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
|
#!/usr/bin/env python3
"""Deterministic equation audits for matched ResNet crossover adapters."""
import json
import os
import sys
import torch
sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
from sdil.conv import CIFARLocalResNet
from sdil.conv_crossover import CIFARDualPropResNet, CIFARPEPITAResNet
def relative_error(actual, expected):
numerator = torch.linalg.vector_norm(actual - expected)
denominator = torch.linalg.vector_norm(expected).clamp_min(1e-30)
return float(numerator / denominator)
def audit_dualprop():
common = dict(
depth=8, base_width=2, seed=71, normalization="batchnorm",
residual_scale=1.0, dtype=torch.float64)
reference = CIFARLocalResNet(**common)
net = CIFARDualPropResNet(
**common, alpha=0.0, dp_beta=0.1, inference_passes=2)
generator = torch.Generator().manual_seed(72)
image = torch.randn(3, 3, 32, 32, generator=generator,
dtype=torch.float64)
labels = torch.tensor([0, 3, 7])
one_hot = torch.nn.functional.one_hot(labels, 10).to(torch.float64)
with torch.no_grad():
reference_output = reference.forward(
image, training=True, update_stats=False)["logits"]
net_output = net.forward(
image, training=True, update_stats=False)["logits"]
clean = net.forward(
image, return_cache=True, training=True, update_stats=False)
plus, minus = net.infer_dual_states(image, one_hot, clean)
forward_error = float(torch.max(torch.abs(
reference_output - net_output)))
parameters = (
net.W + net.gamma + net.beta + [net.W_out, net.b_out])
for parameter in parameters:
parameter.requires_grad_(True)
alpha = net.dp_alpha
beta = net.dp_beta
states = [
(alpha * positive + (1.0 - alpha) * negative).detach()
for positive, negative in zip(plus[:-1], minus[:-1])]
deltas = [
((positive - negative) / beta).detach()
for positive, negative in zip(plus, minus)]
objective = image.new_zeros(())
for index in range(net.n_hidden):
prediction, _ = net._node_prediction(index, states, image)
objective -= torch.sum(deltas[index] * prediction) / image.shape[0]
features = states[-1].mean(dim=(2, 3))
output_prediction = features @ net.W_out.t() + net.b_out
objective -= torch.sum(deltas[-1] * output_prediction) / image.shape[0]
gradients = torch.autograd.grad(objective, parameters)
(directions, gamma_directions, beta_directions,
output_weight, output_bias) = net.dualprop_ascent_directions(
image, plus, minus)
actual = (
directions + gamma_directions + beta_directions
+ [output_weight, output_bias])
errors = [
relative_error(direction, -gradient)
for direction, gradient in zip(actual, gradients)]
for parameter in parameters:
parameter.requires_grad_(False)
if forward_error >= 1e-12 or max(errors) >= 2e-12:
raise AssertionError({
"forward_error": forward_error,
"direction_errors": errors,
})
return {
"matched_forward_max_absolute_error": forward_error,
"contrastive_direction_max_relative_error": max(errors),
"num_audited_parameter_tensors": len(errors),
"uses_symmetric_forward_edge_transposes": True,
"uses_reverse_task_loss_graph": False,
}
def audit_pepita():
common = dict(
depth=8, base_width=2, seed=81, normalization="batchnorm",
residual_scale=1.0, dtype=torch.float64)
reference = CIFARLocalResNet(**common)
net = CIFARPEPITAResNet(
**common, projection_scale=0.05, projection_seed=82)
generator = torch.Generator().manual_seed(83)
image = torch.randn(3, 3, 32, 32, generator=generator,
dtype=torch.float64)
labels = torch.tensor([1, 4, 8])
one_hot = torch.nn.functional.one_hot(labels, 10).to(torch.float64)
with torch.no_grad():
reference_output = reference.forward(
image, training=True, update_stats=False)["logits"]
clean = net.forward(
image, return_cache=True, training=True, update_stats=False)
clean_error = torch.softmax(clean["logits"], dim=1) - one_hot
input_error = torch.einsum(
"bc,cijk->bijk", clean_error, net.input_feedback)
modulated = net.forward(
image + input_error, return_cache=True,
training=True, update_stats=False)
modulated_error = (
torch.softmax(modulated["logits"], dim=1) - one_hot)
forward_error = float(torch.max(torch.abs(
reference_output - clean["logits"])))
parameters = (
net.W + net.gamma + net.beta + [net.W_out, net.b_out])
for parameter in parameters:
parameter.requires_grad_(True)
objective = image.new_zeros(())
batch = image.shape[0]
for index, (clean_hidden, modulated_hidden, cache, spec) in enumerate(zip(
clean["hiddens"], modulated["hiddens"], modulated["caches"],
net.layer_specs)):
field = (clean_hidden - modulated_hidden).detach()
convolution = torch.nn.functional.conv2d(
cache["pre"].detach(), net.W[index],
stride=spec.stride, padding=spec.padding)
normalized, _ = net._normalize(
index, convolution, training=True, update_stats=False)
prediction = spec.branch_scale * normalized
spatial = prediction.shape[2] * prediction.shape[3]
objective += torch.sum(field * prediction) / (batch * spatial)
output_prediction = (
modulated["features"].detach() @ net.W_out.t() + net.b_out)
objective += torch.sum(
modulated_error.detach() * output_prediction) / batch
gradients = torch.autograd.grad(objective, parameters)
(directions, gamma_directions, beta_directions,
output_weight, output_bias) = net.pepita_ascent_directions(
clean, modulated, modulated_error)
actual = (
directions + gamma_directions + beta_directions
+ [output_weight, output_bias])
errors = [
relative_error(direction, -gradient)
for direction, gradient in zip(actual, gradients)]
for parameter in parameters:
parameter.requires_grad_(False)
projection_limit = (6.0 / (3 * 32 * 32)) ** 0.5 * 0.05
observed_limit = float(torch.max(torch.abs(net.input_feedback)))
if (forward_error >= 1e-12 or max(errors) >= 2e-12
or observed_limit > projection_limit):
raise AssertionError({
"forward_error": forward_error,
"direction_errors": errors,
"projection_limit": projection_limit,
"observed_limit": observed_limit,
})
return {
"matched_forward_max_absolute_error": forward_error,
"local_equation_max_relative_error": max(errors),
"input_projection_shape": list(net.input_feedback.shape),
"input_projection_limit": projection_limit,
"observed_input_projection_max": observed_limit,
"uses_reverse_task_loss_graph": False,
}
def main():
print(json.dumps({
"dualprop": audit_dualprop(),
"pepita": audit_pepita(),
},
indent=2, sort_keys=True))
if __name__ == "__main__":
main()
|
