File size: 6,467 Bytes
24c2665 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 |
# for the same model, we first study which prompting strategy is better
import json
import os
import re
from collections import defaultdict
from statistics import mean
import matplotlib.pyplot as plt
from matplotlib.colors import LinearSegmentedColormap
from rich import print
def normalize_to_base_name(name: str) -> str:
name = (
name.replace("Chat", "")
.replace("instruct", "")
.replace("code-llama", "CodeLlama")
.replace("deepseek-coder", "DeepSeek-Coder")
.replace("gpt-4-turbo", "GPT-4-Turbo")
.replace("starcoder", "StarCoder")
.replace("--v0.1", "")
.replace("-base", "")
.replace("-preview", "")
.strip("-")
)
# ${NUM}b -> ${NUM}B
return re.sub(r"(\d+)b", r"\1B", name)
def load_dps_scores(path: str, norm: bool = False):
with open(path) as f:
results = json.load(f)
task2score = {}
for task_id, result in results.items():
# new rpr
new_key = "norm_scores" if norm else "scores"
if result.get(new_key) is not None:
task2score[task_id] = result[new_key]["avg"]
# legacy rpr
legacy_key = "dps_norm" if norm else "dps"
if result.get(legacy_key) is not None:
task2score[task_id] = mean(result[legacy_key])
return task2score
# EvalPerf result file name format: {MODEL}_temp_{TEMP}_ep_{TYPE}_results.json
def parse_model_and_type(result_json: str):
assert "_temp_0.2_" in result_json, f"Invalid result file name: {result_json}"
model_id, rest = result_json.split("_temp_0.2_")
type = rest.split("_")[1]
model_id = normalize_to_base_name(model_id)
# extra 6.7 in "DSCoder-6.7B" and "7B" in "DSCoder-7B"
nb = re.search(r"(\d+(?:\.\d+)?)B", model_id)
if nb:
print(nb)
nb = nb.group(1)
model_id = model_id.replace(f"{nb}B", "").strip("-")
else:
nb = None
return model_id, nb, type
def load_groups_from_directory(result_dir: str, norm: bool = False):
groups = defaultdict(dict) # model -> {type: dps(norm)}
for result_json in os.listdir(result_dir):
if not result_json.endswith(".json"):
continue
model_id, nb, type = parse_model_and_type(result_json)
print(f"{type = :<16}\t{model_id = }")
if nb:
model_id = f"{model_id} ({nb}B)"
groups[model_id][type] = load_dps_scores(
os.path.join(result_dir, result_json), norm
)
return groups
TYPES = ["perf-CoT", "perf-instruct", "instruct", "base"]
def compute_score_matrix(group: dict):
score_matrix = []
for i, type_x in enumerate(TYPES):
score_list = []
for j, type_y in enumerate(TYPES):
if j <= i or type_y not in group or type_x not in group:
score_list.append((0, 0))
continue
task2dps_x = group[type_x]
task2dps_y = group[type_y]
common_tasks = set(task2dps_x.keys()) & set(task2dps_y.keys())
if not common_tasks:
score_list.append(None)
print(f"No common tasks between {type_x} and {type_y}")
continue
dps_x = mean([task2dps_x[task_id] for task_id in common_tasks])
dps_y = mean([task2dps_y[task_id] for task_id in common_tasks])
print(type_x, dps_x, " --- ", type_y, dps_y)
score_list.append((dps_x, dps_y))
score_matrix.append(score_list)
return score_matrix
def main(result_dir: str, norm: bool = False, latex: bool = False):
if latex:
plt.rc("text", usetex=True)
plt.rc("text.latex", preamble=r"\usepackage{xfrac}")
assert os.path.isdir(result_dir), f"{result_dir} is not a directory."
groups = load_groups_from_directory(result_dir, norm=norm)
groups = {k: v for k, v in groups.items() if len(v) >= 2}
# resort by key
groups = dict(sorted(groups.items()))
n_grp = len(groups)
max_grp_per_row = 5
n_row = (n_grp + max_grp_per_row - 1) // max_grp_per_row
fig, axs = plt.subplots(
n_row,
max_grp_per_row,
figsize=(2 * max_grp_per_row, 2 * n_row),
constrained_layout=True,
)
for i, (model, group) in enumerate(groups.items()):
score_matrix = compute_score_matrix(group)
score_matrix_diff = [
[(score[0] - score[1]) for score in score_list]
for score_list in score_matrix
]
ax: plt.Axes = axs[i // max_grp_per_row, i % max_grp_per_row]
cmap = LinearSegmentedColormap.from_list("rg", ["r", "w", "lime"], N=256)
# ax.matshow(score_matrix_diff, cmap=cmap)
cax = ax.matshow(score_matrix_diff, cmap=cmap)
cax.set_clim(-15, 15)
ax.set_xticks(range(len(TYPES)))
ax.set_yticks(range(len(TYPES)))
if i // max_grp_per_row == 0:
ax.set_xticklabels(TYPES, rotation=30, ha="left", rotation_mode="anchor")
else:
ax.set_xticklabels([])
ax.tick_params(top=False)
if i % max_grp_per_row == 0:
ax.set_yticklabels(TYPES)
else:
ax.set_yticklabels([])
ax.tick_params(left=False)
ax.tick_params(bottom=False)
for i in range(len(TYPES)):
for j in range(len(TYPES)):
if j <= i:
continue
x, y = score_matrix[i][j] # i~x, j~y
if x == 0 and y == 0:
continue
gapx = 0.15
gapy = 0.25
ax.text(
j - gapx,
i + gapy,
f"{x:.1f}",
va="center",
ha="center",
color="green" if x > y else "red",
)
ax.text(
j + gapx,
i - gapy,
f"{y:.1f}",
va="center",
ha="center",
color="green" if x < y else "red",
)
xlabel = model
if latex:
xlabel = r"\textbf{" + xlabel + "}"
ax.set_xlabel(xlabel)
imname = "perf_prompt_impact"
if norm:
imname += "_norm"
plt.savefig(f"{imname}.png", dpi=100, bbox_inches="tight")
plt.savefig(f"{imname}.pdf", dpi=100, bbox_inches="tight")
if __name__ == "__main__":
from fire import Fire
Fire(main)
|