def sweep_summary()

in svg/analysis.py [0:0]

• 41 lines of code
• 13 McCabe index (conditional complexity)

def sweep_summary(root):
    configs = {}
    all_summary = []
    for d in glob(f'{root}/*/'):
        eval_df = load_eval(d)
        if eval_df is None:
            continue
#         last_eval_rew = eval_df.episode_reward.values[-10:].mean()
        last_eval_rew = eval_df.episode_reward.values[-1]
        best_eval_rew = eval_df.episode_reward.values.max()
        fname = f'{d}/config.yaml'
        if not os.path.exists(fname):
            fname = f'{d}/.hydra/config.yaml'
            assert os.path.exists(fname)
        config = OmegaConf.load(fname)
        configs[d] = config
        fname = f'{d}/overrides.yaml'
        if not os.path.exists(fname):
            fname = f'{d}/.hydra/overrides.yaml'
            assert os.path.exists(fname)
        overrides = OmegaConf.load(fname)
        summary = dict(x.split('=') for x in overrides)
        summary['best_eval_rew'] = best_eval_rew
        summary['last_eval_rew'] = last_eval_rew
        summary['d'] = d
        summary['env_name'] = config.env_name
        all_summary.append(summary)

    if len(all_summary) == 0:
        print('No experiments with eval data found.')
        return [None]*4

    all_summary = pd.DataFrame(all_summary)
    for col in all_summary.columns:
        if col != 'env_name' and len(all_summary[col].unique()) == 1:
            all_summary.drop(col,inplace=True,axis=1)

    filt = ['env_name', 'seed']
    groups = [x.split('=')[0] for x in overrides]
    groups = [x for x in groups if x not in filt]
    groups = list(set(groups) & set(all_summary.columns))
    groups = ['env_name'] + groups
    groups = all_summary.groupby(groups)
    agg = groups.agg(['mean', 'std'])

    return all_summary, groups, agg, configs