Table 6: Comparing the fine-tuning performance when using different pairs of augmentation for pretraining.

import itertools

import pandas as pd

RENAME = {
    "colorjitter": "Color jitter",
    "timeshift": "Time shift",
    "changertt": "Change RTT",
    "rotate": "Rotate",
    "packetloss": "Packet loss",
}

df = pd.read_csv(
    "./campaigns/ucdavis-icdm19/simclr-other-augmentation-pairs/campaign_summary/simclr-other-augmentation-pairs/summary_flowpic_dim_32.csv",
    header=[0, 1],
    index_col=[0, 1],
)

df = df["acc"][["mean", "ci95"]].round(2)
df = df.reset_index()
df = df.assign(
    aug1=df["level_1"].apply(eval).str[0],
    aug2=df["level_1"].apply(eval).str[1],
)
df = df.drop("level_1", axis=1)
df = df.rename({"level_0": "test_split_name"}, axis=1)
df = df.replace(RENAME)
df = df.pivot(index="test_split_name", columns=["aug1", "aug2"])
df.columns.set_names(["stat", "aug1", "aug2"], inplace=True)
df = df.reorder_levels(["aug1", "aug2", "stat"], axis=1)
df.columns.set_names(["", "", ""], inplace=True)
df.index.name = None

df = df[
    list(itertools.product(["Change RTT"], ["Time shift"], ["mean", "ci95"]))
    + list(
        itertools.product(["Packet loss"], ["Color jitter", "Rotate"], ["mean", "ci95"])
    )
    + list(
        itertools.product(["Change RTT"], ["Color jitter", "Rotate"], ["mean", "ci95"])
    )
    + list(itertools.product(["Color jitter"], ["Rotate"], ["mean", "ci95"]))
]
df = df.loc[["test-script", "test-human"]]

df.to_csv("table5_simclr_other_augmentation_pairs.csv")
df

	Change RTT		Packet loss				Change RTT				Color jitter
	Time shift		Color jitter		Rotate		Color jitter		Rotate		Rotate
	mean	ci95	mean	ci95	mean	ci95	mean	ci95	mean	ci95	mean	ci95
test-script	92.18	0.31	90.17	0.41	91.94	0.3	91.72	0.36	92.38	0.32	91.79	0.34
test-human	74.69	1.13	73.67	1.24	71.22	1.2	75.56	1.23	74.33	1.26	71.64	1.23