catniff 0.8.3 → 0.8.4

package/dist/core.d.ts

@@ -201,6 +201,7 @@ export declare class Tensor {
     triu(diagonal?: number): Tensor;
     tril(diagonal?: number): Tensor;
     maskedFill(mask: Tensor | TensorValue, value: number): Tensor;
+    multinomial(numSamples: number, replacement?: boolean): Tensor;
     static full(shape: number[], num: number, options?: TensorOptions): Tensor;
     static fullLike(tensor: Tensor, num: number, options?: TensorOptions): Tensor;
     static ones(shape?: number[], options?: TensorOptions): Tensor;

package/dist/core.js

@@ -949,14 +949,14 @@ class Tensor {
         // Copy if not contiguous
         const outputSize = this.numel;
         const outputShape = this.shape;
-        let outputValue, outputStrides;
+        const outputValue = new dtype_1.TypedArray[this.dtype](outputSize);
+        const outputStrides = Tensor.getStrides(outputShape);
         if (this.isContiguous()) {
-            outputValue = [...this.value];
-            outputStrides = this.strides;
+            // Fast path: direct copy
+            outputValue.set(this.value.subarray(this.offset, this.offset + outputSize));
         }
         else {
-            outputValue = new dtype_1.TypedArray[this.dtype](outputSize);
-            outputStrides = Tensor.getStrides(outputShape);
+            // Slow path: coordinate conversion
             for (let flatIndex = 0; flatIndex < outputSize; flatIndex++) {
                 const coords = Tensor.indexToCoords(flatIndex, outputStrides);
                 const originalIndex = Tensor.coordsToIndex(coords, this.strides);
@@ -1912,6 +1912,129 @@ class Tensor {
         mask = this.handleOther(mask);
         return this.mul(mask.logicalNot()).add(mask.mul(value));
     }
+    // Multinomial sampling
+    multinomial(numSamples, replacement = false) {
+        // Validate input dimensions (1D or 2D only)
+        if (this.shape.length === 0 || this.shape.length > 2) {
+            throw new Error("multinomial only supports 1D or 2D probability tensors");
+        }
+        const is1D = this.shape.length === 1;
+        const numDist = is1D ? 1 : this.shape[0];
+        const numCategories = is1D ? this.shape[0] : this.shape[1];
+        // Validate numSamples
+        if (numSamples <= 0) {
+            throw new Error("Number of samples must be positive");
+        }
+        if (!replacement && numSamples > numCategories) {
+            throw new Error(`Cannot sample ${numSamples} without replacement from ${numCategories} categories`);
+        }
+        // Make contiguous copy of probabilities
+        const probsSize = this.numel;
+        const probs = new dtype_1.TypedArray[this.dtype](probsSize);
+        if (this.isContiguous()) {
+            // Fast path: direct copy
+            probs.set(this.value.subarray(this.offset, this.offset + probsSize));
+        }
+        else {
+            // Slow path: coordinate conversion
+            const defaultStrides = Tensor.getStrides(this.shape);
+            for (let i = 0; i < probsSize; i++) {
+                const coords = Tensor.indexToCoords(i, defaultStrides);
+                const idx = Tensor.coordsToIndex(coords, this.strides);
+                probs[i] = this.value[idx + this.offset];
+            }
+        }
+        // Output setup
+        const outputShape = is1D ? [numSamples] : [numDist, numSamples];
+        const outputValue = new Int32Array(numDist * numSamples);
+        // Sample from each distribution
+        for (let dist = 0; dist < numDist; dist++) {
+            const offset = dist * numCategories;
+            // Extract this distribution's probabilities
+            const distProbs = probs.slice(offset, offset + numCategories);
+            // Validate and normalize
+            let sum = 0;
+            for (let i = 0; i < numCategories; i++) {
+                if (distProbs[i] < 0) {
+                    throw new Error("Probabilities cannot be negative");
+                }
+                sum += distProbs[i];
+            }
+            if (sum <= 0) {
+                throw new Error("Probabilities must sum to a positive value");
+            }
+            // Normalize
+            for (let i = 0; i < numCategories; i++) {
+                distProbs[i] /= sum;
+            }
+            if (replacement) {
+                // With replacement: use CDF for efficient sampling
+                const cdf = new Array(numCategories);
+                let cumSum = 0;
+                for (let i = 0; i < numCategories; i++) {
+                    cumSum += distProbs[i];
+                    cdf[i] = cumSum;
+                }
+                cdf[numCategories - 1] = 1;
+                for (let s = 0; s < numSamples; s++) {
+                    const r = Math.random();
+                    // Binary search for efficiency
+                    let left = 0;
+                    let right = numCategories - 1;
+                    while (left < right) {
+                        const mid = Math.floor((left + right) / 2);
+                        if (r <= cdf[mid]) {
+                            right = mid;
+                        }
+                        else {
+                            left = mid + 1;
+                        }
+                    }
+                    outputValue[dist * numSamples + s] = left;
+                }
+            }
+            else {
+                // Without replacement: weighted sampling without replacement
+                const available = Array.from({ length: numCategories }, (_, i) => ({
+                    idx: i,
+                    prob: distProbs[i]
+                }));
+                for (let s = 0; s < numSamples; s++) {
+                    // Compute sum of remaining probabilities
+                    let remainingSum = 0;
+                    for (const item of available) {
+                        remainingSum += item.prob;
+                    }
+                    // Sample from remaining
+                    const r = Math.random() * remainingSum;
+                    let cumSum = 0;
+                    let selectedIdx = -1;
+                    for (let i = 0; i < available.length; i++) {
+                        cumSum += available[i].prob;
+                        if (r <= cumSum) {
+                            selectedIdx = i;
+                            break;
+                        }
+                    }
+                    // Handle floating point edge case
+                    if (selectedIdx === -1) {
+                        selectedIdx = available.length - 1;
+                    }
+                    // Store result and remove from available
+                    outputValue[dist * numSamples + s] = available[selectedIdx].idx;
+                    available.splice(selectedIdx, 1);
+                }
+            }
+        }
+        return new Tensor(outputValue, {
+            shape: outputShape,
+            strides: Tensor.getStrides(outputShape),
+            offset: 0,
+            numel: numDist * numSamples,
+            device: this.device,
+            dtype: "int32"
+        });
+    }
     // Utility to create a new tensor filled with a number
     static full(shape, num, options = {}) {
         if (shape.length === 0)
@@ -2094,6 +2217,7 @@ class Tensor {
             shape,
             offset: 0,
             numel: outputSize,
+            dtype: "int32",
             ...options
         });
     }
@@ -2130,6 +2254,7 @@ class Tensor {
             shape: [n],
             offset: 0,
             numel: n,
+            dtype: "int32",
             ...options
         });
     }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "catniff",
-  "version": "0.8.3",
+  "version": "0.8.4",
   "description": "Torch-like deep learning framework for Javascript",
   "main": "index.js",
   "scripts": {