mirror of
https://github.com/illiumst/marl-factory-grid.git
synced 2025-05-23 07:16:44 +02:00
125 lines
4.2 KiB
Python
125 lines
4.2 KiB
Python
import pickle
|
|
from pathlib import Path
|
|
from collections import defaultdict
|
|
|
|
from stable_baselines3.common.callbacks import BaseCallback
|
|
|
|
from environments.logging.plotting import prepare_plot
|
|
|
|
|
|
class FactoryMonitor:
|
|
|
|
def __init__(self, env):
|
|
self._env = env
|
|
self._monitor = defaultdict(lambda: defaultdict(lambda: 0))
|
|
self._last_vals = defaultdict(lambda: 0)
|
|
|
|
def __iter__(self):
|
|
for key, value in self._monitor.items():
|
|
yield key, dict(value)
|
|
|
|
def add(self, key, value, step=None):
|
|
assert step is None or step >= 1 # Is this good practice?
|
|
step = step or self._env.steps
|
|
self._last_vals[key] = self._last_vals[key] + value
|
|
self._monitor[key][step] = self._last_vals[key]
|
|
return self._last_vals[key]
|
|
|
|
def set(self, key, value, step=None):
|
|
assert step is None or step >= 1 # Is this good practice?
|
|
step = step or self._env.steps
|
|
self._last_vals[key] = value
|
|
self._monitor[key][step] = self._last_vals[key]
|
|
return self._last_vals[key]
|
|
|
|
def remove(self, key, value, step=None):
|
|
assert step is None or step >= 1 # Is this good practice?
|
|
step = step or self._env.steps
|
|
self._last_vals[key] = self._last_vals[key] - value
|
|
self._monitor[key][step] = self._last_vals[key]
|
|
return self._last_vals[key]
|
|
|
|
def to_dict(self):
|
|
return dict(self)
|
|
|
|
def to_pd_dataframe(self):
|
|
import pandas as pd
|
|
df = pd.DataFrame.from_dict(self.to_dict())
|
|
df.fillna(0)
|
|
return df
|
|
|
|
def reset(self):
|
|
raise RuntimeError("DO NOT DO THIS! Always initalize a new Monitor per Env-Run.")
|
|
|
|
|
|
class MonitorCallback(BaseCallback):
|
|
|
|
ext = 'png'
|
|
|
|
def __init__(self, env, filepath=Path('debug_out/monitor.pick'), plotting=True):
|
|
super(MonitorCallback, self).__init__()
|
|
self.filepath = Path(filepath)
|
|
self._monitor_list = list()
|
|
self.env = env
|
|
self.plotting = plotting
|
|
self.started = False
|
|
self.closed = False
|
|
|
|
@property
|
|
def monitor_as_df_list(self):
|
|
return [x.to_pd_dataframe() for x in self._monitor_list]
|
|
|
|
def __enter__(self):
|
|
self._on_training_start()
|
|
|
|
def __exit__(self, exc_type, exc_val, exc_tb):
|
|
self._on_training_end()
|
|
|
|
def _on_training_start(self) -> None:
|
|
if self.started:
|
|
pass
|
|
else:
|
|
self.filepath.parent.mkdir(exist_ok=True, parents=True)
|
|
self.started = True
|
|
pass
|
|
|
|
def _on_training_end(self) -> None:
|
|
if self.closed:
|
|
pass
|
|
else:
|
|
# self.out_file.unlink(missing_ok=True)
|
|
with self.filepath.open('wb') as f:
|
|
pickle.dump(self.monitor_as_df_list, f, protocol=pickle.HIGHEST_PROTOCOL)
|
|
if self.plotting:
|
|
print('Monitor files were dumped to disk, now plotting....')
|
|
# %% Imports
|
|
import pandas as pd
|
|
# %% Load MonitorList from Disk
|
|
with self.filepath.open('rb') as f:
|
|
monitor_list = pickle.load(f)
|
|
df = None
|
|
for m_idx, monitor in enumerate(monitor_list):
|
|
monitor['episode'] = m_idx
|
|
if df is None:
|
|
df = pd.DataFrame(columns=monitor.columns)
|
|
for _, row in monitor.iterrows():
|
|
df.loc[df.shape[0]] = row
|
|
if df is None: # The env exited premature, we catch it.
|
|
self.closed = True
|
|
return
|
|
for column in list(df.columns):
|
|
if column != 'episode':
|
|
df[f'{column}_roll'] = df[column].rolling(window=50).mean()
|
|
# result.tail()
|
|
prepare_plot(filepath=self.filepath, results_df=df.filter(regex=(".+_roll")), tag='monitor')
|
|
print('Plotting done.')
|
|
self.closed = True
|
|
|
|
def _on_step(self) -> bool:
|
|
if self.locals['dones'].item():
|
|
self._monitor_list.append(self.env.monitor)
|
|
else:
|
|
pass
|
|
return True
|
|
|