svg update (6f5fcdc7) · Commits · 张泽凯 / VAGEN

vagen/env/README.md

+4 −2

Original line number	Diff line number	Diff line
		@@ -24,6 +24,8 @@ pip install "gymnasium[toy-text]"

		### SVG
		```
		# Download dataset from huggingface

		# Additional dependencies:
		pip install bs4
		pip install svgpathtools
		pip install cairosvg
		```
		No newline at end of file

vagen/env/svg/env.py

+12 −18

Original line number	Diff line number	Diff line
		@@ -10,6 +10,7 @@ import re
		import json
		import logging
		import random
		from PIL import Image
		from typing import Dict, Any, Optional, Tuple
		from pathlib import Path
		from datasets import Dataset
		@@ -32,7 +33,7 @@ class SVGEnv(BaseEnv):
		self.img_id = None
		self.gt_svg_code = None
		self.gt_image = None
		self.gen_svg_code = ""
		self.gen_svg_code = None
		self.gen_image = None

		# Initialize random number generator
		@@ -127,6 +128,12 @@ class SVGEnv(BaseEnv):
		'failure_reason': 'invalid_svg'
		}
		self.failure_logger.info(json.dumps(failure_info))

		done = True
		info["metrics"] = metrics
		self.total_reward += self.reward
		self.gen_svg_code = None
		return self._render(init_obs=False), self.reward, done, info
		else:
		# Valid SVG code - apply format reward and process it
		self.reward += self.config.format_reward
		@@ -225,22 +232,6 @@ class SVGEnv(BaseEnv):
		"""Return the total reward collected so far"""
		return self.total_reward

		def render(self, mode='text'):
		"""Render the current state of the environment

		Args:
		mode: Rendering mode ('text' is the only supported mode)

		Returns:
		String representation of the current state
		"""
		assert mode == 'text', "Only text mode is supported for rendering"

		if not self.gen_svg_code:
		return self.gt_svg_code
		else:
		return self.gen_svg_code

		def close(self):
		"""Close the environment and clean up resources"""
		if hasattr(self, 'failure_logger'):
		@@ -260,8 +251,10 @@ class SVGEnv(BaseEnv):
		# Determine which image to show
		if init_obs:
		img = self.gt_image
		else:
		elif self.gen_svg_code:
		img = self.gen_image
		else:
		img = Image.new('RGB', (256, 256), color='white')

		# Set up multi-modal data with the image
		img_placeholder = self.config.get("image_placeholder", "image")
		@@ -347,6 +340,7 @@ if __name__ == "__main__":
		obs, reward, done, info = env.step(action)
		print(f"Reward: {reward}")
		print(f"Done: {done}")
		print(f"obs:{obs}")
		print(f"Score components: {info.get('scores', {})}")

		# Test with another seed to verify determinism

vagen/env/svg/svg_utils.py

+0 −2

Original line number	Diff line number	Diff line
		@@ -109,7 +109,6 @@ def load_svg_dataset(data_dir, dataset_name, split):
		try:
		from datasets import load_from_disk
		dataset = load_from_disk(local_path)
		print(f"Successfully loaded dataset from {local_path} with {len(dataset)} examples")
		return dataset
		except Exception as e:
		print(f"Error loading from simplified path: {e}")
		@@ -117,7 +116,6 @@ def load_svg_dataset(data_dir, dataset_name, split):
		try:
		print(f"Downloading dataset from HuggingFace: {dataset_name}")
		dataset = load_dataset(dataset_name, split=split)
		print(f"Successfully downloaded dataset with {len(dataset)} examples")

		try:
		os.makedirs(os.path.dirname(local_path), exist_ok=True)

vagen/examples/debug_svg_vision_grpo/env_config.yaml

+2 −2

Original line number	Diff line number	Diff line
		@@ -2,7 +2,7 @@ env1:
		env_name: svg
		env_config:
		split: train
		train_size: 10000
		train_size: 1000
		test_size: 0

		env2:
		@@ -10,6 +10,6 @@ env2:
		env_config:
		split: test
		train_size: 0
		test_size: 100
		test_size: 200

vagen/examples/debug_svg_vision_grpo/run.sh

+4 −4

Original line number	Diff line number	Diff line
		@@ -20,14 +20,14 @@ python3 -m vagen.trainer.main_ppo \
		data.val_files=data/svg-vision-debug/test.parquet \
		data.train_batch_size=16 \
		data.max_prompt_length=1024 \
		data.max_response_length=128 \
		data.max_trajectory_length=1800 \
		data.max_response_length=648 \
		data.max_trajectory_length=3600 \
		data.image_key=images \
		data.truncation=error \
		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_mini_batch_size=4 \
		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 \
		@@ -38,7 +38,7 @@ python3 -m vagen.trainer.main_ppo \
		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=0.3 \
		actor_rollout_ref.rollout.gpu_memory_utilization=0.5 \
		actor_rollout_ref.rollout.enable_chunked_prefill=False \
		actor_rollout_ref.rollout.enforce_eager=False \
		actor_rollout_ref.rollout.free_cache_engine=False \

Admin message