forked from facebookresearch/visual-counterfactuals
-
Notifications
You must be signed in to change notification settings - Fork 0
/
eval_help.py
262 lines (218 loc) · 11 KB
/
eval_help.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
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
import matplotlib.pyplot as plt
from matplotlib.colors import Normalize
from scipy import stats
from functools import reduce
import numpy as np
import pandas as pd
import re
from numpy.lib.stride_tricks import sliding_window_view
def plot_study(study, name, is_resnet=False, print_others=True, print_pareto=True, map=lambda t: 0.8, other_study=None, print_colorbar=True, param_name="", split_simplify=False, labels=[""], label_loc="upper left", switch_label=False):
plt.figure(figsize=(6.4, 4.8), dpi=100)
trials = [t for t in study.trials if t.values is not None and (t.number > 5 or not is_resnet)]
norm = Normalize(0.0, np.ceil(max([map(t) for t in trials])))
scatter = None
other_scatter = None
if print_others:
scatter = plt.scatter([t.values[0] for t in trials], [t.values[1] for t in trials], c=[map(t) for t in trials], cmap='Blues', linewidths=1.0, edgecolors='#000000', norm=norm, label=labels[0])
if other_study is not None:
other_trials = [t for t in other_study.trials if t.values is not None]
other_scatter = plt.scatter([t.values[0] for t in other_trials], [t.values[1] for t in other_trials], c=[map(t) for t in other_trials], cmap='Oranges', linewidths=1.0, edgecolors='#000000', norm=norm, label=labels[1])
plt.legend(loc=label_loc, handles=[other_scatter, scatter] if switch_label else [scatter, other_scatter], fontsize="x-large")
if print_pareto:
plt.scatter([t.values[0] for t in study.best_trials], [t.values[1] for t in study.best_trials], c=[map(t) for t in study.best_trials], cmap='Reds', linewidths=1.0, edgecolors='#000000', norm=norm, label="Pareto Front")
plt.legend(loc=label_loc, fontsize="x-large")
if split_simplify:
simplify_trials = [t for t in trials if t.params["parts_type"] == "minimize_head"]
full_trials = [t for t in trials if t.params["parts_type"] != "minimize_head"]
plt.scatter([t.values[0] for t in simplify_trials], [t.values[1] for t in simplify_trials], c=[0.8] * len(simplify_trials), cmap='Blues', norm=Normalize(0.0, 1.0), linewidths=1.0, edgecolors="black", label="Simplify")
plt.scatter([t.values[0] for t in full_trials], [t.values[1] for t in full_trials], c="white", linewidths=1.0, edgecolors="black", label="Full")
plt.legend(loc=label_loc, fontsize="x-large")
plt.xlabel('KP', fontsize="x-large")
plt.ylabel('Edits', fontsize="x-large")
if print_colorbar:
cb = plt.colorbar(scatter)
cb.set_label(param_name, rotation=0, labelpad=20, y=0.5)
plt.savefig(f"plots/{name}.png", dpi=500, bbox_inches='tight', pad_inches=0)
plt.savefig(f"plots/{name}.pdf", dpi=500, bbox_inches='tight', pad_inches=0)
plt.show()
def get_study_values(study, hyperparams):
string_map = {
"binary": 0,
"distance": 1,
"full": 0,
"minimize_head": 1,
"additive": 0,
"multiplicative": 1,
}
x = [[string_map[t.params[hyperparam]] if isinstance(t.params[hyperparam], str) else t.params[hyperparam] for hyperparam in hyperparams] for t in study.trials if t.values is not None]
y = [[t.values[y_index] for y_index in range(len(t.values))] for t in study.trials if t.values is not None]
return x, y
def print_correlation(correlation, pvalues, hyperparams, offset=0):
for i in range(len(hyperparams)):
x_name = hyperparams[i]
print(f"{x_name} correlation is: {correlation[i][offset]} on KP, {correlation[i][offset + 1]} on edits.")
print(f"{x_name} pvalue is: {pvalues[i][offset]} on KP, {pvalues[i][offset + 1]} on edits.")
def analyze_spearman(study, hyperparams, name):
spearman = study_spearman(study, hyperparams)
dict = {"param": ["kp", "p kp", "edits", "p edits"]}
offset = len(hyperparams)
for i in range(len(hyperparams)):
p = re.sub(r'_', ' ', hyperparams[i])
dict[p] = [
transform_corr(spearman.correlation[i][offset]),
transform_pvalue(spearman.pvalue[i][offset]),
transform_corr(-spearman.correlation[i][offset + 1]),
transform_pvalue(spearman.pvalue[i][offset + 1])]
pd.DataFrame(dict).set_index("param").T.to_csv(f"data/{name}_params_spearman.csv", index_label="param")
print_correlation(spearman.correlation, spearman.pvalue, hyperparams, offset)
def study_spearman(study, hyperparams):
x, y = get_study_values(study, hyperparams)
return stats.spearmanr(x, y)
def get_pearson(x, y):
pearson_results = [[stats.pearsonr(np.transpose(x)[i], np.transpose(y)[j]) for j in range(np.shape(y)[1])] for i in range(np.shape(x)[1])]
pearson = [[cell[0] for cell in row] for row in pearson_results]
pvalue = [[cell[1] for cell in row] for row in pearson_results]
return pearson, pvalue
def analyze_pearson(study, hyperparams):
correlation, pvalues = study_pearson(study, hyperparams)
print_correlation(correlation, pvalues, hyperparams)
def study_pearson(study, hyperparams):
x, y = get_study_values(study, hyperparams)
return get_pearson(x, y)
def transform_corr(r):
return f"{r:.2f}"
def transform_pvalue(p, check=True):
return "0.001" if p < 0.001 and check else f"{p:.3f}"
def evaluate_results_spearman(results, name):
relevant_attributes = [
"avg_edits",
"eval_single_near",
"eval_single_same",
"eval_all_near",
"eval_all_same",
]
x = [[float(r[att]) if att != "avg_edits" else -float(r[att]) for att in relevant_attributes] for r in results]
spearman, pvalue = stats.spearmanr(x)
corr_dict = {"metric": [re.sub(r'_', ' ', a) for a in relevant_attributes]}
for i, att in enumerate(relevant_attributes):
new_att = re.sub(r'_', ' ', att)
corr_dict[new_att] = [transform_corr(r) for r in spearman[i]]
corr_string = reduce(lambda a, b: f"{a}, {b}", corr_dict[new_att])
print(f"{new_att} Spearman correlation is {corr_string}")
pd.DataFrame(corr_dict).to_csv(f"data/{name}_results_spearman.csv", index=False)
print("")
pvalue_dict = {"metric": [re.sub(r'_', ' ', a) for a in relevant_attributes]}
for i, att in enumerate(relevant_attributes):
new_att = re.sub(r'_', ' ', att)
pvalue_dict[new_att] = [transform_pvalue(p, i != j) for j, p in enumerate(pvalue[i])]
pvalue_string = reduce(lambda a, b: f"{a}, {b}", pvalue_dict[new_att])
print(f"{new_att} Spearman pvalue is {pvalue_string}")
pd.DataFrame(pvalue_dict).to_csv(f"data/{name}_results_spearman_pvalue.csv", index=False)
def evaluate_results_pearson(results):
relevant_attributes = [
"avg_edits",
"eval_single_near",
"eval_single_same",
"eval_all_near",
"eval_all_same",
]
x = [[float(r[att]) for att in relevant_attributes] for r in results]
pearson, pvalue = get_pearson(x, x)
for i, att in enumerate(relevant_attributes):
corr_string = reduce(lambda a, b: f"{a}, {b}", pearson[i])
print(f"{att} Pearson correlation is {corr_string}")
print("")
for i, att in enumerate(relevant_attributes):
pvalue_string = reduce(lambda a, b: f"{a}, {b}", pvalue[i])
print(f"{att} Pearson pvalue is {pvalue_string}")
def evaluate_results_average(results):
relevant_attributes = [
"avg_edits",
"eval_single_near",
"eval_single_same",
"eval_all_near",
"eval_all_same",
]
x = [[float(r[att]) for att in relevant_attributes] for r in results]
avg = np.average(x, axis=0)
for i, att in enumerate(relevant_attributes):
print(f"{att} average is {avg[i]}")
def evaluate_results_median(results):
relevant_attributes = [
"avg_edits",
"eval_single_near",
"eval_single_same",
"eval_all_near",
"eval_all_same",
]
x = [[float(r[att]) for att in relevant_attributes] for r in results]
avg = np.median(x, axis=0)
for i, att in enumerate(relevant_attributes):
print(f"{att} median is {avg[i]}")
def evaluate_performance(results):
runtimes = [float(r["time"]) for r in results]
print(f"Total runtime is {np.sum(runtimes)} seconds.")
print(f"Average runtime is {np.average(runtimes)} seconds.")
print(f"Minimum runtime is {np.min(runtimes)} seconds.")
print(f"Maximum runtime is {np.max(runtimes)} seconds.")
print(f"Median runtime is {np.median(runtimes)} seconds.")
def performance_edits_correlation(results):
relevant_attributes = [
"avg_edits",
"time",
]
x = [[float(r[att]) for att in relevant_attributes] for r in results]
# pearson = np.corrcoef(x, rowvar=False)
pearson, pearson_pvalue = stats.pearsonr(np.transpose(x)[0], np.transpose(x)[1])
print(f"Pearson correlation is {pearson}")
print(f"Pearson pvalue is {pearson_pvalue}")
spearman, spearman_pvalue = stats.spearmanr(x)
print(f"Spearman correlation is {spearman}")
print(f"Spearman pvalue is {spearman_pvalue}")
num_iterations = 4411
def average_time_per_edit(results):
time_per_edit = [float(r["time"]) / (float(r["avg_edits"]) * num_iterations) for r in results]
average_time = np.average(time_per_edit)
var_time = np.var(time_per_edit)
print(f"Average time per edit is {average_time} seconds, variance is {var_time}.")
def plot_runtime(results, name):
x = range(len(results))
y = [float(r["time"]) for r in results]
plt.plot(x, y)
window = 15
rolling_average = sliding_window_view(y, window).mean(axis=-1)
rolling_x = range((window - 1) // 2, len(results) - (window - 1) // 2)
plt.plot(rolling_x, rolling_average, color="red")
fit = np.polyfit(x, y, 1)
print(fit)
p = np.poly1d(fit)
diff = y - p(x)
var = np.var(diff)
print(f"Variance of data to best fit is {var}.")
# plt.plot(x, p(x), color="red", linestyle="--", linewidth=2)
plt.xlabel('Iteration', fontsize="x-large")
plt.ylabel('Runtime in seconds', fontsize="x-large")
plt.savefig(f"plots/{name}.png", dpi=500, bbox_inches='tight', pad_inches=0)
plt.savefig(f"plots/{name}.pdf", dpi=500, bbox_inches='tight', pad_inches=0)
plt.show()
def plot_runtime_per_edit(results, name):
x = range(len(results))
y = [float(r["time"]) / (float(r["avg_edits"]) * num_iterations) for r in results]
plt.plot(x, y)
fit = np.polyfit(x, y, 1)
p = np.poly1d(fit)
# plt.plot(x, p(x), color="red", linestyle="--", linewidth=2)
plt.xlabel('Iteration', fontsize="x-large")
plt.ylabel('Runtime per edit in seconds', fontsize="x-large")
plt.savefig(f"plots/{name}.png", dpi=500, bbox_inches='tight', pad_inches=0)
plt.savefig(f"plots/{name}.pdf", dpi=500, bbox_inches='tight', pad_inches=0)
plt.show()
def plot_runtime_edits(results, name):
x = [float(r["avg_edits"]) for r in results]
y = [float(r["time"]) for r in results]
plt.scatter(x, y, linewidths=1.0, edgecolors='#000000')
plt.xlabel('Edits', fontsize="x-large")
plt.ylabel('Runtime in seconds', fontsize="x-large")
plt.savefig(f"plots/{name}.png", dpi=500, bbox_inches='tight', pad_inches=0)
plt.savefig(f"plots/{name}.pdf", dpi=500, bbox_inches='tight', pad_inches=0)
plt.show()