-
Notifications
You must be signed in to change notification settings - Fork 2
/
figures.py
309 lines (238 loc) · 9.38 KB
/
figures.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
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
import os
import fire
import numpy as np
import scipy.stats
import pandas
import matplotlib
import matplotlib.pyplot as plt
import itermplot
from matplotlib.ticker import FuncFormatter
OUTPUT = 'generated'
RNG = np.random.RandomState(0)
# reverse plot color is output is in the terminal
itermplot.THEME = 'rv' if OUTPUT is None else ''
matplotlib.use('module://itermplot')
matplotlib.rcParams['font.sans-serif'] = 'Source Sans Pro'
matplotlib.rcParams['font.family'] = 'sans-serif'
matplotlib.rcParams['font.weight'] = 'semibold'
matplotlib.rcParams['font.size'] = '10'
matplotlib.rcParams['pdf.fonttype'] = 42
def formatter(x, pos):
if x != 0:
if abs(x) > 1: # hacky
x = f'{int(x)}'
else:
x = ('%.2f' % x).lstrip('0').rstrip('0')
else:
x = '0'
return x
def corr(df, x, y, report_p=False):
sel = df[x].notnull() & df[y].notnull()
try:
r, p = scipy.stats.pearsonr(df.loc[sel, x], df.loc[sel, y])
except:
breakpoint()
if p > .05:
r = 'r: n.s.'
else:
r = 'r = ' + f'{r:.2f}'.lstrip('0')
if report_p:
if p < .001:
p_level = f'{p:.0e}'
elif p < .01:
p_level = '.01'
elif p < .05:
p_level = '.05'
else:
p_level = ''
if p_level != '':
r += f' (p < {p_level})'
return r
def output_paper_quality(ax, title=None, xlabel=None, ylabel=None):
ax.set_title(title)
ax.set_xlabel(xlabel, weight='semibold', size=12)
ax.set_ylabel(ylabel, weight='semibold', size=12)
ax.xaxis.set_major_formatter(FuncFormatter(formatter))
ax.yaxis.set_major_formatter(FuncFormatter(formatter))
ax.spines['right'].set_visible(False)
ax.spines['top'].set_visible(False)
ax.tick_params(bottom=False, left=False)
def output(filename):
plt.tight_layout()
if OUTPUT is not None:
plt.savefig(os.path.join(OUTPUT, f'{filename}.png'), bbox_inches='tight', dpi=300)
plt.savefig(os.path.join(OUTPUT, f'{filename}.pdf'), bbox_inches='tight', transparent=True)
else:
plt.show()
def scatterplot(ax, df, x, y, scale=30, annotate=False):
dff = df[df.Model != 'CORnet-S']
jitter = dff.jitter if 'jitter' in dff else 0
ax.scatter(dff[x] + jitter, dff[y], s=scale, color=dff.color, alpha=.7, edgecolors='none')#linewidths=.1)
dff = df[df.Model == 'CORnet-S']
jitter = dff.jitter if 'jitter' in dff else 0
ax.scatter(dff[x] + jitter, dff[y], s=2*scale, color=dff.color, alpha=.7, edgecolors='none')#, linewidths=.1)
if annotate:
for idx, row in df.iterrows():
ax.text(row[x], row[y], row.Model)
return ax
def read_common_data():
df = pandas.read_csv('data/data.csv')
for idx, row in df.iterrows():
if row.Model.lower().startswith('basenet'):
color = 'gray'
elif row.Model.lower().startswith('cornet'):
color = 'crimson'
else:
color = '#078930'
df.loc[idx, 'color'] = color
return df
def _fig1(df, scale=15, inset=False):
ax = plt.subplot(111)
scatterplot(ax, df, x='ImageNet', y='Brain-Score', scale=scale)
if not inset:
df = df[df.ImageNet < .7]
r = corr(df, x='ImageNet', y='Brain-Score')
ax.annotate(r, xy=(.75, .1),
xycoords='axes fraction',
# horizontalalignment='left', verticalalignment='top',
fontsize=10)
output_paper_quality(ax,
xlabel='ImageNet top-1 performance',
ylabel='Brain-Score')
def fig1():
df = read_common_data()
plt.figure(figsize=(4, 4))
_fig1(df, scale=15)
output('fig1')
plt.figure(figsize=(3, 3))
_fig1(df[(df['ImageNet'] >= .7) & (df.Model != 'CORnet-S')], scale=45, inset=True)
output('fig1_inset')
def _fig2(ax, df, x, y, title=None, xlabel=None, ylabel=None, index=0, r=None):
scatterplot(ax, df, x, y)
r = corr(df[df.Model != 'CORnet-S'], x, y)
ax.annotate(r, xy=(.75, .1), xycoords='axes fraction', fontsize=10)
ax.annotate(f"({'abcd'[index]})", xy=(.05 + .24 * index, .9),
xycoords='figure fraction', fontsize=20)
output_paper_quality(ax, title=title,
xlabel=xlabel, ylabel=ylabel)
ax.set_title(title, va='top', pad=20)
def fig2():
df = read_common_data()
dff = df[~df.Model.str.startswith('BaseNet')]
fig, axes = plt.subplots(ncols=4, figsize=(12,3))
_fig2(axes[0], dff, x='IT', y='IT (new data)',
xlabel='IT score (original neurons)', ylabel='IT score (new neurons)',
title='New neural recordings,\nsame images', index=0, r=.93)
_fig2(axes[1], dff, x='IT', y='IT (new images)',
xlabel='IT score (original neurons)', ylabel='IT score (new neurons)',
title='New neural recordings,\nnew images', index=1, r=.76)
_fig2(axes[2], dff, x='Behavior', y='Behavior (new data)',
xlabel='Behavioral score (original)', ylabel='Behavioral score (new)',
title='New behavioral recordings,\nnew images', index=2, r=.83)
_fig2(axes[3], dff, x='Brain-Score', y='CIFAR-100',
xlabel='Brain-Score', ylabel='CIFAR-100 transfer',
title='CIFAR-100 transfer', index=3, r=.69)
output('fig2')
def _fig3(ax, df, y, ylabel=None, title=None):
scatterplot(ax, df, x='Depth', y=y)
ax.set_xscale('log')
ax.set_xticks([10, 25, 50, 100, 200])
output_paper_quality(ax, title=title,
xlabel='Model Depth (log scale)', ylabel=ylabel)
def fig3():
df = read_common_data()
dff = df[~df.Model.str.startswith('BaseNet')].copy()
fig, axes = plt.subplots(ncols=3, figsize=(9,3), sharex=True)
rangex = .05 * (dff['Depth'].max() - dff['Depth'].min())
jitter = []
for idx, row in dff.iterrows():
if row.Model.startswith('MobileNet'):
j = RNG.uniform(-rangex, rangex)
else:
j = 0
jitter.append(j)
dff['jitter'] = jitter
_fig3(axes[0], dff, y='Brain-Score', ylabel='Brain-Score')
_fig3(axes[1], dff, y='ImageNet', ylabel='ImageNet top-1')
_fig3(axes[2], dff, y='CIFAR-100', ylabel='CIFAR-100 transfer')
output('fig3')
def _fig_a1(ax, df, region):
x = f'{region} number of features'
scatterplot(ax, df, x=x, y=region)
r = corr(df[df[x] >= 1000], x, region, report_p=True)
print(f'Fig A1: {r}')
# ax.annotate(r, xy=(.5, .1), xycoords='axes fraction', fontsize=10)
output_paper_quality(ax, xlabel='Number of features', ylabel=f'{region} neural score')
def fig_a1():
df = read_common_data()
dff = df[(df['V4 number of features'] < 20000) & (df['IT number of features'] < 20000)]
fig, axes = plt.subplots(ncols=2, figsize=(6,3))
_fig_a1(axes[0], dff, region='V4')
_fig_a1(axes[1], dff, region='IT')
output('fig_a1')
def fig_a2():
df = pandas.read_csv('data/cornet_search.csv')
plt.figure(figsize=(6, 4))
ax = plt.subplot(111)
ax.scatter(df['ImageNet'], df['Behavior'], s=10, color='gray', alpha=.7, edgecolors='none')
r = corr(df, x='ImageNet', y='Behavior')
ax.annotate(r, xy=(.75, .1), xycoords='axes fraction', fontsize=10)
output_paper_quality(ax, xlabel='ImageNet top-1', ylabel='Behavioral score')
output('fig_a2')
def _fig_a3(ax, df, region, time):
y = f'{region} ({time})'
ms = 100 if time == 'early' else 200
scatterplot(ax, df, x='ImageNet', y=y)
r = corr(df, x='ImageNet', y=y, report_p=True)
ax.annotate(r, xy=(.5, .1), xycoords='axes fraction', fontsize=10)
output_paper_quality(ax, xlabel='ImageNet top-1',
ylabel=f'{region} neural score (at {ms} ms)',
title=f'{region} {time}')
def fig_a3():
df = read_common_data()
fig, axes = plt.subplots(nrows=2, ncols=2, figsize=(6,6))
_fig_a3(axes[0,0], df, region='V4', time='early')
_fig_a3(axes[0,1], df, region='V4', time='late')
_fig_a3(axes[1,0], df, region='IT', time='early')
_fig_a3(axes[1,1], df, region='IT', time='late')
output('fig_a3')
def _highlight_max(data):
if data.dtype.name == 'object': # got a string (model)
return data
max_idx = data.idxmax()
formatted_data = []
for idx, value in data.iteritems():
if value != 0:
value = f'{value:.3f}'.lstrip('0').rstrip('0')
else:
value = '0'
if idx == max_idx:
value = '\\textbf{' + value + '}'
formatted_data.append(value)
return formatted_data
def table_a1():
df = read_common_data()
dff = df[df['Brain-Score'].notnull() & ~df.Model.str.startswith ('BaseNet') & ~df.Model.str.startswith('MobileNet')]
idx = df.loc[df.Model.str.startswith ('BaseNet'), 'Brain-Score'].idxmax()
best_basenet = df.loc[idx].copy()
best_basenet['Model'] = 'Best BaseNet'
dff = dff.append(best_basenet)
idx = df.loc[df.Model.str.startswith ('MobileNet'), 'Brain-Score'].idxmax()
best_mobilenet = df.loc[idx].copy()
best_mobilenet['Model'] = 'Best MobileNet'
dff = dff.append(best_mobilenet)
dff = dff.sort_values(by='Brain-Score', ascending=False)
dff = dff[['Model', 'Brain-Score', 'V4', 'IT', 'OST', 'Behavior']].apply(_highlight_max)
dff.to_latex(os.path.join(OUTPUT, 'table_a1.tex'), escape=False, index=False)
def gen_all():
fig1()
fig2()
fig3()
# data for fig4 is not provided
# fig5 is generated by fig5.py from scratch
table_a1()
fig_a1()
fig_a2()
fig_a3()
if __name__ == '__main__':
fire.Fire()