@huggingface/transformers 3.0.0-alpha.9 → 3.0.0

package/src/ops/registry.js

@@ -1,20 +1,31 @@
 import { createInferenceSession } from "../backends/onnx.js";
 import { Tensor } from "../utils/tensor.js";
 
+/**
+ * Asynchronously creates a wrapper function for running an ONNX inference session.
+ *
+ * @param {number[]} session_bytes The session data in bytes.
+ * @param {import('onnxruntime-common').InferenceSession.SessionOptions} session_options The options for the ONNX session.
+ * @template {string | [string] | string[]} T
+ * @param {T} names The name(s) of the output tensor(s).
+ * 
+ * @returns {Promise<function(Record<string, Tensor>): Promise<T extends string ? Tensor : T extends string[] ? { [K in keyof T]: Tensor } : never>>}
+ * The wrapper function for running the ONNX inference session.
+ */
 const wrap = async (session_bytes, session_options, names) => {
     const session = await createInferenceSession(
         new Uint8Array(session_bytes), session_options,
     );
-    return async (inputs) => {
+    return /** @type {any} */(async (/** @type {Record<string, Tensor>} */ inputs) => {
         const ortFeed = Object.fromEntries(Object.entries(inputs).map(([k, v]) => [k, v.ort_tensor]));
         const outputs = await session.run(ortFeed);
 
         if (Array.isArray(names)) {
             return names.map((n) => new Tensor(outputs[n]));
         } else {
-            return new Tensor(outputs[names]);
+            return new Tensor(outputs[/** @type {string} */(names)]);
         }
-    }
+    })
 }
 
 // In-memory registry of initialized ONNX operators

package/src/pipelines.js

@@ -34,6 +34,7 @@ import {
     AutoModelForImageClassification,
     AutoModelForImageSegmentation,
     AutoModelForSemanticSegmentation,
+    AutoModelForUniversalSegmentation,
     AutoModelForObjectDetection,
     AutoModelForZeroShotObjectDetection,
     AutoModelForDocumentQuestionAnswering,
@@ -3045,7 +3046,7 @@ const SUPPORTED_TASKS = Object.freeze({
     "image-segmentation": {
         // no tokenizer
         "pipeline": ImageSegmentationPipeline,
-        "model": [AutoModelForImageSegmentation, AutoModelForSemanticSegmentation],
+        "model": [AutoModelForImageSegmentation, AutoModelForSemanticSegmentation, AutoModelForUniversalSegmentation],
         "processor": AutoProcessor,
         "default": {
             // TODO: replace with original
@@ -3287,7 +3288,7 @@ async function loadItems(mapping, model, pretrainedOptions) {
 
     /**@type {Promise[]} */
     const promises = [];
-    for (let [name, cls] of mapping.entries()) {
+    for (const [name, cls] of mapping.entries()) {
         if (!cls) continue;
 
         /**@type {Promise} */
@@ -3295,7 +3296,7 @@ async function loadItems(mapping, model, pretrainedOptions) {
         if (Array.isArray(cls)) {
             promise = new Promise(async (resolve, reject) => {
                 let e;
-                for (let c of cls) {
+                for (const c of cls) {
                     if (c === null) {
                         // If null, we resolve it immediately, meaning the relevant
                         // class was not found, but it is optional.
@@ -3333,7 +3334,7 @@ async function loadItems(mapping, model, pretrainedOptions) {
     await Promise.all(promises);
 
     // Then assign to result
-    for (let [name, promise] of Object.entries(result)) {
+    for (const [name, promise] of Object.entries(result)) {
         result[name] = await promise;
     }