File size: 9,438 Bytes
189f45b | 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 | """
Phase 0B: single-target bottleneck trainer for kinematics-vs-mechanics.
Same sender architecture as `_killer_experiment.py` (4-agent × 2-head
Gumbel-Softmax, K=5, HIDDEN_DIM=128) but the receiver is a 3-way classifier
that reads a single message and predicts the binned target class.
Takes `--target <name>` and `--dataset <collision|ramp>` and pulls the
corresponding `_bin` column from the label .npz. Feature file and per-dataset
config are selected by `--backbone <vjepa2|dinov2>`.
Run:
/usr/bin/python3 _kinematics_train.py --dataset collision --backbone vjepa2 \
--target mass --seed 0
"""
import argparse, json, math, os, sys, time, warnings
warnings.filterwarnings("ignore")
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
sys.path.insert(0, os.path.dirname(__file__))
from _killer_experiment import (
TemporalEncoder, DiscreteSender, DiscreteMultiSender
)
DEVICE = torch.device("mps" if torch.backends.mps.is_available() else "cpu")
HIDDEN_DIM = 128
VOCAB_SIZE = 5
N_HEADS = 2
N_AGENTS = 4
MSG_DIM = N_AGENTS * N_HEADS * VOCAB_SIZE # 40
N_POS = N_AGENTS * N_HEADS # 8
BATCH_SIZE = 32
SENDER_LR = 1e-3
RECEIVER_LR = 3e-3
EARLY_STOP_PATIENCE = 50 # matches exp4_faithfulness early-stop style
# ── Classifier receiver ──
class ClassifierReceiver(nn.Module):
"""Reads one message and predicts 3 class logits."""
def __init__(self, msg_dim, hidden_dim, n_classes=3):
super().__init__()
self.net = nn.Sequential(
nn.Linear(msg_dim, hidden_dim), nn.ReLU(),
nn.Linear(hidden_dim, hidden_dim // 2), nn.ReLU(),
nn.Linear(hidden_dim // 2, n_classes),
)
def forward(self, msg):
return self.net(msg)
# ── Dataset / label loading ──
FEATURE_FILES = {
("collision", "vjepa2"): "results/vjepa2_collision_pooled.pt", # [600, 24, 1024]
("collision", "dinov2"): "results/collision_dinov2_features.pt", # [600, 24, 384]
("ramp", "vjepa2"): "results/vjepa2_ramp_temporal.pt", # [300, 16, 1024]
("ramp", "dinov2"): "results/phase54b_dino_features.pt", # [300, 8, 384]
}
LABEL_FILES = {
"collision": "results/kinematics_vs_mechanics/labels_collision.npz",
"ramp": "results/kinematics_vs_mechanics/labels_ramp.npz",
}
def load_features(dataset, backbone):
path = FEATURE_FILES[(dataset, backbone)]
d = torch.load(path, weights_only=False, map_location="cpu")
feat = d["features"].float()
return feat
def load_labels(dataset, target):
"""Return bin labels (int [N])."""
z = np.load(LABEL_FILES[dataset])
key = f"{target}_bin"
if key not in z:
raise ValueError(f"Unknown target '{target}' for dataset '{dataset}'. "
f"Available: {[k.replace('_bin', '') for k in z.files if k.endswith('_bin')]}")
return z[key].astype(np.int64)
# ── Training loop ──
def train_one(dataset, backbone, target, seed,
n_epochs=150, verbose=False):
"""
Returns dict: task_acc, posdis, elapsed_s, seed, dataset, backbone, target.
"""
t0 = time.time()
feat = load_features(dataset, backbone) # [N, T, D]
labels = load_labels(dataset, target) # [N]
N, nf, dim = feat.shape
fpa = max(1, nf // N_AGENTS)
agent_views = [feat[:, (i * fpa):(i + 1) * fpa, :] for i in range(N_AGENTS)]
torch.manual_seed(seed)
np.random.seed(seed)
# 80/20 train / holdout split (stratified by target class)
rng = np.random.RandomState(seed * 1000 + 42)
train_ids = []
holdout_ids = []
for c in np.unique(labels):
ids_c = np.where(labels == c)[0]
rng.shuffle(ids_c)
split = max(1, len(ids_c) // 5)
holdout_ids.extend(ids_c[:split])
train_ids.extend(ids_c[split:])
train_ids = np.array(train_ids)
holdout_ids = np.array(holdout_ids)
n_classes = int(labels.max()) + 1
chance = 1.0 / n_classes
senders = [DiscreteSender(TemporalEncoder(HIDDEN_DIM, dim, fpa),
HIDDEN_DIM, VOCAB_SIZE, N_HEADS)
for _ in range(N_AGENTS)]
sender = DiscreteMultiSender(senders).to(DEVICE)
receivers = [ClassifierReceiver(MSG_DIM, HIDDEN_DIM, n_classes).to(DEVICE)
for _ in range(3)]
so = torch.optim.Adam(sender.parameters(), lr=SENDER_LR)
ros = [torch.optim.Adam(r.parameters(), lr=RECEIVER_LR) for r in receivers]
labels_dev = torch.tensor(labels, dtype=torch.long).to(DEVICE)
me = math.log(VOCAB_SIZE)
n_batches = max(1, len(train_ids) // BATCH_SIZE)
best_acc, best_state, best_ep = 0.0, None, 0
for ep in range(n_epochs):
if ep - best_ep > EARLY_STOP_PATIENCE and best_acc > chance + 0.05:
break
if ep > 0 and ep % 40 == 0:
for i in range(len(receivers)):
receivers[i] = ClassifierReceiver(MSG_DIM, HIDDEN_DIM, n_classes).to(DEVICE)
ros[i] = torch.optim.Adam(receivers[i].parameters(), lr=RECEIVER_LR)
sender.train(); [r.train() for r in receivers]
tau = 3.0 + (1.0 - 3.0) * ep / max(1, n_epochs - 1)
hard = ep >= 30
rng_ep = np.random.RandomState(seed * 10000 + ep)
perm = rng_ep.permutation(train_ids)
for b in range(n_batches):
batch_ids = perm[b * BATCH_SIZE:(b + 1) * BATCH_SIZE]
if len(batch_ids) < 4:
continue
views = [v[batch_ids].to(DEVICE) for v in agent_views]
target_batch = labels_dev[batch_ids]
msg, logits_list = sender(views, tau=tau, hard=hard)
loss = torch.tensor(0.0, device=DEVICE)
for r in receivers:
pred = r(msg)
loss = loss + F.cross_entropy(pred, target_batch)
loss = loss / len(receivers)
# Entropy regularisation on each position
for lg in logits_list:
lp = F.log_softmax(lg, -1)
p = lp.exp().clamp(min=1e-8)
ent = -(p * lp).sum(-1).mean()
if ent / me < 0.1:
loss = loss - 0.03 * ent
if torch.isnan(loss):
so.zero_grad(); [o.zero_grad() for o in ros]
continue
so.zero_grad(); [o.zero_grad() for o in ros]
loss.backward()
torch.nn.utils.clip_grad_norm_(sender.parameters(), 1.0)
so.step(); [o.step() for o in ros]
if ep % 50 == 0 and DEVICE.type == "mps":
torch.mps.empty_cache()
# Evaluation every 10 epochs
if (ep + 1) % 10 == 0 or ep == 0:
sender.eval(); [r.eval() for r in receivers]
with torch.no_grad():
v_ho = [v[holdout_ids].to(DEVICE) for v in agent_views]
msg_ho, _ = sender(v_ho)
target_ho = labels_dev[holdout_ids]
best_per_recv = 0.0
for r in receivers:
preds = r(msg_ho).argmax(-1)
acc = (preds == target_ho).float().mean().item()
best_per_recv = max(best_per_recv, acc)
if verbose and ep % 50 == 0:
print(f" ep={ep} holdout_acc={best_per_recv:.1%}",
flush=True)
if best_per_recv > best_acc:
best_acc = best_per_recv
best_ep = ep
best_state = {k: v.cpu().clone()
for k, v in sender.state_dict().items()}
if best_state:
sender.load_state_dict(best_state)
sender.eval()
# Extract tokens at best state for posdis computation
with torch.no_grad():
toks_list = []
for i in range(0, N, BATCH_SIZE):
vs = [v[i:i + BATCH_SIZE].to(DEVICE) for v in agent_views]
_, logits = sender(vs)
toks_list.append(np.stack([l.argmax(-1).cpu().numpy() for l in logits], 1))
tokens = np.concatenate(toks_list, 0)
# Simple PosDis (target only, MI per position)
try:
from _killer_experiment import positional_disentanglement
attrs = np.stack([labels, labels], axis=1)
posdis, _, _ = positional_disentanglement(tokens, attrs, VOCAB_SIZE)
except Exception as e:
posdis = 0.0
return {
"dataset": dataset,
"backbone": backbone,
"target": target,
"seed": int(seed),
"n_classes": int(n_classes),
"chance": float(chance),
"task_acc": float(best_acc),
"posdis": float(posdis),
"elapsed_s": float(time.time() - t0),
"best_ep": int(best_ep),
"n_train": int(len(train_ids)),
"n_holdout": int(len(holdout_ids)),
}
if __name__ == "__main__":
ap = argparse.ArgumentParser()
ap.add_argument("--dataset", required=True, choices=["collision", "ramp"])
ap.add_argument("--backbone", required=True, choices=["vjepa2", "dinov2"])
ap.add_argument("--target", required=True)
ap.add_argument("--seed", type=int, default=0)
ap.add_argument("--epochs", type=int, default=150)
ap.add_argument("--verbose", action="store_true")
args = ap.parse_args()
r = train_one(args.dataset, args.backbone, args.target, args.seed,
n_epochs=args.epochs, verbose=args.verbose)
print(json.dumps(r, indent=2), flush=True)
|