steffen/FIKS_Entwicklungsumgebung
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

monitor now returning info objects

Browse Source
This commit is contained in:
steffen-illium
2021-05-31 13:58:24 +02:00
parent 7b4e60b0aa
commit 403d38dc24
7 changed files with 61 additions and 59 deletions
Download Patch File Download Diff File Expand all files Collapse all files

66
main.py
View File

@@ -9,6 +9,7 @@ import pandas as pd
from stable_baselines3.common.callbacks import CallbackList
from environments.factory.simple_factory import DirtProperties, SimpleFactory
from environments.helpers import IGNORED_DF_COLUMNS
from environments.logging.monitor import MonitorCallback
from environments.logging.plotting import prepare_plot
from environments.logging.training import TraningMonitor
@@ -22,16 +23,11 @@ def combine_runs(run_path: Union[str, PathLike]):
df_list = list()
for run, monitor_file in enumerate(run_path.rglob('monitor_*.pick')):
with monitor_file.open('rb') as f:
monitor_list = pickle.load(f)
monitor_df = pickle.load(f)
for m_idx in range(len(monitor_list)):
monitor_list[m_idx]['episode'] = m_idx
monitor_list[m_idx]['run'] = run
monitor_df['run'] = run
df = pd.concat(monitor_list, ignore_index=True)
df['train_step'] = range(df.shape[0])
df = df.fillna(0)
monitor_df = monitor_df.fillna(0)
#for column in list(df.columns):
# if column not in ['episode', 'run', 'step', 'train_step']:
@@ -40,20 +36,16 @@ def combine_runs(run_path: Union[str, PathLike]):
# else:
# df[f'{column}_mean_roll'] = df[column].rolling(window=50, min_periods=1).mean()
df_list.append(df)
df_list.append(monitor_df)
df = pd.concat(df_list, ignore_index=True)
df = df.fillna(0).rename(columns={'episode': 'Episode', 'run': 'Run'})
columns = [col for col in df.columns if col not in ['Episode', 'Run', 'train_step', 'step']]
columns = [col for col in df.columns if col not in IGNORED_DF_COLUMNS]
df_group = df.groupby(['Episode', 'Run']).aggregate(
{col: 'mean' if col in ['dirt_amount', 'dirty_tiles'] else 'sum' for col in columns}
)
non_overlapp_window = df.groupby(['Run', df['Episode'] // 20]).mean()
non_overlapp_window = df_group.groupby(['Run', (df_group.index.get_level_values('Episode') // 20)]).mean()
df_melted = non_overlapp_window.reset_index().melt(id_vars=['Episode', 'Run'],
value_vars=columns, var_name="Measurement",
value_name="Score")
df_melted = non_overlapp_window[columns].reset_index().melt(id_vars=['Episode', 'Run'],
value_vars=columns, var_name="Measurement",
value_name="Score")
prepare_plot(run_path / f'{run_path.name}_monitor_lineplot.png', df_melted)
print('Plotting done.')
@@ -61,36 +53,38 @@ def combine_runs(run_path: Union[str, PathLike]):
if __name__ == '__main__':
# combine_runs('debug_out/PPO_1622120377')
# combine_runs('debug_out/PPO_1622399010')
# exit()
from stable_baselines3 import PPO # DQN
from stable_baselines3 import PPO, DQN
dirt_props = DirtProperties()
time_stamp = int(time.time())
out_path = None
for seed in range(5):
for modeL_type in [PPO]:
for seed in range(5):
env = SimpleFactory(n_agents=1, dirt_properties=dirt_props, allow_diagonal_movement=True, allow_no_op=False)
env = SimpleFactory(n_agents=1, dirt_properties=dirt_props,
allow_diagonal_movement=False, allow_no_op=False)
model = PPO("MlpPolicy", env, verbose=1, ent_coef=0.0, seed=seed, device='cpu')
model = modeL_type("MlpPolicy", env, verbose=1, seed=seed, device='cpu')
out_path = Path('debug_out') / f'{model.__class__.__name__}_{time_stamp}'
out_path = Path('debug_out') / f'{model.__class__.__name__}_{time_stamp}'
identifier = f'{seed}_{model.__class__.__name__}_{time_stamp}'
out_path /= identifier
identifier = f'{seed}_{model.__class__.__name__}_{time_stamp}'
out_path /= identifier
callbacks = CallbackList(
[TraningMonitor(out_path / f'train_logging_{identifier}.csv'),
MonitorCallback(env, filepath=out_path / f'monitor_{identifier}.pick', plotting=False)]
)
callbacks = CallbackList(
[TraningMonitor(out_path / f'train_logging_{identifier}.csv'),
MonitorCallback(env, filepath=out_path / f'monitor_{identifier}.pick', plotting=False)]
)
model.learn(total_timesteps=int(2e6), callback=callbacks)
model.learn(total_timesteps=int(5e5), callback=callbacks)
save_path = out_path / f'model_{identifier}.zip'
save_path.parent.mkdir(parents=True, exist_ok=True)
model.save(save_path)
save_path = out_path / f'model_{identifier}.zip'
save_path.parent.mkdir(parents=True, exist_ok=True)
model.save(save_path)
if out_path:
combine_runs(out_path)
if out_path:
combine_runs(out_path.parent)