document-classification-env / server /document_classification_environment.py
TanujInsane
Update inference.py retry loop for server connection
1b64cdd
Raw History Blame Contribute Delete
5.11 kB
"""
Document Classification Environment β€” OpenEnv server-side implementation.
Bridges the existing DocumentClassificationEnv (gym-style) to the OpenEnv
Environment base class. The gym env contains all real RL logic; this adapter
only converts types and wires up reset/step/state for the OpenEnv HTTP server.
Architecture:
OpenEnv HTTP/WebSocket server (server/app.py)
└── DocumentClassificationEnvironment ← this file
└── DocumentClassificationEnv ← environment.py (gym env, unchanged)
"""
from __future__ import annotations
import sys
import os
import uuid
from typing import Any, Optional
# Allow importing from parent directory (environment.py, models.py, tasks.py)
sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
from openenv.core.env_server import Environment
from environment import DocumentClassificationEnv
from models import DocAction, DocObservation, DocState
class DocumentClassificationEnvironment(Environment):
"""
OpenEnv-compliant wrapper around DocumentClassificationEnv.
Satisfies the OpenEnv Environment interface:
- reset(seed, **kwargs) β†’ DocObservation
- step(DocAction) β†’ DocObservation (obs.done, obs.reward set)
- state (property) β†’ DocState
One instance = one difficulty level. The create_app() factory in app.py
instantiates a fresh instance per WebSocket session (for isolation).
"""
def __init__(self, difficulty: str = "easy", seed: int = 42):
self._difficulty = difficulty
self._seed = seed
self._gym_env = DocumentClassificationEnv(
task_difficulty=difficulty, seed=seed
)
self._episode_id: str = str(uuid.uuid4())
self._step_count: int = 0
# ── OpenEnv interface ─────────────────────────────────────────────────────
def reset(
self,
seed: Optional[int] = None,
episode_id: Optional[str] = None,
**kwargs: Any,
) -> DocObservation:
"""Start a new episode and return the first observation."""
obs, _ = self._gym_env.reset(seed=seed)
self._episode_id = episode_id or str(uuid.uuid4())
self._step_count = 0
return self._make_obs(obs, reward=0.0, done=False)
def step(self, action: DocAction, **kwargs: Any) -> DocObservation:
"""Execute one classification action and return the resulting observation."""
obs, reward, done, _, info = self._gym_env.step(int(action.action_id))
self._step_count += 1
return self._make_obs(obs, reward=float(reward), done=bool(done), metadata=info)
@property
def state(self) -> DocState:
"""Return episode-level state metadata (not exposed to agent during training)."""
s = self._gym_env.state()
return DocState(
episode_id=self._episode_id,
step_count=self._step_count,
difficulty=self._difficulty,
accuracy=float(s.get("accuracy", 0.0)),
total_reward=float(s.get("total_reward", 0.0)),
sla_breaches=int(s.get("sla_breaches", 0)),
)
# ── Internal helpers ──────────────────────────────────────────────────────
def _make_obs(
self,
gym_obs: dict,
reward: float,
done: bool,
metadata: Optional[dict] = None,
) -> DocObservation:
"""Convert a raw gymnasium observation dict to a typed DocObservation."""
def _scalar(value, default=0):
"""Unwrap numpy scalars / 1-element arrays to Python primitives."""
if value is None:
return default
if hasattr(value, "__len__") and len(value) == 1:
return type(default)(value[0])
if hasattr(value, "item"):
return value.item()
return value
features_raw = gym_obs.get("features", [])
features = (
features_raw.tolist()
if hasattr(features_raw, "tolist")
else list(features_raw)
)
return DocObservation(
document_id=gym_obs.get("document_id", ""),
content=gym_obs.get("content", ""),
word_count=int(_scalar(gym_obs.get("word_count"), 0)),
has_urgency_markers=bool(_scalar(gym_obs.get("has_urgency_markers"), 0)),
features=[float(f) for f in features],
document_index=int(_scalar(gym_obs.get("document_index"), 0)),
total_documents=int(_scalar(gym_obs.get("total_documents"), 0)),
metadata_response=gym_obs.get(
"metadata_response", "Not requested. Use action to request customer history."
),
sla_remaining=float(_scalar(gym_obs.get("sla_remaining"), 999.0)),
reward=reward,
done=done,
metadata=metadata or {},
)