# pip install wrapt
import collections
import warnings
import numpy as np
import tensorflow as tf
import sonnet as snt
import plotnine as gg
from typing import Any, Callable, Dict, List, Tuple
# Plotnine themes
gg.theme_set(gg.theme_bw(base_size=16, base_family='serif'))
gg.theme_update(figure_size=(12, 8), panel_spacing_x=0.5, panel_spacing_y=0.5)
# Filter meaningless pandas warnings
warnings.filterwarnings(action="ignore", category=UserWarning)
warnings.filterwarnings(action="ignore", category=ImportWarning)
# @title (CODE) Define the _Deep Sea_ environment
TimeStep = collections.namedtuple('TimeStep',
['observation', 'reward', 'pcont'])
class DeepSea(object):
def __init__(self, size: int, seed: int = None, randomize: bool = True):
self._size = size
def upper_interval(x):
return np.mean(x) + 2 * np.std(x)
#############################################################################
# Collating data with Pandas
params_df = config_lib.get_params_df(config)
df = pd.merge(pd.concat(results), params_df, on='unique_id')
plt_df = (df.groupby(['agent', 't']).agg({
'instant_regret': np.mean
}).reset_index())
#############################################################################
# Plotting and analysis (uses plotnine by default)
gg.theme_set(gg.theme_bw(base_size=16, base_family='serif'))
gg.theme_update(figure_size=(12, 8))
p = (gg.ggplot(plt_df) + gg.aes('t', 'instant_regret', colour='agent') +
gg.geom_line())
print(p)
#############################################################################
# Collating data with Pandas
params_df = config_lib.get_params_df(config)
df = pd.merge(pd.concat(results), params_df, on='unique_id')
plt_df = (df.groupby(['agent', 't']).agg({
'cum_reward': np.mean
}).reset_index())
#############################################################################
# Plotting and analysis (uses plotnine by default)
def make_plots():
# Setup plotting
import pandas as pd
import plotnine as gg
import warnings
pd.options.mode.chained_assignment = None
gg.theme_set(gg.theme_bw(base_size=16, base_family='serif'))
gg.theme_update(figure_size=(12, 8),
panel_spacing_x=0.5,
panel_spacing_y=0.5)
warnings.filterwarnings('ignore')
# Load Results
experiments = {
'Random': RANDOM_RESULTS_PATH,
'TRPO': TRPO_RESULTS_PATH,
}
data_frame, sweep_vars = csv_load.load_bsuite(experiments)
bsuite_score = summary_analysis.bsuite_score(data_frame, sweep_vars)
bsuite_summary = summary_analysis.ave_score_by_tag(bsuite_score,
sweep_vars)
# Generate general plots
radar_fig = summary_analysis.bsuite_radar_plot(bsuite_summary, sweep_vars)
radar_fig.savefig(PLOTS_PATH + 'radar_fig.png', bbox_inches='tight')
bar_fig = summary_analysis.bsuite_bar_plot(bsuite_score, sweep_vars)
bar_fig.save(PLOTS_PATH + 'bar_fig.png')
compare_bar_fig = summary_analysis.bsuite_bar_plot_compare(
bsuite_score, sweep_vars)
compare_bar_fig.save(PLOTS_PATH + 'compare_bar_fig.png')
# Generate specific analyses
# Learning performance
from bsuite.experiments.bandit import analysis as bandit_analysis
bandit_df = data_frame[data_frame.bsuite_env == 'bandit'].copy()
bandit_scores = summary_analysis.plot_single_experiment(
bsuite_score, 'bandit', sweep_vars)
bandit_scores.save(PLOTS_PATH + 'bandits_scores.png')
bandit_convergence = bandit_analysis.plot_learning(bandit_df, sweep_vars)
bandit_convergence.save(PLOTS_PATH + 'bandits_convergence.png')
bandit_seeds = bandit_analysis.plot_seeds(bandit_df, sweep_vars)
bandit_seeds.save(PLOTS_PATH + 'bandits_seeds.png')
# Robustness to noise
from bsuite.experiments.bandit_noise import analysis as bandit_noise_analysis
bandit_noise_df = data_frame[data_frame.bsuite_env ==
'bandit_noise'].copy()
bandit_noise_overall = summary_analysis.plot_single_experiment(
bsuite_score, 'bandit_noise', sweep_vars)
bandit_noise_overall.save(PLOTS_PATH + 'bandits_noise_overall.png')
bandit_noise_avg = bandit_noise_analysis.plot_average(
bandit_noise_df, sweep_vars)
bandit_noise_avg.save(PLOTS_PATH + 'bandits_noise_avg.png')
bandit_noise_regret = bandit_noise_analysis.plot_learning(
bandit_noise_df, sweep_vars)
bandit_noise_regret.save(PLOTS_PATH + 'bandits_noise_regret.png')
# Robustness to reward scaling
from bsuite.experiments.bandit_scale import analysis as bandit_scale_analysis
bandit_scale_df = data_frame[data_frame.bsuite_env ==
'bandit_scale'].copy()
bandit_scale_overall = summary_analysis.plot_single_experiment(
bsuite_score, 'bandit_scale', sweep_vars)
bandit_scale_overall.save(PLOTS_PATH + 'bandits_scale_overall.png')
bandit_scale_avg = bandit_scale_analysis.plot_average(
bandit_scale_df, sweep_vars)
bandit_scale_avg.save(PLOTS_PATH + 'bandits_scale_avg.png')
bandit_scale_learn = bandit_scale_analysis.plot_learning(
bandit_scale_df, sweep_vars)
bandit_scale_learn.save(PLOTS_PATH + 'bandits_scale_learn.png')
"""
