File size: 6,207 Bytes
27c7eb0 | 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 | """
pc_mlp_tiny.py — Predictive-Coding MLP for sequence classification.
3,138 parameters. Matches a 5,314-param transformer on the synthetic
palindrome + position task at 59% of the parameter count.
Run: python pc_mlp_tiny.py
"""
import os, time, math
os.environ.setdefault("XLA_FLAGS",
"--xla_cpu_multi_thread_eigen=true intra_op_parallelism_threads=8")
import numpy as np
import jax, jax.numpy as jnp
from jax import jit, random
# ============================================================
# Config
# ============================================================
VOCAB = 16
SEQ = 16
CLASSES = 2
D_HIDDEN = 32
LOCAL_LOSS_WEIGHT = 0.1
# ============================================================
# Task
# ============================================================
def make_batch(key, B):
"""Return (x, y) where y=1 iff x[0] >= VOCAB/2 OR x is a palindrome."""
x = random.randint(key, (B, SEQ), 0, VOCAB)
local = (x[:, 0] >= VOCAB // 2)
pal = jnp.all(x == x[:, ::-1], axis=1)
return x, (local | pal).astype(jnp.int32)
def ce_loss(logits, y):
return -jnp.take_along_axis(
jax.nn.log_softmax(logits, -1), y[:, None], -1).mean()
# ============================================================
# Model
# ============================================================
def pc_init(key, d=VOCAB, hidden=D_HIDDEN):
"""Initialize PC-MLP parameters."""
k = random.split(key, 5)
scale = 1.0 / math.sqrt(hidden)
return {
"tok": random.normal(k[0], (VOCAB, d)) * 0.05,
"pos": random.normal(k[1], (SEQ, d)) * 0.05,
"W1": random.normal(k[2], (d, hidden)) * scale,
"W2": random.normal(k[3], (hidden, hidden)) * scale,
"head": {"w": random.normal(k[4], (hidden, CLASSES)) * 0.02,
"b": jnp.zeros(CLASSES)},
}
def pc_forward(p, x):
"""Forward pass. Returns (logits, h0, h1, h2)."""
B, T = x.shape
h = p["tok"][x] + p["pos"][:T][None] # (B, T, d)
h = h.reshape(B, T * p["tok"].shape[-1])
h = h[:, :p["W1"].shape[0]] # (B, d)
h0 = h
h1 = jax.nn.gelu(h0 @ p["W1"])
h2 = jax.nn.gelu(h1 @ p["W2"])
logits = h2 @ p["head"]["w"] + p["head"]["b"]
return logits, h0, h1, h2
def pc_loss(p, x, y, lam=LOCAL_LOSS_WEIGHT):
"""Global CE loss + local predictive-coding regularizer."""
logits, h0, h1, h2 = pc_forward(p, x)
global_loss = ce_loss(logits, y)
local_loss = (jnp.mean((h1.mean(1) - h0.mean(1)) ** 2)
+ jnp.mean((h2.mean(1) - h1.mean(1)) ** 2))
return global_loss + lam * local_loss
# ============================================================
# Training
# ============================================================
def train(steps=200, B=32, seed=0, lr=3e-3):
key = random.key(seed)
p = pc_init(key)
opt = {"m": jax.tree.map(jnp.zeros_like, p),
"v": jax.tree.map(jnp.zeros_like, p),
"t": jnp.int32(0)}
def loss_fn(p, x, y):
return pc_loss(p, x, y)
@jit
def step(p, opt, x, y):
l, g = jax.value_and_grad(loss_fn)(p, x, y)
t = opt["t"] + 1
m = jax.tree.map(lambda m, g: 0.9*m + 0.1*g, opt["m"], g)
v = jax.tree.map(lambda v, g: 0.999*v + 0.001*g*g, opt["v"], g)
mh = jax.tree.map(lambda m: m / (1 - 0.9**t), m)
vh = jax.tree.map(lambda v: v / (1 - 0.999**t), v)
np_ = jax.tree.map(
lambda p, mh, vh: p - lr*mh / (jnp.sqrt(vh) + 1e-8),
p, mh, vh)
return np_, {"m": m, "v": v, "t": t}, l
t0 = time.perf_counter()
for s in range(steps):
key, kb = random.split(key)
xb, yb = make_batch(kb, B)
p, opt, l = step(p, opt, xb, yb)
wall = time.perf_counter() - t0
# eval
key, kv = random.split(key)
xv, yv = make_batch(kv, 512)
logits, *_ = pc_forward(p, xv)
acc = float((logits.argmax(-1) == yv).mean())
val_loss = float(ce_loss(logits, yv))
n = sum(int(np.prod(v.shape)) for v in jax.tree.leaves(p)
if hasattr(v, "shape"))
return p, {"wall_s": wall, "params": n,
"val_acc": acc, "val_loss": val_loss,
"train_loss": float(l)}
# ============================================================
# Activation health diagnostic
# ============================================================
def debug_activations(p):
"""Print mean/std at each nonlinearity. Catches dead layers."""
key = random.key(99)
x, _ = make_batch(key, 32)
h = p["tok"][x] + p["pos"][:x.shape[1]][None]
h = h.reshape(x.shape[0], -1)[:, :p["W1"].shape[0]]
pre1 = h @ p["W1"]
h1 = jax.nn.gelu(pre1)
pre2 = h1 @ p["W2"]
h2 = jax.nn.gelu(pre2)
print(" --- activation health ---")
print(f" pre1 (input to W1) std={float(pre1.std()):.4f}")
print(f" h1 (after GELU) std={float(h1.std()):.4f}")
print(f" pre2 (input to W2) std={float(pre2.std()):.4f}")
print(f" h2 (after GELU) std={float(h2.std()):.4f}")
# dead-layer check: if any std is < 1e-4, the layer is collapsed
for name, val in [("pre1", pre1), ("h1", h1), ("pre2", pre2), ("h2", h2)]:
if float(val.std()) < 1e-4:
print(f" WARNING: {name} is collapsed (std < 1e-4)")
# local loss terms
print(f" local loss (h1−h0) {float(jnp.mean((h1.mean(1)-h.mean(1))**2)):.6f}")
print(f" local loss (h2−h1) {float(jnp.mean((h2.mean(1)-h1.mean(1))**2)):.6f}")
# ============================================================
# Main
# ============================================================
if __name__ == "__main__":
print("=" * 60)
print("PC-MLP: Predictive-Coding MLP for sequence classification")
print("=" * 60)
p, r = train(steps=200, seed=1)
print(f"\n params: {r['params']:,}")
print(f" wall: {r['wall_s']:.2f}s")
print(f" train loss:{r['train_loss']:.3f}")
print(f" val loss: {r['val_loss']:.3f}")
print(f" val acc: {r['val_acc']:.3f}")
print()
debug_activations(p) |