Benchmark Figures on Selected Tasks

Setup

import ast
import numpy as np
import matplotlib.pyplot as plt
import seaborn as sns

%matplotlib inline

# Define color palette
sns.set_theme(
    color_codes=True, palette="bright", style="ticks", context="talk", font_scale=1.5
)

def load_result(filename):
    """
    Loads results from specified file
    """
    inputs = open(filename, "r")
    lines = inputs.readlines()
    ls = []
    for line in lines:
        ls.append(ast.literal_eval(line))
    return ls


def plot_acc(col, ls, pos, n_train, lw=5):
    if pos == 0:
        for i, l in enumerate(ls[pos]):
            col.plot(
                n_train,
                np.mean(l, axis=0),
                label=legends[i],
                lw=lw,
                color=colors[i],
                linestyle=styles[i],
            )
            qunatiles = np.nanquantile(l, [0.25, 0.75], axis=0)
            col.fill_between(
                n_train,
                qunatiles[0],
                qunatiles[1],
                lw=1,
                facecolor=colors[i],
                linestyle=styles[i],
                alpha=0.3,
            )
    else:
        for i, l in enumerate(ls[pos]):
            col.plot(
                n_train,
                np.mean(l, axis=0),
                lw=lw,
                color=colors[i],
                linestyle=styles[i],
            )
            qunatiles = np.nanquantile(l, [0.25, 0.75], axis=0)
            col.fill_between(
                n_train,
                qunatiles[0],
                qunatiles[1],
                lw=1,
                facecolor=colors[i],
                linestyle=styles[i],
                alpha=0.3,
            )

directory = "../benchmarks/results/"
datasets = ["splice", "pendigits", "cifar10"]
ranges = [23, 74, 500]

Accuracy Plot for Streaming Classifiers

prefixes = ["sdf/", "sdt/", "ht/", "mf/"]
legends = ["SDF", "SDT", "HT", "MF"]
colors = ["r", "r", "g", "y"]
styles = ["-", "--", "--", "-"]

acc_ls = []
for i, dataset in enumerate(datasets):
    acc_l = []

    for prefix in prefixes:
        acc = load_result(directory + prefix + dataset + "_acc.txt")[:10]
        acc_l.append(acc)

    acc_ls.append(acc_l)

fig, ax = plt.subplots(nrows=1, ncols=3, figsize=(17, 6), constrained_layout=True)

fig.text(0.53, -0.07, "Number of Train Samples", ha="center")
xtitles = ["Splice", "Pendigits", "CIFAR-10"]
ytitles = ["Accuracy"]
ylimits = [[0, 1]]
yticks = [[0, 0.5, 1]]

for i, col in enumerate(ax):
    col.set_xscale("log")
    col.set_ylim(ylimits[0])

    n_train = range(100, (ranges[i] + 1) * 100, 100)

    # Label x axis and plot figures
    col.set_title(xtitles[i])
    plot_acc(col, acc_ls, i, n_train)

    # Label y axis
    if i % 3 == 0:
        col.set_yticks(yticks[0])
        col.set_ylabel(ytitles[i])
    else:
        col.set_yticks([])

fig.align_ylabels(
    ax[
        :,
    ]
)

leg = fig.legend(
    bbox_to_anchor=(0.53, -0.22),
    bbox_transform=plt.gcf().transFigure,
    ncol=6,
    loc="lower center",
)
leg.get_frame().set_linewidth(0.0)
for legobj in leg.legendHandles:
    legobj.set_linewidth(5.0)
plt.savefig("../paper/select_acc_stream.pdf", transparent=True, bbox_inches="tight")

Accuracy Plot for Batch Classifiers

prefixes = ["sdf/", "rf/", "sdt/", "dt/"]
legends = ["SDF", "DF", "SDT", "DT"]
colors = ["r", "b", "r", "b"]
styles = ["-", "-", "--", "--"]

acc_ls = []
for i, dataset in enumerate(datasets):
    acc_l = []

    for prefix in prefixes:
        acc = load_result(directory + prefix + dataset + "_acc.txt")[:10]
        acc_l.append(acc)

    acc_ls.append(acc_l)

fig, ax = plt.subplots(nrows=1, ncols=3, figsize=(17, 6), constrained_layout=True)

fig.text(0.53, -0.07, "Number of Train Samples", ha="center")
xtitles = ["Splice", "Pendigits", "CIFAR-10"]
ytitles = ["Accuracy"]
ylimits = [[0, 1]]
yticks = [[0, 0.5, 1]]

for i, col in enumerate(ax):
    col.set_xscale("log")
    col.set_ylim(ylimits[0])

    n_train = range(100, (ranges[i] + 1) * 100, 100)

    # Label x axis and plot figures
    col.set_title(xtitles[i])
    plot_acc(col, acc_ls, i, n_train)

    # Label y axis
    if i % 3 == 0:
        col.set_yticks(yticks[0])
        col.set_ylabel(ytitles[i])
    else:
        col.set_yticks([])

fig.align_ylabels(
    ax[
        :,
    ]
)

leg = fig.legend(
    bbox_to_anchor=(0.53, -0.22),
    bbox_transform=plt.gcf().transFigure,
    ncol=6,
    loc="lower center",
)
leg.get_frame().set_linewidth(0.0)
for legobj in leg.legendHandles:
    legobj.set_linewidth(5.0)
plt.savefig("../paper/select_acc_batch.pdf", transparent=True, bbox_inches="tight")

Time Plot

prefixes = ["sdf/", "rf/", "sdt/", "dt/", "ht/", "mf/"]
legends = ["SDF", "DF", "SDT", "DT", "HT", "MF"]
colors = ["r", "b", "r", "b", "g", "y"]
styles = ["-", "-", "--", "--", "--", "-"]

# Show concatenated mem for batch estimators
concat = True

time_ls = []
for i, dataset in enumerate(datasets):
    time_l = []

    for prefix in prefixes:
        time = load_result(directory + prefix + dataset + "_train_t.txt")[:10]

        if concat and (prefix == "dt/" or prefix == "rf/"):
            for t in time:
                for j in range(1, ranges[i]):
                    t[j] += t[j - 1]

        time_l.append(time)

    time_ls.append(time_l)

fig, ax = plt.subplots(nrows=1, ncols=3, figsize=(17, 6), constrained_layout=True)

fig.text(0.53, -0.07, "Number of Train Samples", ha="center")
xtitles = ["Splice", "Pendigits", "CIFAR-10"]
ytitles = ["Wall Time (s)"]
ylimits = [[1e-4, 1e5]]
yticks = [[1e-4, 1e-1, 1e2, 1e5]]

for i, col in enumerate(ax):
    col.set_xscale("log")
    col.set_yscale("log")
    col.set_ylim(ylimits[0])

    n_train = range(100, (ranges[i] + 1) * 100, 100)

    # Label x axis and plot figures
    col.set_title(xtitles[i])
    plot_acc(col, time_ls, i, n_train)

    # Label y axis
    if i % 3 == 0:
        col.set_yticks(yticks[0])
        col.set_ylabel(ytitles[i])
    else:
        col.set_yticks([])

fig.align_ylabels(
    ax[
        :,
    ]
)

leg = fig.legend(
    bbox_to_anchor=(0.53, -0.22),
    bbox_transform=plt.gcf().transFigure,
    ncol=6,
    loc="lower center",
)
leg.get_frame().set_linewidth(0.0)
for legobj in leg.legendHandles:
    legobj.set_linewidth(5.0)
plt.savefig("../paper/select_time.pdf", transparent=True, bbox_inches="tight")

Memory Plot

# Reorder plots
prefixes = ["sdf/", "sdt/", "ht/", "mf/"]
legends = ["SDF", "SDT", "HT", "MF"]
colors = ["r", "r", "g", "y"]
styles = ["-", "--", "--", "-"]

# Load CIFAR-10 memory records
cifar_data_mem = 8.0 * 56 / 100 * np.ones(len(n_train))
cifar_mem_l = []
for prefix in prefixes:
    cifar_mem = (
        np.mean(load_result(directory + prefix + dataset + "_v_m.txt")[:1], axis=0)
        * 56
        / 100
        #         - cifar_data_mem
    )
    cifar_mem_l.append([cifar_mem])

fig, ax = plt.subplots(nrows=1, ncols=1, figsize=(7, 7), constrained_layout=True)

xtitle = "Number of Train Samples"
ytitle = "Max Virtual Memory (GB)"
ylimit = [0, 32]
ytick = [0, 8, 16, 24, 32]

ax.set_xscale("log")
ax.set_ylim(ylimit)

n_train = range(100, (ranges[2] + 1) * 100, 100)

# Plot dataset size
ax.plot(n_train, cifar_data_mem, label="dataset", color="k", linestyle="dotted", lw=3)

# Label x axis and plot figures
plot_acc(ax, [cifar_mem_l], 0, n_train, lw=3)
ax.set_xlabel(xtitle)

# Label y axis
ax.set_yticks(ytick)
ax.set_ylabel(ytitle)

# Legend
handles, labels = ax.get_legend_handles_labels()
ax.legend(handles[::-1], labels[::-1], frameon=False)

plt.savefig(
    "../paper/select_cifar_mem_stream.pdf", transparent=True, bbox_inches="tight"
)