Comparing Two Learning Curves Side-by-side

Here we compare the training performance versus the performance on the evaluation domain.

We show the training performance in gray, to accentuate the evaluation curve.

Initialize the loader

loader = ML_Logger(root=os.getcwd(), prefix="data/walker-walk/curl")

Check all the files

files = loader.glob(query="**/metrics.pkl", wd=".", recursive=True)
doc.print(files)

['300/metrics.pkl', '400/metrics.pkl', '100/metrics.pkl', '200/metrics.pkl']

A Single Time Series

def group(xKey="step", yKey="train/episode_reward/mean", color=None, bin=10, label=None, dropna=False):
    avg, top, bottom, step = loader.read_metrics(f"{yKey}@mean", f"{yKey}@84%", f"{yKey}@16%", x_key=f"{xKey}@mean",
                                                 path="**/metrics.pkl", bin_size=bin, dropna=dropna)
    plt.plot(step, avg, color=color, label=label)
    plt.fill_between(step, bottom, top, alpha=0.15, color=color)
    return avg

Step 2: Plot

colors = ['#49b8ff', '#444444', '#ff7575', '#66c56c', '#f4b247']

avg = group(yKey="episode_reward/mean", bin=None, color=colors[0], label="Eval")
group(yKey="train/episode_reward/mean", color=colors[1], label="Train")

plt.gca().xaxis.set_major_formatter(ticker.FuncFormatter(lambda x, _: f"{int(x / 1000)}k" if x else "0"))
plt.legend()
plt.title("Walker-walk")
plt.xlabel("Steps")
plt.ylabel("Return")

r.savefig(f"figures/train_vs_eval.png", title="Train VS Eval", dpi=300, zoom="20%")
plt.close()

Train VS Eval

Where does the empty cuts come from?

These cuts are places where the avg is NaN. You can just filter this out in the group function.

doc.print(avg)

step
(-0.001, 5098.039]           20.341932
(5098.039, 10196.078]       159.512312
(10196.078, 20000.0]        296.704378
(20000.0, 30000.0]          343.709928
(30000.0, 40000.0]          440.871364
(40000.0, 50000.0]          440.290806
(50000.0, 60000.0]          464.523289
(60000.0, 70000.0]          512.451430
(70000.0, 75882.353]               NaN
(75882.353, 80980.392]      534.083886
(80980.392, 90000.0]        566.964033
(90000.0, 100000.0]         540.526823
(100000.0, 110000.0]        562.481581
(110000.0, 120000.0]        582.455365
(120000.0, 130000.0]        597.723035
(130000.0, 140000.0]        631.353935
(140000.0, 146666.667]             NaN
(146666.667, 151764.706]    576.559410
(151764.706, 160000.0]      587.484037
(160000.0, 170000.0]        622.087906
(170000.0, 180000.0]        607.022307
(180000.0, 190000.0]        610.693518
(190000.0, 200000.0]        614.979733
(200000.0, 210000.0]        627.987848
(210000.0, 217450.98]              NaN
(217450.98, 222549.02]      621.534480
(222549.02, 230000.0]       635.705141
(230000.0, 240000.0]        594.198112
(240000.0, 250000.0]        638.786110
(250000.0, 260000.0]        666.134860
(260000.0, 270000.0]        637.613542
(270000.0, 280000.0]        697.436256
(280000.0, 290000.0]        642.128056
(290000.0, 303333.333]      636.789103
(303333.333, 318431.373]    637.653942
(318431.373, 330000.0]      660.905099
(330000.0, 340000.0]        673.678899
(340000.0, 350000.0]        645.906549
(350000.0, 360000.0]        627.167255
(360000.0, 373921.569]      696.372121
(373921.569, 389019.608]    654.296411
(389019.608, 400000.0]      662.388210
(400000.0, 410000.0]        706.604360
(410000.0, 420000.0]        675.908748
(420000.0, 430000.0]        682.392554
(430000.0, 444509.804]      647.357963
(444509.804, 459607.843]    675.610188
(459607.843, 470000.0]      682.014805
(470000.0, 480000.0]        676.036846
(480000.0, 490000.0]        688.063507
(490000.0, 500000.0]        682.238766
Name: episode_reward/mean@mean, dtype: float64

avg = group(yKey="episode_reward/mean", bin=None, color=colors[0], label="Eval", dropna=True)

plt.gca().xaxis.set_major_formatter(ticker.FuncFormatter(lambda x, _: f"{int(x / 1000)}k" if x else "0"))
plt.legend()
plt.title("Walker-walk")
plt.xlabel("Steps")
plt.ylabel("Return")

r.savefig(f"figures/train_vs_eval_dropna.png", title="Train VS Eval", dpi=300, zoom="20%")

Train VS Eval