Increase line width & font size for camera-ready

markov_abstr/gridworld/plot_results.py

@@ -157,6 +157,7 @@ def plot(data, x, y, hue, style, col=None):
         col_wrap=col_wrap,
         # col_order=col_order,
         palette=palette,
+        linewidth=2,
         facet_kws={
             'sharey': False,
             'sharex': False
@@ -165,14 +166,19 @@ def plot(data, x, y, hue, style, col=None):
     g.set_titles('{col_name}')
 
     ax = g.axes.flat[0]
-    ax.set_ylim((-100, 0))
+    ax.set_ylim((-90, 0))
     ax.set_xlim((0, 100))
-    ax.axhline(-84.8, dashes=dashes['random'], color=palette['random'])
-    leg = ax.legend(labels, loc='lower center', ncol=4)
+    ax.axhline(-84.8, dashes=dashes['random'], color=palette['random'], linewidth=2)
+    leg = ax.legend(labels, loc='upper center', ncol=4, bbox_to_anchor=(0.43, -0.17), fontsize=12, frameon=False)
     leg.set_draggable(True)
-    ax.set_ylabel('Reward')
-    ax.set_xlabel('Episode')
+    for axis in ['bottom','left']:
+        ax.spines[axis].set_linewidth(2)
+    ax.tick_params(width=2)
+    ax.tick_params(labelsize=16)
+    ax.set_ylabel('Reward',fontsize=18)
+    ax.set_xlabel('Episode',fontsize=18)
     plt.tight_layout()
+    plt.subplots_adjust(bottom=0.25)
     plt.show()
 
 plot(data, x='episode', y='reward', hue='agent', style='agent')

markov_abstr/gridworld/plot_side_by_side_results.py

@@ -0,0 +1,192 @@
+import glob
+import json
+import matplotlib.pyplot as plt
+import numpy as np
+import os
+import pandas as pd
+import seaborn as sns
+
+from visgrid.utils import load_experiment, get_parser
+
+parser = get_parser()
+# yapf: disable
+parser.add_argument('--pretrain-steps', type=str, default='30k',
+                    choices=['3k','30k'], help='Number of pretraining steps')
+parser.add_argument('--smoothing', type=int, default=5,
+                    help='Number of data points for sliding window average')
+# yapf: enable
+args = parser.parse_args()
+
+def load_experiment(path):
+    logfiles = sorted(glob.glob(os.path.join(path, 'scores-*.txt')))
+    agents = [path.split('/')[-2] for f in logfiles]
+    seeds = [int(f.split('-')[-1].split('.')[0]) for f in logfiles]
+    logs = [open(f, 'r').read().splitlines() for f in logfiles]
+
+    def read_log(log):
+        results = [json.loads(item) for item in log]
+        data = smooth(pd.DataFrame(results), args.smoothing)
+        return data
+
+    results = [read_log(log) for log in logs]
+    keys = list(zip(agents, seeds))
+    data = pd.concat(results, join='outer', keys=keys,
+                     names=['agent',
+                            'seed']).sort_values(by='seed',
+                                                 kind='mergesort').reset_index(level=[0, 1])
+    return data  #[data['episode']<=100]
+
+def smooth(data, n):
+    numeric_dtypes = data.dtypes.apply(pd.api.types.is_numeric_dtype)
+    numeric_cols = numeric_dtypes.index[numeric_dtypes]
+    data[numeric_cols] = data[numeric_cols].rolling(n).mean()
+    return data
+
+fig, axes = plt.subplots(1, 2, sharey=False, figsize=(16,6))
+for pretrain_experiments, ax in zip(['pretrain_3k', 'pretrain_30k'], axes):
+# pretrain_experiments = 'pretrain_3k' if args.pretrain_steps == '3k' else 'pretrain_30k'
+    plt.sca(ax)
+    experiments = ['pretrain_0k', pretrain_experiments]
+
+    agents = [
+        'markov',
+        'inv-only',
+        'contr-only',
+        'autoenc',
+        'truestate',
+        'end-to-end',
+        'pixel-pred',
+        # 'random',
+        # 'rearrange_xy',
+    ]
+    root = 'results/scores/'
+    unfiltered_paths = [(root + e + '/' + a + '/', (e, a)) for e in experiments for a in agents]
+    experiments = [experiment for path, experiment in unfiltered_paths if os.path.exists(path)]
+    paths = [path for path, _ in unfiltered_paths if os.path.exists(path)]
+    labels = ['tag', 'features']
+    data = pd.concat([load_experiment(p) for p in paths],
+                    join='outer',
+                    keys=(experiments),
+                    names=labels).reset_index(level=list(range(len(labels))))
+
+    def plot(data, x, y, hue, style, col=None):
+        print("Plotting using hue={hue}, style={style}".format(hue=hue, style=style))
+        assert not data.empty, "DataFrame is empty, please check query"
+
+        # print(data.query('episode==99').groupby('agent', as_index=False)['total_reward'].mean())
+        # print(data.query('episode==99').groupby('agent', as_index=False)['total_reward'].std())
+
+        data = data.replace('markov', 'Markov')
+        data = data.replace('end-to-end', 'visual')
+        data = data.replace('truestate', 'xy-position')
+
+        print(data.groupby('agent', as_index=False)['reward'].mean())
+        print(data.groupby('agent', as_index=False)['reward'].std())
+
+        # If asking for multiple envs, use facetgrid and adjust height
+        height = 4 if col is not None and len(data[col].unique()) > 1 else 5
+        if col:
+            col_wrap = 2 if len(data[col].unique()) > 1 else 1
+        else:
+            col_wrap = None
+
+        # data = data[data['episode'] < 97]
+
+        dashes = {
+            'Markov': '',
+            'inv-only': (1, 1),
+            'contr-only': (1, 2, 5, 2),
+            'autoenc': (2, 2, 1, 2),
+            'visual': (5, 2, 5, 2),
+            'xy-position': (7, 2, 3, 2),
+            'pixel-pred': (7, 1, 1, 1),
+            'random': (1, 2, 3, 2),
+        }
+        algs = [
+            'Markov',
+            'autoenc',
+            'inv-only',
+            'pixel-pred',
+            'contr-only',
+            'visual',
+            'xy-position',
+            'random',
+        ]
+        labels = [
+            'Markov',
+            'Autoenc',
+            'Inverse',
+            'Pixel-Pred',
+            'Ratio',
+            'Visual',
+            'Expert (x,y)',
+            'Random',
+        ]
+        colormap = [
+            'Markov',
+            'inv-only',
+            'autoenc',
+            'visual',
+            'contr-only',
+            'xy-position',
+            'pixel-pred',
+        ]
+        p = sns.color_palette('Set1', n_colors=2)
+        red, _ = p
+
+        p = sns.color_palette('Set1', n_colors=9, desat=0.5)
+        _, blue, green, purple, orange, yellow, brown, pink, gray = p
+
+        palette = [red, blue, brown, purple, orange, yellow, pink]
+        palette = dict(zip(colormap, palette))
+        palette['random'] = gray
+        data = data.append({'agent': 'random', 'reward': -84.8, 'seed': 0, 'episode': 0},
+                        ignore_index=True)# yapf: disable
+
+        g = sns.lineplot(
+            ax=ax,
+            x=x,
+            y=y,
+            data=data,
+            hue=hue,
+            hue_order=algs,
+            style=style,
+            # kind='line',
+            # legend='full',
+            legend=False,
+            dashes=dashes,
+            # height=height,
+            # aspect=1.2,
+            # col=col,
+            # col_wrap=col_wrap,
+            # col_order=col_order,
+            palette=palette,
+            linewidth=2,
+            # facet_kws={
+                # 'sharey': False,
+                # 'sharex': False
+            # }
+            )
+
+        # g.set_titles('{col_name}')
+
+        # ax = g.axes.flat[0]
+        ax.set_ylim((-90, 0))
+        ax.set_xlim((0, 100))
+        ax.axhline(-84.8, dashes=dashes['random'], color=palette['random'], linewidth=2)
+        if pretrain_experiments == 'pretrain_3k':
+            leg = fig.legend(labels, loc='upper center', ncol=4, bbox_to_anchor=(0.5, 0.15), fontsize=16, frameon=False)
+            leg.set_draggable(True)
+        for axis in ['bottom','left']:
+            ax.spines[axis].set_linewidth(2)
+        for axis in ['top','right']:
+            ax.spines[axis].set_visible(False)
+        ax.tick_params(width=2)
+        ax.set_ylabel('Reward',fontsize=18)
+        ax.set_xlabel('Episode',fontsize=18)
+        ax.tick_params(labelsize=16)
+        plt.tight_layout()
+        plt.subplots_adjust(bottom=0.25)
+
+    plot(data, x='episode', y='reward', hue='agent', style='agent')
+plt.show()