catniff 0.3.1 → 0.4.1

package/README.md

@@ -72,6 +72,28 @@ G.backward();
 console.log(X.grad.val(), Y.grad.val());
 ```
 
+## Optimizer
+
+Catniff comes bundled with optimizers as well:
+```js
+const { Tensor, Optim } = require("catniff");
+
+// Define some parameter
+const w = new Tensor([1.0], { requiresGrad: true });
+// Define a fake loss function: L = (w - 3)^2
+const loss = w.sub(3).pow(2);
+// Calculate gradient
+loss.backward();
+// Use Adam optimizer
+const optim = new Optim.Adam([w]);
+// Optimization step
+optim.step();
+
+console.log("Updated weight:", w.data);  // Should move toward 3.0
+```
+
+And it can still do much more, check out the docs mentioned below for more information.
+
 ## Documentation
 
 Full documentation is available in [`./docs/documentation.md`](./docs/documentation.md).
@@ -83,7 +105,6 @@ All available APIs are in [`./src/`](./src/) if you want to dig deeper.
 * Bug fixes.
 * More tensor ops.
 * GPU acceleration.
-* Option to load more backends.
 * Some general neural net APIs.
 * More detailed documentation.
 * Code refactoring.

package/dist/backend.d.ts

@@ -0,0 +1,4 @@
+import { Tensor } from "./core";
+export interface Backend {
+    transfer(tensor: Tensor): Tensor;
+}

package/dist/backend.js

@@ -0,0 +1,2 @@
+"use strict";
+Object.defineProperty(exports, "__esModule", { value: true });

package/dist/core.d.ts

@@ -1,3 +1,4 @@
+import { Backend } from "./backend";
 export type TensorValue = number | TensorValue[];
 export interface TensorOptions {
     shape?: readonly number[];
@@ -6,6 +7,7 @@ export interface TensorOptions {
     requiresGrad?: boolean;
     gradFn?: Function;
     children?: Tensor[];
+    device?: string;
 }
 export declare class Tensor {
     value: number[] | number;
@@ -15,6 +17,7 @@ export declare class Tensor {
     requiresGrad: boolean;
     gradFn: Function;
     children: Tensor[];
+    device: string;
     constructor(value: TensorValue, options?: TensorOptions);
     static flatten(tensor: TensorValue): number[] | number;
     static getShape(tensor: TensorValue): readonly number[];
@@ -161,4 +164,6 @@ export declare class Tensor {
     detach(): Tensor;
     clone(): Tensor;
     replace(other: Tensor, allowShapeMismatch?: boolean): Tensor;
+    static backends: Map<string, Backend>;
+    to(device: string): Tensor;
 }

package/dist/core.js

@@ -10,6 +10,7 @@ class Tensor {
     requiresGrad;
     gradFn;
     children;
+    device;
     constructor(value, options = {}) {
         this.value = Tensor.flatten(value);
         this.shape = options.shape || Tensor.getShape(value);
@@ -18,6 +19,7 @@ class Tensor {
         this.requiresGrad = options.requiresGrad ?? false;
         this.gradFn = options.gradFn || (() => { });
         this.children = options.children || [];
+        this.device = options.device || "cpu";
     }
     // Utility to flatten an nD array to be 1D
     static flatten(tensor) {
@@ -52,6 +54,8 @@ class Tensor {
     }
     // Utility to get strides from shape
     static getStrides(shape) {
+        if (shape.length === 0)
+            return [];
         const strides = new Array(shape.length);
         strides[strides.length - 1] = 1;
         for (let i = strides.length - 2; i >= 0; i--) {
@@ -282,7 +286,8 @@ class Tensor {
         const outValue = outShape.length === 0 ? this.value[0] : this.value;
         const out = new Tensor(outValue, {
             shape: outShape,
-            strides: outStrides
+            strides: outStrides,
+            device: this.device
         });
         // Set up gradient if needed
         if (this.requiresGrad) {
@@ -319,7 +324,7 @@ class Tensor {
             newDimStride = this.strides[dim] * this.shape[dim];
         }
         newStrides.splice(dim, 0, newDimStride);
-        const out = new Tensor(thisValue, { shape: newShape, strides: newStrides });
+        const out = new Tensor(thisValue, { shape: newShape, strides: newStrides, device: this.device });
         // Set up gradient if needed
         if (this.requiresGrad) {
             out.requiresGrad = true;
@@ -981,7 +986,7 @@ class Tensor {
         [newShape[dim1], newShape[dim2]] = [newShape[dim2], newShape[dim1]];
         [newStrides[dim1], newStrides[dim2]] = [newStrides[dim2], newStrides[dim1]];
         // Create new tensor with same data but swapped shape/strides
-        const out = new Tensor(this.value, { shape: newShape, strides: newStrides });
+        const out = new Tensor(this.value, { shape: newShape, strides: newStrides, device: this.device });
         out.requiresGrad = this.requiresGrad;
         // Handle gradient if needed
         if (this.requiresGrad) {
@@ -1160,32 +1165,7 @@ class Tensor {
         if (this.shape.length !== 2 || other.shape.length !== 1) {
             throw new Error("Input is not a 2D and 1D tensor pair");
         }
-        // MM with no grad
-        const thisMat = new Tensor(this.value, { shape: this.shape, strides: this.strides });
-        const otherMat = new Tensor(other.value, { shape: [other.shape[0], 1], strides: [other.strides[0], 1] });
-        const out = thisMat.mm(otherMat).squeeze(1);
-        // Handle grad with original tensors
-        if (this.requiresGrad) {
-            out.requiresGrad = true;
-            out.children.push(this);
-        }
-        if (other.requiresGrad) {
-            out.requiresGrad = true;
-            out.children.push(other);
-        }
-        if (out.requiresGrad) {
-            out.gradFn = () => {
-                // Disable gradient collecting of gradients themselves
-                const outGrad = out.grad.withGrad(false);
-                const selfNoGrad = this.withGrad(false);
-                const otherNoGrad = other.withGrad(false);
-                if (this.requiresGrad)
-                    Tensor.addGrad(this, outGrad.unsqueeze(1).mm(otherNoGrad.unsqueeze(0)));
-                if (other.requiresGrad)
-                    Tensor.addGrad(other, selfNoGrad.t().mv(outGrad));
-            };
-        }
-        return out;
+        return this.mm(other.unsqueeze(1)).squeeze(1);
     }
     // General matrix multiplication with different shapes
     matmul(other) {
@@ -1461,6 +1441,7 @@ class Tensor {
         return new Tensor(this.value, {
             shape: this.shape,
             strides: this.strides,
+            device: this.device,
             requiresGrad
         });
     }
@@ -1469,6 +1450,7 @@ class Tensor {
         return new Tensor(this.value, {
             shape: this.shape,
             strides: this.strides,
+            device: this.device,
             requiresGrad: false
         });
     }
@@ -1493,5 +1475,15 @@ class Tensor {
         this.value = other.value;
         return this;
     }
+    // Holds all available backends
+    static backends = new Map();
+    // Op to transfer tensor to another device
+    to(device) {
+        const backend = Tensor.backends.get(device);
+        if (backend && backend.transfer) {
+            return backend.transfer(this);
+        }
+        throw new Error(`No device found to transfer tensor to or a handler is not implemented for device.`);
+    }
 }
 exports.Tensor = Tensor;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "catniff",
-  "version": "0.3.1",
+  "version": "0.4.1",
   "description": "A small Torch-like deep learning framework for Javascript with tensor and autograd support",
   "main": "index.js",
   "scripts": {