Datasets:

Languages:
English
Size:
< 1K
ArXiv:
Libraries:
Datasets
License:
relentless / experiments /plot_fewshot_landscape.py
asahi417's picture
fix the logit overflow caused by pad_token https://github.com/asahi417/lmppl/issues/5
4a89ca9
raw
history blame contribute delete
No virus
2.44 kB
import os
import matplotlib as mpl
import matplotlib.pyplot as plt
import pandas as pd
from random import shuffle, seed
maps = {"competitor/rival of": "Rival", "friend/ally of": "Ally", "influenced by": "Inf", "known for": "Know", "similar to": "Sim", "average": "Avg"}
# Few-shots + Zero-shot
os.makedirs('figures/fewshots', exist_ok=True)
plt.rcParams.update({'font.size': 16}) # must set in top
# styles = ['o-', 'o--', 'o:', 's-', 's--', 's:', '^-', '^--', '^:', "X-", "X--", "X:"]
styles = ['o', "X", '^', 'P']
seed(1)
colors = list(mpl.colormaps['tab20b'].colors)
shuffle(colors)
for prompt in ['qa', 'lc']:
df = pd.concat([
pd.read_csv(f"results/lm_{prompt}_zeroshot.csv", index_col=0),
pd.read_csv(f"results/lm_{prompt}_fewshots.csv", index_col=0)])
df_full = pd.read_csv(f"results/lm_{prompt}/lm.csv", index_col=0)
for r in maps:
tmp = df[[r, "shot", "seed"]]
tmp[r] = tmp[r] * 100
ax = None
for n, m in enumerate(['Flan-T5\textsubscript{XXL}', 'Flan-UL2', 'OPT\textsubscript{13B}', 'GPT-3\textsubscript{davinci}']):
g = tmp.loc[m]
full_shot = df_full[[r]].loc[[m]] * 100
full_shot["shot"] = 5
full_shot["seed"] = 0
g = pd.concat([g, full_shot])
if "OPT" in m:
# m = "OPT (13B)"
m = "OPT"
if "Flan-T5" in m:
# m = "Flan-T5 (XXL)"
m = "Flan-T5"
if "GPT-3" in m:
# m = "GPT-3 (davinci)"
m = "GPT-3"
ax = g.plot.line(y=r,
x='shot',
# xlabel='Number of Prototypical Examples',
# xlabel='Num. of Proto. Examples',
# ylabel="Correlation",
xlabel="",
ylabel="",
ax=ax,
ms=8,
color=colors[n],
style=styles[n],
label=m,
# figsize=(4, 5),
grid=True)
ax.set_xticks([0, 1, 3, 5])
ax.legend(loc='lower right')
plt.tight_layout()
plt.savefig(f"figures/fewshots/{prompt}.{r.replace(' ', '_').replace('/', '-')}.fewshot.landscape.png", bbox_inches="tight", dpi=600)