mini-jstorch 2.0.1 → 2.0.2

package/Docs/API.md

@@ -0,0 +1,277 @@
+# API Reference
+
+## Model Container
+
+### Sequential
+
+```js
+new Sequential(layers: Layer[])
+```
+
+Container that chains layers sequentially.
+
+**Methods:**
+- forward(x) — Pass input through all layers
+- backward(grad) — Backpropagate gradient through all layers
+- parameters() — Returns [{param, grad}, ...] for all trainable parameters
+- zeroGrad() — Zero all parameter gradients
+- train() — Set all layers to training mode
+- eval() — Set all layers to evaluation mode
+- stateDict() — Get {layer_0.weight, layer_0.bias, ...}
+- loadStateDict(dict) — Load weights from state dict object
+- step(lr) — Apply SGD step to all layers directly
+
+---
+
+## Layers
+
+### Linear
+
+```js
+new Linear(inFeatures: number, outFeatures: number)
+```
+
+Fully connected layer. Weight shape: [inFeatures, outFeatures]. Bias shape: [1, outFeatures].
+
+### Conv2D (experimental)
+
+```js
+new Conv2D(inChannels: number, outChannels: number, kernelSize: number, stride?: number, padding?: number)
+```
+
+2D convolution layer. Input shape: [batch, channels, height, width].
+
+### Flatten
+
+```js
+new Flatten()
+```
+
+Flattens multi-dimensional input per sample. Preserves batch dimension.
+
+---
+
+## Activations
+
+All activations follow the same interface:
+- forward(x) — x: [batch, features], returns: [batch, features]
+- backward(grad) — grad: [batch, features], returns: [batch, features]
+
+| Class | Formula |
+|-------|---------|
+| ReLU() | max(0, x) |
+| Sigmoid() | 1 / (1 + exp(-x)) |
+| Tanh() | tanh(x) |
+| LeakyReLU(alpha?) | x > 0 ? x : alpha * x |
+| GELU() | 0.5 * x * (1 + tanh(sqrt(2/pi) * (x + 0.044715 * x^3))) |
+| ELU(alpha?) | x > 0 ? x : alpha * (exp(x) - 1) |
+| Mish() | x * tanh(ln(1 + exp(x))) |
+| SiLU() | x * sigmoid(x) |
+| Softmax(dim?) | exp(x - max) / sum(exp(x - max)) |
+
+---
+
+## Loss Functions
+
+### MSELoss
+
+```js
+new MSELoss()
+```
+
+loss = mean((pred - target)^2)
+gradient = 2 * (pred - target) / batchSize
+
+### SoftmaxCrossEntropyLoss (recommended for classification)
+
+```js
+new SoftmaxCrossEntropyLoss()
+```
+
+Combines softmax + cross-entropy in a numerically stable way. Input: logits (not probabilities). Do NOT combine with a Softmax layer.
+
+### BCEWithLogitsLoss (recommended for binary classification)
+
+```js
+new BCEWithLogitsLoss()
+```
+
+Combines sigmoid + binary cross-entropy. Numerically stable. Input: logits (not probabilities). Do NOT combine with a Sigmoid layer.
+
+### CrossEntropyLoss (deprecated)
+
+```js
+new CrossEntropyLoss()
+```
+
+Use SoftmaxCrossEntropyLoss instead. Exists for backward compatibility.
+
+---
+
+## Optimizers
+
+### Adam (recommended)
+
+```js
+new Adam(parameters, options)
+// or
+new Adam(parameters, lr, beta1, beta2, eps, maxGradNorm)
+```
+
+Options object: { lr: 0.001, b1: 0.9, b2: 0.999, eps: 1e-8, max_grad_norm: 1.0 }
+
+### AdamW
+
+```js
+new AdamW(parameters, options)
+```
+
+Adam with decoupled weight decay. Options include weight_decay (default: 0.01).
+
+### SGD
+
+```js
+new SGD(parameters, lr?, maxGradNorm?)
+```
+
+### Lion
+
+```js
+new LION(parameters, options)
+```
+
+Memory-efficient optimizer.
+
+**All optimizers have:**
+- step() — Update parameters using accumulated gradients
+- zeroGrad() — Zero all parameter gradients
+
+---
+
+## Learning Rate Schedulers
+
+### StepLR
+
+```js
+new StepLR(optimizer, stepSize, gamma)
+```
+
+Multiplies LR by gamma every stepSize steps.
+
+### LambdaLR
+
+```js
+new LambdaLR(optimizer, fn)
+```
+
+Sets LR = baseLr * fn(epoch) on each step.
+
+### ReduceLROnPlateau
+
+```js
+new ReduceLROnPlateau(optimizer, options)
+```
+
+Reduces LR when loss stops improving. Options: patience, factor, min_lr, threshold, cooldown, verbose.
+
+---
+
+## Regularization
+
+### Dropout
+
+```js
+new Dropout(p?)
+```
+
+Randomly zeros p fraction of neurons during training. Default p = 0.5. Call layer.train() / layer.eval() to toggle.
+
+### BatchNorm2d (experimental)
+
+```js
+new BatchNorm2d(numFeatures, eps?, momentum?, affine?)
+```
+
+2D batch normalization. Input shape: [batch, channels, height, width].
+
+---
+
+## Tokenizer
+
+```js
+new Tokenizer(vocabSize?)
+```
+
+| Method | Description |
+|--------|-------------|
+| fit(texts) | Build vocabulary from text array |
+| transform(texts, maxLength?, padToMax?) | Convert texts to token indices |
+| fitTransform(texts, maxLength?, padToMax?) | Fit + transform in one call |
+| inverseTransform(tokens, skipPad?) | Convert token indices back to text |
+| getVocabulary() | Get vocabulary as string array |
+| getVocabSize() | Get vocabulary size |
+| getWordCounts() | Get word frequency map |
+| getMostCommon(n?) | Get top N most frequent words |
+
+---
+
+## Utilities
+
+zeros(rows, cols) — Matrix filled with 0
+ones(rows, cols) — Matrix filled with 1
+randomMatrix(rows, cols, scale?) — Random matrix (Xavier init if scale omitted)
+transpose(matrix) — Matrix transpose
+dot(a, b) — Matrix multiplication
+addMatrices(a, b) — Element-wise addition
+softmax(vector) — Softmax on 1D array
+crossEntropy(pred, target) — Cross-entropy loss (scalar)
+reshape(tensor, rows, cols) — Reshape to new dimensions
+flattenBatch(batch) — Flatten batch to 2D
+concat(a, b, axis) — Concatenate along axis 0 or 1
+stack(tensors) — Stack tensors
+
+---
+
+## Model Persistence
+
+saveModel(model) — Serialize model to JSON string
+loadModel(model, json) — Load weights from JSON into model
+
+Supports all layer types. Validates layer types and shapes. Logs warnings for mismatches.
+
+---
+
+## Tensor (Advanced)
+
+```js
+new Tensor(data, requiresGrad?)
+```
+
+| Method | Description |
+|--------|-------------|
+| add(tensor) | Element-wise addition |
+| mul(tensor) | Element-wise multiplication |
+| matmul(tensor) | Matrix multiplication |
+| transpose() | Transpose tensor |
+| flatten() | Flatten to 1D array |
+| shape() | Returns [rows, cols] |
+
+Static: Tensor.zeros(r,c), Tensor.ones(r,c), Tensor.random(r,c,scale?)
+
+---
+
+## User-Friendly Utilities (fu_*)
+
+fu_tensor(data, requiresGrad?) — Create tensor from 2D array
+fu_add(a, b) — Element-wise add
+fu_mul(a, b) — Element-wise multiply
+fu_matmul(a, b) — Matrix multiply
+fu_sum(tensor) — Sum all elements
+fu_mean(tensor) — Mean of all elements
+fu_relu(tensor) — ReLU activation
+fu_sigmoid(tensor) — Sigmoid activation
+fu_tanh(tensor) — Tanh activation
+fu_softmax(tensor) — Softmax activation
+fu_flatten(tensor) — Flatten to 1D
+fu_reshape(tensor, rows, cols) — Reshape tensor
+fu_stack(tensors) — Stack tensors

package/README.md

@@ -1,4 +1,4 @@
-## Mini-JSTorch (MAJOR UPDATE)
+## Mini-JSTorch (v2.0.2)
 
 ---
 
@@ -7,15 +7,14 @@ It runs in Node.js and modern browsers, with a simple API inspired by PyTorch-st
 
 This project prioritizes `clarity`, `numerical correctness`, and `accessibility` over performance or large-scale production use.
 
-In this version `2.0.0`, we introduce:
-- **Fixed Linear layer cache** (critical bug fix for training)
-- **Fixed GELU gradient calculation**
-- **Fixed MSELoss gradient scaling**
-- **Optimized Softmax gradient** (O(n²) → O(n))
-- **Improved Tokenizer** with proper PAD/UNK separation
-- **Added Sequential.zeroGrad(), train(), eval(), stateDict() methods**
+### Changelog
 
----
+**v2.0.2:**
+- **Fixed critical training bug:** Optimizers (Adam, SGD, AdamW, Lion) now correctly update Linear and Conv2D layer weights
+- **Fixed BatchNorm2d:** Inference mode no longer produces NaN for multi-channel inputs
+- **Fixed ELU activation:** Backward pass now uses correct derivative formula
+- **Fixed saveModel/loadModel:** Now correctly saves and restores all layer types including Conv2D and BatchNorm2d
+- **Fixed BatchNorm2d gradient zeroing:** gradWeight/gradBias now correctly reset between batches
 
 **⚠️ BREAKING CHANGES in v2.0.0:**
 - Tokenizer API: `tokenizeBatch()` → `transform()`, `detokenizeBatch()` → `inverseTransform()`
@@ -80,7 +79,6 @@ In Browser/Website:
         async function train() {
             const statusEl = document.getElementById('status');
             const logEl = document.getElementById('log');
-            
             try {
                 const model = new Sequential([
                     new Linear(2, 16), new Tanh(),
@@ -107,13 +105,13 @@ In Browser/Website:
                     }
                 }
 
-                statusEl.textContent = '✅ Done';
+                statusEl.textContent = 'Done';
                 const preds = model.forward(X);
                 document.getElementById('res').innerHTML = `<h4>Results:</h4>` + 
                     X.map((input, i) => `[${input}] -> <b>${preds[i][0].toFixed(4)}</b> (Target: ${y[i][0]})`).join('<br>');
 
             } catch (e) {
-                statusEl.textContent = '❌ Error: ' + e.message;
+                statusEl.textContent = 'Error: ' + e.message;
             }
         }
         train();
@@ -319,11 +317,16 @@ console.log(`\nAccuracy: ${(correct / X.length * 100).toFixed(2)}%`);
 # Save & Load Models 
 
 ```javascript
-// WARN: Error/Bug may be expected for this time!
-import { saveModel, loadModel, Sequential } from "./src/jstorch.js";
+import { saveModel, loadModel, Sequential } from "./src/jstorch";
 
+// Save trained model 
 const json = saveModel(model);
-const model2 = new Sequential([...]); // same architecture
+
+// Create fresh model with same architecture and load weights 
+const model2 = new Sequential([
+    new Linear(2, 16), new ReLU(),
+    new Linear(16, 1)
+]);
 loadModel(model2, json);
 ```
 
@@ -360,7 +363,7 @@ node demo/<fileNameInDemo>.js
 
 MIT License
 
-Copyright (c) 2024
+Copyright (c) 2024-2025
 rizal-editors
 
 ---

package/demo/linear_regression.js

@@ -39,4 +39,4 @@ finalPred.forEach((p, i) => {
 });
 console.log(`\nAverage Error: ${(totalError / X.length).toFixed(2)}`);
 console.log(`Weight (slope): ${model.layers[0].W[0][0].toFixed(4)} (expected: 2.0)`);
-console.log(`Bias: ${model.layers[0].b[0].toFixed(4)} (expected: 0.0)`);
+console.log(`Bias: ${model.layers[0].b[0][0].toFixed(4)} (expected: 0.0)`);

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "mini-jstorch",
-  "version": "2.0.1",
+  "version": "2.0.2",
   "type": "module", 
   "description": "A lightweight JavaScript neural network library for learning AI concepts and rapid Frontend experimentation. PyTorch-inspired, zero dependencies, perfect for educational use.",
   "main": "index.js",
@@ -15,8 +15,7 @@
     "tiny-ml",
     "mini-neural-network",
     "mini-ml-library",
-    "mini-js-ml",
-    "educational-ml"
+    "mini-js-ml"
   ],
   "author": "Rizal",
   "license": "MIT"

package/src/jstorch.js

@@ -81,62 +81,26 @@ export function crossEntropy(pred,target){
 
 // ---------------------- USERS FRIENDLY UTILS (USE THIS FOR YOUR UTILS!) ----------------
 export function fu_tensor(data, requiresGrad = false) {
-    if (!Array.isArray(data) || !Array.isArray(data[0])) {
-        throw new Error("fu_tensor: Data must be 2D array");
-    }
-    const tensor = new Tensor(data);
-    tensor.requiresGrad = requiresGrad;
-    return tensor;
+    if (!Array.isArray(data) || !Array.isArray(data[0])){
+		throw new Error("fu_tensor: Data must be 2D array");
+	}
+	return new Tensor(data, requiresGrad);
 }
 
-// fu_add
-export function fu_add(a, b) {
-    if (!(a instanceof Tensor) && !(b instanceof Tensor)) {
-        throw new Error("fu_add: At least one operand must be Tensor");
-    }
-    
-    if (!(a instanceof Tensor)) {
-        a = fu_tensor(Array(b.shape()[0]).fill().map(() => 
-            Array(b.shape()[1]).fill(a)
-        ));
-    }
-    
-    if (!(b instanceof Tensor)) {
-        b = fu_tensor(Array(a.shape()[0]).fill().map(() => 
-            Array(a.shape()[1]).fill(b)
-        ));
-    }
-    
-    if (a.shape()[0] !== b.shape()[0] || a.shape()[1] !== b.shape()[1]) {
-        throw new Error(`fu_add: Shape mismatch ${a.shape()} vs ${b.shape()}`);
-    }
-    
-    return new Tensor(a.data.map((r, i) => r.map((v, j) => v + b.data[i][j])));
+// fu_add 
+export function fu_add(a, b){
+	if (!(a instanceof Tensor)) a = fu_tensor(a);
+	if (!(b instanceof Tensor)) b = fu_tensor(b);
+	
+	return a.add(b);
 }
 
 // fu_mul
-export function fu_mul(a, b) {
-    if (!(a instanceof Tensor) && !(b instanceof Tensor)) {
-        throw new Error("fu_mul: At least one operand must be Tensor");
-    }
-    
-    if (!(a instanceof Tensor)) {
-        a = fu_tensor(Array(b.shape()[0]).fill().map(() => 
-            Array(b.shape()[1]).fill(a)
-        ));
-    }
-    
-    if (!(b instanceof Tensor)) {
-        b = fu_tensor(Array(a.shape()[0]).fill().map(() => 
-            Array(a.shape()[1]).fill(b)
-        ));
-    }
-    
-    if (a.shape()[0] !== b.shape()[0] || a.shape()[1] !== b.shape()[1]) {
-        throw new Error(`fu_mul: Shape mismatch ${a.shape()} vs ${b.shape()}`);
-    }
-    
-    return new Tensor(a.data.map((r, i) => r.map((v, j) => v * b.data[i][j])));
+export function fu_mul(a, b){
+	if (!(a instanceof Tensor)) a = fu_tensor(a);
+	if (!(b instanceof Tensor)) b = fu_tensor(b);
+	
+	return a.mul(b);
 }
 
 // fu_matmul
@@ -144,11 +108,7 @@ export function fu_matmul(a, b) {
     if (!(a instanceof Tensor)) a = fu_tensor(a);
     if (!(b instanceof Tensor)) b = fu_tensor(b);
     
-    if (a.shape()[1] !== b.shape()[0]) {
-        throw new Error(`fu_matmul: Inner dimension mismatch ${a.shape()[1]} vs ${b.shape()[0]}`);
-    }
-    
-    return new Tensor(dot(a.data, b.data));
+    return a.matmul(b);
 }
 
 // fu_sum
@@ -344,276 +304,144 @@ export class Tensor {
 
 // ---------------------- Layers ----------------------
 export class Linear {
-    constructor(inputDim, outputDim){
-        this.W = randomMatrix(inputDim, outputDim);
-        this.b = Array(outputDim).fill(0);
-        this.gradW = zeros(inputDim, outputDim);
-        this.gradb = Array(outputDim).fill(0);
-        this.x = null;
-        this.originalShape = null;
-        
-        this._WFlat = null;
-        this._bFlat = null;
+    constructor(inFeatures, outFeatures) {
+        this.inFeatures = inFeatures;
+        this.outFeatures = outFeatures;
+
+        // Weights: [inFeatures, outFeatures]
+        this.W = randomMatrix(inFeatures, outFeatures);
+        this.gradW = zeros(inFeatures, outFeatures);
+
+        // Bias: [1, outFeatures]
+        this.b = [Array(outFeatures).fill(0)];
+        this.gradB = [Array(outFeatures).fill(0)];
+
+        this.x = null; // cache input
     }
-	
-	_updateCache() {
-		const rows = this.W.length;
-		const cols = this.W[0].length;
-		this._WFlat = new Float32Array(rows * cols);
-		for (let i = 0; i < rows; i++){
-			const offset = i * cols;
-			const row = this.W[i];
-			for (let j = 0; j < cols; j++){
-				this._WFlat[offset + j] = row[j];
-			}
-		}
-		this._bFlat = new Float32Array(this.b);
-	}
 
-    forward(x){
-        this.originalShape = this._getShapeType(x);
-        
-        if (this.originalShape === '3d') {
-            this.x = x.map(sample => sample[0]);
-        } else {
-            this.x = x;
-        }
-		
-		this._updateCache();
-		
-		const m = this.x.length;
-		const k = this.x[0].length;
-		const n = this.W[0].length;
-        
-        if (!this._WFlat) {
-            const rows = this.W.length;
-            const cols = this.W[0].length;
-            this._WFlat = new Float32Array(rows * cols);
-            for (let i = 0; i < rows; i++) {
-                const offset = i * cols;
-                const row = this.W[i];
-                for (let j = 0; j < cols; j++) {
-                    this._WFlat[offset + j] = row[j];
-                }
+    forward(x) {
+        // x: [batch, inFeatures]
+        this.x = x;
+
+        const out = dot(x, this.W); // [batch, outFeatures]
+
+        // add bias
+        for (let i = 0; i < out.length; i++) {
+            for (let j = 0; j < this.outFeatures; j++) {
+                out[i][j] += this.b[0][j];
             }
-            this._bFlat = new Float32Array(this.b);
         }
+
+        return out;
+    }
+
+    backward(grad) {
+        // grad: [batch, outFeatures]
+        const batchSize = grad.length;
         
-        // Flatten input x to Float32Array
-        const xFlat = new Float32Array(m * k);
-        for (let i = 0; i < m; i++) {
-            const row = this.x[i];
-            const offset = i * k;
-            for (let j = 0; j < k; j++) {
-                xFlat[offset + j] = row[j];
+        // mutate in place to preserve optimizer Reference
+        const xT = transpose(this.x);
+        const computedGradW = dot(xT, grad);
+        for (let i=0; i<this.inFeatures; i++){
+            for (let j=0; j<this.outFeatures; j++){
+                this.gradW[i][j] = computedGradW[i][j]
             }
         }
         
-        const outFlat = new Float32Array(m * n);
-        for (let i = 0; i < m; i++) {
-            const xOffset = i * k;
-            for (let j = 0; j < n; j++) {
-                let sum = 0;
-                for (let l = 0; l < k; l++) {
-                    sum += xFlat[xOffset + l] * this._WFlat[l * n + j];
-                }
-                outFlat[i * n + j] = sum + this._bFlat[j];
+        // gradB = sum over batch
+        for (let j=0; j<this.outFeatures; j++){
+            let sum = 0;
+            for (let i=0; i <batchSize; i++){
+                sum+=grad[i][j]
             }
+            this.gradB[0][j] = sum;
         }
         
-        const out = Array(m);
-        for (let i = 0; i < m; i++) {
-            const row = Array(n);
-            const offset = i * n;
-            for (let j = 0; j < n; j++) {
-                row[j] = outFlat[offset + j];
-            }
-            out[i] = row;
-        }
+        const WT = transpose(this.W);
+        const gradInput = dot(grad, WT);
         
-        return out;
+        return gradInput;
     }
 
-    backward(grad){
-		const m = this.x.length;
-		const k = this.W.length;      // input dim
-		const n = this.W[0].length;   // output dim
-    
-		// Convert grad to Float32Array
-		const gradFlat = new Float32Array(m * n);
-		for (let i = 0; i < m; i++) {
-			const row = grad[i];
-			const offset = i * n;
-			for (let j = 0; j < n; j++) {
-				gradFlat[offset + j] = row[j];
-			}
-		}
-    
-		// Convert x to Float32Array
-		const xFlat = new Float32Array(m * k);
-		for (let i = 0; i < m; i++) {
-			const row = this.x[i];
-			const offset = i * k;
-			for (let j = 0; j < k; j++) {
-				xFlat[offset + j] = row[j];
-			}
-		}
-    
-		// Reset gradW
-		for (let i = 0; i < this.gradW.length; i++) {
-			for (let j = 0; j < this.gradW[0].length; j++) {
-				this.gradW[i][j] = 0;
-			}
-		}
-    
-		// Compute gradW = x^T * grad
-		for (let i = 0; i < k; i++) {
-			for (let j = 0; j < n; j++) {
-				let sum = 0;
-				for (let batch = 0; batch < m; batch++) {
-					sum += xFlat[batch * k + i] * gradFlat[batch * n + j];
-				}
-				this.gradW[i][j] = sum;
-			}
-		}
-    
-		// Compute gradb
-		for (let j = 0; j < n; j++) {
-			let sum = 0;
-			for (let batch = 0; batch < m; batch++) {
-				sum += gradFlat[batch * n + j];
-			}
-			this.gradb[j] = sum;
-		}
-    
-		const gradInputFlat = new Float32Array(m * k);
-		for (let i = 0; i < m; i++) {
-			for (let j = 0; j < k; j++) {
-				let sum = 0;
-				for (let l = 0; l < n; l++) {
-					sum += gradFlat[i * n + l] * this.W[j][l];
-				}
-				gradInputFlat[i * k + j] = sum;
-			}
-		}
-    
-		// Convert back to 2D array
-		const gradInput = Array(m);
-		for (let i = 0; i < m; i++) {
-			const row = Array(k);
-			const offset = i * k;
-			for (let j = 0; j < k; j++) {
-				row[j] = gradInputFlat[offset + j];
-			}
-			gradInput[i] = row;
-		}
-    
-		if (this.originalShape === '3d') {
-			return gradInput.map(row => [row]);
-		}
-		return gradInput;
-	}
+    step(lr) {
+        for (let i = 0; i < this.inFeatures; i++) {
+            for (let j = 0; j < this.outFeatures; j++) {
+                this.W[i][j] -= lr * this.gradW[i][j];
+            }
+        }
 
-    _getShapeType(x) {
-        if (Array.isArray(x[0]) && Array.isArray(x[0][0]) && !Array.isArray(x[0][0][0])) {
-            return '3d';
-        } else if (Array.isArray(x[0]) && !Array.isArray(x[0][0])) {
-            return '2d';
-        } else {
-            throw new Error(`Unsupported input shape for Linear layer`);
+        for (let j = 0; j < this.outFeatures; j++) {
+            this.b[0][j] -= lr * this.gradB[0][j];
         }
     }
 
-    parameters(){ 
-        return [ 
-            {param: this.W, grad: this.gradW}, 
-            {param: [this.b], grad: [this.gradb]} 
-        ]; 
+    parameters() {
+        return [
+            { param: this.W, grad: this.gradW },
+            { param: this.b, grad: this.gradB }
+        ];
     }
 }
 
 export class Flatten {
-    constructor() {
-        this.originalShape = null;
-    }
-    
-    forward(x) {
-        // Always convert to [batch, features] format
-        this.originalShape = x.map(sample => this._getShape(sample));
-        
-        return x.map(sample => {
-            const flat = this._flatten(sample);
-            return flat; // Return as 1D array for [batch, features] compatibility
-        });
-    }
-    
-    backward(grad) {
-        // grad is [batch, features], reshape back to original shape
-        return grad.map((flatGrad, batchIdx) => {
-            const shape = this.originalShape[batchIdx];
-            return this._unflatten(flatGrad, shape);
-        });
-    }
-    
-    _getShape(sample) {
-        if (Array.isArray(sample[0]) && Array.isArray(sample[0][0])) {
-            return {
-                type: '3d',
-                dims: [sample.length, sample[0].length, sample[0][0].length]
-            };
-        } else if (Array.isArray(sample[0])) {
-            return {
-                type: '2d', 
-                dims: [sample.length, sample[0].length]
-            };
-        } else {
-            return {
-                type: '1d',
-                dims: [sample.length]
-            };
-        }
-    }
-    
-    _flatten(sample) {
-        if (Array.isArray(sample[0]) && Array.isArray(sample[0][0])) {
-            return sample.flat(2); // [channels, height, width] -> flat
-        } else if (Array.isArray(sample[0])) {
-            return sample.flat(); // [height, width] -> flat
-        } else {
-            return sample; // already flat
-        }
-    }
-    
-    _unflatten(flat, shape) {
-        if (shape.type === '3d') {
-            const [channels, height, width] = shape.dims;
-            const result = [];
-            let index = 0;
-            for (let c = 0; c < channels; c++) {
-                const channel = [];
-                for (let h = 0; h < height; h++) {
-                    const row = [];
-                    for (let w = 0; w < width; w++) {
-                        row.push(flat[index++]);
-                    }
-                    channel.push(row);
-                }
-                result.push(channel);
-            }
-            return result;
-        } else if (shape.type === '2d') {
-            const [height, width] = shape.dims;
-            const result = [];
-            for (let h = 0; h < height; h++) {
-                result.push(flat.slice(h * width, h * width + width));
-            }
-            return result;
-        } else {
-            return flat; // 1d
-        }
-    }
-    
-    parameters() { return []; }
+	constructor(){
+		this.originalShape = null;
+	}
+	
+	forward(x){
+		// Save full shape per sample 
+		this.originalShape = x.map(sample => this._getDims(sample));
+		
+		return x.map(sample => this._flattenDeep(sample));
+	}
+	
+	backward(grad){
+		return grad.map((flat, i) =>
+			this._reshape(flat, this.originalShape[i])
+		);
+	}
+	
+	// Get dimensions recursively 
+	_getDims(arr){
+		const dims = [];
+		let current = arr;
+		while (Array.isArray(current)){
+			dims.push(current.length);
+			current = current[0];
+		}
+		return dims;
+	}
+	
+	// Flatten ANY depth 
+	_flattenDeep(arr){
+		return arr.flat(Infinity);
+	}
+	
+	// Reshape back using saved dims 
+	_reshape(flat, dims){
+		let index = 0;
+		
+		function build(dimIdx){
+			const size = dims[dimIdx];
+			const result = [];
+			
+			if (dimIdx === dims.length - 1){
+				for (let i=0; i<size; i++){
+					result.push(flat[index++]);
+				}
+			} else {
+				for (let i=0; i<size; i++){
+					result.push(build(dimIdx + 1));
+				}
+			}
+			
+			return result;
+		}
+		
+		return build(0);
+	}
+	
+	parameters() { return []; }
 }
 
 // ---------------------- Conv2D (BETA) ----------------------
@@ -632,24 +460,25 @@ export class Conv2D {
         );
         this.x = null;
         
-        // Cache Float32Array untuk kernels
         this._WFlat = null;
-        this._cacheKernels();
+        this._updateCache();
     }
     
-    _cacheKernels() {
-        this._WFlat = this.W.map(oc => 
+    _updateCache(){
+        this._WFlat = this.W.map(oc =>
             oc.map(ic => {
                 const rows = ic.length;
                 const cols = ic[0].length;
                 const flat = new Float32Array(rows * cols);
-                for (let i = 0; i < rows; i++) {
+                
+                for (let i=0; i<rows; i++){
                     const offset = i * cols;
                     const row = ic[i];
-                    for (let j = 0; j < cols; j++) {
-                        flat[offset + j] = row[j];
+                    for (let j=0; j<cols; j++){
+                        flat[offset+j] = row[j];
                     }
                 }
+                
                 return flat;
             })
         );
@@ -697,6 +526,8 @@ export class Conv2D {
     }
 
     forward(batch) {
+        this._updateCache();
+        
         this.x = batch;
         const kH = this.kernel;
         const kW = this.kernel;
@@ -727,43 +558,52 @@ export class Conv2D {
 
     backward(grad) {
         const batchSize = this.x.length;
-        const gradW = this.gradW.map(oc => oc.map(ic => zeros(this.kernel, this.kernel)));
-        const gradInput = this.x.map(sample => 
+        const gradInput = this.x.map(sample =>
             sample.map(chan => zeros(chan.length, chan[0].length))
         );
-
-        for (let b = 0; b < batchSize; b++) {
-            for (let oc = 0; oc < this.outC; oc++) {
-                for (let ic = 0; ic < this.inC; ic++) {
+        
+        // Zero existing gradW in place 
+        for (let oc=0; oc<this.outC; oc++){
+            for (let ic=0; ic<this.inC; ic++){
+                for (let i=0; i<this.kernel; i++){
+                    for (let j=0; j<this.kernel; j++){
+                        this.gradW[oc][ic][i][j] = 0;
+                    }
+                }
+            }
+        }
+        
+        for (let b=0; b<batchSize; b++){
+            for (let oc=0; oc<this.outC; oc++){
+                for (let ic=0; ic<this.inC; ic++){
                     const outGrad = grad[b][oc];
                     
-                    // Compute gradW
-                    for (let i = 0; i < this.kernel; i++) {
-                        for (let j = 0; j < this.kernel; j++) {
+                    // Accumulate gradW in place 
+                    for (let i=0; i<this.kernel; i++){
+                        for (let j=0; j<this.kernel; j++){
                             let sum = 0;
-                            for (let y = 0; y < outGrad.length; y++) {
-                                for (let x = 0; x < outGrad[0].length; x++) {
+                            for (let y=0; y<outGrad.length; y++){
+                                for (let x=0; x<outGrad[0].length; x++){
                                     const inY = y * this.stride + i;
                                     const inX = x * this.stride + j;
-                                    if (inY < this.x[b][ic].length && inX < this.x[b][ic][0].length) {
-                                        sum += this.x[b][ic][inY][inX] * outGrad[y][x];
+                                    if (inY<this.x[b][ic].length && inX < this.x[b][ic][0].length){
+                                        sum+=this.x[b][ic][inY][inX] * outGrad[y][x];
                                     }
                                 }
                             }
-                            gradW[oc][ic][i][j] += sum;
+                            this.gradW[oc][ic][i][j] += sum;
                         }
                     }
-
-                    // Compute gradInput
-                    for (let y = 0; y < outGrad.length; y++) {
-                        for (let x = 0; x < outGrad[0].length; x++) {
-                            for (let ki = 0; ki < this.kernel; ki++) {
-                                for (let kj = 0; kj < this.kernel; kj++) {
+                    
+                    // Compute gradInput 
+                    for (let y=0; y<outGrad.length; y++){
+                        for (let x=0; x<outGrad[0].length; x++){
+                            for (let ki=0; ki<this.kernel; ki++){
+                                for (let kj=0; kj<this.kernel; kj++){
                                     const inY = y * this.stride + ki;
                                     const inX = x * this.stride + kj;
-                                    if (inY < gradInput[b][ic].length && inX < gradInput[b][ic][0].length) {
-                                        gradInput[b][ic][inY][inX] += 
-                                            this.W[oc][ic][ki][kj] * outGrad[y][x];
+                                    if (inY<gradInput[b][ic].length && inX < gradInput[b][ic][0].length){
+                                        gradInput[b][ic][inY][inX] += this.W[oc][ic][ki][kj] * outGrad[y][x];
                                     }
                                 }
                             }
@@ -772,8 +612,7 @@ export class Conv2D {
                 }
             }
         }
-
-        this.gradW = gradW;
+        
         return gradInput;
     }
 
@@ -781,7 +620,7 @@ export class Conv2D {
         return this.W.flatMap((w, oc) => 
             w.map((wc, ic) => ({
                 param: wc, 
-                grad: this.gradW[oc][ic]
+                grad: this.gradW[oc][ic]  // Reference stays valid — gradW is mutated in-place now
             }))
         ); 
     }
@@ -872,6 +711,14 @@ export class Sequential {
 		return state;
 	}
 	
+	step(lr){
+		this.layers.forEach(layer => {
+			if(typeof layer.step === "function"){
+				layer.step(lr);
+			}
+		})
+	}
+	
 	/**
 	* Load state dict 
 	*/
@@ -895,43 +742,35 @@ export class Sequential {
 }
 
 // ---------------------- Activations ----------------------
-export class ReLU{ 
-    constructor(){ this.mask = null; this.originalShape = null; } 
-    
-    forward(x){ 
-        this.originalShape = this._getShapeType(x);
-        
-        if (this.originalShape === '3d') {
-            // Handle [batch, 1, features]
-            this.mask = x.map(sample => sample[0].map(v => v > 0));
-            return x.map(sample => [sample[0].map(v => Math.max(0, v))]);
-        } else {
-            // Handle [batch, features]
-            this.mask = x.map(row => row.map(v => v > 0));
-            return x.map(row => row.map(v => Math.max(0, v)));
-        }
-    } 
-    
-    backward(grad){ 
-        if (this.originalShape === '3d') {
-            return grad.map((sample, i) => 
-                [sample[0].map((v, j) => this.mask[i][j] ? v : 0)]
-            );
-        } else {
-            return grad.map((row, i) => 
-                row.map((v, j) => this.mask[i][j] ? v : 0)
-            );
+export class ReLU {
+    constructor() {
+        this.mask = null;
+    }
+
+    forward(x) {
+        this.mask = this._mapRecursive(x, v => v > 0);
+        return this._mapRecursive(x, v => Math.max(0, v));
+    }
+
+    backward(grad) {
+        return this._mapRecursiveWithMask(grad, this.mask, (g, m) => m ? g : 0);
+    }
+
+    // ===== helper =====
+    _mapRecursive(arr, fn) {
+        if (Array.isArray(arr)) {
+            return arr.map(v => this._mapRecursive(v, fn));
         }
+        return fn(arr);
     }
-    
-    _getShapeType(x) {
-        if (Array.isArray(x[0]) && Array.isArray(x[0][0]) && !Array.isArray(x[0][0][0])) {
-            return '3d';
-        } else if (Array.isArray(x[0]) && !Array.isArray(x[0][0])) {
-            return '2d';
-        } else {
-            throw new Error(`Unsupported input shape for ReLU`);
+
+    _mapRecursiveWithMask(arr, mask, fn) {
+        if (Array.isArray(arr)) {
+            return arr.map((v, i) =>
+                this._mapRecursiveWithMask(v, mask[i], fn)
+            );
         }
+        return fn(arr, mask);
     }
 }
 
@@ -1156,7 +995,10 @@ export class Tokenizer {
             .replace(/\s+/g, ' ')                      
             .trim()                                     
             .split(' ')                                
-            .filter(word => word.length > 0 && !/^[.!?;:,]+$/.test(word) || word.length > 1);
+            .filter(word =>
+				word.length > 0 &&
+				(!/^[.!?;:,]+$/.test(word) || word.length > 1)
+			);
     }
     
     /**
@@ -1812,23 +1654,25 @@ export class ReduceLROnPlateau {
 
 // ---------------------- ELU Activation ----------------------
 export class ELU {
-    constructor(alpha=1.0) {
+    constructor(alpha=1.0){
         this.alpha = alpha;
-        this.out = null;
+        this.x = null; // Cache input for correct derivative
     }
-
-    forward(x) {
-        this.out = x.map(row => 
-            row.map(v => v > 0 ? v : this.alpha * (Math.exp(v) - 1))
+    
+    forward(x){
+        this.x=x; // Store original input for backward
+        return x.map(row =>
+            row.map(v=>v>0 ? v : this.alpha*(Math.exp(v) - 1))
         );
-        return this.out;
     }
-
-    backward(grad) {
-        return grad.map((row, i) => 
-            row.map((v, j) => 
-                v * (this.out[i][j] > 0 ? 1 : this.alpha * Math.exp(this.out[i][j]))
-            )
+    
+    backward(grad){
+        return grad.map((row, i) =>
+            row.map((v, j) => {
+                const xVal = this.x[i][j];
+                // d/dx ELU: 1 if x > 0, else alpha * exp(x)
+                return v * (xVal > 0 ? 1 : this.alpha * Math.exp(xVal));
+            })
         );
     }
 }
@@ -1910,10 +1754,10 @@ export class BatchNorm2d {
         
         // Parameters
         if (affine) {
-            this.weight = Array(numFeatures).fill(1);
-            this.bias = Array(numFeatures).fill(0);
-            this.gradWeight = Array(numFeatures).fill(0);
-            this.gradBias = Array(numFeatures).fill(0);
+            this.weight = [Array(numFeatures).fill(1)];
+			this.bias = [Array(numFeatures).fill(0)];
+			this.gradWeight = [Array(numFeatures).fill(0)];
+			this.gradBias = [Array(numFeatures).fill(0)];
         }
         
         // Running statistics
@@ -1993,7 +1837,7 @@ export class BatchNorm2d {
                             
                             // Apply affine transformation if enabled
                             if (this.affine) {
-                                channelOut[i][j] = channelOut[i][j] * this.weight[c] + this.bias[c];
+								channelOut[i][j] = channelOut[i][j] * this.weight[0][c] + this.bias[0][c];
                             }
                         }
                     }
@@ -2022,7 +1866,7 @@ export class BatchNorm2d {
                             
                             // Apply affine transformation if enabled
                             if (this.affine) {
-                                channelOut[i][j] = channelOut[i][j] * this.weight[c] + this.bias[c];
+                                channelOut[i][j] = channelOut[i][j] * this.weight[0][c] + this.bias[0][c];
                             }
                         }
                     }
@@ -2052,8 +1896,10 @@ export class BatchNorm2d {
         );
         
         if (this.affine) {
-            this.gradWeight.fill(0);
-            this.gradBias.fill(0);
+            for (let c=0; c<channels; c++){
+                this.gradWeight[0][c] = 0;
+                this.gradBias[0][c] = 0;
+            }
         }
         
         for (let c = 0; c < channels; c++) {
@@ -2083,7 +1929,7 @@ export class BatchNorm2d {
                         let grad = channelGrad[i][j];
                         
                         if (this.affine) {
-                            grad *= this.weight[c];
+                            grad *= this.weight[0][c];
                         }
                         
                         grad *= stdInv;
@@ -2093,8 +1939,8 @@ export class BatchNorm2d {
             }
             
             if (this.affine) {
-                this.gradWeight[c] = sumGradWeight / batchSize;
-                this.gradBias[c] = sumGradBias / batchSize;
+                this.gradWeight[0][c] = sumGradWeight / batchSize;
+                this.gradBias[0][c] = sumGradBias / batchSize;
             }
         }
         
@@ -2104,9 +1950,9 @@ export class BatchNorm2d {
     parameters() {
         if (!this.affine) return [];
         return [
-            { param: [this.weight], grad: [this.gradWeight] },
-            { param: [this.bias], grad: [this.gradBias] }
-        ];
+			{ param: this.weight, grad: this.gradWeight },
+			{ param: this.bias, grad: this.gradBias }
+		]
     }
 
     train() { this.training = true; }
@@ -2115,20 +1961,143 @@ export class BatchNorm2d {
 
 // ---------------------- Model Save/Load (BETA) ----------------------
 export function saveModel(model){
-    if(!(model instanceof Sequential)) throw new Error("saveModel supports only Sequential");
-    const weights=model.layers.map(layer=>({weights:layer.W||null,biases:layer.b||null}));
-    return JSON.stringify(weights);
-	/* Didn't expect this to work */
+    if(!(model instanceof Sequential)){
+        throw new Error("saveModel supports only Sequential models");
+    }
+    
+    const state = {
+        version: "2.0.0",
+        layers: model.layers.map((layer, idx) => {
+            const params = layer.parameters ? layer.parameters() : [];
+            
+            if (params.length === 0){
+                return { type: layer.constructor.name, params: [] };
+            }
+            
+            return {
+                type: layer.constructor.name,
+                params: params.map(p => ({
+                    // Deep clone parameter data 
+                    data: p.param.map(row =>
+                        Array.isArray(row) ? [...row] : row
+                    ),
+                    // Preserve shape metadata for validation 
+                    shape: Array.isArray(p.param[0])
+                        ? [p.param.length, p.param[0].length]
+                        : [p.param.length]
+                }))
+            };
+        })
+    };
+    
+    return JSON.stringify(state);
 }
 
-export function loadModel(model,json){
-    if(!(model instanceof Sequential)) throw new Error("loadModel supports only Sequential");
-    const weights=JSON.parse(json);
-    model.layers.forEach((layer,i)=>{
-        if(layer.W && weights[i].weights) layer.W=weights[i].weights;
-        if(layer.b && weights[i].biases) layer.b=weights[i].biases;
-    });
-	/* Didn't expect this to work */
+export function loadModel(model, json){
+    if (!(model instanceof Sequential)){
+        throw new Error("loadModel supports only Sequential models");
+    }
+    
+    const state = JSON.parse(json);
+    
+    // Validate structure 
+    if (!state.layers || !Array.isArray(state.layers)){
+        throw new Error("loadModel: invalid save format - missing 'layers' array");
+    }
+    
+    if (state.layers.length !== model.layers.length){
+        console.warn(
+            `[JST WARN]: Layer count mismatch - saved ${state.layers.length},` +
+            `current model has ${model.layers.length}. Loading what matches.`
+        );
+    }
+    
+    let loadedCount = 0;
+    let skippedCount = 0;
+    
+    for(let i=0; i<Math.min(state.layers.length, model.layers.length); i++){
+        const savedLayer = state.layers[i];
+        const currentLayer = model.layers[i];
+        
+        if(savedLayer.params.length === 0){
+            // Layer with no trainable params - skip 
+            continue 
+        }
+        
+        // Validate layer type 
+        if (savedLayer.type !== currentLayer.constructor.name){
+            console.warn(
+                `[JST WARN]: Layer ${i} type mismatch - ` +
+                `saved: ${savedLayer.type}, current: ${currentLayer.constructor.name}. Skipping.`
+            );
+            skippedCount++;
+            continue;
+        }
+        
+        // Get current layer parameters 
+        const currentParams = currentLayer.parameters ? currentLayer.parameters() : [];
+        
+        if (currentParams.length !== savedLayer.params.length){
+            console.warn(
+                `[JST WARN]: Layer ${i} parameter count mismatch - ` +
+                `saved: ${savedLayer.params.length}, current: ${currentParams.length}. Skipping.`
+            );
+            skippedCount++;
+            continue;
+        }
+        
+        // Load parameters wiht shape validation 
+        for (let j=0; j<savedLayer.params.length; j++){
+            const savedParam = savedLayer.params[j];
+            const currentParam = currentParams[j].param;
+            
+            // Validate shape 
+            const currentRows = currentParam.length;
+            const currentCols = Array.isArray(currentParam[0])
+                ? currentParam[0].length 
+                : 1;
+            
+            const savedRows = savedParam.shape[0];
+            const savedCols = savedParam.shape[1] || 1;
+            
+            if (currentRows !== savedRows || currentCols !== savedCols){
+                console.warn(
+                    `[JST WARN]: Layer ${i} param ${j} shape mismatch - ` +
+                    `saved: [${savedRows}, ${savedCols}],` +
+                    `current: [${currentRows}, ${currentCols}]. Skipping this parameter.`
+                );
+                continue
+            }
+            
+            // Copy parameter data 
+            if (Array.isArray(currentParam[0])){
+                // 2D Parameter 
+                for (let r=0; r<currentRows; r++){
+                    for (let c=0; c<currentCols; c++){
+                        currentParam[r][c] = savedParam.data[r][c];
+                    }
+                }
+            } else {
+                // 1D parameter 
+                for (let r=0; r<currentRows; r++){
+                    currentParam[r] = savedParam.data[r];
+                }
+            }
+        }
+        
+        // Invalidate any cached flat representations 
+        if (typeof currentLayer._updateCache === 'function'){
+            currentLayer._updateCache();
+        }
+        
+        loadedCount++;
+    }
+    
+    console.log(
+        `[JST]: Model loaded: ${loadedCount} layers restored, ${skippedCount} skipped.`
+    );
+    
+    return model;
 }
 
 // ---------------------- Advanced Utils ----------------------