File size: 8,844 Bytes
dd5b02b | 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 | #!/usr/bin/env python3
"""
janus4_temporal_diff.py — Janus 4-way attention reference implementation.
4th attention mechanism: Temporal Diff — attends to CHANGES between positions.
Based on Variant 4 (dedicated wtd+wvd) + Opus analysis fixes:
- Removed distance decay (RoPE handles this in full nanochat)
- Gate init biased against delta (-1.0) — model discovers if/when to use it
- Dedicated projections (no weight sharing with QKV or RRPRAM)
Architecture: QKV (semantic) + RRPRAM (positional) + Echo (self-resonance) + TemporalDiff (change detection)
What TemporalDiff captures that others don't:
- QKV with RoPE: encodes distance between positions, not content of change
- RRPRAM: positional patterns, not transitions
- Echo: self-similarity, not change rate
- TemporalDiff: "where did representation change, and do changes correlate?"
Pure Python, zero deps. For reference/testing. Production = C or PyTorch.
By Arianna Method, 2026-03-25.
"""
import argparse
import math
import random
VOCAB = 256
MAX_T = 48
DIM = 48
HEADS = 4
HD = DIM // HEADS
def bpe_encode(text):
return list(text.encode('utf-8', errors='ignore'))
def bpe_decode(ids):
return bytes([i % 256 for i in ids]).decode('utf-8', errors='ignore')
def rand_mat(r, c, s=0.02):
return [[(random.random() * 2 - 1) * s for _ in range(c)] for _ in range(r)]
def vec_mat(v, m):
out = [0.0] * len(m[0])
for i, vi in enumerate(v):
row = m[i]
for j in range(len(out)):
out[j] += vi * row[j]
return out
def softmax(xs):
mx = max(xs)
ex = [math.exp(x - mx) for x in xs]
s = sum(ex) + 1e-9
return [x / s for x in ex]
class Janus4:
def __init__(self):
self.tok = rand_mat(VOCAB, DIM)
self.pos = rand_mat(MAX_T, DIM)
# QKV — semantic attention
self.wq = rand_mat(DIM, DIM)
self.wk = rand_mat(DIM, DIM)
self.wv = rand_mat(DIM, DIM)
# RRPRAM — positional resonance
self.wr = rand_mat(DIM, MAX_T)
self.wvr = rand_mat(DIM, DIM)
# Echo — self-resonance (W^T * W)
self.wj = rand_mat(DIM, DIM)
# Temporal Diff — dedicated projections (NOT shared with QKV/RRPRAM)
self.wtd = rand_mat(DIM, DIM) # delta key projection
self.wvd = rand_mat(DIM, DIM) # delta value projection
# 4-way gate — delta starts suppressed (Opus recommendation)
self.gate = [0.0, 0.0, 0.0, -1.0] # QKV, RRPRAM, Echo, TemporalDiff
# Output
self.out = rand_mat(DIM, VOCAB)
self.bias = [0.0] * VOCAB
def _dot(self, a, b):
return sum(x * y for x, y in zip(a, b))
def _head(self, v, h):
return v[h * HD:(h + 1) * HD]
def forward(self, ids):
T = len(ids)
x = [[self.tok[ids[t]][e] + self.pos[t][e] for e in range(DIM)] for t in range(T)]
# Precompute all projections
q = [vec_mat(x[t], self.wq) for t in range(T)]
k = [vec_mat(x[t], self.wk) for t in range(T)]
v = [vec_mat(x[t], self.wv) for t in range(T)]
rv = [vec_mat(x[t], self.wvr) for t in range(T)]
je = [vec_mat(x[t], self.wj) for t in range(T)]
# Temporal diff: delta of input
dx = [[0.0] * DIM for _ in range(T)]
for t in range(1, T):
for e in range(DIM):
dx[t][e] = x[t][e] - x[t - 1][e]
# Dedicated projections for delta (not shared!)
dk = [vec_mat(dx[t], self.wtd) for t in range(T)] # delta keys
dv = [vec_mat(dx[t], self.wvd) for t in range(T)] # delta values
g = softmax(self.gate)
cat = [[0.0] * DIM for _ in range(T)]
for h in range(HEADS):
# 1) QKV attention — semantic content matching
a1 = [[-1e9] * T for _ in range(T)]
for i in range(T):
qi = self._head(q[i], h)
for j in range(i + 1):
a1[i][j] = self._dot(qi, self._head(k[j], h)) / math.sqrt(HD)
a1[i] = softmax(a1[i])
ho = [[0.0] * HD for _ in range(T)]
for i in range(T):
for j in range(T):
vv = self._head(v[j], h)
for d in range(HD):
ho[i][d] += a1[i][j] * vv[d]
# 2) RRPRAM — positional resonance
a2 = [[-1e9] * T for _ in range(T)]
for i in range(T):
for j in range(i + 1):
a2[i][j] = sum(x[i][e] * self.wr[e][j] for e in range(DIM)) / math.sqrt(HD)
a2[i] = softmax(a2[i])
ro = [[0.0] * HD for _ in range(T)]
for i in range(T):
for j in range(T):
rvh = self._head(rv[j], h)
for d in range(HD):
ro[i][d] += a2[i][j] * rvh[d]
# 3) Echo — self-resonance (W^T * W)
a3 = [[-1e9] * T for _ in range(T)]
for i in range(T):
ei = self._head(je[i], h)
for j in range(i + 1):
a3[i][j] = self._dot(ei, self._head(je[j], h)) / math.sqrt(HD)
a3[i] = softmax(a3[i])
jo = [[0.0] * HD for _ in range(T)]
for i in range(T):
for j in range(T):
ej = self._head(je[j], h)
for d in range(HD):
jo[i][d] += a3[i][j] * ej[d]
# 4) Temporal Diff — change detection attention
# No distance decay (Opus fix: RoPE handles this in full implementation)
a4 = [[-1e9] * T for _ in range(T)]
for i in range(T):
dki = self._head(dk[i], h)
for j in range(i + 1):
a4[i][j] = self._dot(dki, self._head(dk[j], h)) / math.sqrt(HD)
a4[i] = softmax(a4[i])
to = [[0.0] * HD for _ in range(T)]
for i in range(T):
for j in range(T):
dvh = self._head(dv[j], h)
for d in range(HD):
to[i][d] += a4[i][j] * dvh[d]
# Gate blend — 4-way softmax
for t in range(T):
base = h * HD
for d in range(HD):
cat[t][base + d] = (g[0] * ho[t][d] + g[1] * ro[t][d] +
g[2] * jo[t][d] + g[3] * to[t][d])
logits = [[0.0] * VOCAB for _ in range(T)]
for t in range(T):
for vi in range(VOCAB):
logits[t][vi] = sum(cat[t][e] * self.out[e][vi] for e in range(DIM)) + self.bias[vi]
return logits, cat
def train_step(self, tok, tgt, lr):
logits, cat = self.forward(tok)
loss = 0.0
grad = [[0.0] * VOCAB for _ in range(len(tok))]
for t in range(len(tok)):
p = softmax(logits[t])
loss -= math.log(max(1e-9, p[tgt[t]]))
for vi in range(VOCAB):
grad[t][vi] = p[vi]
grad[t][tgt[t]] -= 1.0
loss /= len(tok)
# Gradient on output layer only (reference impl)
for t in range(len(tok)):
for e in range(DIM):
ce = cat[t][e]
if ce == 0.0:
continue
row = self.out[e]
for vi in range(VOCAB):
row[vi] -= lr * ce * grad[t][vi] / len(tok)
for vi in range(VOCAB):
self.bias[vi] -= lr * grad[t][vi] / len(tok)
return loss
def generate(model, prompt, n=60):
ids = bpe_encode(prompt)[-MAX_T:]
for _ in range(n):
logits, _ = model.forward(ids)
p = softmax(logits[-1])
ids.append(max(range(VOCAB), key=lambda i: p[i]))
ids = ids[-MAX_T:]
return bpe_decode(ids)
def train(model, text, steps, lr):
ids = bpe_encode(text)
losses = []
for step in range(1, steps + 1):
off = random.randint(0, max(0, len(ids) - MAX_T - 2))
tok = ids[off:off + MAX_T]
tgt = ids[off + 1:off + MAX_T + 1]
losses.append(model.train_step(tok, tgt, lr))
if step % 10 == 0:
print(f"step {step:4d}/{steps} loss={losses[-1]:.4f}")
return losses
if __name__ == '__main__':
ap = argparse.ArgumentParser()
ap.add_argument('--train', type=str)
ap.add_argument('--steps', type=int, default=40)
ap.add_argument('--lr', type=float, default=0.05)
ap.add_argument('--generate', type=str)
args = ap.parse_args()
random.seed(42)
m = Janus4()
if args.train:
txt = open(args.train, 'r', encoding='utf-8', errors='ignore').read()
losses = train(m, txt, args.steps, args.lr)
print(f'loss_start={losses[0]:.4f} loss_end={losses[-1]:.4f}')
if args.generate:
print(generate(m, args.generate))
|