catniff 0.5.7 → 0.5.9

package/dist/backend.d.ts

@@ -1,4 +1,5 @@
 import { Tensor } from "./core";
 export interface Backend {
+    create(tensor: Tensor): void;
     transfer(tensor: Tensor): Tensor;
 }

package/dist/core.d.ts

@@ -177,4 +177,5 @@ export declare class Tensor {
     replace(other: Tensor, allowShapeMismatch?: boolean): Tensor;
     static backends: Map<string, Backend>;
     to(device: string): Tensor;
+    to_(device: string): Tensor;
 }

package/dist/core.js

@@ -21,13 +21,8 @@ class Tensor {
         this.gradFn = options.gradFn || (() => { });
         this.children = options.children || [];
         this.device = options.device || "cpu";
-        // Move tensor to device
-        if (this.device !== "cpu") {
-            const backend = Tensor.backends.get(this.device);
-            if (backend && backend.transfer) {
-                backend.transfer(this);
-            }
-        }
+        // Move to device in-place
+        this.to_(this.device);
     }
     // Utility to flatten an nD array to be 1D
     static flatten(tensor) {
@@ -1712,9 +1707,21 @@ class Tensor {
     static backends = new Map();
     // Op to transfer tensor to another device
     to(device) {
+        if (device === "cpu")
+            return this;
         const backend = Tensor.backends.get(device);
         if (backend && backend.transfer) {
-            backend.transfer(this);
+            return backend.transfer(this);
+        }
+        throw new Error(`No device found to transfer tensor to or a handler is not implemented for device.`);
+    }
+    // Op to transfer tensor to another device in-place
+    to_(device) {
+        if (device === "cpu")
+            return this;
+        const backend = Tensor.backends.get(this.device);
+        if (backend && backend.create) {
+            backend.create(this);
             return this;
         }
         throw new Error(`No device found to transfer tensor to or a handler is not implemented for device.`);

package/dist/nn.d.ts

@@ -57,7 +57,7 @@ declare class LayerNorm {
     constructor(normalizedShape: number | number[], eps?: number, elementwiseAffine?: boolean, bias?: boolean, device?: string);
     forward(input: Tensor | TensorValue): Tensor;
 }
-interface StateDict {
+export interface StateDict {
     [key: string]: any;
 }
 export declare const nn: {

package/dist/optim.d.ts

@@ -1,4 +1,9 @@
 import { Tensor } from "./core";
+declare abstract class BaseOptimizer {
+    params: Tensor[];
+    constructor(params: Tensor[]);
+    zeroGrad(): void;
+}
 export interface SGDOptions {
     lr?: number;
     momentum?: number;
@@ -6,8 +11,7 @@ export interface SGDOptions {
     weightDecay?: number;
     nesterov?: boolean;
 }
-declare class SGD {
-    params: Tensor[];
+declare class SGD extends BaseOptimizer {
     momentumBuffers: Map<Tensor, Tensor>;
     lr: number;
     momentum: number;
@@ -23,8 +27,7 @@ export interface AdamOptions {
     eps?: number;
     weightDecay?: number;
 }
-declare class Adam {
-    params: Tensor[];
+declare class Adam extends BaseOptimizer {
     momentumBuffers: Map<Tensor, Tensor>;
     velocityBuffers: Map<Tensor, Tensor>;
     stepCount: number;
@@ -36,6 +39,7 @@ declare class Adam {
     step(): void;
 }
 export declare class Optim {
+    static BaseOptimizer: typeof BaseOptimizer;
     static SGD: typeof SGD;
     static Adam: typeof Adam;
 }

package/dist/optim.js

@@ -2,8 +2,19 @@
 Object.defineProperty(exports, "__esModule", { value: true });
 exports.Optim = void 0;
 const core_1 = require("./core");
-class SGD {
+class BaseOptimizer {
     params;
+    constructor(params) {
+        this.params = params;
+    }
+    zeroGrad() {
+        for (let index = 0; index < this.params.length; index++) {
+            const param = this.params[index];
+            param.grad = core_1.Tensor.zerosLike(param);
+        }
+    }
+}
+class SGD extends BaseOptimizer {
     momentumBuffers = new Map();
     lr;
     momentum;
@@ -11,7 +22,7 @@ class SGD {
     weightDecay;
     nesterov;
     constructor(params, options) {
-        this.params = params;
+        super(params);
         this.lr = options?.lr || 0.001;
         this.momentum = options?.momentum || 0;
         this.dampening = options?.dampening || 0;
@@ -55,8 +66,7 @@ class SGD {
         }
     }
 }
-class Adam {
-    params;
+class Adam extends BaseOptimizer {
     momentumBuffers = new Map(); // First moment (m_t)
     velocityBuffers = new Map(); // Second moment (v_t)
     stepCount = 0;
@@ -65,7 +75,7 @@ class Adam {
     eps;
     weightDecay;
     constructor(params, options) {
-        this.params = params;
+        super(params);
         this.lr = options?.lr || 0.001;
         this.betas = options?.betas || [0.9, 0.999];
         this.eps = options?.eps || 1e-8;
@@ -117,6 +127,7 @@ class Adam {
     }
 }
 class Optim {
+    static BaseOptimizer = BaseOptimizer;
     static SGD = SGD;
     static Adam = Adam;
 }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "catniff",
-  "version": "0.5.7",
+  "version": "0.5.9",
   "description": "A small Torch-like deep learning framework for Javascript",
   "main": "index.js",
   "scripts": {