def plot_scal():
import tex
tex.setup(
width=1,
height=None,
span=False,
l=0.15,
r=0.98,
t=0.98,
b=0.17,
params={"hatch.linewidth": 0.5},
)
per_frame_delay_arr = list()
for i in range(MAX_COTAINER_NUM):
tmp_arr_1 = list()
for r in range(NUM_RUN):
tmp_arr_2 = list()
for j in range(i + 1):
csv_path = os.path.join(
CSV_FILE_DIR,
"%d_%d_%d_yolo_v2_pp_delay.csv" % (i + 1, j + 1, r + 1),
)
with open(csv_path, "r") as f:
total = f.readlines()[-1].strip().split(":")[1]
tmp_arr_2.append(float(total))
tmp_arr_1.append(max(tmp_arr_2) / ((i + 1) * IMAGE_NUM))
per_frame_delay_arr.append(tmp_arr_1[:])
print(per_frame_delay_arr)
fig, ax = plt.subplots()
x = range(MAX_COTAINER_NUM)
x_labels = [a + 1 for a in x]
y = [np.average(a) for a in per_frame_delay_arr]
yerr = [np.std(a) for a in per_frame_delay_arr]
ax.bar(
x,
y,
0.3,
color="blue",
lw=0.6,
yerr=yerr,
ecolor="red",
error_kw={"elinewidth": 1},
alpha=0.8,
edgecolor="black",
)
ax.set_xticks(x)
ax.set_xticklabels(x_labels)
ax.set_ylabel("Average Inference Delay (ms)")
ax.set_xlabel("Number of containers")
tex.save_fig(fig, "./scal_yolov2_pp", fmt="png")
plt.close(fig)
def plot_shared_x():
"""Maybe look nicer"""
import tex
tex.setup(width=1,
height=None,
span=False,
l=0.15,
r=0.98,
t=0.98,
b=0.17,
params={'hatch.linewidth': 0.5})
cmap = plt.get_cmap("tab10")
for m in MODELS:
fig, ax_arr = plt.subplots(2, sharex=True)
ax_arr[0].set_ylabel("Output Size (Bytes)", fontsize=5)
ax_arr[1].set_ylabel("Inference Latency (ms)", fontsize=5)
x, x_labels, out_s = plot_io_sizes(m, just_return=True)
ax_arr[0].bar(x,
out_s,
BAR_WIDTH,
color=cmap(0),
lw=0.6,
hatch="xxx",
alpha=0.8,
edgecolor='black')
ax_arr[0].axhline(y=(604 * 604 * 3),
ls="--",
color="green",
label="Image Size")
x, delay_avg, delay_err = plot_layer_delay(m, just_return=True)
ax_arr[1].bar(x,
delay_avg,
BAR_WIDTH,
color=cmap(1),
lw=0.6,
bottom=0,
hatch="+++",
yerr=delay_err,
ecolor='red',
error_kw={"elinewidth": 1},
alpha=0.8,
edgecolor='black')
ax_arr[1].axhline(y=1000, ls="--", color="green")
for ax in ax_arr:
ax.set_ylim(0)
handles, labels = ax_arr[0].get_legend_handles_labels()
ax_arr[0].legend(handles, labels, loc='best')
ax_arr[0].autoscale_view()
ax_arr[1].set_xticks(x)
ax_arr[1].set_xticklabels(x_labels, rotation="vertical")
fig.align_ylabels(ax_arr)
tex.save_fig(fig, "./combined_%s" % m, fmt="png")
def plot_io_sizes(m, just_return=False):
import tex
tex.setup(
width=1,
height=None,
span=False,
l=0.15,
r=0.98,
t=0.98,
b=0.17,
params={"hatch.linewidth": 0.5},
)
# format: (idx, name, input, output)
label_gen = LABEL_GEN()
data = list()
with open("./net_info_%s.csv" % m, "r") as f:
net_infos = f.readlines()
for info in net_infos:
fields = list(map(str.strip, info.split(",")))
fields = list(map(int, info.split(",")))
data.append((fields[0], label_gen.get_label(fields[1]), fields[2],
fields[3]))
x = [i[0] for i in data]
x_labels = [i[1] for i in data]
out_s = [i[3] for i in data]
if just_return:
return [x, x_labels, out_s]
else:
# Plot io_sizes in separate figures
fig, ax = plt.subplots()
ax.bar(
x,
out_s,
BAR_WIDTH,
color="blue",
lw=0.6,
alpha=0.8,
edgecolor="black",
label="Output Size",
)
# TODO: Improve the looking
ax.set_xticks(x)
ax.set_title("%s" % m)
ax.set_xticklabels(x_labels, rotation="vertical")
ax.set_ylim(0, 1.2 * 10**7)
ax.set_ylabel("Output Size (Bytes)")
tex.save_fig(fig, "./net_info_%s" % m, fmt="png")
plt.close(fig)
def plot_scal():
import tex
tex.setup(width=1,
height=None,
span=False,
l=0.15,
r=0.98,
t=0.98,
b=0.17,
params={'hatch.linewidth': 0.5})
IMAGE_NUM = 25
max_cotainer_num = 3
num_run = 2
per_frame_delay_arr = list()
for i in range(max_cotainer_num):
tmp_arr_1 = list()
for r in range(num_run):
tmp_arr_2 = list()
for j in range(i + 1):
with open(
"./%d_%d_%d_yolo_v2_pp_delay.csv" %
(i + 1, j + 1, r + 1), "r") as f:
total = f.readlines()[-1].strip().split(":")[1]
tmp_arr_2.append(float(total))
tmp_arr_1.append(max(tmp_arr_2) / ((i + 1) * IMAGE_NUM))
per_frame_delay_arr.append(tmp_arr_1[:])
print(per_frame_delay_arr)
fig, ax = plt.subplots()
x = range(max_cotainer_num)
x_labels = [a + 1 for a in x]
y = [np.average(a) for a in per_frame_delay_arr]
yerr = [np.std(a) for a in per_frame_delay_arr]
ax.bar(x,
y,
0.3,
color='blue',
lw=0.6,
yerr=yerr,
ecolor='red',
error_kw={"elinewidth": 1},
alpha=0.8,
edgecolor='black')
ax.set_xticks(x)
ax.set_xticklabels(x_labels)
ax.set_ylabel("Average Inference Delay (ms)")
ax.set_xlabel("Number of containers")
tex.save_fig(fig, "./scal_yolov2_pp", fmt="png")
plt.close(fig)
def plot_layer_delay(m, just_return=False):
"""TODO: Plot pdf instead of average"""
import tex
tex.setup(
width=1,
height=None,
span=False,
l=0.15,
r=0.98,
t=0.98,
b=0.17,
params={"hatch.linewidth": 0.5},
)
label_gen = LABEL_GEN()
data = list()
with open("./per_layer_delay_%s.csv" % m, "r") as f:
lines = f.readlines()
for l in lines:
fields = list(map(str.strip, l.split(",")))
fields = list(map(float, fields))
data.append((int(fields[1]), fields[2], fields[3], fields[0]))
# Avoid duplicated label
idx_layer_map = dict()
layer_len = len(set([int(x[0]) for x in data]))
x_labels = list()
for idx in range(layer_len):
label = label_gen.get_label(data[idx][1])
idx_layer_map[idx] = label
x_labels.append(label)
delay = dict.fromkeys(x_labels)
tl = list()
for k in delay.keys():
delay[k] = copy.deepcopy(tl)
# Calculate delay
for d in data:
delay[idx_layer_map[d[0]]].append(d[2])
delay_avg = list()
delay_err = list()
for l in x_labels:
delay_avg.append(np.average(delay[l]))
delay_err.append(np.std(delay[l]))
x = list(set([int(i[0]) for i in data]))
if just_return:
return [x, delay_avg, delay_err]
else:
fig, ax = plt.subplots()
ax.bar(
x,
delay_avg,
BAR_WIDTH,
color="blue",
lw=0.6,
yerr=delay_err,
ecolor="red",
error_kw={"elinewidth": 1},
alpha=0.8,
edgecolor="black",
label="Latency",
)
ax.set_xticks(x)
ax.set_xticklabels(x_labels, rotation="vertical")
ax.set_ylabel("Inference Latency (ms)")
ax.set_ylim(0)
tex.save_fig(fig, "./per_layer_delay_%s" % m, fmt="png")
plt.close(fig)
}
]
for l in lines:
l["y"] = np.genfromtxt(l["file"], delimiter=",").flatten()
l["y"] = l["y"][~np.isnan(l["y"])]
l["avg"] = np.average(l["y"])
l["std"] = np.std(l["y"])
for l in lines:
y = l["y"]
ax.plot(
x,
np.percentile(y, x, interpolation="nearest"),
label=l["label"],
linestyle=l["ls"],
color=l["color"],
)
ax.set(
title="Latency of YOLO Pre-processing VNF",
xlabel="Percentile",
ylabel="Latency (ms)",
)
handles, labels = ax.get_legend_handles_labels()
ax.legend(handles, labels, loc="upper left")
tex.save_fig(fig, "latency_percentile_50_100", "pdf")
print(lines)