@dniskav/neuron 0.2.3 → 0.2.6

package/dist/index.js

@@ -22,10 +22,20 @@ var index_exports = {};
 __export(index_exports, {
   Adam: () => Adam,
   AttentionHead: () => AttentionHead,
+  BatchNorm: () => BatchNorm,
+  BiasVector: () => BiasVector,
+  ClipOptimizer: () => ClipOptimizer,
+  ClippedOptimizerFactory: () => ClippedOptimizerFactory,
+  Conv1D: () => Conv1D,
+  DataLoader: () => DataLoader,
+  Dropout: () => Dropout,
   EmbeddingMatrix: () => EmbeddingMatrix,
+  GRULayer: () => GRULayer,
+  LRScheduler: () => LRScheduler,
   LSTMLayer: () => LSTMLayer,
   Layer: () => Layer,
   LayerNorm: () => LayerNorm,
+  ModelSaver: () => ModelSaver,
   Momentum: () => Momentum,
   MultiHeadAttention: () => MultiHeadAttention,
   Network: () => Network,
@@ -36,11 +46,13 @@ __export(index_exports, {
   Neuron: () => Neuron,
   NeuronN: () => NeuronN,
   SGD: () => SGD,
+  Trainer: () => Trainer,
   TransformerBlock: () => TransformerBlock,
   WeightMatrix: () => WeightMatrix,
   crossEntropy: () => crossEntropy,
   crossEntropyDelta: () => crossEntropyDelta,
   crossEntropyDeltaRaw: () => crossEntropyDeltaRaw,
+  defaultOptimizer: () => defaultOptimizer,
   elu: () => elu,
   leakyRelu: () => leakyRelu,
   linear: () => linear,
@@ -54,10 +66,82 @@ __export(index_exports, {
   softmax: () => softmax,
   softmaxBackward: () => softmaxBackward,
   tanh: () => tanh,
-  transpose: () => transpose
+  transpose: () => transpose,
+  validate2DArray: () => validate2DArray,
+  validateArray: () => validateArray,
+  validateArrayMinLength: () => validateArrayMinLength,
+  validateNumber: () => validateNumber
 });
 module.exports = __toCommonJS(index_exports);
 
+// src/Validation.ts
+function validateArray(arr, expectedLength, methodName) {
+  if (!Array.isArray(arr)) {
+    throw new Error(`${methodName}: expected array, got ${typeof arr}`);
+  }
+  if (arr.length !== expectedLength) {
+    throw new Error(
+      `${methodName}: expected array of length ${expectedLength}, got ${arr.length}`
+    );
+  }
+  for (let i = 0; i < arr.length; i++) {
+    if (typeof arr[i] !== "number" || !isFinite(arr[i])) {
+      throw new Error(
+        `${methodName}: invalid value at index ${i}: ${arr[i]}`
+      );
+    }
+  }
+}
+function validateArrayMinLength(arr, minLength, methodName) {
+  if (!Array.isArray(arr)) {
+    throw new Error(`${methodName}: expected array, got ${typeof arr}`);
+  }
+  if (arr.length < minLength) {
+    throw new Error(
+      `${methodName}: expected array of at least length ${minLength}, got ${arr.length}`
+    );
+  }
+  for (let i = 0; i < arr.length; i++) {
+    if (typeof arr[i] !== "number" || !isFinite(arr[i])) {
+      throw new Error(
+        `${methodName}: invalid value at index ${i}: ${arr[i]}`
+      );
+    }
+  }
+}
+function validate2DArray(arr, expectedRows, expectedCols, methodName) {
+  if (!Array.isArray(arr)) {
+    throw new Error(`${methodName}: expected 2D array, got ${typeof arr}`);
+  }
+  if (arr.length !== expectedRows) {
+    throw new Error(
+      `${methodName}: expected ${expectedRows} rows, got ${arr.length}`
+    );
+  }
+  for (let i = 0; i < arr.length; i++) {
+    if (!Array.isArray(arr[i])) {
+      throw new Error(`${methodName}: row ${i} is not an array`);
+    }
+    if (arr[i].length !== expectedCols) {
+      throw new Error(
+        `${methodName}: row ${i} expected ${expectedCols} cols, got ${arr[i].length}`
+      );
+    }
+    for (let j = 0; j < arr[i].length; j++) {
+      if (typeof arr[i][j] !== "number" || !isFinite(arr[i][j])) {
+        throw new Error(
+          `${methodName}: invalid value at [${i}][${j}]: ${arr[i][j]}`
+        );
+      }
+    }
+  }
+}
+function validateNumber(value, methodName) {
+  if (typeof value !== "number" || !isFinite(value)) {
+    throw new Error(`${methodName}: expected finite number, got ${value}`);
+  }
+}
+
 // src/Neuron.ts
 function sigmoid(x) {
   return 1 / (1 + Math.exp(-x));
@@ -68,13 +152,18 @@ var Neuron = class {
     this.bias = Math.random() * 0.1;
   }
   predict(input) {
+    validateNumber(input, "Neuron.predict");
     return sigmoid(input * this.weight + this.bias);
   }
   train(input, target, lr) {
+    validateNumber(input, "Neuron.train");
+    validateNumber(target, "Neuron.train");
+    validateNumber(lr, "Neuron.train");
     const prediction = this.predict(input);
     const error = target - prediction;
-    this.weight += lr * error * input;
-    this.bias += lr * error;
+    const grad = error * prediction * (1 - prediction);
+    this.weight += lr * grad * input;
+    this.bias += lr * grad;
   }
 };
 
@@ -114,6 +203,7 @@ function makeElu(alpha = 1) {
 var elu = makeElu(1);
 
 // src/optimizers.ts
+var defaultOptimizer = () => new SGD();
 var SGD = class {
   step(weight, gradient, lr) {
     return weight + lr * gradient;
@@ -129,6 +219,19 @@ var Momentum = class {
     return weight + this.v;
   }
 };
+var ClipOptimizer = class {
+  constructor(inner, clipValue) {
+    this.inner = inner;
+    this.clipValue = clipValue;
+  }
+  step(weight, gradient, lr) {
+    const clipped = Math.max(-this.clipValue, Math.min(this.clipValue, gradient));
+    return this.inner.step(weight, clipped, lr);
+  }
+};
+function ClippedOptimizerFactory(innerFactory, clipValue) {
+  return () => new ClipOptimizer(innerFactory(), clipValue);
+}
 var Adam = class {
   constructor(beta1 = 0.9, beta2 = 0.999, epsilon = 1e-8) {
     this.beta1 = beta1;
@@ -149,7 +252,6 @@ var Adam = class {
 };
 
 // src/NeuronN.ts
-var defaultOptimizer = () => new SGD();
 var NeuronN = class {
   constructor(nInputs, activation = sigmoid2, optimizerFactory = defaultOptimizer) {
     const limit = Math.sqrt(1 / nInputs);
@@ -159,6 +261,7 @@ var NeuronN = class {
     this._opts = Array.from({ length: nInputs + 1 }, optimizerFactory);
   }
   predict(inputs) {
+    validateArray(inputs, this.weights.length, "NeuronN.predict");
     const sum = inputs.reduce((acc, e, i) => acc + e * this.weights[i], this.bias);
     return this.activation.fn(sum);
   }
@@ -171,14 +274,14 @@ var NeuronN = class {
   train(inputs, target, lr) {
     const prediction = this.predict(inputs);
     const error = target - prediction;
-    this._update(inputs.map((inp) => error * inp), error, lr);
+    const grad = error * this.activation.dfn(prediction);
+    this._update(inputs.map((inp) => grad * inp), grad, lr);
   }
 };
 
 // src/Layer.ts
-var defaultOptimizer2 = () => new SGD();
 var Layer = class {
-  constructor(nNeurons, nInputs, activation = sigmoid2, optimizerFactory = defaultOptimizer2) {
+  constructor(nNeurons, nInputs, activation = sigmoid2, optimizerFactory = defaultOptimizer) {
     this.neurons = Array.from(
       { length: nNeurons },
       () => new NeuronN(nInputs, activation, optimizerFactory)
@@ -196,84 +299,233 @@ var Network = class {
     this.outputLayer = new Layer(nOutputs, nHidden);
   }
   predict(inputs) {
+    validateArray(inputs, this.hiddenLayer.neurons[0].weights.length, "Network.predict");
     const hiddenOut = this.hiddenLayer.predict(inputs);
-    return this.outputLayer.predict(hiddenOut)[0];
+    return this.outputLayer.predict(hiddenOut);
   }
   // Trains on a single example. Returns the squared error.
   train(inputs, target, lr) {
+    validateArray(inputs, this.hiddenLayer.neurons[0].weights.length, "Network.train");
+    validateNumber(target, "Network.train");
+    validateNumber(lr, "Network.train");
     const hiddenOut = this.hiddenLayer.predict(inputs);
     const prediction = this.outputLayer.predict(hiddenOut)[0];
-    const outputError = target - prediction;
-    const outputDelta = outputError * prediction * (1 - prediction);
     const outputNeuron = this.outputLayer.neurons[0];
-    outputNeuron.weights = outputNeuron.weights.map(
-      (w, i) => w + lr * outputDelta * hiddenOut[i]
-    );
-    outputNeuron.bias += lr * outputDelta;
-    this.hiddenLayer.neurons.forEach((neuron, i) => {
-      const hiddenOut_i = hiddenOut[i];
+    const outputError = target - prediction;
+    const outputDelta = outputError * outputNeuron.activation.dfn(prediction);
+    const hiddenDeltas = this.hiddenLayer.neurons.map((neuron, i) => {
       const hiddenError = outputDelta * outputNeuron.weights[i];
-      const hiddenDelta = hiddenError * hiddenOut_i * (1 - hiddenOut_i);
-      neuron.weights = neuron.weights.map((w, j) => w + lr * hiddenDelta * inputs[j]);
-      neuron.bias += lr * hiddenDelta;
+      return hiddenError * neuron.activation.dfn(hiddenOut[i]);
+    });
+    this.hiddenLayer.neurons.forEach((neuron, i) => {
+      neuron._update(inputs.map((inp) => hiddenDeltas[i] * inp), hiddenDeltas[i], lr);
     });
+    outputNeuron._update(hiddenOut.map((h) => outputDelta * h), outputDelta, lr);
     return outputError * outputError;
   }
+  // ── Flat weight serialization ─────────────────────────────────────────────
+  // Order: hidden layer (all neurons: weights then bias), then output layer.
+  getWeights() {
+    const w = [];
+    for (const n of this.hiddenLayer.neurons) {
+      w.push(...n.weights, n.bias);
+    }
+    for (const n of this.outputLayer.neurons) {
+      w.push(...n.weights, n.bias);
+    }
+    return w;
+  }
+  setWeights(weights) {
+    let idx = 0;
+    for (const n of this.hiddenLayer.neurons) {
+      for (let j = 0; j < n.weights.length; j++) n.weights[j] = weights[idx++];
+      n.bias = weights[idx++];
+    }
+    for (const n of this.outputLayer.neurons) {
+      for (let j = 0; j < n.weights.length; j++) n.weights[j] = weights[idx++];
+      n.bias = weights[idx++];
+    }
+  }
+};
+
+// src/Dropout.ts
+var Dropout = class {
+  constructor(rate) {
+    this._mask = null;
+    if (rate < 0 || rate >= 1) {
+      throw new Error(`Dropout rate must be in [0, 1), got ${rate}`);
+    }
+    this.rate = rate;
+  }
+  // ── Forward ───────────────────────────────────────────────────────────────
+  // x: number[]  →  number[]
+  // If training, applies inverted dropout mask.
+  // If not training, returns input unchanged.
+  forward(x, training = true) {
+    if (!training || this.rate === 0) {
+      this._mask = null;
+      return [...x];
+    }
+    const scale = 1 / (1 - this.rate);
+    this._mask = x.map(() => Math.random() > this.rate ? scale : 0);
+    return x.map((v, i) => v * this._mask[i]);
+  }
+  // ── Backward ──────────────────────────────────────────────────────────────
+  // dOut: number[]  →  number[]
+  // Applies the same mask (gradient is zeroed where activation was zeroed).
+  backward(dOut) {
+    if (!this._mask) return [...dOut];
+    return dOut.map((d, i) => d * this._mask[i]);
+  }
+  // ── Reset mask between forward passes ─────────────────────────────────────
+  resetMask() {
+    this._mask = null;
+  }
+  // ── No trainable params ───────────────────────────────────────────────────
+  getWeights() {
+    return [];
+  }
+  setWeights(_weights) {
+  }
 };
 
 // src/NetworkN.ts
-var defaultOptimizer3 = () => new SGD();
 var NetworkN = class {
   constructor(structure, options = {}) {
     this.structure = structure;
     const nLayers = structure.length - 1;
     const activations = options.activations ?? Array.from({ length: nLayers }, () => sigmoid2);
-    const optimizer = options.optimizer ?? defaultOptimizer3;
+    const optimizer = options.optimizer ?? defaultOptimizer;
+    const dropoutRate = options.dropoutRate ?? 0;
+    if (activations.length !== nLayers) {
+      throw new Error(`Expected ${nLayers} activations, got ${activations.length}`);
+    }
+    if (dropoutRate < 0 || dropoutRate >= 1) {
+      throw new Error(`Dropout rate must be in [0, 1), got ${dropoutRate}`);
+    }
+    this._residual = options.residual ?? false;
     this.layers = [];
     for (let i = 1; i < structure.length; i++) {
       this.layers.push(new Layer(structure[i], structure[i - 1], activations[i - 1], optimizer));
     }
+    this._dropouts = [];
+    if (dropoutRate > 0) {
+      for (let i = 0; i < nLayers - 1; i++) {
+        this._dropouts.push(new Dropout(dropoutRate));
+      }
+    }
+    const outputLayer = this.layers[this.layers.length - 1];
+    const outputActivation = outputLayer.neurons[0].activation;
+    for (let i = 1; i < outputLayer.neurons.length; i++) {
+      if (outputLayer.neurons[i].activation !== outputActivation) {
+        throw new Error("All output neurons must share the same activation function");
+      }
+    }
   }
-  predict(inputs) {
-    return this.layers.reduce((acc, layer) => layer.predict(acc), inputs);
+  predict(inputs, training = false) {
+    validateArray(inputs, this.structure[0], "NetworkN.predict");
+    let current = [...inputs];
+    for (let i = 0; i < this.layers.length; i++) {
+      const layerInput = [...current];
+      const layerOutput = this.layers[i].predict(current);
+      if (this._shouldResidual(i)) {
+        if (this.structure[i] === this.structure[i + 1]) {
+          current = layerOutput.map((v, j) => v + layerInput[j]);
+        } else {
+          current = [...layerOutput];
+        }
+      } else {
+        current = [...layerOutput];
+      }
+      if (i < this._dropouts.length) {
+        current = this._dropouts[i].forward(current, training);
+      }
+    }
+    return current;
   }
   // Generalized backpropagation across L layers.
   // Returns the mean squared error for the example.
   train(inputs, targets, lr) {
-    const act = [inputs];
-    for (const layer of this.layers) act.push(layer.predict(act[act.length - 1]));
+    validateArray(inputs, this.structure[0], "NetworkN.train");
+    validateArray(targets, this.structure[this.structure.length - 1], "NetworkN.train");
+    const act = this._forwardAll(inputs, true);
     const pred = act[act.length - 1];
     const outAct = this.layers[this.layers.length - 1].neurons[0].activation;
-    let deltas = pred.map((p, i) => (targets[i] - p) * outAct.dfn(p));
-    for (let l = this.layers.length - 1; l >= 0; l--) {
-      const layer = this.layers[l];
-      const layerIn = act[l];
-      const prevAct = l > 0 ? this.layers[l - 1].neurons[0].activation : null;
-      const prevDeltas = layerIn.map((out, j) => {
-        const errProp = layer.neurons.reduce((s, n, k) => s + deltas[k] * n.weights[j], 0);
-        return prevAct ? errProp * prevAct.dfn(out) : errProp;
-      });
-      layer.neurons.forEach((n, k) => {
-        n._update(layerIn.map((inp) => deltas[k] * inp), deltas[k], lr);
-      });
-      deltas = prevDeltas;
-    }
+    const deltas = pred.map((p, i) => (targets[i] - p) * outAct.dfn(p));
+    this._backpropLayers(act, deltas, lr);
     return pred.reduce((s, p, i) => s + (targets[i] - p) ** 2, 0) / pred.length;
   }
   // Backprop with externally provided output-layer deltas.
   // Useful for custom loss functions (e.g. physics-based gradients).
   trainWithDeltas(inputs, outputDeltas, lr) {
+    const act = this._forwardAll(inputs, true);
+    this._backpropLayers(act, outputDeltas, lr);
+  }
+  // ── Flat weight serialization ─────────────────────────────────────────────
+  // Order: layer 0 (all neurons), layer 1, ..., layer N.
+  getWeights() {
+    for (const d of this._dropouts) d.resetMask();
+    const w = [];
+    for (const layer of this.layers) {
+      for (const n of layer.neurons) {
+        w.push(...n.weights, n.bias);
+      }
+    }
+    return w;
+  }
+  setWeights(weights) {
+    for (const d of this._dropouts) d.resetMask();
+    let idx = 0;
+    for (const layer of this.layers) {
+      for (const n of layer.neurons) {
+        for (let j = 0; j < n.weights.length; j++) n.weights[j] = weights[idx++];
+        n.bias = weights[idx++];
+      }
+    }
+  }
+  // ── Private helpers ──────────────────────────────────────────────────────
+  _shouldResidual(layerIndex) {
+    if (typeof this._residual === "function") return this._residual(layerIndex);
+    return this._residual;
+  }
+  // Forward pass storing activations at every layer boundary.
+  // Used by train(), trainWithDeltas(), and predict() shares the same logic.
+  _forwardAll(inputs, training) {
     const act = [inputs];
-    for (const layer of this.layers) act.push(layer.predict(act[act.length - 1]));
+    for (let i = 0; i < this.layers.length; i++) {
+      const layerInput = act[act.length - 1];
+      const layerOutput = this.layers[i].predict(layerInput);
+      let current;
+      if (this._shouldResidual(i) && this.structure[i] === this.structure[i + 1]) {
+        current = layerOutput.map((v, j) => v + layerInput[j]);
+      } else {
+        current = layerOutput;
+      }
+      if (i < this._dropouts.length) {
+        current = this._dropouts[i].forward(current, training);
+      }
+      act.push(current);
+    }
+    return act;
+  }
+  // Backward pass: updates all layer weights given the pre-computed activations
+  // and the initial output-layer deltas.
+  _backpropLayers(act, outputDeltas, lr) {
     let deltas = outputDeltas;
     for (let l = this.layers.length - 1; l >= 0; l--) {
       const layer = this.layers[l];
+      if (l < this._dropouts.length) {
+        deltas = this._dropouts[l].backward(deltas);
+      }
       const layerIn = act[l];
       const prevAct = l > 0 ? this.layers[l - 1].neurons[0].activation : null;
       const prevDeltas = layerIn.map((out, j) => {
         const errProp = layer.neurons.reduce((s, n, k) => s + deltas[k] * n.weights[j], 0);
         return prevAct ? errProp * prevAct.dfn(out) : errProp;
       });
+      if (this._shouldResidual(l) && this.structure[l] === this.structure[l + 1]) {
+        for (let j = 0; j < prevDeltas.length; j++) prevDeltas[j] += deltas[j];
+      }
       layer.neurons.forEach((n, k) => {
         n._update(layerIn.map((inp) => deltas[k] * inp), deltas[k], lr);
       });
@@ -294,7 +546,7 @@ var Gate = class {
   // shape: [hSize]
   constructor(inputSize, hSize, initBias = 0) {
     const n = inputSize + hSize;
-    const limit = Math.sqrt(2 / n);
+    const limit = Math.sqrt(2 / (n + hSize));
     this.W = Array.from(
       { length: hSize },
       () => Array.from({ length: n }, () => (Math.random() * 2 - 1) * limit)
@@ -308,8 +560,11 @@ var Gate = class {
   }
 };
 var LSTMLayer = class {
-  constructor(inputSize, hiddenSize) {
+  constructor(inputSize, hiddenSize, optimizerFactory = () => new SGD()) {
     this._traj = [];
+    if (inputSize <= 0 || hiddenSize <= 0) {
+      throw new Error(`LSTMLayer: inputSize and hiddenSize must be positive, got ${inputSize} and ${hiddenSize}`);
+    }
     this.inputSize = inputSize;
     this.hSize = hiddenSize;
     this.h = new Array(hiddenSize).fill(0);
@@ -318,6 +573,29 @@ var LSTMLayer = class {
     this.inputGate = new Gate(inputSize, hiddenSize);
     this.cellGate = new Gate(inputSize, hiddenSize);
     this.outputGate = new Gate(inputSize, hiddenSize);
+    const combSize = inputSize + hiddenSize;
+    this._optimizers = {
+      forgetW: Array.from(
+        { length: hiddenSize },
+        () => Array.from({ length: combSize }, () => optimizerFactory())
+      ),
+      forgetB: Array.from({ length: hiddenSize }, () => optimizerFactory()),
+      inputW: Array.from(
+        { length: hiddenSize },
+        () => Array.from({ length: combSize }, () => optimizerFactory())
+      ),
+      inputB: Array.from({ length: hiddenSize }, () => optimizerFactory()),
+      cellW: Array.from(
+        { length: hiddenSize },
+        () => Array.from({ length: combSize }, () => optimizerFactory())
+      ),
+      cellB: Array.from({ length: hiddenSize }, () => optimizerFactory()),
+      outputW: Array.from(
+        { length: hiddenSize },
+        () => Array.from({ length: combSize }, () => optimizerFactory())
+      ),
+      outputB: Array.from({ length: hiddenSize }, () => optimizerFactory())
+    };
   }
   // ── Reset state and trajectory (call at episode start) ────────────────────
   reset() {
@@ -327,6 +605,9 @@ var LSTMLayer = class {
   }
   // ── Forward pass ──────────────────────────────────────────────────────────
   predict(inputs) {
+    if (!Array.isArray(inputs) || inputs.length !== this.inputSize) {
+      throw new Error(`LSTMLayer.predict: expected array of length ${this.inputSize}, got ${inputs?.length}`);
+    }
     const combined = [...inputs, ...this.h];
     const c_prev = [...this.c];
     const zf = this.forgetGate.linear(combined);
@@ -401,15 +682,15 @@ var LSTMLayer = class {
     const scale = lr / T;
     for (let k = 0; k < hSize; k++) {
       for (let j = 0; j < combSize; j++) {
-        this.forgetGate.W[k][j] += scale * dWf[k][j];
-        this.inputGate.W[k][j] += scale * dWi[k][j];
-        this.cellGate.W[k][j] += scale * dWg[k][j];
-        this.outputGate.W[k][j] += scale * dWo[k][j];
+        this.forgetGate.W[k][j] = this._optimizers.forgetW[k][j].step(this.forgetGate.W[k][j], dWf[k][j], scale);
+        this.inputGate.W[k][j] = this._optimizers.inputW[k][j].step(this.inputGate.W[k][j], dWi[k][j], scale);
+        this.cellGate.W[k][j] = this._optimizers.cellW[k][j].step(this.cellGate.W[k][j], dWg[k][j], scale);
+        this.outputGate.W[k][j] = this._optimizers.outputW[k][j].step(this.outputGate.W[k][j], dWo[k][j], scale);
       }
-      this.forgetGate.b[k] += scale * dbf[k];
-      this.inputGate.b[k] += scale * dbi[k];
-      this.cellGate.b[k] += scale * dbg[k];
-      this.outputGate.b[k] += scale * dbo[k];
+      this.forgetGate.b[k] = this._optimizers.forgetB[k].step(this.forgetGate.b[k], dbf[k], scale);
+      this.inputGate.b[k] = this._optimizers.inputB[k].step(this.inputGate.b[k], dbi[k], scale);
+      this.cellGate.b[k] = this._optimizers.cellB[k].step(this.cellGate.b[k], dbg[k], scale);
+      this.outputGate.b[k] = this._optimizers.outputB[k].step(this.outputGate.b[k], dbo[k], scale);
     }
     this._traj = [];
   }
@@ -432,10 +713,38 @@ var LSTMLayer = class {
     this.outputGate.W = data.outputGate.W;
     this.outputGate.b = data.outputGate.b;
   }
+  // ── Flat weight serialization ─────────────────────────────────────────────
+  // Order: forgetGate (W, b), inputGate (W, b), cellGate (W, b), outputGate (W, b).
+  getWeightsFlat() {
+    const w = [];
+    for (const row of this.forgetGate.W) w.push(...row);
+    w.push(...this.forgetGate.b);
+    for (const row of this.inputGate.W) w.push(...row);
+    w.push(...this.inputGate.b);
+    for (const row of this.cellGate.W) w.push(...row);
+    w.push(...this.cellGate.b);
+    for (const row of this.outputGate.W) w.push(...row);
+    w.push(...this.outputGate.b);
+    return w;
+  }
+  setWeightsFlat(weights) {
+    let idx = 0;
+    for (let i = 0; i < this.forgetGate.W.length; i++)
+      for (let j = 0; j < this.forgetGate.W[i].length; j++) this.forgetGate.W[i][j] = weights[idx++];
+    for (let i = 0; i < this.forgetGate.b.length; i++) this.forgetGate.b[i] = weights[idx++];
+    for (let i = 0; i < this.inputGate.W.length; i++)
+      for (let j = 0; j < this.inputGate.W[i].length; j++) this.inputGate.W[i][j] = weights[idx++];
+    for (let i = 0; i < this.inputGate.b.length; i++) this.inputGate.b[i] = weights[idx++];
+    for (let i = 0; i < this.cellGate.W.length; i++)
+      for (let j = 0; j < this.cellGate.W[i].length; j++) this.cellGate.W[i][j] = weights[idx++];
+    for (let i = 0; i < this.cellGate.b.length; i++) this.cellGate.b[i] = weights[idx++];
+    for (let i = 0; i < this.outputGate.W.length; i++)
+      for (let j = 0; j < this.outputGate.W[i].length; j++) this.outputGate.W[i][j] = weights[idx++];
+    for (let i = 0; i < this.outputGate.b.length; i++) this.outputGate.b[i] = weights[idx++];
+  }
 };
 
 // src/NetworkLSTM.ts
-var defaultOptimizer4 = () => new SGD();
 var NetworkLSTM = class {
   // [T][layer+1][neuron]
   constructor(inputSize, hiddenSize, denseStructure, options = {}) {
@@ -443,7 +752,7 @@ var NetworkLSTM = class {
     this.hiddenSize = hiddenSize;
     this.lstm = new LSTMLayer(inputSize, hiddenSize);
     const activation = options.denseActivation ?? sigmoid2;
-    const optimizer = options.optimizer ?? defaultOptimizer4;
+    const optimizer = options.optimizer ?? defaultOptimizer;
     this.denseLayers = [];
     const sizes = [hiddenSize, ...denseStructure];
     for (let i = 1; i < sizes.length; i++) {
@@ -458,6 +767,7 @@ var NetworkLSTM = class {
   }
   // ── Forward pass ──────────────────────────────────────────────────────────
   predict(inputs) {
+    validateArray(inputs, this.inputSize, "NetworkLSTM.predict");
     const h = this.lstm.predict(inputs);
     const acts = [h];
     for (const layer of this.denseLayers) {
@@ -533,6 +843,30 @@ var NetworkLSTM = class {
       });
     });
   }
+  // ── Flat weight serialization ─────────────────────────────────────────────
+  // Order: LSTM (flat), then dense layer 0, dense layer 1, ..., dense layer N.
+  getWeightsFlat() {
+    const w = [];
+    w.push(...this.lstm.getWeightsFlat());
+    for (const layer of this.denseLayers) {
+      for (const n of layer.neurons) {
+        w.push(...n.weights, n.bias);
+      }
+    }
+    return w;
+  }
+  setWeightsFlat(weights) {
+    let idx = 0;
+    const lstmLen = this.lstm.getWeightsFlat().length;
+    this.lstm.setWeightsFlat(weights.slice(idx, idx + lstmLen));
+    idx += lstmLen;
+    for (const layer of this.denseLayers) {
+      for (const n of layer.neurons) {
+        for (let j = 0; j < n.weights.length; j++) n.weights[j] = weights[idx++];
+        n.bias = weights[idx++];
+      }
+    }
+  }
 };
 
 // src/MatMul.ts
@@ -540,6 +874,9 @@ function matMul(A, B) {
   const rows = A.length;
   const inner = B.length;
   const cols = B[0].length;
+  if (A[0].length !== B.length) {
+    throw new Error(`Incompatible dimensions for matrix multiplication: A cols (${A[0].length}) !== B rows (${B.length})`);
+  }
   const C = Array.from({ length: rows }, () => new Array(cols).fill(0));
   for (let i = 0; i < rows; i++)
     for (let k = 0; k < inner; k++) {
@@ -590,6 +927,33 @@ var WeightMatrix = class {
         this.W[i][j] = this.opts[i][j].step(this.W[i][j], g, lr);
       }
   }
+  // ── Flat weight serialization ─────────────────────────────────────────────
+  getWeights() {
+    const w = [];
+    for (const row of this.W) w.push(...row);
+    return w;
+  }
+  setWeights(weights) {
+    let idx = 0;
+    for (let i = 0; i < this.W.length; i++)
+      for (let j = 0; j < this.W[i].length; j++) this.W[i][j] = weights[idx++];
+  }
+};
+var BiasVector = class {
+  constructor(size) {
+    this.values = new Array(size).fill(0);
+    this.opts = Array.from({ length: size }, () => new Adam());
+  }
+  update(grad, lr) {
+    for (let i = 0; i < this.values.length; i++)
+      this.values[i] = this.opts[i].step(this.values[i], grad[i], lr);
+  }
+  getWeights() {
+    return [...this.values];
+  }
+  setWeights(weights) {
+    for (let i = 0; i < this.values.length; i++) this.values[i] = weights[i];
+  }
 };
 var EmbeddingMatrix = class {
   constructor(vocabSize, d_model) {
@@ -606,15 +970,29 @@ var EmbeddingMatrix = class {
     for (let m = 0; m < this.W[idx].length; m++)
       this.W[idx][m] += lr * grad[m];
   }
+  // ── Serializable interface ─────────────────────────────────────────────────
+  // Flattened order: row 0, row 1, ... row (vocabSize-1)
+  getWeights() {
+    const w = [];
+    for (const row of this.W) w.push(...row);
+    return w;
+  }
+  setWeights(weights) {
+    let idx = 0;
+    for (let i = 0; i < this.W.length; i++)
+      for (let j = 0; j < this.W[i].length; j++)
+        this.W[i][j] = weights[idx++];
+  }
 };
 
 // src/AttentionHead.ts
 var AttentionHead = class {
-  constructor(d_model, d_k, d_v) {
+  constructor(d_model, d_k, d_v, causal = false) {
     // d_v × d_model
     this.cache = null;
     this.d_k = d_k;
     this.d_v = d_v;
+    this.causal = causal;
     this.Wq = new WeightMatrix(d_k, d_model);
     this.Wk = new WeightMatrix(d_k, d_model);
     this.Wv = new WeightMatrix(d_v, d_model);
@@ -635,10 +1013,10 @@ var AttentionHead = class {
     );
     const scores = Array.from(
       { length: seqLen },
-      (_, i) => Array.from(
-        { length: seqLen },
-        (_2, j) => Q[i].reduce((s, q, k) => s + q * K[j][k], 0) * scale
-      )
+      (_, i) => Array.from({ length: seqLen }, (_2, j) => {
+        if (this.causal && j > i) return -Infinity;
+        return Q[i].reduce((s, q, k) => s + q * K[j][k], 0) * scale;
+      })
     );
     const attn = scores.map((row) => softmax(row));
     const out = Array.from(
@@ -662,6 +1040,7 @@ var AttentionHead = class {
   //   5. dWq  = dQ^T @ X,      dWk = dK^T @ X,  dWv = dV^T @ X
   //   6. dX   = dQ @ Wq  +  dK @ Wk  +  dV @ Wv
   backward(dOut, lr) {
+    if (!this.cache) throw new Error("AttentionHead.backward() called before predict()");
     const { X, Q, K, V, attn } = this.cache;
     const seqLen = X.length;
     const d_model = X[0].length;
@@ -734,21 +1113,40 @@ var AttentionHead = class {
   getAttentionWeights() {
     return this.cache ? this.cache.attn : null;
   }
+  // ── Flat weight serialization ─────────────────────────────────────────────
+  // Order: Wq, Wk, Wv.
+  getWeights() {
+    const w = [];
+    for (const row of this.Wq.W) w.push(...row);
+    for (const row of this.Wk.W) w.push(...row);
+    for (const row of this.Wv.W) w.push(...row);
+    return w;
+  }
+  setWeights(weights) {
+    let idx = 0;
+    for (let i = 0; i < this.Wq.W.length; i++)
+      for (let j = 0; j < this.Wq.W[i].length; j++) this.Wq.W[i][j] = weights[idx++];
+    for (let i = 0; i < this.Wk.W.length; i++)
+      for (let j = 0; j < this.Wk.W[i].length; j++) this.Wk.W[i][j] = weights[idx++];
+    for (let i = 0; i < this.Wv.W.length; i++)
+      for (let j = 0; j < this.Wv.W[i].length; j++) this.Wv.W[i][j] = weights[idx++];
+  }
 };
 
 // src/MultiHeadAttention.ts
 var MultiHeadAttention = class {
   // seqLen × (nHeads * d_k)
-  constructor(d_model, nHeads) {
+  constructor(d_model, nHeads, causal = false) {
     // d_model × (nHeads * d_k)
     // Cached for backward
     this._concat = null;
     this.nHeads = nHeads;
     this.d_model = d_model;
     this.d_k = Math.floor(d_model / nHeads);
+    this.causal = causal;
     this.heads = Array.from(
       { length: nHeads },
-      () => new AttentionHead(d_model, this.d_k, this.d_k)
+      () => new AttentionHead(d_model, this.d_k, this.d_k, causal)
     );
     this.Wo = new WeightMatrix(d_model, nHeads * this.d_k);
   }
@@ -770,6 +1168,7 @@ var MultiHeadAttention = class {
   // ── Backward ──────────────────────────────────────────────────────────────
   // dOut: seqLen × d_model  →  dX: seqLen × d_model
   backward(dOut, lr) {
+    if (!this._concat) throw new Error("MultiHeadAttention.backward() called before predict()");
     const seqLen = dOut.length;
     const concatD = this.nHeads * this.d_k;
     const d_model = this.d_model;
@@ -807,6 +1206,31 @@ var MultiHeadAttention = class {
   getAttentionWeights() {
     return this.heads.map((h) => h.getAttentionWeights());
   }
+  // ── Flat weight serialization ─────────────────────────────────────────────
+  // Order: head0 (Wq, Wk, Wv), head1, ..., headN, then Wo.
+  getWeights() {
+    const w = [];
+    for (const head of this.heads) {
+      for (const row of head.Wq.W) w.push(...row);
+      for (const row of head.Wk.W) w.push(...row);
+      for (const row of head.Wv.W) w.push(...row);
+    }
+    for (const row of this.Wo.W) w.push(...row);
+    return w;
+  }
+  setWeights(weights) {
+    let idx = 0;
+    for (const head of this.heads) {
+      for (let i = 0; i < head.Wq.W.length; i++)
+        for (let j = 0; j < head.Wq.W[i].length; j++) head.Wq.W[i][j] = weights[idx++];
+      for (let i = 0; i < head.Wk.W.length; i++)
+        for (let j = 0; j < head.Wk.W[i].length; j++) head.Wk.W[i][j] = weights[idx++];
+      for (let i = 0; i < head.Wv.W.length; i++)
+        for (let j = 0; j < head.Wv.W[i].length; j++) head.Wv.W[i][j] = weights[idx++];
+    }
+    for (let i = 0; i < this.Wo.W.length; i++)
+      for (let j = 0; j < this.Wo.W[i].length; j++) this.Wo.W[i][j] = weights[idx++];
+  }
 };
 
 // src/LayerNorm.ts
@@ -849,20 +1273,32 @@ var LayerNorm = class {
   backwardOne(dOut, pos, lr) {
     const { x_norm, std } = this._cache[pos];
     const N = dOut.length;
+    const gammaOld = this.gamma.slice();
     for (let i = 0; i < N; i++) {
       this.gamma[i] += lr * dOut[i] * x_norm[i];
       this.beta[i] += lr * dOut[i];
     }
-    const D = dOut.map((d, i) => d * this.gamma[i]);
+    const D = dOut.map((d, i) => d * gammaOld[i]);
     const mD = D.reduce((s, v) => s + v, 0) / N;
     const mDxn = D.reduce((s, d, i) => s + d * x_norm[i], 0) / N;
     return D.map((d, i) => (d - mD - x_norm[i] * mDxn) / std);
   }
+  // ── Flat weight serialization ─────────────────────────────────────────────
+  // Order: gamma, beta.
+  getWeights() {
+    return [...this.gamma, ...this.beta];
+  }
+  setWeights(weights) {
+    const dim = this.gamma.length;
+    for (let i = 0; i < dim; i++) this.gamma[i] = weights[i];
+    for (let i = 0; i < dim; i++) this.beta[i] = weights[dim + i];
+  }
 };
 
 // src/TransformerBlock.ts
 var TransformerBlock = class {
-  constructor({ d_model, nHeads, d_ff }) {
+  constructor({ d_model, nHeads, d_ff, causal = false }) {
+    // d_model
     // Forward caches (needed for backprop)
     this._X = null;
     this._attnOut = null;
@@ -874,15 +1310,13 @@ var TransformerBlock = class {
     this._ff2Out = null;
     this.d_model = d_model;
     this.d_ff = d_ff;
-    this.attn = new MultiHeadAttention(d_model, nHeads);
+    this.attn = new MultiHeadAttention(d_model, nHeads, causal);
     this.norm1 = new LayerNorm(d_model);
     this.norm2 = new LayerNorm(d_model);
     this.ff1 = new WeightMatrix(d_ff, d_model);
     this.ff2 = new WeightMatrix(d_model, d_ff);
-    this.b1 = new Array(d_ff).fill(0);
-    this.b2 = new Array(d_model).fill(0);
-    this.b1Opts = Array.from({ length: d_ff }, () => new Adam());
-    this.b2Opts = Array.from({ length: d_model }, () => new Adam());
+    this.b1 = new BiasVector(d_ff);
+    this.b2 = new BiasVector(d_model);
   }
   // ── Forward ───────────────────────────────────────────────────────────────
   // X: seqLen × d_model  →  out: seqLen × d_model
@@ -895,11 +1329,11 @@ var TransformerBlock = class {
       return this.norm1.predictOne(added, i);
     });
     const ff1Pre = h1.map(
-      (h) => this.ff1.W.map((row, k) => row.reduce((s, w, m) => s + w * h[m], this.b1[k]))
+      (h) => this.ff1.W.map((row, k) => row.reduce((s, w, m) => s + w * h[m], this.b1.values[k]))
     );
     const ff1Out = ff1Pre.map((pre) => pre.map((v) => Math.max(0, v)));
     const ff2Out = ff1Out.map(
-      (h) => this.ff2.W.map((row, k) => row.reduce((s, w, m) => s + w * h[m], this.b2[k]))
+      (h) => this.ff2.W.map((row, k) => row.reduce((s, w, m) => s + w * h[m], this.b2.values[k]))
     );
     this.norm2.resetCache(seqLen);
     const out = h1.map((h, i) => {
@@ -917,6 +1351,9 @@ var TransformerBlock = class {
   // ── Backward ──────────────────────────────────────────────────────────────
   // dOut: seqLen × d_model  →  dX: seqLen × d_model
   backward(dOut, lr) {
+    if (!this._h1 || !this._ff1Out || !this._ff1Pre) {
+      throw new Error("TransformerBlock.backward() called before predict()");
+    }
     const seqLen = dOut.length;
     const d_model = this.d_model;
     const h1 = this._h1;
@@ -941,8 +1378,7 @@ var TransformerBlock = class {
       (_, m) => dAdded2.reduce((s, da) => s + da[m], 0)
     );
     this.ff2.update(dW2, lr);
-    for (let m = 0; m < d_model; m++)
-      this.b2[m] = this.b2Opts[m].step(this.b2[m], db2[m], lr);
+    this.b2.update(db2, lr);
     const dFf1Pre = dFf1Out.map(
       (d, i) => d.map((v, k) => ff1Pre[i][k] > 0 ? v : 0)
     );
@@ -964,8 +1400,7 @@ var TransformerBlock = class {
       (_, k) => dFf1Pre.reduce((s, dp) => s + dp[k], 0)
     );
     this.ff1.update(dW1, lr);
-    for (let k = 0; k < this.d_ff; k++)
-      this.b1[k] = this.b1Opts[k].step(this.b1[k], db1[k], lr);
+    this.b1.update(db1, lr);
     const dH1 = Array.from(
       { length: seqLen },
       (_, i) => dH1_fromFf[i].map((v, m) => v + dAdded2[i][m])
@@ -987,6 +1422,36 @@ var TransformerBlock = class {
   getAttentionWeights() {
     return this.attn.getAttentionWeights();
   }
+  // ── Flat weight serialization ─────────────────────────────────────────────
+  // Order: attn (MHA), norm1 (gamma, beta), ff1, b1, ff2, b2, norm2 (gamma, beta).
+  getWeights() {
+    const w = [];
+    w.push(...this.attn.getWeights());
+    w.push(...this.norm1.gamma, ...this.norm1.beta);
+    for (const row of this.ff1.W) w.push(...row);
+    w.push(...this.b1.values);
+    for (const row of this.ff2.W) w.push(...row);
+    w.push(...this.b2.values);
+    w.push(...this.norm2.gamma, ...this.norm2.beta);
+    return w;
+  }
+  setWeights(weights) {
+    let idx = 0;
+    const attnLen = this.attn.getWeights().length;
+    this.attn.setWeights(weights.slice(idx, idx + attnLen));
+    idx += attnLen;
+    this.norm1.setWeights(weights.slice(idx, idx + this.norm1.getWeights().length));
+    idx += this.norm1.getWeights().length;
+    this.ff1.setWeights(weights.slice(idx, idx + this.ff1.getWeights().length));
+    idx += this.ff1.getWeights().length;
+    this.b1.setWeights(weights.slice(idx, idx + this.b1.values.length));
+    idx += this.b1.values.length;
+    this.ff2.setWeights(weights.slice(idx, idx + this.ff2.getWeights().length));
+    idx += this.ff2.getWeights().length;
+    this.b2.setWeights(weights.slice(idx, idx + this.b2.values.length));
+    idx += this.b2.values.length;
+    this.norm2.setWeights(weights.slice(idx, idx + this.norm2.getWeights().length));
+  }
 };
 
 // src/NetworkTransformer.ts
@@ -1011,8 +1476,7 @@ var NetworkTransformer = class {
       () => new TransformerBlock({ d_model, nHeads, d_ff })
     );
     this.outputProj = new WeightMatrix(nClasses, d_model);
-    this.outputBias = new Array(nClasses).fill(0);
-    this.outBiasOpts = Array.from({ length: nClasses }, () => new Adam());
+    this.outputBias = new BiasVector(nClasses);
   }
   // ── Forward pass ──────────────────────────────────────────────────────────
   // tokens: seqLen integer ids  →  seqLen * nClasses logits (flattened)
@@ -1020,7 +1484,7 @@ var NetworkTransformer = class {
     const h = this._forward(tokens);
     return h.flatMap(
       (hi) => this.outputProj.W.map(
-        (row, c) => row.reduce((s, w, m) => s + w * hi[m], this.outputBias[c])
+        (row, c) => row.reduce((s, w, m) => s + w * hi[m], this.outputBias.values[c])
       )
     );
   }
@@ -1034,7 +1498,7 @@ var NetworkTransformer = class {
     const h = this._forward(tokens);
     const logits = h.map(
       (hi) => this.outputProj.W.map(
-        (row, c) => row.reduce((s, w, m) => s + w * hi[m], this.outputBias[c])
+        (row, c) => row.reduce((s, w, m) => s + w * hi[m], this.outputBias.values[c])
       )
     );
     let loss = 0;
@@ -1069,8 +1533,7 @@ var NetworkTransformer = class {
       (_, c) => dLogits.reduce((s, dl) => s + dl[c], 0)
     );
     this.outputProj.update(dWout, lr);
-    for (let c = 0; c < this.nClasses; c++)
-      this.outputBias[c] = this.outBiasOpts[c].step(this.outputBias[c], dBout[c], lr);
+    this.outputBias.update(dBout, lr);
     let dX = dH;
     for (let b = this.blocks.length - 1; b >= 0; b--)
       dX = this.blocks[b].backward(dX, lr);
@@ -1085,6 +1548,35 @@ var NetworkTransformer = class {
   getAttentionWeights() {
     return this.blocks.map((b) => b.getAttentionWeights());
   }
+  // ── Flat weight serialization ─────────────────────────────────────────────
+  // Order: tokenEmb, posEmb, block0, block1, ..., blockN, outputProj, outputBias.
+  getWeights() {
+    const w = [];
+    w.push(...this.tokenEmb.getWeights());
+    w.push(...this.posEmb.getWeights());
+    for (const block of this.blocks) w.push(...block.getWeights());
+    w.push(...this.outputProj.getWeights());
+    w.push(...this.outputBias.getWeights());
+    return w;
+  }
+  setWeights(weights) {
+    let idx = 0;
+    const tokenEmbLen = this.tokenEmb.getWeights().length;
+    this.tokenEmb.setWeights(weights.slice(idx, idx + tokenEmbLen));
+    idx += tokenEmbLen;
+    const posEmbLen = this.posEmb.getWeights().length;
+    this.posEmb.setWeights(weights.slice(idx, idx + posEmbLen));
+    idx += posEmbLen;
+    for (const block of this.blocks) {
+      const blockLen = block.getWeights().length;
+      block.setWeights(weights.slice(idx, idx + blockLen));
+      idx += blockLen;
+    }
+    const outProjLen = this.outputProj.getWeights().length;
+    this.outputProj.setWeights(weights.slice(idx, idx + outProjLen));
+    idx += outProjLen;
+    this.outputBias.setWeights(weights.slice(idx, idx + this.outputBias.values.length));
+  }
   // ── Internal ──────────────────────────────────────────────────────────────
   // Shared embedding + block forward pass.
   _forward(tokens) {
@@ -1104,25 +1596,28 @@ var NetworkTransformerRL = class {
   constructor(seqLen, inputDim, options = {}) {
     // Forward caches para backprop
     this._projected = null;
+    // For max pooling backward: argmax per dimension across all positions
+    this._argmax = null;
     const {
       d_model = 32,
       nHeads = 2,
       d_ff = 64,
       nBlocks = 2,
-      nActions = 2
+      nActions = 2,
+      pooling = "weighted"
     } = options;
     this.seqLen = seqLen;
     this.inputDim = inputDim;
     this.d_model = d_model;
     this.nActions = nActions;
+    this._pooling = pooling;
     this.inputProj = new WeightMatrix(d_model, inputDim);
     this.blocks = Array.from(
       { length: nBlocks },
-      () => new TransformerBlock({ d_model, nHeads, d_ff })
+      () => new TransformerBlock({ d_model, nHeads, d_ff, causal: true })
     );
     this.outputProj = new WeightMatrix(nActions, d_model);
-    this.outputBias = new Array(nActions).fill(0);
-    this.outBiasOpts = Array.from({ length: nActions }, () => new Adam());
+    this.outputBias = new BiasVector(nActions);
   }
   // ── Forward ────────────────────────────────────────────────────────────────
   // sequence: seqLen × inputDim → nActions Q-values
@@ -1130,7 +1625,7 @@ var NetworkTransformerRL = class {
     const h = this._forward(sequence);
     const pooled = this._pool(h);
     return this.outputProj.W.map(
-      (row, c) => row.reduce((s, w, m) => s + w * pooled[m], this.outputBias[c])
+      (row, c) => row.reduce((s, w, m) => s + w * pooled[m], this.outputBias.values[c])
     );
   }
   // ── Training ────────────────────────────────────────────────────────────────
@@ -1142,7 +1637,7 @@ var NetworkTransformerRL = class {
     const h = this._forward(sequence);
     const pooled = this._pool(h);
     const pred = this.outputProj.W.map(
-      (row, c) => row.reduce((s, w, m) => s + w * pooled[m], this.outputBias[c])
+      (row, c) => row.reduce((s, w, m) => s + w * pooled[m], this.outputBias.values[c])
     );
     const n = this.nActions;
     let loss = 0;
@@ -1165,13 +1660,8 @@ var NetworkTransformerRL = class {
     );
     const dBout = dPred.slice();
     this.outputProj.update(dWout, lr);
-    for (let c = 0; c < this.nActions; c++)
-      this.outputBias[c] = this.outBiasOpts[c].step(this.outputBias[c], dBout[c], lr);
-    let dH = Array.from(
-      { length: this.seqLen },
-      (_, i) => dPooled.map((v) => v / this.seqLen)
-      // Gradiente dividido entre posiciones
-    );
+    this.outputBias.update(dBout, lr);
+    let dH = this._distributePoolGradient(dPooled);
     for (let b = this.blocks.length - 1; b >= 0; b--)
       dH = this.blocks[b].backward(dH, lr);
     for (let i = 0; i < this.seqLen; i++) {
@@ -1190,8 +1680,32 @@ var NetworkTransformerRL = class {
   getAttentionWeights() {
     return this.blocks.map((b) => b.getAttentionWeights());
   }
-  // ── Serialization ──────────────────────────────────────────────────────────
-  getWeights() {
+  // ── Flat weight serialization ─────────────────────────────────────────────
+  // Order: inputProj, block0, block1, ..., blockN, outputProj, outputBias.
+  getWeightsFlat() {
+    const w = [];
+    w.push(...this.inputProj.getWeights());
+    for (const block of this.blocks) w.push(...block.getWeights());
+    w.push(...this.outputProj.getWeights());
+    w.push(...this.outputBias.getWeights());
+    return w;
+  }
+  setWeightsFlat(weights) {
+    let idx = 0;
+    const inputProjLen = this.inputProj.getWeights().length;
+    this.inputProj.setWeights(weights.slice(idx, idx + inputProjLen));
+    idx += inputProjLen;
+    for (const block of this.blocks) {
+      const blockLen = block.getWeights().length;
+      block.setWeights(weights.slice(idx, idx + blockLen));
+      idx += blockLen;
+    }
+    const outProjLen = this.outputProj.getWeights().length;
+    this.outputProj.setWeights(weights.slice(idx, idx + outProjLen));
+    idx += outProjLen;
+    this.outputBias.setWeights(weights.slice(idx, idx + this.outputBias.values.length));
+  }
+  getWeightsStructured() {
     return {
       inputProj: this.inputProj.W.map((r) => [...r]),
       blocks: this.blocks.map((b) => ({
@@ -1207,17 +1721,15 @@ var NetworkTransformerRL = class {
         norm2: { gamma: [...b.norm2.gamma], beta: [...b.norm2.beta] },
         ff1: b.ff1.W.map((r) => [...r]),
         ff2: b.ff2.W.map((r) => [...r]),
-        b1: [...b.b1],
-        b2: [...b.b2]
+        b1: [...b.b1.values],
+        b2: [...b.b2.values]
       })),
       outputProj: this.outputProj.W.map((r) => [...r]),
-      outputBias: [...this.outputBias]
+      outputBias: [...this.outputBias.values]
     };
   }
-  setWeights(data) {
-    data.inputProj.forEach((row, i) => {
-      this.inputProj.W[i] = [...row];
-    });
+  setWeightsStructured(data) {
+    this.inputProj.setWeights(data.inputProj.flat());
     data.blocks.forEach((bd, b) => {
       const blk = this.blocks[b];
       bd.attn.heads.forEach((hd, h) => {
@@ -1232,11 +1744,20 @@ var NetworkTransformerRL = class {
       blk.norm2.beta = [...bd.norm2.beta];
       blk.ff1.W = bd.ff1.map((r) => [...r]);
       blk.ff2.W = bd.ff2.map((r) => [...r]);
-      blk.b1 = [...bd.b1];
-      blk.b2 = [...bd.b2];
+      blk.b1.setWeights(bd.b1);
+      blk.b2.setWeights(bd.b2);
     });
     this.outputProj.W = data.outputProj.map((r) => [...r]);
-    this.outputBias = [...data.outputBias];
+    this.outputBias.setWeights(data.outputBias);
+  }
+  // ── Serializable interface (flat array) ────────────────────────────────────
+  // These satisfy the Serializable interface from ModelSaver, which requires
+  // getWeights(): number[] and setWeights(weights: number[]): void.
+  getWeights() {
+    return this.getWeightsFlat();
+  }
+  setWeights(weights) {
+    this.setWeightsFlat(weights);
   }
   // ── Internal ────────────────────────────────────────────────────────────────
   _forward(sequence) {
@@ -1251,6 +1772,44 @@ var NetworkTransformerRL = class {
     return h;
   }
   _pool(h) {
+    switch (this._pooling) {
+      case "avg":
+        return this._poolAvg(h);
+      case "max":
+        return this._poolMax(h);
+      case "last":
+        return this._poolLast(h);
+      case "weighted":
+      default:
+        return this._poolWeighted(h);
+    }
+  }
+  _poolAvg(h) {
+    const n = h.length;
+    return Array.from({ length: this.d_model }, (_, m) => {
+      let sum = 0;
+      for (let i = 0; i < n; i++)
+        sum += h[i][m];
+      return sum / n;
+    });
+  }
+  _poolMax(h) {
+    this._argmax = new Array(this.d_model).fill(0);
+    return Array.from({ length: this.d_model }, (_, m) => {
+      let maxVal = -Infinity;
+      for (let i = 0; i < h.length; i++) {
+        if (h[i][m] > maxVal) {
+          maxVal = h[i][m];
+          this._argmax[m] = i;
+        }
+      }
+      return maxVal;
+    });
+  }
+  _poolLast(h) {
+    return [...h[h.length - 1]];
+  }
+  _poolWeighted(h) {
     const weights = Array.from(
       { length: this.seqLen },
       (_, i) => i === this.seqLen - 1 ? 2 : 1
@@ -1263,6 +1822,55 @@ var NetworkTransformerRL = class {
       return sum / totalWeight;
     });
   }
+  /** Returns the current pooling type for inspection. */
+  getPoolingType() {
+    return this._pooling;
+  }
+  // ── Helper: distribute pooled gradient back to each position ────────────────
+  // Must match the same distribution as _pool() used during forward.
+  _distributePoolGradient(dPooled) {
+    switch (this._pooling) {
+      case "avg": {
+        const n = this.seqLen;
+        return Array.from(
+          { length: n },
+          () => dPooled.map((v) => v / n)
+        );
+      }
+      case "max": {
+        if (!this._argmax) {
+          const n = this.seqLen;
+          return Array.from(
+            { length: n },
+            () => dPooled.map((v) => v / n)
+          );
+        }
+        const argmax = this._argmax;
+        return Array.from(
+          { length: this.seqLen },
+          (_, i) => dPooled.map((v, m) => i === argmax[m] ? v : 0)
+        );
+      }
+      case "last": {
+        return Array.from(
+          { length: this.seqLen },
+          (_, i) => i === this.seqLen - 1 ? [...dPooled] : new Array(this.d_model).fill(0)
+        );
+      }
+      case "weighted":
+      default: {
+        const weights = Array.from(
+          { length: this.seqLen },
+          (_, i) => i === this.seqLen - 1 ? 2 : 1
+        );
+        const totalWeight = weights.reduce((a, b) => a + b, 0);
+        return Array.from(
+          { length: this.seqLen },
+          (_, i) => dPooled.map((v) => v * weights[i] / totalWeight)
+        );
+      }
+    }
+  }
 };
 
 // src/losses.ts
@@ -1287,14 +1895,803 @@ function crossEntropyDeltaRaw(predicted, actual) {
   const p = Math.max(eps, Math.min(1 - eps, predicted));
   return actual / p - (1 - actual) / (1 - p);
 }
+
+// src/GRU.ts
+function sigmoid4(x) {
+  return 1 / (1 + Math.exp(-x));
+}
+function tanhFn(x) {
+  const e = Math.exp(2 * x);
+  return (e - 1) / (e + 1);
+}
+var Gate2 = class {
+  constructor(inputSize, hSize, initBias = 0) {
+    const n = inputSize + hSize;
+    const limit = Math.sqrt(2 / (n + hSize));
+    this.W = Array.from(
+      { length: hSize },
+      () => Array.from({ length: n }, () => (Math.random() * 2 - 1) * limit)
+    );
+    this.b = new Array(hSize).fill(initBias);
+  }
+  linear(combined) {
+    return this.W.map(
+      (row, i) => row.reduce((s, w, j) => s + w * combined[j], this.b[i])
+    );
+  }
+};
+var GRULayer = class {
+  constructor(inputSize, hiddenSize, optimizerFactory = () => new SGD()) {
+    this._traj = [];
+    if (inputSize <= 0 || hiddenSize <= 0) {
+      throw new Error(`GRULayer: inputSize and hiddenSize must be positive, got ${inputSize} and ${hiddenSize}`);
+    }
+    this.inputSize = inputSize;
+    this.hSize = hiddenSize;
+    this.h = new Array(hiddenSize).fill(0);
+    this.resetGate = new Gate2(inputSize, hiddenSize);
+    this.updateGate = new Gate2(inputSize, hiddenSize);
+    this.newGate = new Gate2(inputSize, hiddenSize);
+    const combSize = inputSize + hiddenSize;
+    this._optimizers = {
+      resetW: Array.from(
+        { length: hiddenSize },
+        () => Array.from({ length: combSize }, () => optimizerFactory())
+      ),
+      resetB: Array.from({ length: hiddenSize }, () => optimizerFactory()),
+      updateW: Array.from(
+        { length: hiddenSize },
+        () => Array.from({ length: combSize }, () => optimizerFactory())
+      ),
+      updateB: Array.from({ length: hiddenSize }, () => optimizerFactory()),
+      newW: Array.from(
+        { length: hiddenSize },
+        () => Array.from({ length: combSize }, () => optimizerFactory())
+      ),
+      newB: Array.from({ length: hiddenSize }, () => optimizerFactory())
+    };
+  }
+  reset() {
+    this.h = new Array(this.hSize).fill(0);
+    this._traj = [];
+  }
+  predict(inputs) {
+    if (!Array.isArray(inputs) || inputs.length !== this.inputSize) {
+      throw new Error(`GRULayer.predict: expected array of length ${this.inputSize}, got ${inputs?.length}`);
+    }
+    const combined = [...inputs, ...this.h];
+    const h_prev = [...this.h];
+    const r_pre = this.resetGate.linear(combined);
+    const z_pre = this.updateGate.linear(combined);
+    const r_a = r_pre.map(sigmoid4);
+    const z_a = z_pre.map(sigmoid4);
+    const combined_r = [...inputs, ...r_a.map((r, i) => r * h_prev[i])];
+    const n_pre = this.newGate.linear(combined_r);
+    const n_a = n_pre.map(tanhFn);
+    const h = n_a.map((n, i) => (1 - z_a[i]) * n + z_a[i] * h_prev[i]);
+    this._traj.push({ combined, h_prev, r: r_pre, r_a, z: z_pre, z_a, combined_r, n_pre, n_a, h });
+    this.h = h;
+    return h;
+  }
+  backprop(dh_seq, lr) {
+    const T = this._traj.length;
+    if (T === 0 || dh_seq.length !== T) return;
+    const hSize = this.hSize;
+    const combSize = this.inputSize + hSize;
+    const dWr = Array.from({ length: hSize }, () => new Array(combSize).fill(0));
+    const dWz = Array.from({ length: hSize }, () => new Array(combSize).fill(0));
+    const dWn = Array.from({ length: hSize }, () => new Array(combSize).fill(0));
+    const dbr = new Array(hSize).fill(0);
+    const dbz = new Array(hSize).fill(0);
+    const dbn = new Array(hSize).fill(0);
+    let dh_next = new Array(hSize).fill(0);
+    for (let t = T - 1; t >= 0; t--) {
+      const s = this._traj[t];
+      const dh = dh_seq[t].map((d, i) => d + dh_next[i]);
+      const dz_a = dh.map((d, i) => (s.h_prev[i] - s.n_a[i]) * d);
+      const dn_a = dh.map((d, i) => (1 - s.z_a[i]) * d);
+      const dn_pre = dn_a.map((d, i) => d * (1 - s.n_a[i] ** 2));
+      const dz_pre = dz_a.map((d, i) => d * s.z_a[i] * (1 - s.z_a[i]));
+      const dr_hprev = Array.from(
+        { length: hSize },
+        (_, i) => this.newGate.W.reduce((sum, row, k) => sum + dn_pre[k] * row[this.inputSize + i], 0)
+      );
+      const dr_a = dr_hprev.map((d, i) => d * s.h_prev[i]);
+      const dr_pre = dr_a.map((d, i) => d * s.r_a[i] * (1 - s.r_a[i]));
+      for (let k = 0; k < hSize; k++) {
+        for (let j = 0; j < combSize; j++) {
+          dWr[k][j] += dr_pre[k] * s.combined[j];
+          dWz[k][j] += dz_pre[k] * s.combined[j];
+          dWn[k][j] += dn_pre[k] * s.combined_r[j];
+        }
+        dbr[k] += dr_pre[k];
+        dbz[k] += dz_pre[k];
+        dbn[k] += dn_pre[k];
+      }
+      dh_next = new Array(hSize).fill(0);
+      for (let k = 0; k < hSize; k++) {
+        for (let j = this.inputSize; j < combSize; j++) {
+          dh_next[j - this.inputSize] += dr_pre[k] * this.resetGate.W[k][j] + dz_pre[k] * this.updateGate.W[k][j];
+        }
+        dh_next[k] += dr_hprev[k] * s.r_a[k];
+        dh_next[k] += dh[k] * s.z_a[k];
+      }
+    }
+    const scale = lr / T;
+    for (let k = 0; k < hSize; k++) {
+      for (let j = 0; j < combSize; j++) {
+        this.resetGate.W[k][j] = this._optimizers.resetW[k][j].step(this.resetGate.W[k][j], dWr[k][j], scale);
+        this.updateGate.W[k][j] = this._optimizers.updateW[k][j].step(this.updateGate.W[k][j], dWz[k][j], scale);
+        this.newGate.W[k][j] = this._optimizers.newW[k][j].step(this.newGate.W[k][j], dWn[k][j], scale);
+      }
+      this.resetGate.b[k] = this._optimizers.resetB[k].step(this.resetGate.b[k], dbr[k], scale);
+      this.updateGate.b[k] = this._optimizers.updateB[k].step(this.updateGate.b[k], dbz[k], scale);
+      this.newGate.b[k] = this._optimizers.newB[k].step(this.newGate.b[k], dbn[k], scale);
+    }
+    this._traj = [];
+  }
+  // ── Flat weight serialization ─────────────────────────────────────────────
+  // Order: resetGate (W, b), updateGate (W, b), newGate (W, b).
+  getWeightsFlat() {
+    const w = [];
+    for (const row of this.resetGate.W) w.push(...row);
+    w.push(...this.resetGate.b);
+    for (const row of this.updateGate.W) w.push(...row);
+    w.push(...this.updateGate.b);
+    for (const row of this.newGate.W) w.push(...row);
+    w.push(...this.newGate.b);
+    return w;
+  }
+  setWeightsFlat(weights) {
+    let idx = 0;
+    for (let i = 0; i < this.resetGate.W.length; i++)
+      for (let j = 0; j < this.resetGate.W[i].length; j++) this.resetGate.W[i][j] = weights[idx++];
+    for (let i = 0; i < this.resetGate.b.length; i++) this.resetGate.b[i] = weights[idx++];
+    for (let i = 0; i < this.updateGate.W.length; i++)
+      for (let j = 0; j < this.updateGate.W[i].length; j++) this.updateGate.W[i][j] = weights[idx++];
+    for (let i = 0; i < this.updateGate.b.length; i++) this.updateGate.b[i] = weights[idx++];
+    for (let i = 0; i < this.newGate.W.length; i++)
+      for (let j = 0; j < this.newGate.W[i].length; j++) this.newGate.W[i][j] = weights[idx++];
+    for (let i = 0; i < this.newGate.b.length; i++) this.newGate.b[i] = weights[idx++];
+  }
+  getWeights() {
+    return {
+      resetGate: { W: this.resetGate.W, b: this.resetGate.b },
+      updateGate: { W: this.updateGate.W, b: this.updateGate.b },
+      newGate: { W: this.newGate.W, b: this.newGate.b }
+    };
+  }
+  setWeights(data) {
+    this.resetGate.W = data.resetGate.W;
+    this.resetGate.b = data.resetGate.b;
+    this.updateGate.W = data.updateGate.W;
+    this.updateGate.b = data.updateGate.b;
+    this.newGate.W = data.newGate.W;
+    this.newGate.b = data.newGate.b;
+  }
+};
+
+// src/BatchNorm.ts
+var BatchNorm = class {
+  constructor(dim, momentum = 0.1) {
+    this._xNorm = null;
+    this._std = null;
+    this.dim = dim;
+    this.momentum = momentum;
+    this.gamma = new Array(dim).fill(1);
+    this.beta = new Array(dim).fill(0);
+    this.runningMean = new Array(dim).fill(0);
+    this.runningVar = new Array(dim).fill(1);
+  }
+  // ── Forward ───────────────────────────────────────────────────────────────
+  forward(x) {
+    if (x.length !== this.dim) {
+      throw new Error(`BatchNorm.forward: expected array of length ${this.dim}, got ${x.length}`);
+    }
+    const eps = 1e-5;
+    for (let i = 0; i < this.dim; i++) {
+      this.runningMean[i] = this.momentum * this.runningMean[i] + (1 - this.momentum) * x[i];
+      const diff = x[i] - this.runningMean[i];
+      this.runningVar[i] = this.momentum * this.runningVar[i] + (1 - this.momentum) * diff * diff;
+    }
+    this._std = this.runningVar.map((v) => Math.sqrt(v + eps));
+    this._xNorm = x.map((v, i) => (v - this.runningMean[i]) / this._std[i]);
+    return this._xNorm.map((xn, i) => this.gamma[i] * xn + this.beta[i]);
+  }
+  // ── Backward ──────────────────────────────────────────────────────────────
+  backward(dOut) {
+    if (!this._xNorm || !this._std) {
+      throw new Error("BatchNorm.backward: call forward() first");
+    }
+    for (let i = 0; i < this.dim; i++) {
+    }
+    return dOut.map((d, i) => d * this.gamma[i] / this._std[i]);
+  }
+  // ── Train gamma and beta (call after backward) ────────────────────────────
+  trainParams(dOut, lr) {
+    if (!this._xNorm) return;
+    for (let i = 0; i < this.dim; i++) {
+      this.gamma[i] += lr * dOut[i] * this._xNorm[i];
+      this.beta[i] += lr * dOut[i];
+    }
+  }
+  // ── Flat weight serialization ─────────────────────────────────────────────
+  // Order: gamma, beta.
+  getWeights() {
+    return [...this.gamma, ...this.beta];
+  }
+  setWeights(weights) {
+    for (let i = 0; i < this.dim; i++) this.gamma[i] = weights[i];
+    for (let i = 0; i < this.dim; i++) this.beta[i] = weights[this.dim + i];
+  }
+};
+
+// src/Conv1D.ts
+var Conv1D = class {
+  constructor(inputLength, kernelSize, filters, stride = 1, padding = "valid", optimizerFactory = () => new SGD(), inputChannels = 1) {
+    // [filters]
+    this._input = null;
+    this._paddedInput = null;
+    if (inputLength <= 0 || kernelSize <= 0 || filters <= 0) {
+      throw new Error("Conv1D: inputLength, kernelSize, and filters must be positive");
+    }
+    if (kernelSize > inputLength && padding === "valid") {
+      throw new Error("Conv1D: kernelSize cannot exceed inputLength with valid padding");
+    }
+    if (inputChannels < 1) {
+      throw new Error("Conv1D: inputChannels must be >= 1");
+    }
+    this.inputLength = inputLength;
+    this.kernelSize = kernelSize;
+    this.filters = filters;
+    this.stride = stride;
+    this.padding = padding;
+    this.inputChannels = inputChannels;
+    const limit = Math.sqrt(2 / (kernelSize * inputChannels));
+    this.kernels = Array.from(
+      { length: filters },
+      () => Array.from(
+        { length: kernelSize },
+        () => Array.from({ length: inputChannels }, () => (Math.random() * 2 - 1) * limit)
+      )
+    );
+    this.biases = new Array(filters).fill(0);
+    this._kOpts = Array.from(
+      { length: filters },
+      () => Array.from(
+        { length: kernelSize },
+        () => Array.from({ length: inputChannels }, () => optimizerFactory())
+      )
+    );
+    this._bOpts = Array.from({ length: filters }, () => optimizerFactory());
+  }
+  // ── Forward ───────────────────────────────────────────────────────────────
+  // Accepts either number[] (when inputChannels=1) or number[][] (multi-channel).
+  forward(input) {
+    const input2D = this._normalizeInput(input);
+    this._input = input2D.map((row) => [...row]);
+    let padded;
+    if (this.padding === "same") {
+      const padSize = Math.floor((this.kernelSize - 1) / 2);
+      const padRow = new Array(this.inputChannels).fill(0);
+      padded = new Array(padSize).fill(null).map(() => [...padRow]).concat(input2D).concat(new Array(padSize).fill(null).map(() => [...padRow]));
+    } else {
+      padded = input2D;
+    }
+    this._paddedInput = padded;
+    const outputLength = Math.floor((padded.length - this.kernelSize) / this.stride) + 1;
+    const output = Array.from(
+      { length: this.filters },
+      () => new Array(outputLength).fill(0)
+    );
+    for (let f = 0; f < this.filters; f++) {
+      for (let pos = 0; pos < outputLength; pos++) {
+        const start = pos * this.stride;
+        let sum = this.biases[f];
+        for (let k = 0; k < this.kernelSize; k++) {
+          for (let c = 0; c < this.inputChannels; c++) {
+            sum += this.kernels[f][k][c] * padded[start + k][c];
+          }
+        }
+        output[f][pos] = sum;
+      }
+    }
+    return output;
+  }
+  // ── Backward ──────────────────────────────────────────────────────────────
+  backward(dOut, lr = 1e-3) {
+    if (!this._paddedInput || !this._input) {
+      throw new Error("Conv1D.backward: call forward() first");
+    }
+    const padded = this._paddedInput;
+    const outputLength = dOut[0].length;
+    const dKernels = Array.from(
+      { length: this.filters },
+      () => Array.from(
+        { length: this.kernelSize },
+        () => new Array(this.inputChannels).fill(0)
+      )
+    );
+    const dBiases = new Array(this.filters).fill(0);
+    const dPadded = padded.map((row) => new Array(this.inputChannels).fill(0));
+    for (let f = 0; f < this.filters; f++) {
+      for (let pos = 0; pos < outputLength; pos++) {
+        const start = pos * this.stride;
+        dBiases[f] += dOut[f][pos];
+        for (let k = 0; k < this.kernelSize; k++) {
+          for (let c = 0; c < this.inputChannels; c++) {
+            dKernels[f][k][c] += dOut[f][pos] * padded[start + k][c];
+            dPadded[start + k][c] += dOut[f][pos] * this.kernels[f][k][c];
+          }
+        }
+      }
+    }
+    for (let f = 0; f < this.filters; f++) {
+      for (let k = 0; k < this.kernelSize; k++) {
+        for (let c = 0; c < this.inputChannels; c++) {
+          this.kernels[f][k][c] = this._kOpts[f][k][c].step(this.kernels[f][k][c], dKernels[f][k][c], lr);
+        }
+      }
+      this.biases[f] = this._bOpts[f].step(this.biases[f], dBiases[f], lr);
+    }
+    if (this.padding === "same") {
+      const padSize = Math.floor((this.kernelSize - 1) / 2);
+      return dPadded.slice(padSize, padSize + this.inputLength);
+    }
+    return dPadded.slice(0, this.inputLength);
+  }
+  // ── Output length ─────────────────────────────────────────────────────────
+  getOutputLength() {
+    if (this.padding === "same") {
+      return Math.ceil(this.inputLength / this.stride);
+    }
+    return Math.floor((this.inputLength - this.kernelSize) / this.stride) + 1;
+  }
+  // ── Flat weight serialization ─────────────────────────────────────────────
+  // Order: kernels (flattened), biases.
+  getWeights() {
+    const w = [];
+    for (const kernel of this.kernels)
+      for (const k of kernel)
+        for (const c of k)
+          w.push(c);
+    w.push(...this.biases);
+    return w;
+  }
+  setWeights(weights) {
+    let idx = 0;
+    for (let f = 0; f < this.filters; f++)
+      for (let k = 0; k < this.kernelSize; k++)
+        for (let c = 0; c < this.inputChannels; c++)
+          this.kernels[f][k][c] = weights[idx++];
+    for (let f = 0; f < this.filters; f++)
+      this.biases[f] = weights[idx++];
+  }
+  // ── Normalize input to 2D format ─────────────────────────────────────────
+  _normalizeInput(input) {
+    if (input.length === 0) {
+      throw new Error("Conv1D.forward: input cannot be empty");
+    }
+    if (typeof input[0] === "number") {
+      if (this.inputChannels !== 1) {
+        throw new Error(`Conv1D.forward: expected 2D input with ${this.inputChannels} channels, got 1D`);
+      }
+      const input1D = input;
+      if (input1D.length !== this.inputLength) {
+        throw new Error(`Conv1D.forward: expected input of length ${this.inputLength}, got ${input1D.length}`);
+      }
+      return input1D.map((v) => [v]);
+    }
+    const input2D = input;
+    if (input2D.length !== this.inputLength) {
+      throw new Error(`Conv1D.forward: expected input of length ${this.inputLength}, got ${input2D.length}`);
+    }
+    for (let i = 0; i < input2D.length; i++) {
+      if (input2D[i].length !== this.inputChannels) {
+        throw new Error(`Conv1D.forward: expected ${this.inputChannels} channels at position ${i}, got ${input2D[i].length}`);
+      }
+    }
+    return input2D;
+  }
+};
+
+// src/Trainer.ts
+var Trainer = class {
+  constructor(network, options = {}) {
+    this._history = [];
+    this._bestLoss = Infinity;
+    this._patienceCounter = 0;
+    this._stopReason = "maxEpochs";
+    this._metrics = [];
+    this.network = network;
+    this.epochs = options.epochs ?? 1e3;
+    this.lrInitial = options.lr ?? 0.1;
+    this.lrDecay = options.lrDecay ?? 1;
+    this.verbose = options.verbose ?? false;
+    this.weightDecay = options.weightDecay ?? 0;
+    this._earlyStopping = options.earlyStopping;
+    this._computeMetrics = options.computeMetrics ?? false;
+    this.clipValue = options.clipValue ?? 0;
+  }
+  // ── Set external validation data (for early stopping) ────────────────────
+  setValidationData(dataset) {
+    if (dataset.inputs.length !== dataset.targets.length) {
+      throw new Error(
+        "Trainer.setValidationData: inputs and targets must have the same length"
+      );
+    }
+    this._validationData = dataset;
+  }
+  // ── Get best validation loss during training ─────────────────────────────
+  getBestLoss() {
+    return this._bestLoss === Infinity ? -1 : this._bestLoss;
+  }
+  // ── Why did training stop? ───────────────────────────────────────────────
+  getStopReason() {
+    return this._stopReason;
+  }
+  // ── Get per-epoch classification metrics ─────────────────────────────────
+  getMetrics() {
+    return [...this._metrics];
+  }
+  // ── Train on dataset ──────────────────────────────────────────────────────
+  train(dataset) {
+    const { inputs, targets } = dataset;
+    if (inputs.length !== targets.length) {
+      throw new Error(
+        "Trainer.train: inputs and targets must have the same length"
+      );
+    }
+    const n = inputs.length;
+    let lr = this.lrInitial;
+    this._history = [];
+    this._bestLoss = Infinity;
+    this._patienceCounter = 0;
+    this._stopReason = "maxEpochs";
+    this._metrics = [];
+    const netExt = this._hasWeights(this.network);
+    if (this.weightDecay > 0 && !netExt) {
+      console.warn(
+        "Trainer: weightDecay requires a network with getWeights/setWeights/predict. Skipping weight decay."
+      );
+    }
+    if (this._earlyStopping && !netExt) {
+      console.warn(
+        "Trainer: earlyStopping requires a network with predict(). Skipping early stopping."
+      );
+    }
+    if (this._computeMetrics && !netExt) {
+      console.warn(
+        "Trainer: computeMetrics requires a network with predict(). Skipping metrics."
+      );
+    }
+    const canDecay = this.weightDecay > 0 && netExt;
+    const canValidate = !!this._earlyStopping && netExt && !!this._validationData;
+    const canMetric = this._computeMetrics && netExt;
+    const isClass = canMetric && this._isClassification(targets);
+    if (canMetric && !isClass) {
+      console.warn(
+        "Trainer: computeMetrics is set but targets do not appear to be one-hot or single-class. Metrics will be skipped."
+      );
+    }
+    for (let epoch = 0; epoch < this.epochs; epoch++) {
+      const indices = Array.from({ length: n }, (_, i) => i);
+      for (let i = n - 1; i > 0; i--) {
+        const j = Math.floor(Math.random() * (i + 1));
+        [indices[i], indices[j]] = [indices[j], indices[i]];
+      }
+      let epochLoss = 0;
+      for (const i of indices) {
+        if (canDecay) {
+          const w = netExt.getWeights();
+          for (let j = 0; j < w.length; j++) {
+            w[j] *= 1 - lr * this.weightDecay;
+          }
+          netExt.setWeights(w);
+        }
+        epochLoss += this.network.train(inputs[i], targets[i], lr);
+      }
+      epochLoss /= n;
+      this._history.push(epochLoss);
+      if (canMetric && isClass) {
+        this._metrics.push(this._computeMetricsArray(netExt, inputs, targets));
+      }
+      if (canValidate && this._validationData) {
+        const valLoss = this._computeLoss(netExt, this._validationData);
+        const minDelta = this._earlyStopping.minDelta;
+        if (valLoss < this._bestLoss - minDelta) {
+          this._bestLoss = valLoss;
+          this._patienceCounter = 0;
+        } else {
+          this._patienceCounter++;
+        }
+        if (this._patienceCounter >= this._earlyStopping.patience) {
+          this._stopReason = "earlyStopping";
+          break;
+        }
+      }
+      lr *= this.lrDecay;
+      if (this.verbose && (epoch + 1) % 100 === 0) {
+        console.log(
+          `Epoch ${epoch + 1}/${this.epochs}, loss: ${epochLoss.toFixed(6)}, lr: ${lr.toFixed(6)}`
+        );
+      }
+    }
+    return this._history;
+  }
+  // ── Get loss history ──────────────────────────────────────────────────────
+  getHistory() {
+    return [...this._history];
+  }
+  // ── Private helpers ───────────────────────────────────────────────────────
+  /** Type guard: does this network support getWeights/setWeights/predict? */
+  _hasWeights(network) {
+    if ("getWeights" in network && "setWeights" in network && "predict" in network && typeof network.getWeights === "function" && typeof network.setWeights === "function" && typeof network.predict === "function") {
+      return network;
+    }
+    return null;
+  }
+  /** Mean squared error on a dataset (used for validation loss). */
+  _computeLoss(net, data) {
+    let totalLoss = 0;
+    for (let i = 0; i < data.inputs.length; i++) {
+      const pred = net.predict(data.inputs[i]);
+      const target = data.targets[i];
+      let sampleLoss = 0;
+      for (let j = 0; j < pred.length; j++) {
+        sampleLoss += (target[j] - pred[j]) ** 2;
+      }
+      totalLoss += sampleLoss / pred.length;
+    }
+    return totalLoss / data.inputs.length;
+  }
+  /** Heuristic: are targets classification-style (one-hot or single-class)? */
+  _isClassification(targets) {
+    if (targets.length === 0) return false;
+    const first = targets[0];
+    if (first.length === 1) return true;
+    for (const t of targets) {
+      let sum = 0;
+      for (const v of t) {
+        sum += v;
+        if (v < -0.01 || v > 0.01 && v < 0.99 && Math.abs(v - 1) > 0.01)
+          return false;
+      }
+      if (Math.abs(sum - 1) > 0.01) return false;
+    }
+    return true;
+  }
+  /** Compute classification metrics from predictions vs targets. */
+  _computeMetricsArray(net, inputs, targets) {
+    const targetLen = targets[0].length;
+    const nClasses = targetLen === 1 ? 2 : targetLen;
+    const confusion = Array.from(
+      { length: nClasses },
+      () => Array(nClasses).fill(0)
+    );
+    for (let i = 0; i < inputs.length; i++) {
+      const pred = net.predict(inputs[i]);
+      const target = targets[i];
+      let predClass;
+      let trueClass;
+      if (targetLen === 1) {
+        trueClass = target[0] >= 0.5 ? 1 : 0;
+        if (pred.length === 1) {
+          predClass = pred[0] >= 0.5 ? 1 : 0;
+        } else {
+          predClass = pred.indexOf(Math.max(...pred));
+        }
+      } else {
+        predClass = pred.indexOf(Math.max(...pred));
+        trueClass = target.indexOf(Math.max(...target));
+      }
+      predClass = Math.max(0, Math.min(nClasses - 1, predClass));
+      trueClass = Math.max(0, Math.min(nClasses - 1, trueClass));
+      confusion[trueClass][predClass]++;
+    }
+    let totalCorrect = 0;
+    let totalSamples = 0;
+    const precisions = [];
+    const recalls = [];
+    for (let c = 0; c < nClasses; c++) {
+      const tp = confusion[c][c];
+      totalCorrect += tp;
+      let colSum = 0;
+      let rowSum = 0;
+      for (let r = 0; r < nClasses; r++) {
+        colSum += confusion[r][c];
+        rowSum += confusion[c][r];
+      }
+      totalSamples += rowSum;
+      precisions.push(colSum > 0 ? tp / colSum : 0);
+      recalls.push(rowSum > 0 ? tp / rowSum : 0);
+    }
+    const accuracy = totalSamples > 0 ? totalCorrect / totalSamples : 0;
+    const macroPrecision = precisions.reduce((a, b) => a + b, 0) / nClasses;
+    const macroRecall = recalls.reduce((a, b) => a + b, 0) / nClasses;
+    const f1 = macroPrecision + macroRecall > 0 ? 2 * macroPrecision * macroRecall / (macroPrecision + macroRecall) : 0;
+    return {
+      accuracy,
+      precision: macroPrecision,
+      recall: macroRecall,
+      f1
+    };
+  }
+};
+
+// src/DataLoader.ts
+var DataLoader = class _DataLoader {
+  constructor(data, batchSize = 1, validationSplit = 0) {
+    if (data.inputs.length !== data.targets.length) {
+      throw new Error("DataLoader: inputs and targets must have the same length");
+    }
+    if (validationSplit < 0 || validationSplit >= 1) {
+      throw new Error(`DataLoader: validationSplit must be in [0, 1), got ${validationSplit}`);
+    }
+    this.data = data;
+    this.batchSize = batchSize;
+    this._validationSplit = validationSplit;
+    const fullIndices = Array.from({ length: data.inputs.length }, (_, i) => i);
+    for (let i = fullIndices.length - 1; i > 0; i--) {
+      const j = Math.floor(Math.random() * (i + 1));
+      [fullIndices[i], fullIndices[j]] = [fullIndices[j], fullIndices[i]];
+    }
+    if (validationSplit > 0) {
+      const valSize = Math.round(data.inputs.length * validationSplit);
+      const trainSize = data.inputs.length - valSize;
+      this._trainIndices = fullIndices.slice(0, trainSize);
+      this._valIndices = fullIndices.slice(trainSize);
+    } else {
+      this._trainIndices = [...fullIndices];
+      this._valIndices = [];
+    }
+    this._indices = [...this._trainIndices];
+    this._pos = 0;
+  }
+  // ── Shuffle the training data ──────────────────────────────────────────────
+  shuffle() {
+    for (let i = this._trainIndices.length - 1; i > 0; i--) {
+      const j = Math.floor(Math.random() * (i + 1));
+      [this._trainIndices[i], this._trainIndices[j]] = [this._trainIndices[j], this._trainIndices[i]];
+    }
+    this._indices = [...this._trainIndices];
+    this._pos = 0;
+  }
+  // ── Check if more batches are available ───────────────────────────────────
+  hasNext() {
+    return this._pos < this._indices.length;
+  }
+  // ── Get next batch ────────────────────────────────────────────────────────
+  next() {
+    const end = Math.min(this._pos + this.batchSize, this._indices.length);
+    const batchIndices = this._indices.slice(this._pos, end);
+    this._pos = end;
+    return {
+      inputs: batchIndices.map((i) => this.data.inputs[i]),
+      targets: batchIndices.map((i) => this.data.targets[i])
+    };
+  }
+  // ── Reset iteration ───────────────────────────────────────────────────────
+  reset() {
+    this._pos = 0;
+  }
+  // ── Get total number of training samples ───────────────────────────────────
+  get length() {
+    return this._trainIndices.length;
+  }
+  // ── Get validation data as a DataPair ──────────────────────────────────────
+  // Returns the validation samples (inputs + targets) in their shuffled order.
+  // Returns empty arrays if no validation split was configured.
+  getValidationData() {
+    return {
+      inputs: this._valIndices.map((i) => this.data.inputs[i]),
+      targets: this._valIndices.map((i) => this.data.targets[i])
+    };
+  }
+  // ── Get number of validation samples ───────────────────────────────────────
+  get validationLength() {
+    return this._valIndices.length;
+  }
+  // ── Create sequence windows from a time series ────────────────────────────
+  static sequences(data, seqLen, validationSplit = 0) {
+    if (data.length < seqLen + 1) {
+      throw new Error("DataLoader.sequences: data length must be >= seqLen + 1");
+    }
+    const inputs = [];
+    const targets = [];
+    for (let i = 0; i <= data.length - seqLen - 1; i++) {
+      inputs.push(data.slice(i, i + seqLen).flat());
+      targets.push(data[i + seqLen]);
+    }
+    return new _DataLoader({ inputs, targets }, 1, validationSplit);
+  }
+};
+
+// src/LRScheduler.ts
+var LRScheduler = class {
+  // ── Step Decay ────────────────────────────────────────────────────────────
+  // lr = initialLr * dropRate^floor(epoch / epochsDrop)
+  stepDecay(lr, epoch, dropRate, epochsDrop) {
+    return lr * Math.pow(dropRate, Math.floor(epoch / epochsDrop));
+  }
+  // ── Exponential Decay ─────────────────────────────────────────────────────
+  // lr = initialLr * decayRate^epoch
+  exponentialDecay(lr, epoch, decayRate) {
+    return lr * Math.pow(decayRate, epoch);
+  }
+  // ── Plateau Decay ─────────────────────────────────────────────────────────
+  // If loss hasn't improved for `patience` epochs, multiply lr by `factor`.
+  // Returns the new lr. Call this after each epoch with the current loss.
+  //
+  // Usage:
+  //   let patience_counter = 0
+  //   let best_loss = Infinity
+  //   for (let epoch = 0; epoch < 1000; epoch++) {
+  //     const loss = train(...)
+  //     lr = scheduler.plateauDecay(lr, loss, history, 10, 0.5)
+  //   }
+  plateauDecay(lr, currentLoss, history, patience, factor) {
+    if (history.length < patience) return lr;
+    const recentLosses = history.slice(-patience);
+    const minRecentLoss = Math.min(...recentLosses);
+    if (currentLoss >= minRecentLoss) {
+      return lr * factor;
+    }
+    return lr;
+  }
+  // ── Cosine Annealing ──────────────────────────────────────────────────────
+  // lr = minLr + 0.5 * (maxLr - minLr) * (1 + cos(π * epoch / maxEpochs))
+  cosineAnnealing(lr, epoch, maxEpochs, minLr = 0) {
+    return minLr + 0.5 * (lr - minLr) * (1 + Math.cos(Math.PI * epoch / maxEpochs));
+  }
+};
+
+// src/ModelSaver.ts
+var ModelSaver = class _ModelSaver {
+  // ── Serialize to JSON string ──────────────────────────────────────────────
+  static toJSON(model) {
+    return JSON.stringify({
+      weights: model.getWeights(),
+      timestamp: Date.now()
+    });
+  }
+  // ── Deserialize from JSON string ──────────────────────────────────────────
+  static fromJSON(model, json) {
+    const data = JSON.parse(json);
+    if (!data.weights || !Array.isArray(data.weights)) {
+      throw new Error("ModelSaver.fromJSON: invalid model data");
+    }
+    model.setWeights(data.weights);
+  }
+  // ── Save to file (requires write function) ────────────────────────────────
+  static saveToFile(model, path, writeFn) {
+    const json = _ModelSaver.toJSON(model);
+    writeFn(path, json);
+  }
+  // ── Load from file (requires read function) ───────────────────────────────
+  static loadFromFile(model, path, readFn) {
+    const json = readFn(path);
+    _ModelSaver.fromJSON(model, json);
+  }
+};
 // Annotate the CommonJS export names for ESM import in node:
 0 && (module.exports = {
   Adam,
   AttentionHead,
+  BatchNorm,
+  BiasVector,
+  ClipOptimizer,
+  ClippedOptimizerFactory,
+  Conv1D,
+  DataLoader,
+  Dropout,
   EmbeddingMatrix,
+  GRULayer,
+  LRScheduler,
   LSTMLayer,
   Layer,
   LayerNorm,
+  ModelSaver,
   Momentum,
   MultiHeadAttention,
   Network,
@@ -1305,11 +2702,13 @@ function crossEntropyDeltaRaw(predicted, actual) {
   Neuron,
   NeuronN,
   SGD,
+  Trainer,
   TransformerBlock,
   WeightMatrix,
   crossEntropy,
   crossEntropyDelta,
   crossEntropyDeltaRaw,
+  defaultOptimizer,
   elu,
   leakyRelu,
   linear,
@@ -1323,5 +2722,9 @@ function crossEntropyDeltaRaw(predicted, actual) {
   softmax,
   softmaxBackward,
   tanh,
-  transpose
+  transpose,
+  validate2DArray,
+  validateArray,
+  validateArrayMinLength,
+  validateNumber
 });