-
Notifications
You must be signed in to change notification settings - Fork 3
/
glue_performance.py
185 lines (150 loc) · 6.52 KB
/
glue_performance.py
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
import os
import sys
import json
import matplotlib.pyplot as plt
import util.nethook as nethook
from transformers import AutoModelForCausalLM
from attrdict import AttrDict
import torch
sys.path.append('/home/akshatgupta/KnowledgeEditing_local/unified-model-editing')
from useful_functions import save_data
if __name__ == '__main__':
model_name = 'gpt2-xl'#'/data/akshat/models/Llama-2-7b-hf'
model = AutoModelForCausalLM.from_pretrained(model_name)
algo = 'FT'
run = 'run_003'
hparams_filename = 'hparams/FT/gpt2-xl_unconstr.json'
#hparams_filename = 'hparams/' + algo + '/gpt2-xl.json'
f = open(hparams_filename)
hparams = AttrDict(json.load(f))
layers_edited = list(hparams.layers)
original_norms = {}
for layer_num in layers_edited:
edited_layer = nethook.get_module(model, hparams.rewrite_module_tmp.format(layer_num))
original_norms[str(layer_num)] = torch.norm(edited_layer.weight).item()
x_tick_size = 22
y_tick_size = 22
x_lim = 1000
y_lim = 100
axis_fontsize = 24
legend_fontsize = 16
metric_names = ['correct', 'f1', 'mcc', 'invalid']
task_names = ['sst', 'mmmlu', 'nli', 'rte']
glue_eval = {'distance':{}}
for task in task_names:
glue_eval[task] = {}
for metric in metric_names:
glue_eval[task][metric] = {}
save_location = 'downstream_eval/plots/' + algo + '_' + run + '/'
os.makedirs(save_location, exist_ok=True)
data_location = 'results/' + algo + '/' + run + '/glue_eval/'
for filename in os.listdir(data_location):
file_loc = data_location + filename
if 'glue' in filename:
with open(file_loc, "r") as f:
data = json.load(f)
if 'base' in filename:
edit_num = 0
else:
edit_num = data['edit_num'] + 1
#plot distance data
for layer in data['distance_from_original']:
if layer not in glue_eval['distance']:
glue_eval['distance'][layer] = {}
glue_eval['distance'][layer][0] = original_norms[layer]
glue_eval['distance'][layer][edit_num] = data['objective_distances'][layer]['new_weights_norm']
for task in task_names:
if task in data:
for metric in metric_names:
glue_eval[task][metric][int(edit_num)] = data[task][metric]
task_dict = {'sst':'SST2', 'mmmlu':'MMLU', 'rte':'COLA', 'nli':'NLI'}
run_title = {}
task_colors = {'sst':'r', 'mmmlu':'b', 'rte':'g', 'nli':'k'}
#plot metrics individual with number of edits
for metric in metric_names:
plt.figure(figsize=(6.5, 5.5))
for task in task_names:
sorted_dict = sorted(glue_eval[task][metric].items(), key=lambda item: item[0])
x, y = [], []
for edit_num, correct in sorted_dict:
x.append(edit_num)
if metric in ['f1']:
y.append(correct * 100)
else:
y.append(correct)
plt.plot(x,y, label = task_dict[task], linewidth =3, color=task_colors[task])
plt.legend(fontsize=legend_fontsize)
plt.xlabel('Number of Edits', fontsize=axis_fontsize)
if metric == 'correct':
metric = 'accuracy'
plt.ylabel(metric.upper(), fontsize=axis_fontsize)
#plt.xlim(0, x_lim)
plt.ylim(0, y_lim)
plt.tick_params(axis='x', labelsize=x_tick_size)
plt.tick_params(axis='y', labelsize=y_tick_size)
plt.tight_layout()
if run in run_title:
plt.savefig(save_location + algo + '_' + 'glue_' + metric + '_' + run_title[run] + '.png')
else:
plt.savefig(save_location + algo + '_' + 'glue_' + metric + '.png')
plt.close()
#plot distance as a function of number of edits
metric = 'distance'
x_store = []
y_store = []
for l, layer in enumerate(glue_eval[metric]):
sorted_dict = sorted(glue_eval[metric][layer].items(), key=lambda item: item[0])
x, y = [], []
for edit_num, correct in sorted_dict:
x.append(edit_num)
y.append(correct)
x_store.append(x)
y_store.append(y)
if 'transformer' in layer:
layer = layer.split('.')[2]
if l == 0:
plt.plot(x,y, linewidth =3, color = 'r', label = 'Layer ' + str(int(layer) + 1))
else:
plt.plot(x,y, linewidth =3, label = 'Layer ' + str(int(layer) + 1))
plt.legend(fontsize=legend_fontsize)
plt.xlabel('Number of Edits', fontsize=axis_fontsize)
plt.ylabel('Normalized Distance', fontsize=axis_fontsize)
#plt.ylim(0, 1000)
plt.tick_params(axis='x', labelsize=x_tick_size)
plt.tick_params(axis='y', labelsize=y_tick_size)
plt.tight_layout()
if run in run_title:
plt.savefig(save_location + algo + '_' + 'distance_' + run_title[run] + '.png')
else:
plt.savefig(save_location + algo + '_' + 'distance.png')
plt.close()
#print(len(y))
#print(y)
#ave_data(algo + sample_num + '_distance.pkl', [x_store,y_store])
#plot glue performance as a function of number of edits
#plot metrics individual with number of edits
for layer_num in original_norms:
for metric in metric_names:
plt.figure(figsize=(6.5, 5.5))
for task in task_names:
sorted_dict = sorted(glue_eval[task][metric].items(), key=lambda item: item[0])
x, y = [], []
for index, (edit_num, correct) in enumerate(sorted_dict):
x.append(glue_eval['distance'][str(layer_num)][edit_num])
if metric in ['f1', 'accuracy']:
y.append(correct * 100)
else:
y.append(correct)
plt.plot(x,y, label = task_dict[task], linewidth =3, color=task_colors[task])
plt.legend(fontsize=legend_fontsize)
plt.xlabel('Matrix Norm', fontsize=axis_fontsize)
if metric == 'correct':
metric = 'accuracy'
plt.ylabel(metric.upper(), fontsize=axis_fontsize)
#plt.xlim(100, 200)
plt.ylim(0, y_lim)
plt.tick_params(axis='x', labelsize=x_tick_size)
plt.tick_params(axis='y', labelsize=y_tick_size)
plt.tight_layout()
plt.savefig(save_location + algo + '_' + 'glue_' + metric + '_distance_' + layer_num + '.png')
plt.close()