-
Notifications
You must be signed in to change notification settings - Fork 0
/
plot_connectomes_bsds.py
203 lines (182 loc) · 8.88 KB
/
plot_connectomes_bsds.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
import numpy as np
from matplotlib import pyplot as plt
import matplotlib.pylab as pl
from glob import glob
import os
from argparse import ArgumentParser
from matplotlib.font_manager import FontProperties
from skimage.transform import rotate
from scipy.signal import convolve
def process_h(hiddens, model):
"""Process the hidden state using model weights to flip signs if needed."""
# h2 bn via c1_bn_gamma_ and c1_bn_beta_
nh = (hiddens - hiddens.mean(axis=(1, 2), keepdims=True)) / hiddens.std(axis=(1, 2), keepdims=True)
nh = model["c1_bn_beta_0"] + nh * model["c1_bn_gamma_0"]
# c1 = h2 conv with horizontal_kernels_inh
c1 = convolve(nh, model["symm_horizontal_kernels_inh_0"], mode="same")
# Sup-drive = (alpha * h2 + mu) * c1
drive = (model["alpha_0"] * nh + model["recurrent_vgg16_mu"]) * c1
return drive
def main(path, channel=0):
"""Plot connectomes."""
connectome_files = glob(path)
performance_files = glob(path.replace("optim.npy", "perf.npy"))
tuning_exp_files = glob(path.replace("optim.npy", "curves.npy"))
tuning_gt_files = glob(path.replace("optim.npy", "label.npy"))
model_vars = np.load("gn_vars.npz")
assert len(connectome_files), "Couldnt find anythin {}".format(path)
connectome_files = sorted(connectome_files, key=os.path.getmtime)
performance_files = sorted(performance_files, key=os.path.getmtime)
tuning_exp_files = sorted(tuning_exp_files, key=os.path.getmtime)
tuning_gt_files = sorted(tuning_gt_files, key=os.path.getmtime)
results_dir = path.split(os.path.sep)[0]
# How many distinct bins have we finished?
bins = np.asarray([int(x.split("_")[-2].replace(".py", "")) for x in connectome_files])
num_bins = len(bins)
# Plot performance + tuning in subplots on the same figure
f, axs = plt.subplots(3, num_bins, figsize=(18, 5)) # noqa perf/tuning, num bins TODO: Add images
plt.subplots_adjust(wspace=0.7, hspace=0.3)
axs = axs.ravel()
perf_tuning = zip(performance_files, tuning_exp_files, tuning_gt_files)
fontP = FontProperties()
fontP.set_size('xx-small')
rcs = []
for idx, pt in enumerate(perf_tuning):
perf, tex, tgt = pt
# Plot performance
ax = axs[idx]
perf = np.load(perf).squeeze()
ax.plot(perf, "k", linewidth=1)
ax.set_xlabel("Optim steps", fontsize=3)
ax.set_title("${}^\circ$".format(bins[idx] - 91), fontsize=3) # noqa
ax.set_ylabel("$L^2$ loss", fontsize=3)
ax.tick_params(axis='both', which='major', labelsize=6)
ax.tick_params(axis='both', which='minor', labelsize=6)
# Plot tuning
ax = axs[idx + num_bins]
gt = np.load(tgt).squeeze()[-6:]
ex = np.load(tex).squeeze()[:, :6]
n = ex.shape[0]
colors = pl.cm.RdBu_r(np.linspace(0, 1, n))
for i in range(n):
ax.plot(np.concatenate((ex[i], [ex[i][0]])), color=colors[i], label=i)
ax.plot(np.concatenate((gt, [gt[0]])), "k--", label="GT")
ax.tick_params(axis='both', which='major', labelsize=6)
ax.tick_params(axis='both', which='minor', labelsize=6)
ax.set_xlabel("Orientation", fontsize=3)
ax.set_xticks(np.arange(7))
ax.set_xticklabels([-90, -60, -30, 0, 30, 60, 90], fontsize=3)
ax.set_ylabel("Population response", fontsize=3)
# ax.legend(
# loc="upper left",
# prop=fontP,
# fancybox=True,
# shadow=False)
# Plot connectome
ax = axs[idx + num_bins * 2]
con = np.load(connectome_files[idx])[0][channel] # [0] - np.load(connectome)[channel][1] # additive/mulitiplicative
mcon = con.squeeze().mean(-1)
# minmax = max(np.abs(con[40:45, 40:45].min()), con[40:45, 40:45].max())
minmax = 20 # max(np.abs(con.min()), con.max())
minmax = minmax + minmax * 2
# mask = 1. - (con == 0).astype(np.float32)
tuning = np.load("bsds_sel_def.npy")
tuning = np.ma.masked_where(mcon != 0, tuning)
# ax.imshow(mcon, cmap="RdBu_r", vmin=minmax * np.sign(mcon.min()), vmax=minmax) # , alpha=mask)
ax.imshow(mcon, cmap="RdBu_r", vmin=-4, vmax=4) # , alpha=mask)
ax.imshow(tuning, cmap="Greys") # , alpha=1. - mask)
ax.set_xticks([])
ax.set_yticks([])
theta = bins[idx]
ax.set_xlabel(r"$\theta = {}$".format(theta))
# rcs.append(rotate(con, bins[idx], preserve_range=True, order=0))
plt.savefig(os.path.join(results_dir, "performance.pdf"))
# plt.show()
plt.close(f)
f = plt.figure(dpi=300)
minmax = 20 # max(np.abs(con.min()), con.max())
if "circuit_exc" in path:
minmax = minmax + minmax * 1
elif "plaid_exc" in path:
minmax = minmax + minmax
else:
minmax = minmax + minmax * 1
# minmax = max(np.abs(rc.min()), rc.max())
tuning = np.load("bsds_sel_def.npy")
tuning = np.ma.masked_where(mcon != 0, tuning)
# plt.imshow(mcon, cmap="RdBu_r", vmin=minmax * np.sign(mcon.min()), vmax=minmax) # , alpha=mask)
plt.imshow(mcon, cmap="RdBu_r", vmin=-4, vmax=4) # , alpha=mask)
plt.colorbar()
plt.imshow(tuning, cmap="Greys") # , alpha=1. - mask)
plt.axis("off")
plt.savefig(os.path.join(results_dir, "mean_connectome.pdf"))
plt.show()
plt.close(f)
# Visualize per feature suppression
filters = glob("vgg_filters/*.npy")
filters = np.asarray(filters)
filters = np.sort(filters)
filter_data = []
for f in filters:
filter_data.append(np.load(f))
filter_data = np.asarray(filter_data).squeeze()
# Get top E and top I
k = 5
loading = con.squeeze().reshape(-1, 128).mean(0) / con.squeeze().reshape(-1, 128).std(0)
# loading = con.squeeze().max(0).max(0)
top_e = np.argsort(loading)[:k] # Negative
top_i = np.argsort(loading)[::-1][:k] # Positive
f = plt.figure(dpi=300)
"""
for idx in range(128):
plt.subplot(8, 16, idx + 1)
plt.axis("off")
plt.imshow(
con.squeeze()[..., idx],
cmap="RdBu_r",
vmin=minmax * np.sign(con.min()),
vmax=minmax)
"""
for idx in range(k):
plt.subplot(4, k, idx + 1)
plt.axis("off")
plt.imshow(con.squeeze()[..., top_e[idx]], cmap="RdBu_r", vmin=minmax * np.sign(con.min()), vmax=minmax)
plt.title("Exc Connectome {}".format(idx), fontsize=3)
plt.subplot(4, k, k + idx + 1)
plt.axis("off")
plt.imshow(filter_data[top_e[idx]], cmap="Greys") # , vmin=minmax * np.sign(con.min()), vmax=minmax)
plt.title("Exc Feature {}".format(idx), fontsize=3)
plt.subplot(4, k, 2 * k + idx + 1)
plt.axis("off")
plt.imshow(con.squeeze()[..., top_i[idx]], cmap="RdBu_r", vmin=minmax * np.sign(con.min()), vmax=minmax)
plt.title("Inh Connectome {}".format(idx), fontsize=3)
plt.subplot(4, k, 3 * k + idx + 1)
plt.axis("off")
plt.imshow(filter_data[top_i[idx]], cmap="Greys") # , vmin=minmax * np.sign(con.min()), vmax=minmax)
plt.title("Inh Feature {}".format(idx), fontsize=3)
plt.savefig(os.path.join(results_dir, "per_feature_connectome.pdf"))
plt.show()
plt.close(f)
if __name__ == '__main__':
parser = ArgumentParser()
parser.add_argument(
"--path",
type=str,
dest="path",
default="circuits_BSDS_exc_perturb/BSDS_exc_perturb_circuit_circuit_exc_*_optim.npy",
# default="circuits_BSDS_inh_perturb/BSDS_inh_perturb_circuit_circuit_inh_*_optim.npy",
# default="circuits_BSDS_exc_perturb/exc_1.5/BSDS_exc_perturb_circuit_circuit_exc_*_optim.npy",
# default="circuits_BSDS_exc_perturb/exc_2.0/BSDS_exc_perturb_circuit_circuit_exc_*_optim.npy",
# default="circuits_BSDS_inh_perturb/inh_0.001/BSDS_inh_perturb_circuit_circuit_inh_*_optim.npy",
# default="circuits_BSDS_inh_perturb/inh_0.01/BSDS_inh_perturb_circuit_circuit_inh_*_optim.npy",
# default="circuits_BSDS_inh_perturb/inh_0.1/BSDS_inh_perturb_circuit_circuit_inh_*_optim.npy",
# default="circuits_BSDS_exc_perturb/BSDS_exc_perturb_circuit_circuit_exc_full_field_*_optim.npy",
# default="circuits_BSDS_inh_perturb/BSDS_inh_perturb_circuit_circuit_inh_full_field_*_optim.npy",
# default="circuits_BSDS_exc_phase_perturb/BSDS_exc_phase_perturb_circuit_circuit_exc_*_optim.npy",
# default="circuits_BSDS_exc_phase_perturb_phase_180/BSDS_exc_phase_perturb_circuit_circuit_exc_*_optim.npy",
# default="circuits_BSDS_inh_phase_perturb/BSDS_inh_phase_perturb_circuit_circuit_inh_*_optim.npy",
# default="circuits_BSDS_inh_perturb/BSDS_inh_perturb_circuit_circuit_inh_*_optim.npy",
# default="circuits_BSDS_exc_perturb/BSDS_exc_perturb_circuit_circuit_plaid_exc_*_optim.npy",
# default="circuits_BSDS_inh_perturb/BSDS_inh_perturb_circuit_circuit_plaid_inh_*_optim.npy",
help="Name of experiment with model responses.")
main(**vars(parser.parse_args()))