-
Notifications
You must be signed in to change notification settings - Fork 0
/
basin_ellipsoid.py
183 lines (136 loc) · 6.87 KB
/
basin_ellipsoid.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
import meta_poisoning_typical as mp
import mlp
Params = mlp.Params
MLP = mlp.MLP
ellipsoid_norm = mlp.ellipsoid_norm
import sys
import jax
import jax.numpy as jnp
import numpy as np
import matplotlib.pyplot as plt
from tqdm import trange, tqdm
from einops import einsum, rearrange, repeat, reduce
def find_radius(center, vec, cutoff, fn, rtol=1e-1, high=None, low=0, init_mult=1, iters=10, jump=2.0):
center_loss = fn(center)
vec_loss = fn(center + init_mult * vec)
if iters == 0 or jnp.abs(vec_loss - center_loss - cutoff) < cutoff * rtol:
return init_mult, vec_loss - center_loss
if vec_loss - center_loss < cutoff: # too low
low = init_mult
if high is None:
new_init_mult = init_mult * jump
else:
new_init_mult = (high + low) / 2
else: # too high
high = init_mult
new_init_mult = (high + low) / 2
return find_radius(center, vec, cutoff, fn=fn, high=high, low=low, init_mult=new_init_mult, iters=iters - 1)
def experiment(split_id, params_path):
chonk = 4810 // 7
indices_start = chonk * int(split_id)
indices_end = indices_start + chonk
if split_id == 6:
indices_end = 4810
indices = jnp.arange(indices_start, indices_end)
train_size = 64 if "_64" in params_path else 128 if "_128" in params_path else 256
cfg = mp.MetaConfig(num_layers=1, spherical=True,
train_size=train_size,
meta_constrain=True, mesa_constrain=True)
X_train, Y_train, X_untrain, Y_untrain, X_test, Y_test = mp.get_digits_splits(cfg)
XY = {'train': (X_train, Y_train), 'untrain': (X_untrain, Y_untrain), 'test': (X_test, Y_test)}
model, params_init = mp.get_model(cfg, X_train)
with open(params_path, 'rb') as f:
params_spher = mlp.Params(jnp.load(f, allow_pickle=True), params_init.unravel)
def train_fn(params_raveled):
params_raveled = params_raveled * jnp.linalg.norm(params_spher.raveled) / jnp.linalg.norm(params_raveled)
params = Params(params_raveled, params_spher.unravel)
apply_fn = mp.make_apply_full(model, params.unravel)
_, _, state = mp.train(
params_raveled,
X_train, Y_train, X_untrain, Y_untrain, X_test, Y_test,
apply_fn, cfg,
target_norm=ellipsoid_norm(params_spher, spherical=True), unravel=params_spher.unravel,
return_state=True,
)
return state.params['p']
def dataset_loss(params_raveled, X, Y):
params_raveled = params_raveled * jnp.linalg.norm(params_spher.raveled) / jnp.linalg.norm(params_raveled)
logits = model.apply(params_spher.unravel(params_raveled), X)
preds = jnp.argmax(logits, axis=-1)
loss = mp.sparse_xent(logits, Y).mean()
return loss
def train_loss_fn(params_raveled):
return dataset_loss(params_raveled, X_train, Y_train)
def final_dataset_loss(init_params_raveled, X, Y):
init_params_raveled = init_params_raveled * jnp.linalg.norm(params_spher.raveled) / jnp.linalg.norm(init_params_raveled)
params_raveled = train_fn(init_params_raveled)
logits = model.apply(params_spher.unravel(params_raveled), X)
preds = jnp.argmax(logits, axis=-1)
loss = mp.sparse_xent(logits, Y).mean()
return loss
def final_train_loss_fn(params_raveled):
return final_dataset_loss(params_raveled, X_train, Y_train)
def final_untrain_loss_fn(params_raveled):
return final_dataset_loss(params_raveled, X_untrain, Y_untrain)
hess_fn = jax.hessian(train_loss_fn)
H = hess_fn(train_fn(params_spher.raveled))
evals, evecs = jnp.linalg.eigh(H)
jac_fn = jax.jacfwd(train_fn)
J = jac_fn(params_spher.raveled)
params_unit = params_spher.raveled / jnp.linalg.norm(params_spher.raveled)
proj_param = einsum(params_unit, params_unit, 'i, j -> i j')
J_unorth = J + proj_param
u, s, vt = jnp.linalg.svd(J_unorth)
final_untrain_radii = []
final_train_radii = []
final_untrain_neg_radii = []
final_train_neg_radii = []
init_mult = 1
rtol = 0.1
train_cutoff = 0.5 - final_train_loss_fn(params_spher.raveled)
untrain_cutoff = 1e-2
jump = 2
train_iters = 20
untrain_iters = 20
my_params = params_spher.raveled
my_train_fn = final_train_loss_fn
my_untrain_fn = lambda x: -final_untrain_loss_fn(x)
def Jgen_direction(i):
return vt.T @ (u.T @ evecs[:, i])
for i in indices:
vec = Jgen_direction(i)
final_untrain_radii.append(find_radius(my_params, vec, untrain_cutoff, rtol=rtol, init_mult=init_mult,
fn=my_untrain_fn, iters=untrain_iters, jump=jump))
final_train_radii.append(find_radius(my_params, vec, train_cutoff, rtol=rtol, init_mult=init_mult,
fn=my_train_fn, iters=train_iters, jump=jump))
final_untrain_neg_radii.append(find_radius(my_params, -vec, untrain_cutoff, rtol=rtol, init_mult=init_mult,
fn=my_untrain_fn, iters=untrain_iters, jump=jump))
final_train_neg_radii.append(find_radius(my_params, -vec, train_cutoff, rtol=rtol, init_mult=init_mult,
fn=my_train_fn, iters=train_iters, jump=jump))
final_untrain_radii_jnp = jnp.array(final_untrain_radii)
final_train_radii_jnp = jnp.array(final_train_radii)
final_untrain_neg_radii_jnp = jnp.array(final_untrain_neg_radii)
final_train_neg_radii_jnp = jnp.array(final_train_neg_radii)
final_untrain_diameters = final_untrain_radii_jnp[:, 0] + final_untrain_neg_radii_jnp[:, 0]
final_train_diameters = final_train_radii_jnp[:, 0] + final_train_neg_radii_jnp[:, 0]
final_untrain_deltas = (final_untrain_radii_jnp[:, 1], final_untrain_neg_radii_jnp[:, 1])
final_train_deltas = (final_train_radii_jnp[:, 1], final_train_neg_radii_jnp[:, 1])
min_radii = jnp.min(jnp.array([final_train_radii_jnp[:, 0], final_untrain_radii_jnp[:, 0]]), axis=0)
min_neg_radii = jnp.min(jnp.array([final_train_neg_radii_jnp[:, 0], final_untrain_neg_radii_jnp[:, 0]]), axis=0)
min_diameters = min_radii + min_neg_radii
# final_logvols = logvol_estimate(indices, min_diameters)
return min_diameters, (final_untrain_radii_jnp, final_train_radii_jnp, final_untrain_neg_radii_jnp, final_train_neg_radii_jnp)
if __name__ == '__main__':
# get split ID from command line
split_id = sys.argv[1]
PARAM_PATHS = [
'pinit_0930_beta09_128.npy',
'pinit_0928_beta097_128.npy',
]
for path in PARAM_PATHS:
diameters, radii = experiment(split_id, path)
# save stuff
diameters_jnp = jnp.array(diameters)
radii_jnp = jnp.array(radii)
np.save(f'out0930_{path}_split{split_id}_diameters.npy', diameters_jnp)
np.save(f'out0930_{path}_split{split_id}_radii.npy', radii_jnp)