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
|
#!/usr/bin/env python3
"""Run the frozen no-KP layerwise causal-bootstrap capture screen."""
import argparse
import json
import math
import os
import subprocess
import sys
import time
import torch
sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
from sdil.conv import (CIFARHierarchicalFAResNet,
conv_hierarchical_alignment_report,
layerwise_causal_bootstrap_sweep)
from sdil.data import DATA_DIR, get_cifar_image_splits
ROOT = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
def provenance():
def run(command):
return subprocess.run(
command, cwd=ROOT, check=True, capture_output=True,
text=True).stdout.strip()
return {
"git_commit": run(["git", "rev-parse", "HEAD"]),
"git_tracked_dirty": bool(run(
["git", "status", "--porcelain", "--untracked-files=no"])),
}
def summarize_alignment(report):
values = report["teaching_negative_gradient_cosine"]
early = max(1, len(values) // 3)
ratios = report["feedback_forward_norm_ratio"]
cosines = report["feedback_forward_cosine"]
return {
"per_layer": values,
"early_third_alignment": sum(values[:early]) / early,
"all_layer_alignment": sum(values) / len(values),
"mean_feedback_forward_cosine": sum(cosines) / len(cosines),
"min_feedback_forward_norm_ratio": min(ratios),
"max_feedback_forward_norm_ratio": max(ratios),
"feedback_forward_cosine": cosines,
"feedback_forward_norm_ratio": ratios,
}
def forward_state(net):
return [value.clone() for value in (
net.W + net.gamma + net.beta + net.running_mean + net.running_var
+ net.mW + net.mgamma + net.mbeta
+ [net.W_out, net.b_out, net.mW_out, net.mb_out])]
def main():
parser = argparse.ArgumentParser()
parser.add_argument("--device", default="cuda")
parser.add_argument("--data_dir", default=DATA_DIR)
parser.add_argument("--out", default="results/oral_a_v5_calibration/result.json")
args = parser.parse_args()
settings = {
"depth": 20, "width": 16, "seed": 0, "loader_seed": 0,
"batch_size": 128, "train_limit": 10000,
"val_examples": 5000, "split_seed": 2027,
"normalization": "batchnorm", "residual_scale": 1.0,
"feedback_scale": 1.0, "sigma": 0.01, "eta_A": 0.1,
"perturb_seed": 5000, "sweeps": 20, "alignment_probe": 64,
"calibration_augmentation": False,
}
torch.manual_seed(settings["seed"])
if str(args.device).startswith("cuda"):
if not torch.cuda.is_available():
raise RuntimeError("CUDA requested but unavailable")
torch.cuda.manual_seed_all(settings["seed"])
torch.cuda.reset_peak_memory_stats(torch.device(args.device))
train, _, _, input_shape, n_out, split = get_cifar_image_splits(
batch_size=settings["batch_size"], data_dir=args.data_dir,
device=args.device, train_limit=settings["train_limit"],
val_examples=settings["val_examples"], split_seed=settings["split_seed"],
loader_seed=settings["loader_seed"], augment_train=False)
if input_shape != (3, 32, 32) or n_out != 10:
raise AssertionError("unexpected CIFAR dimensions")
net = CIFARHierarchicalFAResNet(
depth=settings["depth"], base_width=settings["width"],
n_classes=10, device=args.device, seed=settings["seed"],
residual_scale=settings["residual_scale"],
normalization=settings["normalization"],
feedback_scale=settings["feedback_scale"])
audit_x = train.x[:settings["alignment_probe"]]
audit_y = train.y[:settings["alignment_probe"]]
fixed = summarize_alignment(
conv_hierarchical_alignment_report(net, audit_x, audit_y))
state_before = forward_state(net)
generator = torch.Generator(device=torch.device(args.device)).manual_seed(
settings["perturb_seed"])
if str(args.device).startswith("cuda"):
torch.cuda.synchronize(torch.device(args.device))
start = time.time()
sweeps = []
for sweep_index in range(settings["sweeps"]):
start_index = sweep_index * settings["batch_size"]
stop_index = start_index + settings["batch_size"]
metric = layerwise_causal_bootstrap_sweep(
net, train.x[start_index:stop_index], train.y[start_index:stop_index],
sigma=settings["sigma"], eta=settings["eta_A"],
generator=generator)
sweeps.append({
key: value for key, value in metric.items() if key != "edges"})
print(json.dumps({"sweep": sweep_index + 1, **sweeps[-1]}), flush=True)
if str(args.device).startswith("cuda"):
torch.cuda.synchronize(torch.device(args.device))
wall_seconds = time.time() - start
state_after = forward_state(net)
forward_state_max_difference = max(
float((before - after).abs().max())
for before, after in zip(state_before, state_after))
learned = summarize_alignment(
conv_hierarchical_alignment_report(net, audit_x, audit_y))
total_events = sum(value["events"] for value in sweeps)
total_queries = sum(value["logical_batch_loss_queries"] for value in sweeps)
total_observations = sum(
value["per_example_causal_observations"] for value in sweeps)
batch = settings["batch_size"]
clean_forward_examples = settings["sweeps"] * batch
perturbation_forward_examples = 2 * total_events * batch
teaching_macs = total_events * batch * net.apical_macs_per_example
work = {
"edge_events": total_events,
"logical_batch_loss_queries": total_queries,
"per_example_causal_observations": total_observations,
"per_example_cross_entropy_terms": 2 * total_observations,
"clean_forward_examples": clean_forward_examples,
"perturbation_forward_examples": perturbation_forward_examples,
"forward_macs": ((clean_forward_examples + perturbation_forward_examples)
* net.forward_macs_per_example),
"hierarchical_teaching_and_local_correlation_macs_estimate": teaching_macs,
}
work["total_macs_estimate"] = (
work["forward_macs"]
+ work["hierarchical_teaching_and_local_correlation_macs_estimate"])
finite_values = [
fixed["early_third_alignment"], fixed["all_layer_alignment"],
learned["early_third_alignment"], learned["all_layer_alignment"],
learned["min_feedback_forward_norm_ratio"],
learned["max_feedback_forward_norm_ratio"],
] + [value[key] for value in sweeps for key in (
"mean_field_prediction_target_cosine",
"mean_parameter_update_rms", "max_parameter_update_rms")]
output = {
"schema_version": 1,
"protocol": "oral_a_v5_layerwise_causal_bootstrap_capture_v1",
"settings": settings,
"provenance": provenance(),
"split": split,
"architecture": {
"family": "CIFAR 6n+2 ResNet, option-A shortcuts",
"forward_parameters": net.n_forward_parameters,
"adaptive_feedback_parameters": net.n_fixed_feedback_parameters,
"forward_macs_per_example": net.forward_macs_per_example,
"feedback_macs_per_example": net.apical_macs_per_example,
},
"method_audit": {
"forward_weight_reads_in_feedback_update": 0,
"reverse_mode_learning_operations": 0,
"causal_query_normalization_state": "evaluation_running_statistics",
"ordinary_task_normalization_state": "not_run_forward_frozen",
"forward_state_max_absolute_difference": (
forward_state_max_difference),
},
"fixed_hfa": fixed,
"learned_lcb": learned,
"sweeps": sweeps,
"work": work,
"wall_seconds": wall_seconds,
"finite": all(math.isfinite(value) for value in finite_values),
"test_examples_touched": 0,
"validation_endpoints_observed": 0,
"hardware": {
"device": str(args.device), "torch_version": torch.__version__,
"cuda_visible_devices": os.environ.get("CUDA_VISIBLE_DEVICES"),
"cuda_device_name": (torch.cuda.get_device_name(torch.device(args.device))
if str(args.device).startswith("cuda") else None),
"peak_memory_allocated_bytes": (
torch.cuda.max_memory_allocated(torch.device(args.device))
if str(args.device).startswith("cuda") else None),
},
}
os.makedirs(os.path.dirname(os.path.abspath(args.out)), exist_ok=True)
with open(args.out, "w") as handle:
json.dump(output, handle, indent=2, sort_keys=True)
handle.write("\n")
print(json.dumps({
"fixed_hfa": fixed, "learned_lcb": learned, "work": work,
"finite": output["finite"], "wall_seconds": wall_seconds,
"out": args.out,
}, indent=2), flush=True)
if __name__ == "__main__":
main()
|
