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
|
"""
SDIL main training / diagnostics driver.
Trains one of {bp, dfa, sdil} (sdil with ablation flags) on MNIST/FashionMNIST,
logging the quantities that actually test the hypothesis:
- train loss, test accuracy
- per-hidden-layer cos(innovation r_l, -grad h_l) <- the headline metric
- cos(raw apical a_l, -grad) and cos(A_l c, -grad) <- residualization ablation
- single-step loss-decrease ratio vs exact GD
Everything is JSON-logged for later plotting.
"""
import argparse
import json
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.core import SDILNet, SDILConfig, sdil_step, neutral_p_update
from sdil.baselines import BPNet, dfa_config, evaluate
from sdil.local_baselines import FANet
from sdil import probes
from sdil.data import (get_dataset, onehot, make_hierarchical,
make_teacher_student, make_tentmap)
REAL_DATASETS = ("mnist", "fmnist", "cifar10")
SYNTHETIC_DATASETS = ("teacher", "hierarchical", "tentmap")
def code_provenance():
"""Best-effort source revision metadata for reproducible result files."""
root = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
try:
commit = subprocess.run(
["git", "rev-parse", "HEAD"], cwd=root, check=True,
capture_output=True, text=True).stdout.strip()
dirty = bool(subprocess.run(
["git", "status", "--porcelain", "--untracked-files=no"], cwd=root,
check=True, capture_output=True, text=True).stdout.strip())
return {"git_commit": commit, "git_dirty": dirty}
except (OSError, subprocess.CalledProcessError):
return {"git_commit": None, "git_dirty": None}
def build(args, device):
sizes = [args.n_in] + [args.width] * args.depth + [args.n_out]
if args.mode == "bp":
net = BPNet(sizes, act=args.act, device=device, seed=args.seed,
w_scale=args.w_scale, nuis_rho=0.0, residual=bool(args.residual),
predictor_mode=args.predictor_mode)
cfg = SDILConfig(eta=args.eta, momentum=args.momentum)
return net, cfg
if args.mode == "fa":
net = FANet(sizes, act=args.act, device=device, seed=args.seed,
w_scale=args.w_scale, nuis_rho=0.0,
residual=bool(args.residual),
predictor_mode=args.predictor_mode,
b_scale=args.feedback_scale)
return net, SDILConfig(eta=args.eta, momentum=args.momentum)
net = SDILNet(sizes, act=args.act, device=device, seed=args.seed,
w_scale=args.w_scale, a_scale=args.a_scale,
nuis_rho=args.nuis_rho, feedback=args.feedback,
residual=bool(args.residual), predictor_mode=args.predictor_mode)
if args.mode == "dfa":
cfg = dfa_config(eta=args.eta, momentum=args.momentum)
elif args.mode == "sdil":
cfg = SDILConfig(
eta=args.eta, eta_A=args.eta_A, eta_P=args.eta_P,
use_residual=bool(args.use_residual), learn_A=bool(args.learn_A),
learn_P=bool(args.learn_P), pert_sigma=args.pert_sigma,
pert_every=args.pert_every, pert_ndirs=args.pert_ndirs,
pert_mode=args.pert_mode,
momentum=args.momentum, settle_steps=args.settle_steps,
kappa=args.kappa, feedback=args.feedback,
p_update_on_neutral=bool(args.p_neutral),
normalize_delta=bool(args.normalize_delta),
raw_scale_control=args.raw_scale_control)
else:
raise ValueError(args.mode)
return net, cfg
def load_task(args, device):
"""Load a real dataset or construct one fixed synthetic task.
``task_seed`` controls examples and the target function; ``seed`` controls
only the student initialization. A depth sweep therefore compares models
on exactly the same compositional problem instead of silently changing the
teacher between model seeds.
"""
if args.dataset in REAL_DATASETS:
return get_dataset(
args.dataset, batch_size=args.batch_size, device=device,
train_limit=args.train_examples or None)
common = dict(
n_train=args.task_train_examples,
n_test=args.task_test_examples,
seed=args.task_seed,
batch_size=args.batch_size,
device=device,
)
if args.dataset == "teacher":
return make_teacher_student(
n_in=args.task_n_in, n_classes=args.task_classes,
t_depth=args.teacher_depth, t_width=args.teacher_width,
residual=bool(args.teacher_residual), **common)
if args.dataset == "hierarchical":
return make_hierarchical(
levels=args.task_levels, n_classes=args.task_classes, **common)
if args.dataset == "tentmap":
return make_tentmap(
levels=args.task_levels, n_in=args.task_n_in, **common)
raise ValueError(f"unknown dataset: {args.dataset}")
def train(args):
device = args.device
torch.manual_seed(args.seed)
train_loader, test_loader, n_in, n_out = load_task(args, device)
args.n_in, args.n_out = n_in, n_out
net, cfg = build(args, device)
# a fixed probe batch for stable alignment tracking
px, py = next(iter(test_loader))
px, py = px[:args.probe_bs].to(device), py[:args.probe_bs].to(device)
poh = onehot(py, n_out, device=device)
log = {"args": vars(args), "provenance": code_provenance(), "steps": [], "final": {}}
step = 0
prev_error = None
t0 = time.time()
# predictor warmup on neutral-period (c=0) drive, so P cancels the apical
# nuisance before task plasticity relies on the residual (no-op when rho=0).
if args.mode == "sdil" and args.learn_P and args.p_warmup_steps > 0 and args.nuis_rho > 0:
it = iter(train_loader)
for _ in range(args.p_warmup_steps):
try:
wx, _ = next(it)
except StopIteration:
it = iter(train_loader)
wx, _ = next(it)
neutral_p_update(net, wx.to(device), args.p_warmup_eta)
for epoch in range(args.epochs):
for x, y in train_loader:
x, y = x.to(device), y.to(device)
yoh = onehot(y, n_out, device=device)
if args.mode == "bp":
loss = net.bp_step(x, y, cfg.eta, momentum=cfg.momentum)
elif args.mode == "fa":
loss = net.fa_step(x, y, yoh, cfg.eta, momentum=cfg.momentum)
else:
loss, aux = sdil_step(net, x, y, yoh, cfg, step, prev_error=prev_error)
prev_error = aux["error"]
if step % args.log_every == 0:
rec = {"step": step, "epoch": epoch, "train_loss": float(loss)}
if args.mode == "fa":
rec.update(probes.fa_alignment_report(net, px, py, poh))
elif args.mode != "bp":
al = probes.alignment_report(net, px, py, poh, cfg)
rec["cos_r_negg"] = al["cos_r_negg"]
rec["cos_innovation_negg"] = al["cos_innovation_negg"]
rec["cos_apical_negg"] = al["cos_apical_negg"]
rec["cos_Ac_negg"] = al["cos_Ac_negg"]
rec["r_norm"] = al["r_norm"]
if args.mode == "sdil" and step % (args.log_every * 5) == 0:
rec["ldr"] = probes.loss_decrease_ratio(net, px, py, poh, cfg, step)
log["steps"].append(rec)
step += 1
if args.max_steps and step >= args.max_steps:
break
if args.max_steps and step >= args.max_steps:
break
acc, tloss = evaluate(net, test_loader)
msg = f"[{args.tag}] epoch {epoch} step {step} loss {loss:.4f} test_acc {acc:.4f}"
if args.mode == "fa":
al = probes.fa_alignment_report(net, px, py, poh)
meancos = sum(al["cos_fa_negg"]) / len(al["cos_fa_negg"])
msg += f" mean_cos(fa,-g) {meancos:+.3f} per-layer {['%.2f'%v for v in al['cos_fa_negg']]}"
elif args.mode != "bp":
al = probes.alignment_report(net, px, py, poh, cfg)
meancos = sum(al["cos_r_negg"]) / len(al["cos_r_negg"])
msg += f" mean_cos(r,-g) {meancos:+.3f} per-layer {['%.2f'%v for v in al['cos_r_negg']]}"
print(msg, flush=True)
log["steps"].append({"epoch_end": epoch, "step": step, "test_acc": acc, "test_loss": tloss})
acc, tloss = evaluate(net, test_loader)
log["final"] = {"test_acc": acc, "test_loss": tloss, "wall_s": time.time() - t0}
if args.mode == "fa":
log["final"].update(probes.fa_alignment_report(net, px, py, poh))
elif args.mode != "bp":
al = probes.alignment_report(net, px, py, poh, cfg)
log["final"]["cos_r_negg"] = al["cos_r_negg"]
log["final"]["cos_innovation_negg"] = al["cos_innovation_negg"]
log["final"]["cos_apical_negg"] = al["cos_apical_negg"]
log["final"]["cos_Ac_negg"] = al["cos_Ac_negg"]
os.makedirs(args.outdir, exist_ok=True)
outpath = os.path.join(args.outdir, f"{args.tag}.json")
with open(outpath, "w") as f:
json.dump(log, f)
print(f"[{args.tag}] DONE test_acc={acc:.4f} -> {outpath}", flush=True)
return log
def get_args():
p = argparse.ArgumentParser()
p.add_argument("--mode", default="sdil", choices=["bp", "fa", "dfa", "sdil"])
p.add_argument("--dataset", default="mnist",
choices=list(REAL_DATASETS + SYNTHETIC_DATASETS))
p.add_argument("--depth", type=int, default=3) # hidden layers
p.add_argument("--width", type=int, default=256)
p.add_argument("--act", default="tanh", choices=["tanh", "gelu", "silu", "relu"])
p.add_argument("--residual", type=int, default=0) # skip connections (deep no-BN)
p.add_argument("--epochs", type=int, default=15)
p.add_argument("--batch_size", type=int, default=128)
p.add_argument("--train_examples", type=int, default=0,
help="0 uses the full training split")
p.add_argument("--task_seed", type=int, default=0,
help="fixed target/data seed, separate from student --seed")
p.add_argument("--task_train_examples", type=int, default=50000)
p.add_argument("--task_test_examples", type=int, default=10000)
p.add_argument("--task_levels", type=int, default=8)
p.add_argument("--task_n_in", type=int, default=128)
p.add_argument("--task_classes", type=int, default=10)
p.add_argument("--teacher_depth", type=int, default=8)
p.add_argument("--teacher_width", type=int, default=64)
p.add_argument("--teacher_residual", type=int, default=1)
p.add_argument("--eta", type=float, default=0.05)
p.add_argument("--eta_A", type=float, default=0.02)
p.add_argument("--eta_P", type=float, default=0.002)
p.add_argument("--momentum", type=float, default=0.9)
p.add_argument("--w_scale", type=float, default=1.0)
p.add_argument("--a_scale", type=float, default=1.0)
p.add_argument("--feedback_scale", type=float, default=1.0)
p.add_argument("--pert_sigma", type=float, default=1e-2)
p.add_argument("--pert_every", type=int, default=4)
p.add_argument("--pert_ndirs", type=int, default=4)
p.add_argument("--pert_mode", default="layerwise", choices=["layerwise", "simultaneous"])
p.add_argument("--use_residual", type=int, default=1)
p.add_argument("--raw_scale_control", default="none",
choices=["none", "match_innovation_norm"])
p.add_argument("--learn_A", type=int, default=1)
p.add_argument("--learn_P", type=int, default=1)
p.add_argument("--p_neutral", type=int, default=1) # P update on neutral (c=0) drive
p.add_argument("--p_warmup_steps", type=int, default=200) # pre-task neutral P warmup
p.add_argument("--p_warmup_eta", type=float, default=0.05)
p.add_argument("--nuis_rho", type=float, default=0.0)
p.add_argument("--predictor_mode", default="diagonal", choices=["diagonal", "full"])
p.add_argument("--normalize_delta", type=int, default=0)
p.add_argument("--settle_steps", type=int, default=0)
p.add_argument("--kappa", type=float, default=0.0)
p.add_argument("--feedback", default="error", choices=["error", "error_deriv"])
p.add_argument("--seed", type=int, default=0)
p.add_argument("--device", default="cuda" if torch.cuda.is_available() else "cpu")
p.add_argument("--log_every", type=int, default=50)
p.add_argument("--max_steps", type=int, default=0) # 0 = no cap (smoke only)
p.add_argument("--probe_bs", type=int, default=512)
p.add_argument("--outdir", default="results")
p.add_argument("--tag", default="sdil_run")
return p.parse_args()
if __name__ == "__main__":
train(get_args())
|
