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
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
//! The node type — pure metadata (ports, kind, dim, activation).
//!
//! Nodes hold no tensors; execution happens in
//! [`Network`](crate::graph::network::Network). The NAS knobs live here:
//! `hidden_dim`, `activation`, `combine_op`, `standardize`.
use flodl::Variable;
use serde::{Deserialize, Serialize};
/// How multiple incoming tensors are combined before the node transforms them.
#[derive(Clone, Copy, Debug, PartialEq, Eq, Hash, Serialize, Deserialize)]
pub enum CombineOp {
/// Sum of incoming tensors.
Add,
/// Average of incoming tensors.
Mean,
/// Element-wise product of incoming tensors.
Multiply,
/// Element-wise subtraction: first - second - third - ...
Subtract,
/// Element-wise division: first / second / third - ...
Divide,
/// Element-wise maximum across incoming tensors.
Max,
/// Element-wise minimum across incoming tensors.
Min,
}
impl CombineOp {
/// Apply this combine operation to a slice of tensors.
pub fn apply(&self, tensors: &[Variable]) -> flodl::tensor::Result<Variable> {
match self {
CombineOp::Add => {
let mut result = tensors[0].clone();
for t in &tensors[1..] {
result = result.add(t)?;
}
Ok(result)
}
CombineOp::Mean => {
let sum = CombineOp::Add.apply(tensors)?;
sum.mul_scalar(1.0 / tensors.len() as f64)
}
CombineOp::Multiply => {
let mut result = tensors[0].clone();
for t in &tensors[1..] {
result = result.mul(t)?;
}
Ok(result)
}
CombineOp::Subtract => {
let mut result = tensors[0].clone();
for t in &tensors[1..] {
result = result.sub(t)?;
}
Ok(result)
}
CombineOp::Divide => {
let mut result = tensors[0].clone();
for t in &tensors[1..] {
result = result.div(t)?;
}
Ok(result)
}
CombineOp::Max => {
let mut result = tensors[0].clone();
for t in &tensors[1..] {
result = result.maximum(t)?;
}
Ok(result)
}
CombineOp::Min => {
let mut result = tensors[0].clone();
for t in &tensors[1..] {
result = result.minimum(t)?;
}
Ok(result)
}
}
}
}
/// Activation applied after a node's linear transform.
#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash, Default, Serialize, Deserialize)]
pub enum Activation {
/// No activation -- pure linear.
#[default]
Identity,
/// max(0, x) -- sparse, efficient.
ReLU,
/// x * Phi(x) -- smooth ReLU approximation.
GeLU,
/// x * sigmoid(x) -- self-gated, smooth.
SiLU,
/// Self-normalizing ELU -- preserves mean/variance.
SELU,
/// Hyperbolic tangent -- output in (-1, 1).
Tanh,
/// 1/(1+e^-x) -- output in (0, 1), for gating.
Sigmoid,
/// x * tanh(softplus(x)) -- smooth, non-monotonic.
Mish,
/// max(0.01x, x) -- leaky ReLU with slope 0.01.
LeakyReLU,
/// x if x>0, else e^x - 1 -- smooth negative branch.
ELU,
/// GeLU via tanh approximation -- faster than exact GeLU.
GeluTanh,
/// log(1 + e^x) -- smooth ReLU, always positive.
Softplus,
/// x * ReLU6(x+3)/6 -- efficient GeLU variant.
HardSwish,
/// ReLU6(x+3)/6 -- efficient Sigmoid approximation.
HardSigmoid,
/// sin(x) -- periodic, bounded in (-1, 1).
Sin,
/// cos(x) -- periodic, bounded in (-1, 1).
Cos,
/// softmax(x) over the feature dim -- outputs sum to 1 per sample.
/// Gating-style transform: turns a node into a competitive mixer over
/// its channels (winner-take-most signal reweighting).
Softmax,
/// log(softmax(x)) over the feature dim -- numerically stable log of
/// Softmax. Pairs with NLL-style downstream losses; strictly negative.
LogSoftmax,
}
impl Activation {
/// Apply this activation to a tensor, propagating flodl errors.
pub fn apply(&self, x: &Variable) -> flodl::tensor::Result<Variable> {
match self {
Activation::Identity => Ok(x.clone()),
Activation::ReLU => x.relu(),
Activation::GeLU => x.gelu(),
Activation::SiLU => x.silu(),
Activation::SELU => x.selu(),
Activation::Tanh => x.tanh(),
Activation::Sigmoid => x.sigmoid(),
Activation::Mish => x.mish(),
Activation::LeakyReLU => x.leaky_relu(0.01),
Activation::ELU => x.elu(1.0),
Activation::GeluTanh => x.gelu_tanh(),
Activation::Softplus => x.softplus(1.0, 20.0),
Activation::HardSwish => x.hardswish(),
Activation::HardSigmoid => x.hardsigmoid(),
Activation::Sin => x.sin(),
Activation::Cos => x.cos(),
// Dim -1 = feature dim: matches LayerNorm's axis convention, so a
// node's normalization/activation both operate per-sample.
Activation::Softmax => x.softmax(-1),
Activation::LogSoftmax => x.log_softmax(-1),
}
}
}
/// Normalization after linear, before activation. Part of per-node NAS knobs.
#[derive(Debug, Clone, Copy, Default, PartialEq, Eq, Hash, Serialize, Deserialize)]
pub enum StandardizeOp {
/// No normalization — the linear output passes straight to activation.
#[default]
Identity,
/// Layer normalization over the feature dimension.
LayerNorm,
/// RMS normalization over the feature dimension -- scales by the root
/// mean square (no mean subtraction, no learnable params). Cheaper than
/// LayerNorm, often equal quality on tabular. Stateless: replay-safe.
RmsNorm,
/// Instance normalization over the feature dimension -- per-sample z-score
/// (the affine/running-stats variant is deliberately NOT used: learnable
/// params + stateful stats would break the replay contract). Stateless:
/// replay-safe.
InstanceNorm,
}
impl StandardizeOp {
/// Apply this standardize op. All variants are pure `f(x)` — stateless,
/// no learnable params — so the replay/catch-up contract holds.
pub fn apply(&self, x: &Variable) -> flodl::tensor::Result<Variable> {
match self {
StandardizeOp::Identity => Ok(x.clone()),
StandardizeOp::LayerNorm => {
// z-score across feature dim: (x - mean) / sqrt(var + eps)
let mean = x.mean_dim(-1, true)?; // [batch, 1]
let centered = x.sub(&mean)?;
let var = centered.mul(¢ered)?.mean_dim(-1, true)?; // [batch, 1]
let std = var.add_scalar(1e-5)?.sqrt()?; // [batch, 1]
let normed = centered.div(&std)?;
Ok(normed)
}
StandardizeOp::RmsNorm => {
// x / sqrt(mean(x²) + eps) — no centering, scale-only
let ms = x.mul(x)?.mean_dim(-1, true)?; // [batch, 1]
let rms = ms.add_scalar(1e-5)?.sqrt()?;
Ok(x.div(&rms)?)
}
StandardizeOp::InstanceNorm => {
// LayerNorm without the centering-vs-variance asymmetry... it
// IS the z-score here, per-sample over the feature dim. The
// flodl nn::instance_norm module carries running stats (state!) —
// hand-rolled instead, matching LayerNorm's pure-f(x) shape.
let mean = x.mean_dim(-1, true)?;
let centered = x.sub(&mean)?;
let var = centered.mul(¢ered)?.mean_dim(-1, true)?;
let std = var.add_scalar(1e-5)?.sqrt()?;
Ok(centered.div(&std)?)
}
}
}
}
/// A node in the computational graph . Receives tensors, transforms,
/// applies activation, exposes outputs. Invariants enforced by `Topology::validate`.
#[derive(Debug, Clone, PartialEq, Eq, Hash, Serialize, Deserialize)]
pub struct Node {
pub id: usize, // unique id; also execution order (0 runs first)
pub num_inputs: usize, // how many input ports this node has
pub num_outputs: usize, // how many output ports this node has
pub kind: NodeKind, // role: Input / Hidden / Output
/// Per-node feature-dimension override for the layer's *output*
/// (`None` = inherit the graph's `hidden_dim`). The layer's *input* dim
/// is derived from its sources at build time. This is the knob NAS
/// evolution will mutate to grow/shrink the network channel-wise.
pub hidden_dim: Option<usize>,
/// Activation applied after this node's linear transform.
pub activation: Activation,
/// Per-node combine override: how this node merges its incoming tensors
/// (`None` = inherit the graph's `combine_op`). `#[serde(default)]` keeps
/// older topology JSON (no field) loadable.
#[serde(default)]
pub combine_op: Option<CombineOp>,
/// Per-node standardize op: normalization applied after linear, before
/// activation (`None` = inherit the graph's `standardize_op`).
#[serde(default)]
pub standardize: Option<StandardizeOp>,
}
#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash, Serialize, Deserialize)]
pub enum NodeKind {
Input, // start of the network: no inputs, feeds the rest
Hidden, // middle of the network: combine -> transform -> pass on
Output, // end of the network: its output becomes the network output
}
impl Node {
/// Create an input node: 0 inputs, `num_outputs` outputs.
pub fn new_input(id: usize, num_outputs: usize) -> Self {
Node {
id,
num_inputs: 0,
num_outputs,
kind: NodeKind::Input,
hidden_dim: None,
activation: Activation::Identity,
combine_op: None,
standardize: None,
}
}
/// Create a hidden node.
pub fn new_hidden(id: usize, num_inputs: usize, num_outputs: usize) -> Self {
Node {
id,
num_inputs,
num_outputs,
kind: NodeKind::Hidden,
hidden_dim: None,
activation: Activation::Identity,
combine_op: None,
standardize: None,
}
}
/// Create an output node.
pub fn new_output(id: usize, num_inputs: usize, num_outputs: usize) -> Self {
Node {
id,
num_inputs,
num_outputs,
kind: NodeKind::Output,
hidden_dim: None,
activation: Activation::Identity,
combine_op: None,
standardize: None,
}
}
/// Set activation (builder style).
pub fn with_activation(mut self, activation: Activation) -> Self {
self.activation = activation;
self
}
/// Set combine-op override (builder style).
pub fn with_combine_op(mut self, combine_op: CombineOp) -> Self {
self.combine_op = Some(combine_op);
self
}
/// Set per-node channel-width override (builder style).
pub fn with_hidden_dim(mut self, hidden_dim: usize) -> Self {
self.hidden_dim = Some(hidden_dim);
self
}
}
#[cfg(test)]
mod tests {
use super::*;
use crate::graph::topology::{Topology, TopologyOptions};
use proptest::prelude::*;
#[test]
fn new_pool_ops_are_numerically_sound() {
// [batch=2, features=4] input with known structure.
let x = flodl::Tensor::from_f32(
&[1.0, 2.0, 3.0, 4.0, -1.0, 0.0, 1.0, 2.0],
&[2, 4],
flodl::Device::CPU,
)
.unwrap();
let x = flodl::Variable::new(x, false);
// Softmax over features: rows sum to 1, all positive.
let s = Activation::Softmax.apply(&x).unwrap();
let sums = s.data().sum().unwrap().item().unwrap() as f32;
assert!((sums - 2.0).abs() < 1e-4, "softmax rows must sum to 1 each (got {sums})");
// LogSoftmax: log of softmax — strictly non-positive, exp sums to 1.
let ls = Activation::LogSoftmax.apply(&x).unwrap();
let recon = ls.data().exp().unwrap().sum().unwrap().item().unwrap() as f32;
assert!((recon - 2.0).abs() < 1e-4, "exp(log_softmax) must sum to 1 per row (got {recon})");
// RMSNorm: per-row RMS of output ≈ 1.
let r = StandardizeOp::RmsNorm.apply(&x).unwrap();
let ms = r
.data()
.mul(&r.data())
.unwrap()
.mean()
.unwrap()
.item()
.unwrap() as f32;
let rms = ms.sqrt();
assert!((rms - 1.0).abs() < 1e-3, "rmsnorm output RMS ≈ 1 (got {rms})");
// InstanceNorm (hand-rolled, stateless): per-row mean 0, std 1.
let i = StandardizeOp::InstanceNorm.apply(&x).unwrap();
let m = i.data().mean().unwrap().item().unwrap() as f32;
assert!(m.abs() < 1e-5, "instancenorm per-row mean ≈ 0 (got {m})");
}
/// Arbitrary valid node metadata (port counts, kind, id).
fn node_strategy() -> impl Strategy<Value = Node> {
(0usize..100, 0usize..8, 0usize..8, 0usize..3).prop_map(
|(id, num_inputs, num_outputs, kind)| Node {
id,
num_inputs,
num_outputs,
kind: match kind {
0 => NodeKind::Input,
1 => NodeKind::Hidden,
_ => NodeKind::Output,
},
hidden_dim: None,
activation: Activation::Identity,
combine_op: None,
standardize: None,
},
)
}
#[test]
fn test_new_node_inputs() {
let node: Node = Node::new_input(1, 2);
assert_eq!(node.num_inputs, 0);
assert_eq!(node.num_outputs, 2);
assert_eq!(node.hidden_dim, None);
assert_eq!(node.activation, Activation::Identity);
}
#[test]
fn test_new_node_hidden() {
let node: Node = Node::new_hidden(1, 3, 2);
assert_eq!(node.num_inputs, 3);
assert_eq!(node.num_outputs, 2);
assert_eq!(node.activation, Activation::Identity);
}
#[test]
fn test_node_builders() {
let node = Node::new_hidden(1, 3, 2)
.with_activation(Activation::GeLU)
.with_hidden_dim(32);
assert_eq!(node.activation, Activation::GeLU);
assert_eq!(node.hidden_dim, Some(32));
// Chaining doesn't disturb the port counts / kind
assert_eq!(node.num_inputs, 3);
assert_eq!(node.num_outputs, 2);
assert_eq!(node.kind, NodeKind::Hidden);
// Order-independent: with_hidden_dim before with_activation
let wide = Node::new_hidden(1, 3, 2)
.with_hidden_dim(16)
.with_activation(Activation::ReLU);
assert_eq!(wide.hidden_dim, Some(16));
assert_eq!(wide.activation, Activation::ReLU);
}
#[test]
fn test_new_node_outputs() {
let node: Node = Node::new_output(1, 3, 2);
assert_eq!(node.num_inputs, 3);
assert_eq!(node.num_outputs, 2);
assert_eq!(node.hidden_dim, None);
}
#[test]
fn test_activation_default_and_display() {
assert_eq!(Activation::default(), Activation::Identity);
assert_eq!(Activation::ReLU.to_string(), "relu");
assert_eq!(Activation::GeLU.to_string(), "gelu");
assert_eq!(Activation::SELU.to_string(), "selu");
assert_eq!(Activation::LeakyReLU.to_string(), "leaky_relu");
assert_eq!(Activation::ELU.to_string(), "elu");
assert_eq!(Activation::GeluTanh.to_string(), "gelu_tanh");
assert_eq!(Activation::Softplus.to_string(), "softplus");
assert_eq!(Activation::HardSwish.to_string(), "hardswish");
assert_eq!(Activation::HardSigmoid.to_string(), "hardsigmoid");
}
// ── property tests (proptest) ───────────────────────────────────────────
proptest! {
/// The builder methods only touch their target field: id, kind and
/// port counts must survive chaining, in either order.
#[test]
fn prop_node_builders_preserve_identity(
node in node_strategy(),
hidden_dim in 1usize..128,
relu in any::<bool>(),
) {
let activation = if relu { Activation::ReLU } else { Activation::GeLU };
let built = node
.clone()
.with_hidden_dim(hidden_dim)
.with_activation(activation);
prop_assert_eq!(built.id, node.id);
prop_assert_eq!(built.kind, node.kind);
prop_assert_eq!(built.num_inputs, node.num_inputs);
prop_assert_eq!(built.num_outputs, node.num_outputs);
prop_assert_eq!(built.activation, activation);
prop_assert_eq!(built.hidden_dim, Some(hidden_dim));
// With hidden_dim 0 the graph rejects the node at validate() time.
let bad = node.with_hidden_dim(0);
prop_assert_eq!(bad.hidden_dim, Some(0));
}
/// A random hidden node's port counts always land inside the options
/// ranges, and its id/kind follow the "append" contract.
#[test]
fn prop_random_hidden_node_respects_port_ranges(
min_inputs in 1usize..5,
min_outputs in 1usize..5,
span_in in 0usize..4,
span_out in 0usize..4,
) {
let opts = TopologyOptions {
topology_seed: 16,
min_hidden_num_nodes: 2,
max_hidden_num_nodes: 5,
min_hidden_inputs_per_node: min_inputs,
max_hidden_inputs_per_node: min_inputs + span_in,
min_hidden_outputs_per_node: min_outputs,
max_hidden_outputs_per_node: min_outputs + span_out,
input_dim: Some(4),
output_dim: Some(2),
dropout_prob: 0.0,
};
let mut graph = Topology::new(0, Some(opts));
graph.create_random_hidden_node();
let node = &graph.nodes[0];
prop_assert!(node.id == 0);
prop_assert!(node.kind == NodeKind::Hidden);
prop_assert!(
node.num_inputs >= min_inputs && node.num_inputs <= min_inputs + span_in,
"num_inputs {} outside [{}, {}]",
node.num_inputs,
min_inputs,
min_inputs + span_in
);
prop_assert!(
node.num_outputs >= min_outputs && node.num_outputs <= min_outputs + span_out,
"num_outputs {} outside [{}, {}]",
node.num_outputs,
min_outputs,
min_outputs + span_out
);
}
}
}