: sokoban # or frozenlake
		env_config:
		@@ -42,29 +42,23 @@ def create_dataset_from_yaml(yaml_file_path: str, force_gen=False):
		else:
		yaml_config = yaml_file_path

		train_path = yaml_config.get('train_path')
		test_path = yaml_config.get('test_path')

		os.makedirs(os.path.dirname(train_path), exist_ok=True)
		os.makedirs(os.path.dirname(test_path), exist_ok=True)

		if not force_gen and os.path.exists(train_path) and os.path.exists(test_path):
		print(f"Dataset files already exist at {train_path} and {test_path}. Skipping generation.")
		print(f"Use --force-gen to override and regenerate the dataset.")
		return
		return train_path, test_path


		train_instances = []
		test_instances = []

		global_seed = yaml_config.get('seed', 42)
		global_seed = seed
		permanent_seed(global_seed)


		for key, value in yaml_config.items():
		if key in ['train_path', 'test_path','seed']:
		continue

		env_name = value.get('env_name')
		custom_env_config = value.get('env_config', {})
		train_size,test_size = (value.get('train_size', 100), value.get('test_size', 100))
		@@ -132,9 +126,12 @@ if __name__ == "__main__":
		parser = argparse.ArgumentParser()
		parser.add_argument("--yaml_path", type=str, required=True, help="Path to YAML configuration file")
		parser.add_argument("--force_gen", action="store_true", help="Force regenerate dataset even if exists")
		parser.add_argument("--train_path", type=str, default="./train.parquet", help="Path to save the training dataset")
		parser.add_argument("--test_path", type=str, default="./test.parquet", help="Path to save the testing dataset")
		parser.add_argument("--seed", type=int, default=42, help="Seed for random number generation")
		args = parser.parse_args()

		train_path, test_path = create_dataset_from_yaml(args.yaml_path, args.force_gen)
		print(args)
		train_path, test_path = create_dataset_from_yaml(args.yaml_path, args.force_gen, args.seed, args.train_path, args.test_path)

		# Optionally load the dataset and print examples
		train_dataset = load_dataset('parquet', data_files={"train": train_path}, split="train")

vagen/env/frozenlake/env.py

+6 −6

Original line number	Diff line number	Diff line
		@@ -137,11 +137,11 @@ class FrozenLakeEnv(BaseEnv):

		def _render(self, init_obs=False):
		"""Render the environment"""
		multi_modal_inputs = None
		multi_modal_data = None

		if self.config.render_mode == 'vision':
		img_placeholder = self.config.image_placeholder
		multi_modal_inputs = {
		multi_modal_data = {
		img_placeholder: [convert_numpy_to_PIL(self.gym_env._render_gui(mode='rgb_array'))]
		}
		img_str = img_placeholder
		@@ -160,10 +160,10 @@ class FrozenLakeEnv(BaseEnv):
		done=self._finished(),
		)

		if multi_modal_inputs is not None:
		if multi_modal_data is not None:
		return {
		"obs_str": obs_str,
		"multi_modal_inputs": multi_modal_inputs,
		"multi_modal_data": multi_modal_data,
		}
		else:
		return {
		@@ -216,7 +216,7 @@ if __name__ == "__main__":
		import os
		if config.render_mode == 'vision':
		os.makedirs("./test_frozenlake", exist_ok=True)
		img = obs["multi_modal_inputs"][config.image_placeholder][0]
		img = obs["multi_modal_data"][config.image_placeholder][0]
		img.save(f"./test_frozenlake/frozenlake_{i}.png")
		while True:
		i += 1
		@@ -226,7 +226,7 @@ if __name__ == "__main__":
		print(obs["obs_str"])
		if config.render_mode == 'vision':
		# save the image
		img = obs["multi_modal_inputs"][config.image_placeholder][0]
		img = obs["multi_modal_data"][config.image_placeholder][0]
		img.save(f"./test_frozenlake/frozenlake_{i}.png")
		if done:
		break

vagen/env/sokoban/config.py

+1 −1

Original line number	Diff line number	Diff line
		@@ -6,7 +6,7 @@ class SokobanConfig(BaseConfig):
		dim_room: tuple = (6, 6)
		max_steps: int = 100
		num_boxes: int = 1
		render_mode: str = "vision"
		render_mode: str = "vision" # "vision" or "text"
		min_actions_to_succeed: int = 5
		max_actions_per_step: int = 3

Admin message