@kubohiroya/turbowarp-teachable-machine 4.0.1 → 4.0.2

package/README.ja.md

@@ -46,35 +46,51 @@ flowchart LR
 
 - 公開済みのTeachable Machineのポーズ・画像・音声モデルのURL（エディターのURLではなく、モデルフォルダーのURL）
 - カメラまたはマイクと、ブラウザーでのその使用許可
-- Teachable Machineのブラウザー実行環境とモデルファイルを取得するためのネットワーク接続
-- TurboWarpの**サンドボックスなしで実行**オプション。これはファイルから読み込んだ場合にのみ選べます（[導入方法](#導入方法)を参照）
+- モデルファイルを取得するためのネットワーク接続（軽量版ではライブラリーの取得にも使用）
+- TurboWarpの**サンドボックスなしで実行**オプション。これは**テキスト**または**ファイル**から読み込んだ場合に選べます（[導入方法](#導入方法)を参照）
 
 > [!IMPORTANT]
 > この機能拡張は、カメラ・マイク・ステージにアクセスするためサンドボックスなしで動作します。信頼できる提供元のコードだけを読み込んでください。またカメラAPIには、HTTPSや`localhost`などの安全なブラウザーコンテキストが必要です。
 
 ## 導入方法
 
-[`dist/tm.js`](dist/tm.js)をダウンロードし、TurboWarpのカスタム機能拡張ダイアログで**ファイル**タブを開いてそのファイルを選び、**サンドボックスなしで実行**を有効のままにしてください。
+標準のTurboWarp Editorで、次のどれか1つを読み込みます。
+
+| ファイル | 初期計算モード | 依存ライブラリーとWASMバイナリ |
+|---|---|---|
+| [`dist/teachable-machine-webgpu.js`](dist/teachable-machine-webgpu.js) | `webgpu` | すべて内包 |
+| [`dist/teachable-machine-wasm.js`](dist/teachable-machine-wasm.js) | `wasm` | すべて内包 |
+| [`dist/teachable-machine.js`](dist/teachable-machine.js) | `auto` | 必要時にネット取得する軽量版 |
+
+**テキスト**タブではファイルの全文を貼り付け、**サンドボックスなしで実行**を有効にします。
+**ファイル**タブから同じJavaScriptファイルを選んでも利用できます。
+
+内包版はTensorFlow.jsを1系統だけ含み、Teachable Machine Pose/Image、Speech Commands、
+WebGPU/WebGL/WASM/CPUの全バックエンド、3種類のWASMバイナリ、並列ワーカーを同梱します。
+初期方式以外への切り替えやフォールバックでも、ライブラリーやWASMバイナリの外部取得は発生しません。
+モデルファイルとPoseNetの重みは別途取得するため、完全オフラインのモデルパッケージではありません。
+実際の方式は`active compute backend`で確認できます。
 
 > [!IMPORTANT]
-> ダイアログの**URL**タブからは読み込めません。信頼されていないURLから読み込まれた機能拡張をTurboWarpは常にサンドボックス内で実行し、URLタブには**サンドボックスなしで実行**の選択肢自体がありません。この機能拡張はサンドボックス内では起動を拒否します。下のjsDelivrのリンクをURLタブに貼ると、ダイアログは何も表示しないまま待ち続け、ブラウザーのコンソールに`TM must run without the extension sandbox.`とだけ出ます。ファイルをダウンロードして**ファイル**タブから読み込んでください。
+> URLタブは信頼されていない拡張をサンドボックス化します。**テキスト**または**ファイル**で、
+> **サンドボックスなしで実行**を有効にしてください。
 
-ダウンロード先、および合成ランタイムが取得する成果物は、jsDelivr上のバージョン固定ビルドです。
+軽量版のバージョン固定CDN URLは、対応する成果物の公開後に利用できます。
 
 ```text
-https://cdn.jsdelivr.net/npm/@kubohiroya/turbowarp-teachable-machine@4.0.1/dist/tm.js
+https://cdn.jsdelivr.net/npm/@kubohiroya/turbowarp-teachable-machine@4.0.2/dist/teachable-machine.js
 ```
 
 合成ランタイムは、縮小済みの第三者バンドルを書き換えることなく、同じ検証済み成果物を読み込んだり同梱したりできます。
 
 ```text
-https://cdn.jsdelivr.net/npm/@kubohiroya/turbowarp-teachable-machine@4.0.1/dist/runtime.js
+https://cdn.jsdelivr.net/npm/@kubohiroya/turbowarp-teachable-machine@4.0.2/dist/runtime.js
 ```
 
 公開済みパッケージを別のプロジェクトに追加するには次のようにします。
 
 ```sh
-pnpm add --save-exact @kubohiroya/turbowarp-teachable-machine@4.0.1
+pnpm add --save-exact @kubohiroya/turbowarp-teachable-machine@4.0.2
 ```
 
 ### 名前について
@@ -417,17 +433,24 @@ pnpm install --frozen-lockfile
 pnpm check
 ```
 
-`pnpm check`は、型検査・テスト・本番ビルド・生成ドキュメントの検証・Pagesのリンク検証・配布物の再現性・npmパッケージのドライランをまとめて実行します。ビルドは`dist/tm.js`・`dist/composition.js`・`dist/runtime.js`・`dist/posenet.js`と、`dist/posenet/`以下の3つの生のPoseNetモデル資産を生成します。
+`pnpm check`は、型検査・テスト・本番ビルド・生成ドキュメントの検証・Pagesのリンク検証・配布物の再現性・npmパッケージのドライランをまとめて実行します。ビルドは`dist/teachable-machine.js`・`dist/teachable-machine-webgpu.js`・`dist/teachable-machine-wasm.js`・`dist/composition.js`・`dist/runtime.js`・`dist/posenet.js`と、`dist/posenet/`以下の3つの生のPoseNetモデル資産を生成します。
 
 [英語版README](README.md#blocks)のブロック一覧は`src/block-definitions.json`から`pnpm docs`で生成されます。生成された節ではなく定義のほうを編集してください。リリースの基準は`package.json`のバージョンです。実行時のレポーターはそこに`-typescript`を付けた値を返し、リリース整合性チェックが、生成されたバンドル・READMEのバージョン固定の例・Pagesのバッジ・リリースタグをその値と揃えます。
 
+### 標準Editorでの検証
+
+WebGPU版・WASM版をテキストとファイルから読み込む4通りの認識を確認しました。
+[検証条件・再実行手順](docs/standalone-validation.ja.md)を参照してください。
+
 ### npmへの公開
 
 npmへの公開には、長期間有効な書き込みトークンではなくGitHub Actionsのtrusted publishingを使います。npmのパッケージ設定で、`kubohiroya/turbowarp-teachable-machine`のワークフローファイル`publish-npm.yml`を`npm publish`について信頼する設定にしてください。注釈付きのリリースタグとGitHub Releaseを作ったうえで、そのタグを指定して**Publish npm package**を起動します。ワークフローはタグをチェックアウトし、`package.json`と完全に一致することを確認し、すべてのチェックを実行してから、短命のOIDC資格情報で公開します。自動化を戻したい場合は、npmのパッケージ設定からtrusted publisherを外してください。リポジトリーに書き込みトークンを追加しないでください。
 
 ## 外部ライブラリー
 
-バージョン固定の`dist/runtime.js`は、TensorFlow.js 4.22.0、Teachable Machine Pose 0.8.6、Teachable Machine Image 0.8.5、TensorFlow.js Speech Commands 0.5.4、PoseNet 2.2.2を、1つの検証済みモジュールグラフとして同梱しています。ランタイムが注入されない場合、標準の機能拡張はこの単一の成果物をjsDelivrから読み込みます。任意の`dist/backend-wasm.js`と`dist/backend-webgpu.js`は、その実行環境がすでに公開したインスタンスに対してTensorFlow.jsを解決するので、1つのページでTensorFlow.jsのコアが二重になることはありません。`dist/wasm/`以下のWebAssemblyバイナリーと、`dist/posenet/`以下の固定オフラインPoseNetモデルデータは、元のバイナリーファイルのまま配布されます。
+バージョン固定の`dist/runtime.js`は、TensorFlow.js 4.22.0、Teachable Machine Pose 0.8.6、Teachable Machine Image 0.8.5、TensorFlow.js Speech Commands 0.5.4、PoseNet 2.2.2を、1つの検証済みモジュールグラフとして同梱しています。ランタイムが注入されない場合、軽量版はこの単一の成果物をjsDelivrから読み込みます。任意の`dist/backend-wasm.js`と`dist/backend-webgpu.js`は、その実行環境がすでに公開したインスタンスに対してTensorFlow.jsを解決するので、1つのページでTensorFlow.jsのコアが二重になることはありません。`dist/wasm/`以下のWebAssemblyバイナリーと、`dist/posenet/`以下の固定オフラインPoseNetモデルデータは、元のバイナリーファイルのまま配布されます。
+
+内包版は上記のライブラリー・バックエンド・WASMバイナリと、第三者ライセンス全文をJavaScript内に含みます。
 
 第三者ライセンスは[THIRD_PARTY_NOTICES.md](THIRD_PARTY_NOTICES.md)にまとめてあります。
 

package/README.md

@@ -43,12 +43,11 @@ runtime libraries and the published model are fetched.
 
 ## Requirements
 
-- A published Teachable Machine pose, image, or audio model URL — the model folder URL, not the
-  editor URL.
-- A camera or microphone, and browser permission to use it.
-- Network access for the reviewed Teachable Machine browser runtime and the model files.
-- TurboWarp's **Run without sandbox** option, which only exists when the extension is loaded from a
-  file. See [Installation](#installation).
+- a published Teachable Machine Pose, Image, or Audio model URL;
+- a camera or microphone and permission to use it in the browser;
+- network access for model files (and, for the lightweight build, runtime libraries);
+- TurboWarp's **Run without sandbox** option, which is only available when the extension is
+  loaded from **Text** or **Files** (see [Installation](#installation)).
 
 > [!IMPORTANT]
 > This extension is unsandboxed because it needs camera, microphone, and stage access. Load
@@ -57,33 +56,61 @@ runtime libraries and the published model are fetched.
 
 ## Installation
 
-Download [`dist/tm.js`](dist/tm.js). In TurboWarp's custom extension dialog, open the **Files** tab,
-select that file, and keep **Run without sandbox** enabled.
+Choose **one** of these files in a standard TurboWarp Editor:
+
+| File | Initial compute mode | Bundled dependencies |
+|---|---|---|
+| [`dist/teachable-machine-webgpu.js`](dist/teachable-machine-webgpu.js) | `webgpu` | All runtime libraries, all backends, and all WASM binaries |
+| [`dist/teachable-machine-wasm.js`](dist/teachable-machine-wasm.js) | `wasm` | All runtime libraries, all backends, and all WASM binaries |
+| [`dist/teachable-machine.js`](dist/teachable-machine.js) | `auto` | Lightweight loader; downloads runtime libraries and optional backends |
+
+For **Text**, copy the entire file into the custom extension dialog's **Text** tab,
+enable **Run without sandbox**, and load it. Alternatively, select the JavaScript file
+in the **Files** tab and enable the same option. A custom TurboWarp build is not required.
+
+The WebGPU and WASM files each contain one TensorFlow.js module graph, Teachable Machine
+Pose/Image, Speech Commands, WebGPU/WebGL/WASM/CPU kernels, all three WASM binaries as data
+URLs, and the WASM worker. They do not fetch JavaScript libraries, workers, or WASM binaries
+from a CDN. Both include every backend so fallback and later mode changes also work without
+fetching library assets. They differ only in the initial compute mode; load one variant per editor.
+
+These are not complete offline model bundles: your Teachable Machine model files are still
+loaded from the URL you specify, and pose recognition also loads the PoseNet model weights.
+Camera/microphone permissions and HTTPS or localhost are still required.
+
+Set `set compute mode to [webgpu]` or `[wasm]` before loading a model to change the requested
+backend. If it is unavailable or fails the convolution check, the extension falls back;
+`active compute backend` reports the actual backend, and `last error` explains a fallback.
 
 > [!IMPORTANT]
-> The dialog's **URL** tab cannot load this extension. TurboWarp always sandboxes an extension
-> loaded from an untrusted URL — the URL tab has no **Run without sandbox** option at all — and this
-> extension refuses to start sandboxed. Pasting the jsDelivr link below into the URL tab leaves the
-> dialog waiting with nothing on screen and only `TM must run without the extension sandbox.` in the
-> browser console. Download the file and use the **Files** tab.
+> The **URL** tab sandboxes untrusted extension URLs and cannot provide the access this extension
+> needs. Use **Text** or **Files** with **Run without sandbox** enabled.
+
+All extension distribution filenames use the `teachable-machine` prefix.
+Existing Scratch extension IDs, opcodes, and runtime API names remain compatible.
 
-The download, and the artifact a composite runtime fetches, is the version-pinned build on jsDelivr:
+The version-pinned lightweight build uses this CDN URL after publication:
 
 ```text
-https://cdn.jsdelivr.net/npm/@kubohiroya/turbowarp-teachable-machine@4.0.1/dist/tm.js
+https://cdn.jsdelivr.net/npm/@kubohiroya/turbowarp-teachable-machine@4.0.2/dist/teachable-machine.js
 ```
 
-A composite runtime can load or embed the same reviewed artifact without rewriting a minified
-third-party bundle:
+The lightweight extension loads one reviewed browser runtime that contains one TensorFlow.js 4.22.0
+module graph together with Teachable Machine Pose 0.8.6, Teachable Machine Image 0.8.5, and
+TensorFlow.js Speech Commands 0.5.4. That runtime ships the WebGL and CPU kernels; the WebGPU and
+WebAssembly backends live in separate bundles next to it and are fetched only when a project uses
+them. See [Compute backends](#compute-backends).
+Composite runtimes can load or embed the same artifact without rewriting a minified third-party
+bundle:
 
 ```text
-https://cdn.jsdelivr.net/npm/@kubohiroya/turbowarp-teachable-machine@4.0.1/dist/runtime.js
+https://cdn.jsdelivr.net/npm/@kubohiroya/turbowarp-teachable-machine@4.0.2/dist/runtime.js
 ```
 
 To add the published package to another project:
 
 ```sh
-pnpm add --save-exact @kubohiroya/turbowarp-teachable-machine@4.0.1
+pnpm add --save-exact @kubohiroya/turbowarp-teachable-machine@4.0.2
 ```
 
 ### Names
@@ -958,8 +985,9 @@ pnpm check
 
 `pnpm check` runs type checking, tests, the production build, generated-documentation validation,
 Pages link validation, distribution reproducibility, and an npm package dry run. The build produces
-`dist/tm.js`, `dist/composition.js`, `dist/runtime.js`, `dist/posenet.js`, and the three raw PoseNet
-model assets under `dist/posenet/`.
+`dist/teachable-machine.js`, `dist/teachable-machine-webgpu.js`, `dist/teachable-machine-wasm.js`,
+`dist/composition.js`, `dist/runtime.js`, `dist/posenet.js`, and the three raw
+PoseNet model assets under `dist/posenet/`.
 
 The block reference above is generated from `src/block-definitions.json` by `pnpm docs`; edit the
 definitions, not the generated section. The version in `package.json` is the release source of
@@ -967,6 +995,15 @@ truth: the runtime reporter appends `-typescript` to it, and the release consist
 generated bundles, the exact-version README examples, the Pages badges, and the release tag aligned
 with it.
 
+### Standalone editor validation
+
+The [standalone validation guide (Japanese)](docs/standalone-validation.ja.md) records the
+standard-editor model-recognition checks for both **Text** and **Files**, including WebGPU
+and multithreaded WASM. To repeat them against a dedicated Chrome profile with a fake camera,
+follow that guide and run `node scripts/check-standalone-editor.mjs`. This optional browser
+check requires network access to the editor and the public Pose model; its JSON report records
+the browser version and SHA-256 of each tested JavaScript file.
+
 ### npm publishing
 
 npm publication uses GitHub Actions trusted publishing rather than a long-lived write token. The npm
@@ -981,12 +1018,15 @@ not add a write token to the repository.
 
 The version-pinned `dist/runtime.js` artifact bundles TensorFlow.js 4.22.0, Teachable Machine Pose
 0.8.6, Teachable Machine Image 0.8.5, TensorFlow.js Speech Commands 0.5.4, and PoseNet 2.2.2 in one
-reviewed module graph. The standalone extension loads that single artifact from jsDelivr when no
+reviewed module graph. The lightweight extension loads that single artifact from jsDelivr when no
 runtime is injected. The optional `dist/backend-wasm.js` and `dist/backend-webgpu.js` bundles
 resolve TensorFlow.js against the instance that runtime already published, so a page never ends up
 with two TensorFlow.js cores. The WebAssembly binaries under `dist/wasm/` and the fixed offline
 PoseNet model data under `dist/posenet/` are distributed as their original binary files.
 
+The self-contained WebGPU/WASM extensions also embed the runtime libraries, backend code,
+and WASM binaries in their JavaScript file, together with the complete third-party notice.
+
 Third-party licences are listed in [THIRD_PARTY_NOTICES.md](THIRD_PARTY_NOTICES.md).
 
 ## Fixed input editors

package/THIRD_PARTY_NOTICES.md

@@ -1,6 +1,7 @@
 # Third-party notices
 
-The reviewed browser runtime (`dist/runtime.js`), the optional compute backend
+The self-contained extensions (`dist/teachable-machine-webgpu.js` and
+`dist/teachable-machine-wasm.js`), the reviewed browser runtime (`dist/runtime.js`), the optional compute backend
 bundles (`dist/backend-wasm.js`, `dist/backend-webgpu.js`, `dist/wasm/`), and the
 fixed offline PoseNet model supply (`dist/posenet/`) include the following
 software and model data:

package/dist/THIRD_PARTY_NOTICES.md

@@ -1,6 +1,7 @@
 # Third-party notices
 
-The reviewed browser runtime (`dist/runtime.js`), the optional compute backend
+The self-contained extensions (`dist/teachable-machine-webgpu.js` and
+`dist/teachable-machine-wasm.js`), the reviewed browser runtime (`dist/runtime.js`), the optional compute backend
 bundles (`dist/backend-wasm.js`, `dist/backend-webgpu.js`, `dist/wasm/`), and the
 fixed offline PoseNet model supply (`dist/posenet/`) include the following
 software and model data:

package/dist/backend-wasm.js

@@ -1,4 +1,4 @@
-/*! @license Includes TensorFlow.js WASM backend 4.22.0 (Apache-2.0). See https://github.com/kubohiroya/turbowarp-teachable-machine/blob/v4.0.1/THIRD_PARTY_NOTICES.md. */
+/*! @license Includes TensorFlow.js WASM backend 4.22.0 (Apache-2.0). See https://github.com/kubohiroya/turbowarp-teachable-machine/blob/v4.0.2/THIRD_PARTY_NOTICES.md. */
 (function(i){var ki=Object.create,wr=Object.defineProperty,Si=Object.getOwnPropertyDescriptor,Mi=Object.getOwnPropertyNames,Ai=Object.getPrototypeOf,Fr=Object.prototype.hasOwnProperty,dn=(t,a)=>()=>(a||(t((a={exports:{}}).exports,a),t=null),a.exports),xi=(t,a,u,d)=>{if(a&&typeof a=="object"||typeof a=="function")for(var r=Mi(a),n=0,o=r.length,p;n<o;n++)p=r[n],!Fr.call(t,p)&&p!==u&&wr(t,p,{get:(l=>a[l]).bind(null,p),enumerable:!(d=Si(a,p))||d.enumerable});return t},Nr=(t,a,u)=>(u=t!=null?ki(Ai(t)):{},xi(a||!t||!t.__esModule||!Fr.call(t,"default")?wr(u,"default",{value:t,enumerable:!0}):u,t)),G;(function(t){t[t.float32=0]="float32",t[t.int32=1]="int32",t[t.bool=2]="bool",t[t.string=3]="string",t[t.complex64=4]="complex64"})(G||(G={}));var qe;(function(t){t[t.linear=0]="linear",t[t.relu=1]="relu",t[t.relu6=2]="relu6",t[t.prelu=3]="prelu",t[t.leakyrelu=4]="leakyrelu",t[t.sigmoid=5]="sigmoid",t[t.elu=6]="elu"})(qe||(qe={}));var Cr;function Di(t){Cr=t.wasm.cwrap(i._FusedMatMul,null,["number","array","number","number","array","number","number","number","number","number","number","number","number"])}function wi(t){const{inputs:a,backend:u,attrs:d}=t,{a:r,b:n,bias:o,preluActivationWeights:p}=a;if(r.dtype!=="float32"||n.dtype!=="float32")throw new Error("_FusedMatMul for non non-float32 tensors not yet supported.");const{transposeA:l,transposeB:s,activation:e,leakyreluAlpha:m}=d,f=u.dataIdMap.get(r.dataId).id,g=u.dataIdMap.get(n.dataId).id;let h=0;if(o!=null){const B=u.dataIdMap.get(o.dataId);if(B.shape.length!==1)throw new Error(`_FusedMatMul only supports rank-1 bias but got rank ${B.shape.length}.`);h=B.id}const b=p==null?0:u.dataIdMap.get(p.dataId).id,y=qe[e];if(y==null)throw new Error(`${e} activation not yet supported for FusedConv2D in the wasm backend.`);const k=l?r.shape[2]:r.shape[1],I=s?n.shape[1]:n.shape[2],M=i.broadcast_util.assertAndGetBroadcastShape(r.shape.slice(0,-2),n.shape.slice(0,-2)),D=u.makeOutput([...M,k,I],r.dtype),F=u.dataIdMap.get(D.dataId).id,E=new Uint8Array(new Int32Array(r.shape).buffer),O=new Uint8Array(new Int32Array(n.shape).buffer);return Cr(f,E,r.shape.length,g,O,n.shape.length,l,s,y,h,b,m||0,F),D}var Fi={kernelName:i._FusedMatMul,backendName:"wasm",setupFunc:Di,kernelFunc:wi};function W(t,a){let u;function d(n){u=n.wasm.cwrap(t,null,["number","number","number"])}function r(n){const{backend:o,inputs:{x:p}}=n,l=o.dataIdMap.get(p.dataId).id,s=o.makeOutput(p.shape,a||p.dtype),e=o.dataIdMap.get(s.dataId).id;return i.util.sizeFromShape(s.shape)===0||u(l,G[p.dtype],e),s}return{kernelName:t,backendName:"wasm",setupFunc:d,kernelFunc:r}}var Ni=W(i.Abs),Ci=W(i.Acos),Ri=W(i.Acosh);function Y(t,a,u){let d;function r(o){d=o.wasm.cwrap(t,null,["number","array","number","number","array","number","number","number"])}function n(o){const{backend:p,inputs:l}=o,{a:s,b:e}=l,m=p.dataIdMap.get(s.dataId).id,f=p.dataIdMap.get(e.dataId).id,g=u??s.dtype,h=i.backend_util.assertAndGetBroadcastShape(s.shape,e.shape),b=p.makeOutput(h,g);if(i.util.sizeFromShape(h)===0)return b;const y=new Uint8Array(new Int32Array(s.shape).buffer),k=new Uint8Array(new Int32Array(e.shape).buffer),I=p.dataIdMap.get(b.dataId).id;return d(m,y,s.shape.length,f,k,e.shape.length,G[s.dtype],I),b}return{kernelName:t,backendName:"wasm",setupFunc:r,kernelFunc:n}}var _i=!0,Pi=Y(i.Add,_i),Rr;function Ei(t){Rr=t.wasm.cwrap(i.AddN,null,["array","number","number","number"])}function Bi(t){const{inputs:a,backend:u}=t,d=u.makeOutput(a[0].shape,a[0].dtype);if(i.util.sizeFromShape(d.shape)===0)return d;const r=a.map(p=>u.dataIdMap.get(p.dataId).id),n=new Uint8Array(new Int32Array(r).buffer),o=u.dataIdMap.get(d.dataId).id;return Rr(n,r.length,G[d.dtype],o),d}var Ti={kernelName:i.AddN,backendName:"wasm",setupFunc:Ei,kernelFunc:Bi};function on(t){const{inputs:{x:a},backend:u}=t;if(a.dtype==="string")return(0,i.tensor)(u.readSync(a.dataId),a.shape,a.dtype);const d=u.makeOutput(a.shape,a.dtype),r=u.typedArrayFromHeap(a);return u.typedArrayFromHeap(d).set(r),d}var Oi={kernelName:i.Identity,backendName:"wasm",kernelFunc:on},_r;function Wi(t){_r=t.wasm.cwrap(i.Transpose,null,["number","array","number","number","number","array","number"])}function Ie(t){const{inputs:a,backend:u,attrs:d}=t,[r,n]=Hi(a.x.shape,d.perm);let o=!0;for(let h=0;h<n.length;h++)n[h]!==h&&(o=!1);const p=zi(a.x.shape,d.perm),l={dataId:a.x.dataId,shape:r,dtype:a.x.dtype};if(o){const h=on({inputs:a,backend:u});return h.shape=p,h}const s=u.makeOutput(p,l.dtype),e=u.dataIdMap.get(l.dataId).id,m=u.dataIdMap.get(s.dataId).id,f=new Uint8Array(new Int32Array(n).buffer),g=new Uint8Array(new Int32Array(l.shape).buffer);return _r(e,g,l.shape.length,G[l.dtype],m,f,n.length),s}function zi(t,a){const u=new Array(t.length);for(let d=0;d<u.length;d++)u[d]=t[a[d]];return u}function Hi(t,a){const u=[],d=[];for(let r=0;r<t.length;++r)t[r]!==1&&u.push(t[r]),t[a[r]]!==1&&d.push(a[r]);for(let r=0;r<d.length;++r){let n=-1;for(let o=0;o<d.length;++o)d[o]>=r&&(n===-1||d[n]>d[o])&&(n=o);d[n]=r}return[u,d]}var Li={kernelName:i.Transpose,backendName:"wasm",kernelFunc:Ie,setupFunc:Wi};function ke(t,a,u){const d=t.shape,r=t.shape.length,n=i.util.parseAxisParam(a,d);let o=n;const p=i.backend_util.getAxesPermutation(o,r);let l=null,s=!1;if(p!=null){const e=new Array(r);for(let f=0;f<e.length;f++)e[f]=d[p[f]];o=i.backend_util.getInnerMostAxes(o.length,r),l=Ie({inputs:{x:t},attrs:{perm:p},backend:u});const m=u.dataIdMap.get(t.dataId).id;u.dataIdMap.get(l.dataId).id!==m&&(s=!0)}return{transposed:l,originalAxes:n,axes:o,inputWasTransposed:s}}var Pr;function Gi(t){Pr=t.wasm.cwrap(i.All,null,["number, number, number"])}function $i(t){const{backend:a,inputs:u,attrs:d}=t,{axis:r,keepDims:n}=d,{x:o}=u;let p=a.dataIdMap.get(o.dataId).id,l=o;const{transposed:s,axes:e,originalAxes:m,inputWasTransposed:f}=ke(o,r,a);if(f){const I=a.dataIdMap.get(s.dataId).id;l=s,p=I}const g=l.shape.length;i.backend_util.assertAxesAreInnerMostDims("all",e,g);const[h,b]=i.backend_util.computeOutAndReduceShapes(l.shape,e),y=i.util.sizeFromShape(b),k=a.makeOutput(h,o.dtype);if(i.util.sizeFromShape(l.shape)!==0){const I=a.dataIdMap.get(k.dataId).id;Pr(p,y,I)}return f&&a.disposeData(s.dataId),n&&(k.shape=i.backend_util.expandShapeToKeepDim(k.shape,m)),k}var Vi={kernelName:i.All,backendName:"wasm",setupFunc:Gi,kernelFunc:$i},Er;function Ui(t){Er=t.wasm.cwrap(i.Any,null,["number, number, number"])}function qi(t){const{backend:a,inputs:u,attrs:d}=t,{axis:r,keepDims:n}=d,{x:o}=u;let p=a.dataIdMap.get(o.dataId).id,l=o;const{transposed:s,axes:e,originalAxes:m,inputWasTransposed:f}=ke(o,r,a);if(f){const I=a.dataIdMap.get(s.dataId).id;l=s,p=I}const g=l.shape.length;i.backend_util.assertAxesAreInnerMostDims("any",e,g);const[h,b]=i.backend_util.computeOutAndReduceShapes(l.shape,e),y=i.util.sizeFromShape(b),k=a.makeOutput(h,o.dtype);if(i.util.sizeFromShape(l.shape)!==0){const I=a.dataIdMap.get(k.dataId).id;Er(p,y,I)}return f&&a.disposeData(s.dataId),n&&(k.shape=i.backend_util.expandShapeToKeepDim(k.shape,m)),k}var Ki={kernelName:i.Any,backendName:"wasm",setupFunc:Ui,kernelFunc:qi};function Br(t){let a;function u(r){a=r.wasm.cwrap(t,null,["number","number","number","number","number"])}function d(r){const{backend:n,inputs:o,attrs:p}=r,{axis:l}=p,{x:s}=o,e=n.dataIdMap.get(s.dataId).id;let m=e,f=s;const{transposed:g,axes:h,inputWasTransposed:b}=ke(s,l,n);if(b){const F=n.dataIdMap.get(g.dataId).id;F!==e&&(f=g,m=F)}const y=f.shape.slice(0,-1),k=n.makeOutput(y,"int32"),I=n.dataIdMap.get(k.dataId).id,M=i.util.sizeFromShape(k.shape),D=f.shape[h[0]];return a(m,G[f.dtype],M,D,I),b&&n.disposeData(g.dataId),k}return{kernelName:t,backendName:"wasm",setupFunc:u,kernelFunc:d}}var Xi=Br(i.ArgMax),Ji=Br(i.ArgMin),Qi=W(i.Asin),Yi=W(i.Asinh),Zi=W(i.Atan),ji=Y(i.Atan2,!1),es=W(i.Atanh),Tr;function ns(t){Tr=t.wasm.cwrap(i.AvgPool,null,["number","number","number","number","number","number","number","number","number","number","number","number","number","number"])}function ts(t){const{inputs:a,attrs:u,backend:d}=t,r=a.x,n=d.dataIdMap.get(r.dataId).id,{filterSize:o,strides:p,pad:l,dimRoundingMode:s}=u,e=i.backend_util.computePool2DInfo(r.shape,o,p,1,l,s),m=e.filterHeight,f=e.filterWidth,g=e.padInfo.top,h=e.padInfo.right,b=e.padInfo.bottom,y=e.padInfo.left,k=e.strideHeight,I=e.strideWidth,M=e.inChannels;if(e.dataFormat!=="channelsLast")throw new Error(`wasm backend does not support dataFormat:'${e.dataFormat}'. Please use 'channelsLast'.`);if(e.dilationWidth!==1||e.dilationHeight!==1)throw new Error(`was backend only supports average pooling with dilation = [1, 1], got [${e.dilationHeight}, ${e.dilationWidth}].`);const D=d.makeOutput(e.outShape,"float32"),F=d.dataIdMap.get(D.dataId).id;return Tr(n,r.shape[0],r.shape[1],r.shape[2],m,f,g,h,b,y,k,I,M,F),D}var as={kernelName:i.AvgPool,backendName:"wasm",setupFunc:ns,kernelFunc:ts},Or;function rs(t){Or=t.wasm.cwrap("AvgPool3D",null,["number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number"])}function us(t){const{inputs:a,backend:u,attrs:d}=t,{x:r}=a,{filterSize:n,strides:o,pad:p,dimRoundingMode:l,dataFormat:s}=d,e=i.backend_util.computePool3DInfo(r.shape,n,o,1,p,l,s),m=u.makeOutput(e.outShape,r.dtype);return Or(u.dataIdMap.get(r.dataId).id,u.dataIdMap.get(m.dataId).id,e.batchSize,e.inChannels,e.inDepth,e.inHeight,e.inWidth,e.outDepth,e.outHeight,e.outWidth,e.strideDepth,e.strideHeight,e.strideWidth,e.dilationDepth,e.dilationHeight,e.dilationWidth,e.effectiveFilterDepth,e.effectiveFilterHeight,e.effectiveFilterWidth,e.padInfo.front,e.padInfo.top,e.padInfo.left),m}var is={kernelName:i.AvgPool3D,backendName:"wasm",setupFunc:rs,kernelFunc:us},Wr;function ss(t){Wr=t.wasm.cwrap("AvgPool3DGrad",null,["number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number"])}function ds(t){const{inputs:a,backend:u,attrs:d}=t,{dy:r,input:n}=a,{filterSize:o,strides:p,pad:l,dimRoundingMode:s}=d,e=i.backend_util.computePool3DInfo(n.shape,o,p,1,l,s),m=u.makeOutput(n.shape,n.dtype);return Wr(u.dataIdMap.get(r.dataId).id,u.dataIdMap.get(m.dataId).id,e.batchSize,e.inChannels,e.inDepth,e.inHeight,e.inWidth,e.outDepth,e.outHeight,e.outWidth,e.strideDepth,e.strideHeight,e.strideWidth,e.dilationDepth,e.dilationHeight,e.dilationWidth,e.effectiveFilterDepth,e.effectiveFilterHeight,e.effectiveFilterWidth,e.padInfo.front,e.padInfo.top,e.padInfo.left,e.filterDepth,e.filterHeight,e.filterWidth),m}var os={kernelName:i.AvgPool3DGrad,backendName:"wasm",setupFunc:ss,kernelFunc:ds},zr;function ps(t){zr=t.wasm.cwrap("AvgPoolGrad",null,["number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number"])}function ls(t){const{inputs:a,backend:u,attrs:d}=t,{dy:r,input:n}=a,{filterSize:o,strides:p,pad:l}=d,s=i.backend_util.computePool2DInfo(n.shape,o,p,1,l),e=u.makeOutput(n.shape,n.dtype);return zr(u.dataIdMap.get(r.dataId).id,u.dataIdMap.get(e.dataId).id,s.batchSize,s.inChannels,s.inHeight,s.inWidth,s.outHeight,s.outWidth,s.strideHeight,s.strideWidth,s.dilationHeight,s.dilationWidth,s.effectiveFilterHeight,s.effectiveFilterWidth,s.padInfo.top,s.padInfo.left,s.filterHeight,s.filterWidth),e}var cs={kernelName:i.AvgPoolGrad,backendName:"wasm",setupFunc:ps,kernelFunc:ls};function ue(t){const{inputs:a,attrs:u}=t,{x:d}=a,{shape:r}=u,n=i.util.sizeFromShape(d.shape),o=i.util.inferFromImplicitShape(r,n);return i.util.assert(n===i.util.sizeFromShape(o),()=>`new shape: ${o}, old shape: ${d.shape}. New shape and old shape must have the same number of elements.`),t.backend.incRef(d.dataId),{dataId:d.dataId,shape:o,dtype:d.dtype}}var ms={kernelName:i.Reshape,backendName:"wasm",kernelFunc:ue},Hr;function hs(t){Hr=t.wasm.cwrap(i.BatchMatMul,null,["number","array","number","number","array","number","number","number","number"])}function gs(t){const{inputs:a,backend:u,attrs:d}=t,{a:r,b:n}=a,{transposeA:o,transposeB:p}=d;if(r.dtype!=="float32"||n.dtype!=="float32")throw new Error("BatchMatMul for non non-float32 tensors not yet supported.");const l=r.shape.length,s=n.shape.length,e=o?r.shape[l-2]:r.shape[l-1],m=p?n.shape[s-1]:n.shape[s-2],f=o?r.shape[l-1]:r.shape[l-2],g=p?n.shape[s-2]:n.shape[s-1],h=r.shape.slice(0,-2),b=n.shape.slice(0,-2),y=i.util.sizeFromShape(h),k=i.util.sizeFromShape(b),I=i.broadcast_util.assertAndGetBroadcastShape(r.shape.slice(0,-2),n.shape.slice(0,-2)).concat([f,g]);i.util.assert(e===m,()=>`Error in matMul: inner shapes (${e}) and (${m}) of Tensors with shapes ${r.shape} and ${n.shape} and transposeA=${o} and transposeB=${p} must match.`);const M=o?[y,e,f]:[y,f,e],D=p?[k,g,m]:[k,m,g],F=ue({inputs:{x:r},backend:u,attrs:{shape:M}}),E=ue({inputs:{x:n},backend:u,attrs:{shape:D}}),O=u.dataIdMap.get(F.dataId).id,B=u.dataIdMap.get(E.dataId).id,K=o?F.shape[2]:F.shape[1],X=p?E.shape[1]:E.shape[2],U=Math.max(y,k),q=u.makeOutput([U,K,X],F.dtype),Z=u.dataIdMap.get(q.dataId).id,ie=new Uint8Array(new Int32Array(F.shape).buffer),le=new Uint8Array(new Int32Array(E.shape).buffer);return Hr(O,ie,F.shape.length,B,le,E.shape.length,o,p,Z),u.disposeData(F.dataId),u.disposeData(E.dataId),q.shape=I,q}var fs={kernelName:i.BatchMatMul,backendName:"wasm",setupFunc:hs,kernelFunc:gs};function bs(t,a,u,d){const r=i.util.getArrayFromDType(u,i.util.sizeFromShape(a));if(d&&u!=="string"){let n=0;t.forEach(o=>{const p=i.util.sizeFromShape(o.shape);r.set(o.vals,n),n+=p})}else{let n=0;t.forEach(o=>{const p=u==="string"?i.backend_util.fromUint8ToStringArray(o.vals):o.vals;let l=0;for(let s=0;s<o.shape[0];++s){const e=s*a[1]+n;for(let m=0;m<o.shape[1];++m)r[e+m]=p[l++]}n+=o.shape[1]})}return r}function vs(t,a,u,d){if(t===a||t<a&&u<0||a<t&&u>1)return i.util.makeZerosTypedArray(0,d);const r=Math.abs(Math.ceil((a-t)/u)),n=i.util.makeZerosTypedArray(r,d);a<t&&u===1&&(u=-1),n[0]=t;for(let o=1;o<n.length;o++)n[o]=n[o-1]+u;return n}function Lr(t,a,u,d,r){const n=i.slice_util.isSliceContinous(d,a,u),o=i.util.sizeFromShape(u),p=i.util.computeStrides(d);if(n){const m=i.slice_util.computeFlatOffset(a,p);return r==="string"?t.slice(m,m+o):t.subarray(m,m+o)}const l=r==="string"?i.backend_util.fromUint8ToStringArray(t):t,s=(0,i.buffer)(d,r,l),e=(0,i.buffer)(u,r);for(let m=0;m<e.size;++m){const f=e.indexToLoc(m),g=f.map((h,b)=>h+a[b]);e.set(s.get(...g),...f)}return r==="string"?i.backend_util.fromStringArrayToUint8(e.values):e.values}var ys=class{constructor(t,a,u,d,r,n){this.separator=i.util.encodeString(t),this.nGramWidths=a,this.leftPad=i.util.encodeString(u),this.rightPad=i.util.encodeString(d),this.padWidth=r,this.preserveShort=n}getPadWidth(t){return Math.min(this.padWidth<0?t-1:this.padWidth,t-1)}getNumNGrams(t,a){const u=this.getPadWidth(a);return Math.max(0,t+2*u-a+1)}createNGrams(t,a,u,d,r,n){for(let o=0;o<r;++o){const p=this.getPadWidth(n),l=Math.max(0,p-o),s=Math.max(0,p-(r-(o+1))),e=n-(l+s),m=a+(l>0?0:o-p);let f=0;f+=l*this.leftPad.length;for(let k=0;k<e;++k)f+=t[m+k].length;f+=s*this.rightPad.length;const g=l+s+e-1;f+=g*this.separator.length,u[d+o]=new Uint8Array(f);const h=u[d+o];let b=0;const y=k=>k.forEach(I=>h[b++]=I);for(let k=0;k<l;++k)y(this.leftPad),y(this.separator);for(let k=0;k<e-1;++k)y(t[m+k]),y(this.separator);if(e>0){y(t[m+e-1]);for(let k=0;k<s;++k)y(this.separator),y(this.rightPad)}else{for(let k=0;k<s-1;++k)y(this.rightPad),y(this.separator);y(this.rightPad)}}}compute(t,a){const u=t.length,d=a.length;if(d>0){let p=a[0];if(p!==0)throw new Error(`First split value must be 0, got ${p}`);for(let l=1;l<d;++l){let s=a[l]>=p;if(s=s&&a[l]<=u,!s)throw new Error(`Invalid split value ${a[l]}, must be in [${p}, ${u}]`);p=a[l]}if(p!==u)throw new Error(`Last split value must be data size. Expected ${u}, got ${p}`)}const r=d-1,n=i.util.getArrayFromDType("int32",d);if(u===0||d===0){const p=new Array(u);for(let l=0;l<=r;++l)n[l]=0;return[p,n]}n[0]=0;for(let p=1;p<=r;++p){const l=a[p]-a[p-1];let s=0;this.nGramWidths.forEach(e=>{s+=this.getNumNGrams(l,e)}),this.preserveShort&&l>0&&s===0&&(s=1),n[p]=n[p-1]+s}const o=new Array(n[r]);for(let p=0;p<r;++p){const l=a[p];let s=n[p];if(this.nGramWidths.forEach(e=>{const m=a[p+1]-a[p],f=this.getNumNGrams(m,e);this.createNGrams(t,l,o,s,f,e),s+=f}),this.preserveShort&&s===n[p]){const e=a[p+1]-a[p];if(e===0)continue;const m=e+2*this.padWidth;this.createNGrams(t,l,o,s,1,m)}}return[o,n]}};function Is(t,a,u,d,r,n,o,p){return new ys(u,d,r,n,o,p).compute(t,a)}function ks(t,a,u,d){if(!t.length)return;if(a.length===0){for(let n=0;n<t.length;++n)d.push(t.subarray(n,n+1));return}if(a.length===1){const n=a[0];let o=t.indexOf(n);for(;o!==-1;){const p=t.subarray(0,o);(!u||p.length!==0)&&d.push(p),t=t.subarray(o+1),o=t.indexOf(n)}(!u||t.length!==0)&&d.push(t);return}let r=0;for(let n=0;n<t.length+1;n++)if(n===t.length||a.indexOf(t[n])!==-1){const o=t.subarray(r,n);(!u||o.length!==0)&&d.push(o),r=n+1}}function Ss(t,a,u){const d=t.length,r=[];let n=0,o=0;const p=new Array(d);for(let f=0;f<d;++f){const g=r.length;ks(t[f],a,u,r);const h=r.length-g;p[f]=h,n+=h,o=Math.max(o,h)}const l=i.util.getArrayFromDType("int32",n*2),s=new Array(n),e=[d,o];let m=0;for(let f=0;f<d;++f)for(let g=0;g<p[f];++g)l[m*2]=f,l[m*2+1]=g,s[m]=r[m],++m;return[l,s,e]}function Ms(t,a){const u=i.util.getArrayFromDType("int32",t.length);for(let d=0;d<t.length;++d)u[d]=i.util.fingerPrint64(t[d]).modulo(a).getLowBitsUnsigned();return u}function As(t,a,u,d){const r=i.util.parseAxisParam(a,u)[0],n=[1,u[0],1];for(let h=0;h<r;h++)n[0]*=u[h];n[1]=u[r];for(let h=r+1;h<u.length;h++)n[2]*=u[h];const o=new Map,p=new Int32Array(u[r]),l=new i.TensorBuffer(n,d,t),s=[],e=n[0]===1&&n[2]===1;for(let h=0;h<u[r];h++){let b;if(e)b=t[h].toString();else{const k=[];for(let I=0;I<n[0];I++)for(let M=0;M<n[2];M++)k.push(l.get(I,h,M));b=k.join(",")}const y=o.get(b);if(y!=null)p[h]=y;else{const k=o.size;o.set(b,k),p[h]=k,s.push(h)}}const m=n.slice();m[1]=o.size;const f=new i.TensorBuffer(m,d);s.forEach((h,b)=>{for(let y=0;y<n[0];y++)for(let k=0;k<n[2];k++)f.set(l.get(y,h,k),y,b,k)});const g=u.slice();return g[r]=m[1],{outputValues:f.values,outputShape:g,indices:p}}function De(t){const{inputs:{x:a},attrs:{begin:u,size:d},backend:r}=t,[n,o]=i.slice_util.parseSliceParams(a,u,d),p=i.slice_util.isSliceContinous(a.shape,n,o),l=r.readSync(a.dataId),s=r.makeOutput(o,a.dtype),e=i.util.computeStrides(a.shape),m=r.dataIdMap.get(s.dataId);if(p){const h=i.slice_util.computeFlatOffset(n,e);return a.dtype==="string"?m.stringBytes=l.slice(h,h+i.util.sizeFromShape(o)):r.typedArrayFromHeap(s).set(l.subarray(h,h+i.util.sizeFromShape(o))),s}if(a.dtype==="string")return m.stringBytes=Lr(l,n,o,a.shape,a.dtype),s;const f=r.typedArrayFromHeap(s),g=a.shape.length;if(g===2)xs(l,e[0],f,n,o);else if(g===3)Ds(l,e[0],e[1],f,n,o);else if(g===4)ws(l,e[0],e[1],e[2],f,n,o);else{const h=Lr(l,n,o,a.shape,a.dtype);f.set(h)}return s}function xs(t,a,u,d,r){let n=0;const o=d[0],p=d[1],l=o+r[0];for(let s=o;s<l;s++){const e=s*a+p;u.set(t.subarray(e,e+r[1]),n),n+=r[1]}}function Ds(t,a,u,d,r,n){let o=0;const p=r[0],l=r[1],s=r[2],e=p+n[0],m=l+n[1];for(let f=p;f<e;f++)for(let g=l;g<m;g++){const h=f*a+g*u+s;d.set(t.subarray(h,h+n[2]),o),o+=n[2]}}function ws(t,a,u,d,r,n,o){let p=0;const l=n[0],s=n[1],e=n[2],m=l+o[0],f=s+o[1],g=e+o[2],h=n[3];for(let b=l;b<m;b++)for(let y=s;y<f;y++)for(let k=e;k<g;k++){const I=b*a+y*u+k*d+h;r.set(t.subarray(I,I+o[3]),p),p+=o[3]}}var Fs={kernelName:i.Slice,backendName:"wasm",kernelFunc:De};function Ns(t){const{inputs:a,backend:u,attrs:d}=t,{x:r}=a,{blockShape:n,crops:o}=d,p=n.reduce((k,I)=>k*I),l=i.backend_util.getReshaped(r.shape,n,p),s=i.backend_util.getPermuted(l.length,n.length),e=i.backend_util.getReshapedPermuted(r.shape,n,p),m=i.backend_util.getSliceBeginCoords(o,n.length),f=i.backend_util.getSliceSize(e,o,n.length),g=ue({inputs:{x:r},backend:u,attrs:{shape:l}}),h=Ie({inputs:{x:g},backend:u,attrs:{perm:s}}),b=ue({inputs:{x:h},backend:u,attrs:{shape:e}}),y=De({inputs:{x:b},backend:u,attrs:{begin:m,size:f}});return u.disposeData(g.dataId),u.disposeData(h.dataId),u.disposeData(b.dataId),y}var Cs={kernelName:i.BatchToSpaceND,backendName:"wasm",kernelFunc:Ns},Gr;function Rs(t){Gr=t.wasm.cwrap(i.Bincount,null,["number","number","boolean","number","number","number"])}function _s(t){const{backend:a,inputs:u,attrs:d}=t,{x:r,weights:n}=u,{size:o}=d,p=n.shape.reduce((m,f)=>m*f,1)!==0,l=r.shape.length===1?[o]:[r.shape[0],o],s=a.makeOutput(l,n.dtype);function e(m){return a.dataIdMap.get(m.dataId).id}return Gr(e(r),o,p,e(n),G[n.dtype],e(s)),s}var Ps={kernelName:i.Bincount,backendName:"wasm",setupFunc:Rs,kernelFunc:_s},Es=!0,Bs=Y(i.BitwiseAnd,Es);function Ts(t){const{inputs:a,backend:u}=t,{s0:d,s1:r}=a,n=u.typedArrayFromHeap(d),o=u.typedArrayFromHeap(r),p=i.backend_util.assertAndGetBroadcastShape(Array.from(n),Array.from(o));return u.makeOutput([p.length],"int32",void 0,new Int32Array(p))}var Os={kernelName:i.BroadcastArgs,backendName:"wasm",kernelFunc:Ts};function Se(t){const{inputs:{x:a},attrs:{dtype:u},backend:d}=t,r=d.makeOutput(a.shape,u),n=d.typedArrayFromHeap(a);return d.typedArrayFromHeap(r).set(n),r}var Ws={kernelName:i.Cast,backendName:"wasm",kernelFunc:Se},zs=W(i.Ceil),$r;function Hs(t){$r=t.wasm.cwrap(i.ClipByValue,null,["number","number","number","number"])}function Ls(t){const{inputs:a,backend:u,attrs:d}=t,{x:r}=a,{clipValueMin:n,clipValueMax:o}=d,p=u.dataIdMap.get(r.dataId).id,l=u.makeOutput(r.shape,r.dtype),s=u.dataIdMap.get(l.dataId).id;return $r(p,n,o,s),l}var Gs={kernelName:i.ClipByValue,backendName:"wasm",setupFunc:Hs,kernelFunc:Ls};function Vr(t){const{inputs:a,backend:u}=t,d=i.util.parseAxisParam(t.attrs.axis,a[0].shape)[0],r=a.map(g=>g.shape);i.backend_util.assertParamsConsistent(r,d);let n=i.backend_util.computeOutShape(a.map(g=>g.shape),d);const o=a.filter(g=>i.util.sizeFromShape(g.shape)>0);if(o.length===1)return on({inputs:{x:o[0]},backend:u});const p=u.makeOutput(n,a[0].dtype);if(i.util.sizeFromShape(n)===0)return p;if(o[0].dtype==="string"){const g=o.map(I=>{const M=[-1,i.util.sizeFromShape(I.shape.slice(d))];return ue({inputs:{x:I},backend:u,attrs:{shape:M}})}),h=g.map(I=>({vals:u.readSync(I.dataId),shape:I.shape}));n=i.backend_util.computeOutShape(g.map(I=>I.shape),1);const b=g[0].shape[0]===1,y=bs(h,n,a[0].dtype,b);p.shape=i.backend_util.computeOutShape(o.map(I=>I.shape),d);const k=u.dataIdMap.get(p.dataId);return k.stringBytes=i.backend_util.fromStringArrayToUint8(y),g.forEach(I=>u.disposeData(I.dataId)),p}const l=i.util.sizeFromShape(o[0].shape.slice(0,d));let s=0;const e=o.map(g=>{const h=i.util.sizeFromShape(g.shape.slice(d));return s+=h,h}),m=o.map(g=>u.typedArrayFromHeap(g)),f=u.typedArrayFromHeap(p);for(let g=0;g<l;g++){let h=g*s;for(let b=0;b<m.length;b++){const y=e[b],k=g*y,I=m[b].subarray(k,k+y);f.set(I,h),h+=y}}return p}var $s={kernelName:i.Concat,backendName:"wasm",kernelFunc:Vr},Ur;function Vs(t){Ur=t.wasm.cwrap(i.Conv2D,null,["number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number"])}function Us(t){const{inputs:a,attrs:u,backend:d}=t,{x:r,filter:n}=a,o=d.dataIdMap.get(r.dataId).id,p=d.dataIdMap.get(n.dataId).id,{strides:l,dilations:s,pad:e,dimRoundingMode:m,dataFormat:f}=u,g=i.backend_util.convertConv2DDataFormat(f),h=i.backend_util.computeConv2DInfo(r.shape,n.shape,l,s,e,m,!1,g),b=h.filterHeight,y=h.filterWidth,k=h.padInfo.top,I=h.padInfo.right,M=h.padInfo.bottom,D=h.padInfo.left,F=h.dilationHeight,E=h.dilationWidth,O=h.strideHeight,B=h.strideWidth,K=h.inChannels,X=h.outChannels,U=h.padInfo.type==="SAME"?1:0;if(h.dataFormat!=="channelsLast")throw new Error(`wasm backend Conv2D does not support dataFormat:'${h.dataFormat}'. Please use 'channelsLast'.`);const q=d.makeOutput(h.outShape,"float32"),Z=d.dataIdMap.get(q.dataId).id;return Ur(o,r.shape[0],r.shape[1],r.shape[2],p,b,y,k,I,M,D,U,F,E,O,B,K,X,Z),q}var qs={kernelName:i.Conv2D,backendName:"wasm",setupFunc:Vs,kernelFunc:Us},qr;function Ks(t){qr=t.wasm.cwrap(i.Conv2DBackpropInput,null,["number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number"])}function Xs(t){const{backend:a,inputs:u,attrs:d}=t,{dy:r,filter:n}=u,{strides:o,pad:p,dataFormat:l,dimRoundingMode:s,inputShape:e}=d,m=1,f=i.backend_util.convertConv2DDataFormat(l),g=i.backend_util.computeConv2DInfo(e,n.shape,o,m,p,s,!1,f),{batchSize:h,filterHeight:b,filterWidth:y,inChannels:k,inHeight:I,inWidth:M,outChannels:D,outHeight:F,outWidth:E,strideHeight:O,strideWidth:B}=g,K=b-1-g.padInfo.top,X=y-1-g.padInfo.left,U=g.dataFormat==="channelsLast",q=i.util.computeStrides(g.inShape),Z=i.util.computeStrides(r.shape),[ie,le,Q]=i.util.computeStrides(n.shape),j=q[0],ge=U?q[1]:q[2],$=U?q[2]:1,ce=U?1:q[1],me=Z[0],se=U?Z[1]:Z[2],ee=U?Z[2]:1,Te=U?1:Z[1],Fe=a.makeOutput(g.inShape,"float32"),Oe=a.dataIdMap.get(Fe.dataId).id,Je=a.dataIdMap.get(r.dataId).id,Qe=a.dataIdMap.get(n.dataId).id;return qr(Je,Qe,h,b,y,I,M,k,F,E,D,O,B,K,X,ie,le,Q,j,ge,$,ce,me,se,ee,Te,Oe),Fe}var Js={kernelName:i.Conv2DBackpropInput,backendName:"wasm",setupFunc:Ks,kernelFunc:Xs},Kr;function Qs(t){Kr=t.wasm.cwrap(i.Conv3D,null,["number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number"])}function Ys(t){const{inputs:a,backend:u,attrs:d}=t,{x:r,filter:n}=a,{strides:o,pad:p,dilations:l}=d;if(r.dtype!=="float32")throw new Error(`Tensor x must have dtype float32, got ${r.dtype}`);if(n.dtype!=="float32")throw new Error(`Tensor filter must have dtype float32, got ${n.dtype}`);const s=i.backend_util.computeConv3DInfo(r.shape,n.shape,o,l,p),e=u.makeOutput(s.outShape,r.dtype);return Kr(u.dataIdMap.get(r.dataId).id,u.dataIdMap.get(n.dataId).id,u.dataIdMap.get(e.dataId).id,s.batchSize,s.inDepth,s.inHeight,s.inWidth,s.inChannels,s.outDepth,s.outHeight,s.outWidth,s.outChannels,s.strideDepth,s.strideHeight,s.strideWidth,s.dilationDepth,s.dilationHeight,s.dilationWidth,s.filterDepth,s.filterHeight,s.filterWidth,s.padInfo.front,s.padInfo.top,s.padInfo.left),e}var Zs={kernelName:i.Conv3D,backendName:"wasm",setupFunc:Qs,kernelFunc:Ys},Xr;function js(t){Xr=t.wasm.cwrap(i.Conv3DBackpropFilterV2,null,["number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number"])}function ed(t){const{inputs:a,backend:u,attrs:d}=t,{x:r,dy:n}=a,{strides:o,pad:p,filterShape:l}=d;if(r.dtype!=="float32")throw new Error(`Tensor dy must have dtype float32, got ${r.dtype}`);if(n.dtype!=="float32")throw new Error(`Tensor filter must have dtype float32, got ${n.dtype}`);const s=i.backend_util.computeConv3DInfo(r.shape,l,o,1,p),e=u.makeOutput(s.filterShape,n.dtype);return Xr(u.dataIdMap.get(r.dataId).id,u.dataIdMap.get(n.dataId).id,u.dataIdMap.get(e.dataId).id,s.batchSize,s.inDepth,s.inHeight,s.inWidth,s.inChannels,s.outDepth,s.outHeight,s.outWidth,s.outChannels,s.strideDepth,s.strideHeight,s.strideWidth,s.dilationDepth,s.dilationHeight,s.dilationWidth,s.filterDepth,s.filterHeight,s.filterWidth,s.padInfo.front,s.padInfo.top,s.padInfo.left),e}var nd={kernelName:i.Conv3DBackpropFilterV2,backendName:"wasm",setupFunc:js,kernelFunc:ed},Jr;function td(t){Jr=t.wasm.cwrap(i.Conv3DBackpropInputV2,null,["number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number"])}function ad(t){const{inputs:a,backend:u,attrs:d}=t,{dy:r,filter:n}=a,{pad:o,strides:p,inputShape:l}=d;if(r.dtype!=="float32")throw new Error(`Tensor dy must have dtype float32, got ${r.dtype}`);if(n.dtype!=="float32")throw new Error(`Tensor filter must have dtype float32, got ${n.dtype}`);const s=i.backend_util.computeConv3DInfo(l,n.shape,p,1,o),e=u.makeOutput(s.inShape,r.dtype);return Jr(u.dataIdMap.get(n.dataId).id,u.dataIdMap.get(r.dataId).id,u.dataIdMap.get(e.dataId).id,s.batchSize,s.inDepth,s.inHeight,s.inWidth,s.inChannels,s.outDepth,s.outHeight,s.outWidth,s.outChannels,s.strideDepth,s.strideHeight,s.strideWidth,s.dilationDepth,s.dilationHeight,s.dilationWidth,s.filterDepth,s.filterHeight,s.filterWidth,s.padInfo.front,s.padInfo.top,s.padInfo.left),e}var rd={kernelName:i.Conv3DBackpropInputV2,backendName:"wasm",setupFunc:td,kernelFunc:ad},ud=W(i.Cos),id=W(i.Cosh),Un;(function(t){t[t.bilinear=0]="bilinear",t[t.nearest=1]="nearest"})(Un||(Un={}));var Qr;function sd(t){Qr=t.wasm.cwrap(i.CropAndResize,null,["number","number","number","number","array","number","number","number","number","number"])}function dd(t){const{backend:a,inputs:u,attrs:d}=t,{method:r,extrapolationValue:n,cropSize:o}=d,{image:p,boxes:l,boxInd:s}=u,e=l.shape[0],[m,f]=o,g=[e,m,f,p.shape[3]];let h=a.dataIdMap.get(p.dataId),b;p.dtype!=="float32"&&(b=Se({backend:a,inputs:{x:p},attrs:{dtype:"float32"}}),h=a.dataIdMap.get(b.dataId));const y=h.id,k=a.dataIdMap.get(l.dataId).id,I=a.dataIdMap.get(s.dataId).id,M=a.makeOutput(g,"float32"),D=a.dataIdMap.get(M.dataId).id,F=new Uint8Array(new Int32Array(p.shape).buffer);return Qr(y,k,I,e,F,m,f,Un[r],n,D),b!=null&&a.disposeData(b.dataId),M}var od={kernelName:i.CropAndResize,backendName:"wasm",setupFunc:sd,kernelFunc:dd},Yr;function pd(t){Yr=t.wasm.cwrap(i.Cumprod,null,["number","number","number","number","number","number"])}function ld(t){const{inputs:a,backend:u,attrs:d}=t,{x:r}=a,{axis:n,exclusive:o,reverse:p}=d,l=r.shape.length;i.util.assert(r.dtype==="float32"||r.dtype==="int32",()=>`cumprod does not support ${r.dtype} tensors in the WASM backend`);const s=i.backend_util.getAxesPermutation([n],l);let e=r;s!==null&&(e=Ie({inputs:{x:r},attrs:{perm:s},backend:u}));const m=i.backend_util.getInnerMostAxes(1,l)[0];i.backend_util.assertAxesAreInnerMostDims("cumprod",[m],l);const f=u.makeOutput(e.shape,e.dtype),g=e.shape[m],h=u.dataIdMap.get(e.dataId).id,b=u.dataIdMap.get(f.dataId).id;Yr(h,o?1:0,p?1:0,g,b,G[r.dtype]);let y=f;if(s!==null){const k=i.backend_util.getUndoAxesPermutation(s);y=Ie({inputs:{x:f},attrs:{perm:k},backend:u}),u.disposeData(e.dataId),u.disposeData(f.dataId)}return y}var cd={kernelName:i.Cumprod,backendName:"wasm",setupFunc:pd,kernelFunc:ld},Zr;function md(t){Zr=t.wasm.cwrap(i.Cumsum,null,["number","number","number","number","number","number"])}function hd(t){const{inputs:a,backend:u,attrs:d}=t,{x:r}=a,{axis:n,exclusive:o,reverse:p}=d,l=r.shape.length;i.util.assert(r.dtype==="float32"||r.dtype==="int32",()=>`cumsum does not support ${r.dtype} tensors in the WASM backend`);const s=i.backend_util.getAxesPermutation([n],l);let e=r;s!==null&&(e=Ie({inputs:{x:r},attrs:{perm:s},backend:u}));const m=i.backend_util.getInnerMostAxes(1,l)[0];i.backend_util.assertAxesAreInnerMostDims("cumsum",[m],l);const f=u.makeOutput(e.shape,e.dtype),g=e.shape[m],h=u.dataIdMap.get(e.dataId).id,b=u.dataIdMap.get(f.dataId).id;Zr(h,o?1:0,p?1:0,g,b,G[r.dtype]);let y=f;if(s!==null){const k=i.backend_util.getUndoAxesPermutation(s);y=Ie({inputs:{x:f},attrs:{perm:k},backend:u}),u.disposeData(e.dataId),u.disposeData(f.dataId)}return y}var gd={kernelName:i.Cumsum,backendName:"wasm",setupFunc:md,kernelFunc:hd},jr;function fd(t){jr=t.wasm.cwrap("DenseBincount",null,["number","array","number","number","boolean","number","number","boolean","number"])}function bd(t){const{backend:a,inputs:u,attrs:d}=t,{x:r,weights:n}=u,{size:o,binaryOutput:p}=d,l=n.shape.reduce((f,g)=>f*g,1)!==0,s=r.shape.length===1?[o]:[r.shape[0],o],e=a.makeOutput(s,n.dtype);function m(f){return a.dataIdMap.get(f.dataId).id}return jr(m(r),new Uint8Array(new Int32Array(r.shape).buffer),r.shape.length,o,l,m(n),G[n.dtype],p,m(e)),e}var vd={kernelName:i.DenseBincount,backendName:"wasm",setupFunc:fd,kernelFunc:bd},eu;function yd(t){eu=t.wasm.cwrap(i.DepthToSpace,null,["number","number","number","array","number","array","array","number","number"])}function Id(t){const{backend:a,inputs:u,attrs:d}=t,{x:r}=u,{blockSize:n,dataFormat:o}=d,p=r.shape[0],l=o==="NHWC"?r.shape[1]:r.shape[2],s=o==="NHWC"?r.shape[2]:r.shape[3],e=o==="NHWC"?r.shape[3]:r.shape[1],m=l*n,f=s*n,g=e/(n*n),h=o==="NHWC"?[p,m,f,g]:[p,g,m,f],b=a.makeOutput(h,"float32"),y=a.dataIdMap.get(r.dataId).id,k=new Uint8Array(new Int32Array(i.util.computeStrides(r.shape)).buffer),I=new Uint8Array(new Int32Array(h).buffer),M=new Uint8Array(new Int32Array(i.util.computeStrides(h)).buffer),D=a.dataIdMap.get(b.dataId).id;return eu(y,n,o==="NHWC"?1:0,k,r.shape.length-1,I,M,h.length,D),b}var kd={kernelName:i.DepthToSpace,backendName:"wasm",setupFunc:yd,kernelFunc:Id},nu;function Sd(t){nu=t.wasm.cwrap(i.DepthwiseConv2dNative,null,["number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number"])}function Md(t){const{inputs:a,attrs:u,backend:d}=t,{x:r,filter:n}=a,o=d.dataIdMap.get(r.dataId).id,p=d.dataIdMap.get(n.dataId).id,{strides:l,dilations:s,pad:e,dimRoundingMode:m}=u,f=s??[1,1],g=i.backend_util.computeConv2DInfo(r.shape,n.shape,l,f,e,m,!0),h=g.filterHeight,b=g.filterWidth,y=g.padInfo.top,k=g.padInfo.right,I=g.padInfo.bottom,M=g.padInfo.left,D=g.dilationHeight,F=g.dilationWidth,E=g.strideHeight,O=g.strideWidth,B=g.inChannels,K=g.outChannels,X=g.padInfo.type==="SAME"?1:0;if(g.dataFormat!=="channelsLast")throw new Error(`wasm backend DepthwiseConv2dNative does not support dataFormat:'${g.dataFormat}'. Please use 'channelsLast'.`);const U=d.makeOutput(g.outShape,"float32"),q=d.dataIdMap.get(U.dataId).id;return nu(o,r.shape[0],r.shape[1],r.shape[2],p,h,b,y,k,I,M,X,D,F,E,O,B,K,q),U}var Ad={kernelName:i.DepthwiseConv2dNative,backendName:"wasm",setupFunc:Sd,kernelFunc:Md},tu;function xd(t){tu=t.wasm.cwrap("Diag",null,["number","number","number","number"])}function Dd(t){const{inputs:a,backend:u}=t,{x:d}=a,r=i.util.sizeFromShape(d.shape),n=u.makeOutput([...d.shape,...d.shape],d.dtype);return tu(u.dataIdMap.get(d.dataId).id,G[d.dtype],r,u.dataIdMap.get(n.dataId).id),n}var wd={kernelName:i.Diag,backendName:"wasm",setupFunc:xd,kernelFunc:Dd},au;function Fd(t){au=t.wasm.cwrap(i.Dilation2D,null,["number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number"])}function Nd(t){const{inputs:a,backend:u,attrs:d}=t,{x:r,filter:n}=a,{strides:o,pad:p,dilations:l}=d;if(r.dtype!==n.dtype)throw new Error(`Dilation2D error: x must have the same dtype as filter. Got ${r.dtype} and ${n.dtype}`);const s=i.backend_util.computeDilation2DInfo(r.shape,n.shape,o,p,"NHWC",l),e=u.makeOutput(s.outShape,r.dtype);return au(u.dataIdMap.get(r.dataId).id,u.dataIdMap.get(n.dataId).id,u.dataIdMap.get(e.dataId).id,G[r.dtype],s.batchSize,s.inChannels,s.inHeight,s.inWidth,s.outHeight,s.outWidth,s.strideHeight,s.strideWidth,s.dilationHeight,s.dilationWidth,s.filterHeight,s.filterWidth,s.padInfo.top,s.padInfo.left),e}var Cd={kernelName:i.Dilation2D,backendName:"wasm",setupFunc:Fd,kernelFunc:Nd},ru;function Rd(t){ru=t.wasm.cwrap(i.Dilation2DBackpropFilter,null,["number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number"])}function _d(t){const{inputs:a,backend:u,attrs:d}=t,{x:r,filter:n,dy:o}=a,{strides:p,pad:l,dilations:s}=d;if(r.dtype!==n.dtype||r.dtype!==o.dtype)throw new Error(`Dilation2DBackpropFilter error: x must have the same dtype as filter and dy. Got ${r.dtype}, ${n.dtype}, and ${o.dtype}`);const e=i.backend_util.computeDilation2DInfo(r.shape,n.shape,p,l,"NHWC",s),m=u.makeOutput(n.shape,n.dtype);return ru(u.dataIdMap.get(r.dataId).id,u.dataIdMap.get(n.dataId).id,u.dataIdMap.get(o.dataId).id,u.dataIdMap.get(m.dataId).id,G[r.dtype],e.batchSize,e.inChannels,e.inHeight,e.inWidth,e.outHeight,e.outWidth,e.strideHeight,e.strideWidth,e.dilationHeight,e.dilationWidth,e.filterHeight,e.filterWidth,e.padInfo.top,e.padInfo.left),m}var Pd={kernelName:i.Dilation2DBackpropFilter,backendName:"wasm",setupFunc:Rd,kernelFunc:_d},uu;function Ed(t){uu=t.wasm.cwrap(i.Dilation2DBackpropInput,null,["number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number"])}function Bd(t){const{inputs:a,backend:u,attrs:d}=t,{x:r,filter:n,dy:o}=a,{strides:p,pad:l,dilations:s}=d;if(r.dtype!==n.dtype||r.dtype!==o.dtype)throw new Error(`Dilation2DBackpropInput error: x must have the same dtype as filter and dy. Got ${r.dtype}, ${n.dtype}, and ${o.dtype}`);const e=i.backend_util.computeDilation2DInfo(r.shape,n.shape,p,l,"NHWC",s),m=u.makeOutput(r.shape,r.dtype);return uu(u.dataIdMap.get(r.dataId).id,u.dataIdMap.get(n.dataId).id,u.dataIdMap.get(o.dataId).id,u.dataIdMap.get(m.dataId).id,G[r.dtype],e.batchSize,e.inChannels,e.inHeight,e.inWidth,e.outHeight,e.outWidth,e.strideHeight,e.strideWidth,e.dilationHeight,e.dilationWidth,e.filterHeight,e.filterWidth,e.padInfo.top,e.padInfo.left),m}var Td={kernelName:i.Dilation2DBackpropInput,backendName:"wasm",setupFunc:Ed,kernelFunc:Bd},Od=W(i.Elu),iu;function Wd(t){iu=t.wasm.cwrap(i.EluGrad,null,["number","number","number"])}function zd(t){const{inputs:a,backend:u}=t,{dy:d,y:r}=a,n=u.makeOutput(r.shape,"float32"),o=p=>u.dataIdMap.get(p.dataId).id;return iu(o(r),o(d),o(n)),n}var Hd={kernelName:i.EluGrad,backendName:"wasm",setupFunc:Wd,kernelFunc:zd},Ld=!1,Gd=Y(i.Equal,Ld,"bool"),$d=W(i.Erf),Vd=W(i.Exp,"float32");function qn(t){const{inputs:a,attrs:u,backend:d}=t,{input:r}=a,{dim:n}=u,o=r.shape.length,p=r.shape.slice();let l=n;return n<0&&(i.util.assert(-(o+1)<=n,()=>`Axis must be in the interval [${-(o+1)}, ${o}]`),l=o+n+1),p.splice(l,0,1),ue({inputs:{x:r},backend:d,attrs:{shape:p}})}var Ud={kernelName:i.ExpandDims,backendName:"wasm",kernelFunc:qn},qd=W(i.Expm1,"float32");function su(t){const{attrs:{shape:a,value:u},backend:d}=t;let{attrs:{dtype:r}}=t;r=r||i.util.inferDtype(u);const n=d.makeOutput(a,r);return d.typedArrayFromHeap(n).fill(u),n}var Kd={kernelName:i.Fill,backendName:"wasm",kernelFunc:su},du;function Xd(t){du=t.wasm.cwrap(i.FlipLeftRight,null,["number","number","number","number","number","number"])}function Jd(t){const{inputs:a,backend:u}=t,{image:d}=a,r=u.makeOutput(d.shape,d.dtype),n=u.dataIdMap.get(d.dataId).id,o=u.dataIdMap.get(r.dataId).id,[p,l,s,e]=d.shape;return du(n,p,l,s,e,o),r}var Qd={kernelName:i.FlipLeftRight,backendName:"wasm",kernelFunc:Jd,setupFunc:Xd},Yd=W(i.Floor),Zd=!1,jd=Y(i.FloorDiv,Zd),ou;function eo(t){ou=t.wasm.cwrap(i.FusedBatchNorm,null,["number","number","number","number","number","number","number"])}function no(t){const{backend:a,inputs:u,attrs:d}=t,{varianceEpsilon:r}=d,{x:n,mean:o,variance:p,offset:l,scale:s}=u,e=a.dataIdMap.get(n.dataId).id,m=a.dataIdMap.get(o.dataId).id,f=a.dataIdMap.get(p.dataId).id,g=l!=null?a.dataIdMap.get(l.dataId).id:0,h=s!=null?a.dataIdMap.get(s.dataId).id:0,b=a.makeOutput(n.shape,n.dtype);if(i.util.sizeFromShape(n.shape)===0)return b;const y=a.dataIdMap.get(b.dataId).id;return ou(e,m,f,g,h,r,y),b}var to={kernelName:i.FusedBatchNorm,backendName:"wasm",setupFunc:eo,kernelFunc:no},pu;function ao(t){pu=t.wasm.cwrap(i.FusedConv2D,null,["number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number"])}function ro(t){const{inputs:a,attrs:u,backend:d}=t,{x:r,filter:n,bias:o,preluActivationWeights:p}=a,{strides:l,pad:s,dilations:e,dataFormat:m,dimRoundingMode:f,activation:g,leakyreluAlpha:h}=u,b=i.backend_util.computeConv2DInfo(r.shape,n.shape,l,e,s,f),y=qe[g];if(y==null)throw new Error(`${g} activation not yet supported for FusedConv2D in the wasm backend.`);const k=d.dataIdMap.get(r.dataId).id,I=d.dataIdMap.get(n.dataId).id,M=b.outChannels;let D=0;if(o!=null){const ee=d.dataIdMap.get(o.dataId);if(ee.shape.length!==1)throw new Error(`FusedConv2D only supports rank-1 bias but got rank ${ee.shape.length}.`);if(ee.shape[0]!==M)throw new Error(`FusedConv2D bias shape (${ee.shape}) does not match the number of output channels (${M})`);D=ee.id}const F=b.filterHeight,E=b.filterWidth,O=b.padInfo.top,B=b.padInfo.right,K=b.padInfo.bottom,X=b.padInfo.left,U=b.dilationHeight,q=b.dilationWidth,Z=b.strideHeight,ie=b.strideWidth,le=b.inChannels,Q=b.padInfo.type==="SAME"?1:0,j=b.batchSize,ge=b.inHeight,$=b.inWidth;if(m!=="NHWC")throw new Error(`wasm backend FusedConv2D does not support dataFormat:'${m}'. Please use 'NHWC'.`);const ce=d.makeOutput(b.outShape,"float32"),me=d.dataIdMap.get(ce.dataId).id,se=p==null?0:d.dataIdMap.get(p.dataId).id;return pu(k,j,ge,$,I,F,E,D,O,B,K,X,Q,U,q,Z,ie,le,M,y,se,h||0,me),ce}var uo={kernelName:i.FusedConv2D,backendName:"wasm",setupFunc:ao,kernelFunc:ro},lu;function io(t){lu=t.wasm.cwrap(i.FusedDepthwiseConv2D,null,["number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number"])}function so(t){const{inputs:a,attrs:u,backend:d}=t,{x:r,filter:n,bias:o,preluActivationWeights:p}=a,{strides:l,pad:s,dilations:e,dataFormat:m,dimRoundingMode:f,activation:g,leakyreluAlpha:h}=u,b=i.backend_util.computeConv2DInfo(r.shape,n.shape,l,e,s,f,!0),y=qe[g];if(y==null)throw new Error(`${g} activation not yet supported for FusedDepthwiseConv2D in the wasm backend.`);const k=d.dataIdMap.get(r.dataId).id,I=d.dataIdMap.get(n.dataId).id,M=b.outChannels;let D=0;if(o!=null){const ee=d.dataIdMap.get(o.dataId);if(ee.shape.length!==1)throw new Error(`FusedDepthwiseConv2D only supports rank-1 bias but got rank ${ee.shape.length}.`);if(ee.shape[0]!==M)throw new Error(`FusedDepthwiseConv2D bias shape (${ee.shape}) does not match the number of output channels (${M})`);D=ee.id}const F=b.filterHeight,E=b.filterWidth,O=b.padInfo.top,B=b.padInfo.right,K=b.padInfo.bottom,X=b.padInfo.left,U=b.dilationHeight,q=b.dilationWidth,Z=b.strideHeight,ie=b.strideWidth,le=b.inChannels,Q=b.padInfo.type==="SAME"?1:0,j=b.batchSize,ge=b.inHeight,$=b.inWidth;if(m!=="NHWC")throw new Error(`wasm backend FusedDepthwiseConv2D does not support dataFormat:'${m}'. Please use 'NHWC'.`);const ce=d.makeOutput(b.outShape,"float32"),me=d.dataIdMap.get(ce.dataId).id,se=p==null?0:d.dataIdMap.get(p.dataId).id;return lu(k,j,ge,$,I,F,E,D,O,B,K,X,Q,U,q,Z,ie,le,M,y,se,h||0,me),ce}var oo={kernelName:i.FusedDepthwiseConv2D,backendName:"wasm",setupFunc:io,kernelFunc:so},cu;function po(t){cu=t.wasm.cwrap(i.GatherNd,null,["number","number","number","number","number","number","array","number"])}function lo(t){const{backend:a,inputs:u}=t,{params:d,indices:r}=u,[n,o,p,l]=i.gather_util.prepareAndValidate(d,r),s=a.makeOutput(n,d.dtype);if(o===0)return s;const e=r.shape,m=e[e.length-1],f=a.dataIdMap.get(d.dataId).id,g=a.dataIdMap.get(r.dataId).id,h=new Uint8Array(new Int32Array(l).buffer),b=a.dataIdMap.get(s.dataId).id;return cu(f,G[d.dtype],g,o,m,p,h,b),s}var co={kernelName:i.GatherNd,backendName:"wasm",setupFunc:po,kernelFunc:lo},mu;function mo(t){mu=t.wasm.cwrap("Gather",null,["number","number","array","number","number","number","array","number"])}function ho(t){const{backend:a,inputs:u,attrs:d}=t,{x:r,indices:n}=u,{axis:o,batchDims:p}=d,l=i.util.parseAxisParam(o,r.shape)[0],s=a.readSync(n.dataId),e=r.shape[l];for(let O=0;O<s.length;++O){const B=s[O];i.util.assert(B<=e-1&&B>=0,()=>`GatherV2: the index value ${B} is not in [0, ${e-1}]`)}const m=i.backend_util.segment_util.collectGatherOpShapeInfo(r,n,l,p),f=ue({inputs:{x:r},attrs:{shape:[m.batchSize,m.outerSize,m.dimSize,m.sliceSize]},backend:a}),g=i.util.sizeFromShape(n.shape),h=ue({inputs:{x:n},attrs:{shape:[m.batchSize,g/m.batchSize]},backend:a}),b=[m.batchSize,m.outerSize,g/m.batchSize,m.sliceSize],y=a.makeOutput(b,r.dtype);if(i.util.sizeFromShape(r.shape)===0)return y;const k=f.shape.length-1,I=a.dataIdMap.get(f.dataId).id,M=a.dataIdMap.get(h.dataId).id,D=a.dataIdMap.get(y.dataId).id,F=new Uint8Array(new Int32Array(i.util.computeStrides(f.shape)).buffer),E=new Uint8Array(new Int32Array(i.util.computeStrides(b)).buffer);return mu(I,G[r.dtype],F,k,M,m.batchSize,E,D),a.disposeData(f.dataId),a.disposeData(h.dataId),y.shape=m.outputShape,y}var go={kernelName:i.GatherV2,backendName:"wasm",setupFunc:mo,kernelFunc:ho},fo=!1,bo=Y(i.Greater,fo,"bool"),vo=!1,yo=Y(i.GreaterEqual,vo,"bool"),Io=W(i.IsFinite,"bool"),ko=W(i.IsInf,"bool"),So=W(i.IsNan,"bool"),hu;function Mo(t){hu=t.wasm.cwrap(i.LeakyRelu,null,["number","number","number","number"])}function Ao(t){const{inputs:{x:a},attrs:{alpha:u},backend:d}=t,r=d.dataIdMap.get(a.dataId).id,n=d.makeOutput(a.shape,"float32");if(i.util.sizeFromShape(a.shape)!==0){const o=d.dataIdMap.get(n.dataId).id;hu(r,G[a.dtype],u,o)}return n}var xo={kernelName:i.LeakyRelu,backendName:"wasm",setupFunc:Mo,kernelFunc:Ao},Do=!1,wo=Y(i.Less,Do,"bool"),Fo=!1,No=Y(i.LessEqual,Fo,"bool"),gu;function Co(t){gu=t.wasm.cwrap(i.LinSpace,null,["number","number","number","number"])}function Ro(t){const{attrs:a,backend:u}=t,{start:d,stop:r,num:n}=a,o=Math.floor(n),p=u.makeOutput([o],"float32");return gu(u.dataIdMap.get(p.dataId).id,d,r,o),p}var _o={kernelName:i.LinSpace,backendName:"wasm",setupFunc:Co,kernelFunc:Ro},Po=W(i.Log),Eo=W(i.Log1p),Bo=!1,To=Y(i.LogicalAnd,Bo,"bool"),Oo=W(i.LogicalNot),Wo=!1,zo=Y(i.LogicalOr,Wo,"bool"),Ho=!1,Lo=Y(i.LogicalXor,Ho,"bool"),fu;function Go(t){fu=t.wasm.cwrap(i.LRN,null,["number","number","number","number","number","number","number"])}function $o(t){const{inputs:a,backend:u,attrs:d}=t,{x:r}=a,{depthRadius:n,bias:o,alpha:p,beta:l}=d;if(r.dtype!=="float32")throw new Error("LRN error: x must have dtype float32");const s=u.makeOutput(r.shape,r.dtype);return fu(u.dataIdMap.get(r.dataId).id,u.dataIdMap.get(s.dataId).id,r.shape[3],n,o,p,l),s}var Vo={kernelName:i.LRN,backendName:"wasm",setupFunc:Go,kernelFunc:$o},bu;function Uo(t){bu=t.wasm.cwrap(i.LRNGrad,null,["number","number","number","number","number","number","number","number","number"])}function qo(t){const{inputs:a,backend:u,attrs:d}=t,{x:r,y:n,dy:o}=a,{depthRadius:p,bias:l,alpha:s,beta:e}=d;if(r.dtype!=="float32"||n.dtype!=="float32"||o.dtype!=="float32")throw new Error("LRNGrad error: x, y, and dy must have dtype float32");const m=u.makeOutput(r.shape,r.dtype);return bu(u.dataIdMap.get(r.dataId).id,u.dataIdMap.get(n.dataId).id,u.dataIdMap.get(o.dataId).id,u.dataIdMap.get(m.dataId).id,o.shape[3],p,l,s,e),m}var Ko={kernelName:i.LRNGrad,backendName:"wasm",setupFunc:Uo,kernelFunc:qo},vu;function Xo(t){vu=t.wasm.cwrap(i.Max,null,["number","number","number","number"])}function Jo(t){const{backend:a,inputs:u,attrs:d}=t,{reductionIndices:r,keepDims:n}=d,{x:o}=u;let p=a.dataIdMap.get(o.dataId).id,l=o;const{transposed:s,axes:e,originalAxes:m,inputWasTransposed:f}=ke(o,r,a);if(f){const I=a.dataIdMap.get(s.dataId).id;l=s,p=I}const g=l.shape.length;i.backend_util.assertAxesAreInnerMostDims("max",e,g);const[h,b]=i.backend_util.computeOutAndReduceShapes(l.shape,e),y=i.util.sizeFromShape(b),k=a.makeOutput(h,o.dtype);if(i.util.sizeFromShape(l.shape)!==0){const I=a.dataIdMap.get(k.dataId).id;vu(p,G[o.dtype],y,I)}return f&&a.disposeData(s.dataId),n&&(k.shape=i.backend_util.expandShapeToKeepDim(k.shape,m)),k}var Qo={kernelName:i.Max,backendName:"wasm",setupFunc:Xo,kernelFunc:Jo},Yo=!1,Zo=Y(i.Maximum,Yo),yu;function jo(t){yu=t.wasm.cwrap(i.MaxPool,null,["number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number"])}function ep(t){const{inputs:a,attrs:u,backend:d}=t,r=a.x,n=d.dataIdMap.get(r.dataId).id;i.util.assert(r.dtype==="float32",()=>`Error in MaxPool: only float32 input is supported. Got ${r.dtype}.`);const{filterSize:o,strides:p,pad:l,dimRoundingMode:s}=u,e=i.backend_util.computePool2DInfo(r.shape,o,p,1,l,s),m=e.filterHeight,f=e.filterWidth,g=e.padInfo.top,h=e.padInfo.right,b=e.padInfo.bottom,y=e.padInfo.left,k=e.dilationHeight,I=e.dilationWidth,M=e.strideHeight,D=e.strideWidth,F=e.inChannels,E=e.outChannels;if(e.dataFormat!=="channelsLast")throw new Error(`wasm backend does not support dataFormat:'${e.dataFormat}'. Please use 'channelsLast'.`);const O=d.makeOutput(e.outShape,"float32"),B=d.dataIdMap.get(O.dataId).id;return yu(n,r.shape[0],r.shape[1],r.shape[2],m,f,g,h,b,y,k,I,M,D,F,E,B),O}var np={kernelName:i.MaxPool,backendName:"wasm",setupFunc:jo,kernelFunc:ep},Iu;function tp(t){Iu=t.wasm.cwrap("MaxPool3D",null,["number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number"])}function ap(t){const{inputs:a,backend:u,attrs:d}=t,{x:r}=a,{filterSize:n,strides:o,pad:p,dimRoundingMode:l,dataFormat:s}=d,e=i.backend_util.computePool3DInfo(r.shape,n,o,1,p,l,s),m=u.makeOutput(e.outShape,r.dtype);return Iu(u.dataIdMap.get(r.dataId).id,u.dataIdMap.get(m.dataId).id,e.batchSize,e.inChannels,e.inDepth,e.inHeight,e.inWidth,e.outDepth,e.outHeight,e.outWidth,e.strideDepth,e.strideHeight,e.strideWidth,e.dilationDepth,e.dilationHeight,e.dilationWidth,e.effectiveFilterDepth,e.effectiveFilterHeight,e.effectiveFilterWidth,e.padInfo.front,e.padInfo.top,e.padInfo.left),m}var rp={kernelName:i.MaxPool3D,backendName:"wasm",setupFunc:tp,kernelFunc:ap},ku;function up(t){ku=t.wasm.cwrap("MaxPool3DGrad",null,["number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number"])}function ip(t){const{inputs:a,backend:u,attrs:d}=t,{dy:r,input:n}=a,{filterSize:o,strides:p,pad:l,dimRoundingMode:s}=d,e=i.backend_util.computePool3DInfo(n.shape,o,p,1,l,s),m=u.makeOutput(n.shape,n.dtype);return ku(u.dataIdMap.get(n.dataId).id,u.dataIdMap.get(r.dataId).id,u.dataIdMap.get(m.dataId).id,e.batchSize,e.inChannels,e.inDepth,e.inHeight,e.inWidth,e.outDepth,e.outHeight,e.outWidth,e.strideDepth,e.strideHeight,e.strideWidth,e.dilationDepth,e.dilationHeight,e.dilationWidth,e.effectiveFilterDepth,e.effectiveFilterHeight,e.effectiveFilterWidth,e.padInfo.front,e.padInfo.top,e.padInfo.left),m}var sp={kernelName:i.MaxPool3DGrad,backendName:"wasm",setupFunc:up,kernelFunc:ip},Su;function dp(t){Su=t.wasm.cwrap("MaxPoolGrad",null,["number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number"])}function op(t){const{inputs:a,backend:u,attrs:d}=t,{dy:r,input:n}=a,{filterSize:o,strides:p,pad:l,dimRoundingMode:s}=d,e=i.backend_util.computePool2DInfo(n.shape,o,p,1,l,s),m=u.makeOutput(n.shape,n.dtype);return Su(u.dataIdMap.get(n.dataId).id,u.dataIdMap.get(r.dataId).id,u.dataIdMap.get(m.dataId).id,e.batchSize,e.inChannels,e.inHeight,e.inWidth,e.outHeight,e.outWidth,e.strideHeight,e.strideWidth,e.dilationHeight,e.dilationWidth,e.effectiveFilterHeight,e.effectiveFilterWidth,e.padInfo.top,e.padInfo.left),m}var pp={kernelName:i.MaxPoolGrad,backendName:"wasm",setupFunc:dp,kernelFunc:op},Mu;function lp(t){Mu=t.wasm.cwrap("MaxPoolWithArgmax",null,["number","number","number","number","boolean","number","number","number","number","number","number","number","number","number","number","number","number","number","number"])}function cp(t){const{inputs:a,backend:u,attrs:d}=t,{x:r}=a,{filterSize:n,strides:o,pad:p,includeBatchInIndex:l}=d;i.util.assert(r.shape.length===4,()=>`Error in maxPool: input must be rank 4 but got rank ${r.shape.length}.`);const s=[1,1];i.util.assert(i.backend_util.eitherStridesOrDilationsAreOne(o,s),()=>`Error in maxPool: Either strides or dilations must be 1. Got strides ${o} and dilations '${s}'`);const e=i.backend_util.computePool2DInfo(r.shape,n,o,[1,1],p),m=u.makeOutput(e.outShape,r.dtype),f=u.makeOutput(e.outShape,"int32");return Mu(u.dataIdMap.get(r.dataId).id,u.dataIdMap.get(m.dataId).id,u.dataIdMap.get(f.dataId).id,G[r.dtype],l,e.batchSize,e.inChannels,e.inHeight,e.inWidth,e.outHeight,e.outWidth,e.strideHeight,e.strideWidth,e.dilationHeight,e.dilationWidth,e.effectiveFilterHeight,e.effectiveFilterWidth,e.padInfo.top,e.padInfo.left),[m,f]}var mp={kernelName:i.MaxPoolWithArgmax,backendName:"wasm",setupFunc:lp,kernelFunc:cp},Au;function hp(t){Au=t.wasm.cwrap(i.Mean,null,["number, number, number"])}function gp(t){const{backend:a,inputs:u,attrs:d}=t,{axis:r,keepDims:n}=d,{x:o}=u,p=a.dataIdMap.get(o.dataId).id;let l=p,s=o;const{transposed:e,axes:m,originalAxes:f,inputWasTransposed:g}=ke(o,r,a);let h=m;if(g){const D=a.dataIdMap.get(e.dataId).id;D!==p&&(s=e,l=D,h=i.backend_util.getInnerMostAxes(h.length,s.shape.length))}i.backend_util.assertAxesAreInnerMostDims("mean",h,s.shape.length);const[b,y]=i.backend_util.computeOutAndReduceShapes(s.shape,h),k=i.util.sizeFromShape(y);let I=s;s.dtype!=="float32"&&(I=Se({backend:a,inputs:{x:s},attrs:{dtype:"float32"}}),l=a.dataIdMap.get(I.dataId).id);const M=a.makeOutput(b,"float32");if(i.util.sizeFromShape(s.shape)!==0){const D=a.dataIdMap.get(M.dataId).id;Au(l,k,D)}return g&&a.disposeData(e.dataId),n&&(M.shape=i.backend_util.expandShapeToKeepDim(M.shape,f)),s.dtype!=="float32"&&a.disposeData(I.dataId),M}var fp={kernelName:i.Mean,backendName:"wasm",setupFunc:hp,kernelFunc:gp},xu;function bp(t){xu=t.wasm.cwrap(i.Min,null,["number","number","number","number"])}function vp(t){const{backend:a,inputs:u,attrs:d}=t,{axis:r,keepDims:n}=d,{x:o}=u,p=a.dataIdMap.get(o.dataId).id;let l=p,s=o;const{transposed:e,axes:m,originalAxes:f,inputWasTransposed:g}=ke(o,r,a);if(g){const M=a.dataIdMap.get(e.dataId).id;M!==p&&(s=e,l=M)}const h=s.shape.length;i.backend_util.assertAxesAreInnerMostDims("min",m,h);const[b,y]=i.backend_util.computeOutAndReduceShapes(s.shape,m),k=i.util.sizeFromShape(y),I=a.makeOutput(b,s.dtype);if(i.util.sizeFromShape(s.shape)!==0){const M=a.dataIdMap.get(I.dataId).id;xu(l,G[o.dtype],k,M)}return g&&a.disposeData(e.dataId),n&&(I.shape=i.backend_util.expandShapeToKeepDim(I.shape,f)),I}var yp={kernelName:i.Min,backendName:"wasm",setupFunc:bp,kernelFunc:vp},Ip=!1,kp=Y(i.Minimum,Ip),Kn;(function(t){t[t.reflect=0]="reflect",t[t.symmetric=1]="symmetric"})(Kn||(Kn={}));var Du;function Sp(t){Du=t.wasm.cwrap(i.MirrorPad,null,["number","array","number","number","array","array","number","number"])}function Mp(t){const{inputs:{x:a},backend:u,attrs:{paddings:d,mode:r}}=t,n=d.map((h,b)=>h[0]+a.shape[b]+h[1]),o=u.dataIdMap.get(a.dataId).id,p=u.makeOutput(n,a.dtype),l=u.dataIdMap.get(p.dataId).id,s=new Uint8Array(new Int32Array(a.shape).buffer),e=d.map(h=>h[0]),m=d.map(h=>h[1]),f=new Uint8Array(new Int32Array(e).buffer),g=new Uint8Array(new Int32Array(m).buffer);return Du(o,s,a.shape.length,G[a.dtype],f,g,Kn[r],l),p}var Ap={kernelName:i.MirrorPad,backendName:"wasm",kernelFunc:Mp,setupFunc:Sp},wu;function xp(t){wu=t.wasm.cwrap(i.Softmax,null,["number","number","number","number"])}function Fu(t){const{backend:a,inputs:{logits:u},attrs:{dim:d}}=t,r=a.dataIdMap.get(u.dataId).id,n=a.makeOutput(u.shape,u.dtype),o=a.dataIdMap.get(n.dataId).id,p=u.shape[d],l=i.util.sizeFromShape(u.shape)/p;return i.util.sizeFromShape(n.shape)===0||wu(r,o,p,l),n}var Dp={kernelName:i.Softmax,backendName:"wasm",setupFunc:xp,kernelFunc:Fu},Nu;function wp(t){Nu=t.wasm.cwrap(i.Multinomial,null,["number","number","number","number","number","number"])}function Fp(t){const{inputs:a,backend:u,attrs:d}=t,{logits:r}=a,{numSamples:n,seed:o,normalized:p}=d;if(r.dtype!=="float32")throw new Error(`Tensor logits must have dtype float32, got ${r.dtype}`);const l=p?r:Fu({inputs:{logits:r},backend:u,attrs:{dim:r.shape.length-1}}),[s,e]=l.shape,m=u.makeOutput([s,n],"int32");return Nu(u.dataIdMap.get(l.dataId).id,s,e,n,o,u.dataIdMap.get(m.dataId).id),p||u.disposeData(l.dataId),m}var Np={kernelName:i.Multinomial,backendName:"wasm",setupFunc:wp,kernelFunc:Fp},Cp=Y(i.Mod,!0),Rp=!0,_p=Y(i.Multiply,Rp),Pp=W(i.Neg);function Xn(t,a){const u=new Int32Array(t.wasm.HEAPU8.buffer,a,4),d=u[0],r=u[1],n=u[2],o=u[3];return t.wasm._free(a),{pSelectedIndices:d,selectedSize:r,pSelectedScores:n,pValidOutputs:o}}var Cu;function Ep(t){Cu=t.wasm.cwrap(i.NonMaxSuppressionV3,"number",["number","number","number","number","number"])}function Bp(t){const{backend:a,inputs:u,attrs:d}=t,{iouThreshold:r,maxOutputSize:n,scoreThreshold:o}=d,{boxes:p,scores:l}=u,s=a.dataIdMap.get(p.dataId).id,e=a.dataIdMap.get(l.dataId).id,m=Cu(s,e,n,r,o),{pSelectedIndices:f,selectedSize:g,pSelectedScores:h,pValidOutputs:b}=Xn(a,m);return a.wasm._free(h),a.wasm._free(b),a.makeOutput([g],"int32",f)}var Tp={kernelName:i.NonMaxSuppressionV3,backendName:"wasm",setupFunc:Ep,kernelFunc:Bp},Ru;function Op(t){Ru=t.wasm.cwrap(i.NonMaxSuppressionV4,"number",["number","number","number","number","number","bool"])}function Wp(t){const{backend:a,inputs:u,attrs:d}=t,{iouThreshold:r,maxOutputSize:n,scoreThreshold:o,padToMaxOutputSize:p}=d,{boxes:l,scores:s}=u,e=a.dataIdMap.get(l.dataId).id,m=a.dataIdMap.get(s.dataId).id,f=Ru(e,m,n,r,o,p),{pSelectedIndices:g,selectedSize:h,pSelectedScores:b,pValidOutputs:y}=Xn(a,f);return a.wasm._free(b),[a.makeOutput([h],"int32",g),a.makeOutput([],"int32",y)]}var zp={kernelName:i.NonMaxSuppressionV4,backendName:"wasm",setupFunc:Op,kernelFunc:Wp},_u;function Hp(t){_u=t.wasm.cwrap(i.NonMaxSuppressionV5,"number",["number","number","number","number","number","number"])}function Lp(t){const{backend:a,inputs:u,attrs:d}=t,{iouThreshold:r,maxOutputSize:n,scoreThreshold:o,softNmsSigma:p}=d,{boxes:l,scores:s}=u,e=a.dataIdMap.get(l.dataId).id,m=a.dataIdMap.get(s.dataId).id,f=_u(e,m,n,r,o,p),{pSelectedIndices:g,selectedSize:h,pSelectedScores:b,pValidOutputs:y}=Xn(a,f);return a.wasm._free(y),[a.makeOutput([h],"int32",g),a.makeOutput([h],"float32",b)]}var Gp={kernelName:i.NonMaxSuppressionV5,backendName:"wasm",setupFunc:Hp,kernelFunc:Lp},$p=!1,Vp=Y(i.NotEqual,$p,"bool"),Pu;function Up(t){Pu=t.wasm.cwrap(i.OneHot,null,["number","number","number","number","number"])}function qp(t){const{inputs:a,backend:u,attrs:d}=t,{indices:r}=a,{dtype:n,depth:o,onValue:p,offValue:l}=d,s=u.makeOutput([...r.shape,o],n),e=u.dataIdMap.get(s.dataId).id,m=u.dataIdMap.get(r.dataId).id;return Pu(m,o,p,l,e),s}var Kp={kernelName:i.OneHot,backendName:"wasm",setupFunc:Up,kernelFunc:qp};function Xp(t){const{inputs:{x:a},backend:u}=t,d=u.makeOutput(a.shape,a.dtype);return u.typedArrayFromHeap(d).fill(1),d}var Jp={kernelName:i.OnesLike,backendName:"wasm",kernelFunc:Xp};function Qp(t){const{inputs:a,backend:u,attrs:d}=t,{axis:r}=d;if(a.length===1)return qn({inputs:{input:a[0]},backend:u,attrs:{dim:r}});const n=a[0].shape,o=a[0].dtype;a.forEach(e=>{i.util.assertShapesMatch(n,e.shape,"All tensors passed to stack must have matching shapes"),i.util.assert(o===e.dtype,()=>"All tensors passed to stack must have matching dtypes")});const p=[],l=a.map(e=>{const m=qn({inputs:{input:e},backend:u,attrs:{dim:r}});return p.push(m),m}),s=Vr({inputs:l,backend:u,attrs:{axis:r}});return p.forEach(e=>u.disposeData(e.dataId)),s}var Yp={kernelName:i.Pack,backendName:"wasm",kernelFunc:Qp},Eu;function Zp(t){Eu=t.wasm.cwrap(i.PadV2,null,["number","array","number","number","array","array","number","number"])}function jp(t){const{inputs:{x:a},backend:u,attrs:{paddings:d,constantValue:r}}=t,n=d.map((h,b)=>h[0]+a.shape[b]+h[1]);if(i.util.sizeFromShape(a.shape)===0)return su({backend:u,attrs:{shape:n,value:r,dtype:a.dtype}});const o=u.dataIdMap.get(a.dataId).id,p=u.makeOutput(n,a.dtype),l=u.dataIdMap.get(p.dataId).id,s=new Uint8Array(new Int32Array(a.shape).buffer),e=d.map(h=>h[0]),m=d.map(h=>h[1]),f=new Uint8Array(new Int32Array(e).buffer),g=new Uint8Array(new Int32Array(m).buffer);return Eu(o,s,a.shape.length,G[a.dtype],f,g,r,l),p}var Bu={kernelName:i.PadV2,backendName:"wasm",kernelFunc:jp,setupFunc:Zp},el=!1,nl=Y(i.Pow,el),Tu;function tl(t){Tu=t.wasm.cwrap(i.Prelu,null,["number","number","number"])}function al(t){const{inputs:a,backend:u}=t,{x:d,alpha:r}=a,n=u.dataIdMap.get(d.dataId).id,o=u.dataIdMap.get(r.dataId).id;let p=n;const l=d;let s=l;l.dtype!=="float32"&&(s=Se({backend:u,inputs:{x:d},attrs:{dtype:"float32"}}),p=u.dataIdMap.get(s.dataId).id);const e=u.makeOutput(d.shape,"float32"),m=u.dataIdMap.get(e.dataId).id;return Tu(p,o,m),l.dtype!=="float32"&&u.disposeData(s.dataId),e}var rl={kernelName:i.Prelu,backendName:"wasm",setupFunc:tl,kernelFunc:al},Ou;function ul(t){Ou=t.wasm.cwrap(i.Prod,null,["number","number","number","number"])}function il(t){const{backend:a,inputs:u,attrs:d}=t,{axis:r,keepDims:n}=d,{x:o}=u,p=a.dataIdMap.get(o.dataId).id;let l=p,s=o;const{transposed:e,axes:m,originalAxes:f,inputWasTransposed:g}=ke(o,r,a);let h=m;if(g){const M=a.dataIdMap.get(e.dataId).id;M!==p&&(s=e,l=M,h=i.backend_util.getInnerMostAxes(h.length,s.shape.length))}i.backend_util.assertAxesAreInnerMostDims("prod",h,s.shape.length);const[b,y]=i.backend_util.computeOutAndReduceShapes(s.shape,h),k=i.util.sizeFromShape(y),I=a.makeOutput(b,s.dtype);if(i.util.sizeFromShape(s.shape)!==0){const M=a.dataIdMap.get(I.dataId).id;Ou(l,k,G[I.dtype],M)}return g&&a.disposeData(e.dataId),n&&(I.shape=i.backend_util.expandShapeToKeepDim(I.shape,f)),I}var sl={kernelName:i.Prod,backendName:"wasm",setupFunc:ul,kernelFunc:il},dl=t=>{const{backend:a,attrs:u}=t,{start:d,stop:r,step:n,dtype:o}=u,p=vs(d,r,n,o),l=a.makeOutput([p.length],o);return a.typedArrayFromHeap(l).set(p),l},ol={kernelName:i.Range,backendName:"wasm",kernelFunc:dl},pl=!0,ll=Y(i.RealDiv,pl),cl=W(i.Reciprocal),ml=W(i.Relu),hl=W(i.Relu6),Wu;function gl(t){Wu=t.wasm.cwrap(i.ResizeBilinear,null,["number","number","number","number","number","number","number","number","number","number"])}function fl(t){const{backend:a,inputs:u,attrs:d}=t,{images:r}=u,{alignCorners:n,halfPixelCenters:o,size:p}=d,[l,s]=p,[e,m,f,g]=r.shape,h=[e,l,s,g];let b=a.dataIdMap.get(r.dataId),y;b.dtype!=="float32"&&(y=Se({backend:a,inputs:{x:r},attrs:{dtype:"float32"}}),b=a.dataIdMap.get(y.dataId));const k=b.id,I=a.makeOutput(h,"float32");if(i.util.sizeFromShape(r.shape)===0)return I;const M=a.dataIdMap.get(I.dataId).id;return Wu(k,e,m,f,g,l,s,n?1:0,o?1:0,M),y!=null&&a.disposeData(y.dataId),I}var bl={kernelName:i.ResizeBilinear,backendName:"wasm",setupFunc:gl,kernelFunc:fl},zu;function vl(t){zu=t.wasm.cwrap(i.ResizeBilinearGrad,null,["number","number","number","array","array","boolean"])}function yl(t){const{inputs:a,backend:u,attrs:d}=t,{images:r,dy:n}=a,{alignCorners:o}=d,p=u.makeOutput(r.shape,"float32");let l=u.dataIdMap.get(r.dataId),s;return l.dtype!=="float32"&&(s=Se({backend:u,inputs:{x:r},attrs:{dtype:"float32"}}),l=u.dataIdMap.get(s.dataId)),zu(u.dataIdMap.get(r.dataId).id,u.dataIdMap.get(n.dataId).id,u.dataIdMap.get(p.dataId).id,new Uint8Array(new Int32Array(r.shape).buffer),new Uint8Array(new Int32Array(n.shape).buffer),o),s!=null&&u.disposeData(s.dataId),p}var Il={kernelName:i.ResizeBilinearGrad,backendName:"wasm",setupFunc:vl,kernelFunc:yl},Hu;function kl(t){Hu=t.wasm.cwrap(i.ResizeNearestNeighbor,null,["number","number","number","number","number","number","number","number","number","number"])}function Sl(t){const{backend:a,inputs:u,attrs:d}=t,{images:r}=u,{alignCorners:n,halfPixelCenters:o,size:p}=d,[l,s]=p,[e,m,f,g]=r.shape,h=[e,l,s,g],b=a.makeOutput(h,"float32");if(i.util.sizeFromShape(r.shape)===0)return b;let y=a.dataIdMap.get(r.dataId),k;y.dtype!=="float32"&&(k=Se({backend:a,inputs:{x:r},attrs:{dtype:"float32"}}),y=a.dataIdMap.get(k.dataId));const I=y.id,M=a.dataIdMap.get(b.dataId).id;return Hu(I,e,m,f,g,l,s,n?1:0,o?1:0,M),k!=null&&a.disposeData(k.dataId),b}var Ml={kernelName:i.ResizeNearestNeighbor,backendName:"wasm",setupFunc:kl,kernelFunc:Sl},Lu;function Al(t){Lu=t.wasm.cwrap(i.ResizeNearestNeighborGrad,null,["number","number","number","array","array","boolean"])}function xl(t){const{inputs:a,backend:u,attrs:d}=t,{images:r,dy:n}=a,{alignCorners:o}=d,p=u.makeOutput(r.shape,"float32");let l=u.dataIdMap.get(r.dataId),s;return l.dtype!=="float32"&&(s=Se({backend:u,inputs:{x:r},attrs:{dtype:"float32"}}),l=u.dataIdMap.get(s.dataId)),Lu(u.dataIdMap.get(r.dataId).id,u.dataIdMap.get(n.dataId).id,u.dataIdMap.get(p.dataId).id,new Uint8Array(new Int32Array(r.shape).buffer),new Uint8Array(new Int32Array(n.shape).buffer),o),s!=null&&u.disposeData(s.dataId),p}var Dl={kernelName:i.ResizeNearestNeighborGrad,backendName:"wasm",setupFunc:Al,kernelFunc:xl},Gu;function wl(t){Gu=t.wasm.cwrap(i.Reverse,null,["number","array","number","array","number","number"])}function Fl(t){const{inputs:a,backend:u,attrs:d}=t,{x:r}=a,{dims:n}=d,o=i.util.parseAxisParam(n,r.shape);if(r.shape.length===0)return on({inputs:{x:r},backend:u});const p=u.makeOutput(r.shape,r.dtype),l=u.dataIdMap.get(r.dataId).id,s=u.dataIdMap.get(p.dataId).id,e=new Uint8Array(new Int32Array(o).buffer),m=new Uint8Array(new Int32Array(r.shape).buffer);Gu(l,e,o.length,m,r.shape.length,s);const f=ue({inputs:{x:p},attrs:{shape:r.shape},backend:u});return u.disposeData(p.dataId),f}var Nl={kernelName:i.Reverse,backendName:"wasm",kernelFunc:Fl,setupFunc:wl},$u;function Cl(t){$u=t.wasm.cwrap(i.RotateWithOffset,null,["number","number","number","number","number","number","number","number","array","number","number"])}function Rl(t){const{inputs:a,backend:u,attrs:d}=t,{image:r}=a,{radians:n,fillValue:o,center:p}=d,l=u.makeOutput(r.shape,r.dtype),s=u.dataIdMap.get(r.dataId).id,e=u.dataIdMap.get(l.dataId).id,[m,f,g,h]=r.shape,[b,y]=i.backend_util.getImageCenter(p,f,g),k=o===0,I=255,M=typeof o=="number"?[o,o,o,k?0:I]:[...o,I],D=new Uint8Array(new Int32Array(M).buffer);return $u(s,m,f,g,h,n,b,y,D,M.length,e),l}var _l={kernelName:i.RotateWithOffset,backendName:"wasm",kernelFunc:Rl,setupFunc:Cl},Pl=W(i.Round),El=W(i.Rsqrt),Vu;function Bl(t){Vu=t.wasm.cwrap(i.ScatterNd,null,["number","number","number","number","number","number","array","number","number"])}function Tl(t){const{backend:a,inputs:u,attrs:d}=t,{indices:r,updates:n}=u,{shape:o}=d,p=a.makeOutput(o,n.dtype);if(i.util.sizeFromShape(o)===0)return p;const{sliceRank:l,numUpdates:s,sliceSize:e,strides:m,outputSize:f}=i.scatter_util.calculateShapes(n,r,o),g=a.dataIdMap.get(r.dataId).id,h=a.dataIdMap.get(n.dataId).id,b=new Uint8Array(new Int32Array(m).buffer),y=a.dataIdMap.get(p.dataId).id;return Vu(g,h,G[n.dtype],l,s,e,b,f,y),p}var Ol={kernelName:i.ScatterNd,backendName:"wasm",setupFunc:Bl,kernelFunc:Tl},Uu;function Wl(t){Uu=t.wasm.cwrap(i.SearchSorted,null,["number","number","number","number","number","number","bool","number"])}function zl(t){const{inputs:a,backend:u,attrs:d}=t,{sortedSequence:r,values:n}=a,{side:o}=d;if(r.dtype!==n.dtype)throw new Error(`SearchSorted error: sorted_sequence must have the same dtype as values. Got ${r.dtype} and ${n.dtype}`);const p=u.makeOutput(n.shape,"int32");function l(s){return u.dataIdMap.get(s.dataId).id}return Uu(l(r),l(n),r.shape[0],r.shape[1],n.shape[1],G[r.dtype],o==="left",l(p)),p}var Hl={kernelName:i.SearchSorted,backendName:"wasm",setupFunc:Wl,kernelFunc:zl},qu;function Ll(t){qu=t.wasm.cwrap("SelectV2",null,["number","number","number","number","number"])}function Gl(t){const{inputs:a,backend:u}=t,{condition:d,t:r,e:n}=a,o=u.dataIdMap.get(d.dataId).id,p=u.dataIdMap.get(r.dataId).id,l=u.dataIdMap.get(n.dataId).id,s=u.makeOutput(r.shape,r.dtype),e=u.dataIdMap.get(s.dataId).id,m=d.shape.length,f=r.shape.length,g=m===0||m>1||f===1?1:i.util.sizeFromShape(r.shape.slice(1));return qu(o,p,l,g,e),s}var $l={kernelName:i.Select,backendName:"wasm",kernelFunc:Gl,setupFunc:Ll},Vl=W(i.Selu),Ku;function Ul(t){Ku=t.wasm.cwrap(i.Sigmoid,null,["number","number"])}function ql(t){const{backend:a,inputs:{x:u}}=t,d=a.dataIdMap.get(u.dataId).id,r=a.makeOutput(u.shape,u.dtype),n=a.dataIdMap.get(r.dataId).id;return i.util.sizeFromShape(r.shape)===0||Ku(d,n),r}var Kl={kernelName:"Sigmoid",backendName:"wasm",setupFunc:Ul,kernelFunc:ql},Xl=W(i.Sign),Jl=W(i.Sin),Ql=W(i.Sinh),Yl=W(i.Softplus);function Zl(t){const{inputs:a,backend:u,attrs:d}=t,{x:r}=a,{blockShape:n,paddings:o}=d,p=i.util.sizeFromShape(n),l=[[0,0]];l.push(...o);for(let y=1+n.length;y<r.shape.length;++y)l.push([0,0]);const s=Bu.kernelFunc({inputs:{x:r},backend:u,attrs:{paddings:l,constantValue:0}}),e=i.backend_util.getReshaped(s.shape,n,p,!1),m=i.backend_util.getPermuted(e.length,n.length,!1),f=i.backend_util.getReshapedPermuted(s.shape,n,p,!1),g=ue({inputs:{x:s},backend:u,attrs:{shape:e}}),h=Ie({inputs:{x:g},backend:u,attrs:{perm:m}}),b=ue({inputs:{x:h},backend:u,attrs:{shape:f}});return u.disposeData(s.dataId),u.disposeData(g.dataId),u.disposeData(h.dataId),b}var jl={kernelName:i.SpaceToBatchND,backendName:"wasm",kernelFunc:Zl},Xu;function ec(t){Xu=t.wasm.cwrap("SparseFillEmptyRows","number",["number","number","number","number","number","number","number","number","number","number","number","number"])}function nc(t){const{backend:a,inputs:u}=t,{indices:d,values:r,denseShape:n,defaultValue:o}=u,p=d.shape[0],l=d.shape[1],s=a.readSync(n.dataId)[0],e=[p+s,l],m=a.dataIdMap.get(d.dataId).id,f=a.dataIdMap.get(r.dataId).id,g=a.dataIdMap.get(o.dataId).id,h=a.makeOutput(e,d.dtype),b=a.dataIdMap.get(h.dataId).id,y=a.makeOutput(e.slice(0,1),r.dtype),k=a.dataIdMap.get(y.dataId).id,I=a.makeOutput([s],"bool"),M=a.dataIdMap.get(I.dataId).id,D=a.makeOutput([p],d.dtype),F=a.dataIdMap.get(D.dataId).id,E=a.makeOutput([4],"int32"),O=a.dataIdMap.get(E.dataId).id,B=Xu(m,f,G[r.dtype],p,s,l,g,b,k,M,F,O),K=a.readSync(E.dataId);let X;switch(K[0]){case 1:X=i.backend_util.getSparseFillEmptyRowsIndicesDenseShapeMismatch(K[1]);break;case 2:X=i.backend_util.getSparseFillEmptyRowsNegativeIndexErrorMessage(K[1],K[2]);break;case 3:X=i.backend_util.getSparseFillEmptyRowsOutOfRangeIndexErrorMessage(K[1],K[2],K[3]);break;default:X=""}if(a.disposeData(E.dataId),X)throw a.disposeData(h.dataId),a.disposeData(y.dataId),a.disposeData(I.dataId),a.disposeData(D.dataId),new Error(X);let U=h,q=y;return B!==e[0]&&(U=De({inputs:{x:h},attrs:{begin:0,size:[B,l]},backend:a}),q=De({inputs:{x:y},attrs:{begin:0,size:B},backend:a}),a.disposeData(h.dataId),a.disposeData(y.dataId)),[U,q,I,D]}var tc={kernelName:i.SparseFillEmptyRows,backendName:"wasm",setupFunc:ec,kernelFunc:nc},Ju;function ac(t){Ju=t.wasm.cwrap(i.SparseReshape,null,["number","number","number","number","number","number","number"])}function rc(t){const{backend:a,inputs:u}=t,{inputIndices:d,inputShape:r,newShape:n}=u;if(d.shape.length!==2)throw new Error(`Input indices should be a matrix but received shape
         ${d.shape}`);if(r.shape.length!==1)throw new Error(`Input shape should be a vector but received shape
         ${r.shape}`);if(n.shape.length!==1)throw new Error(`Target shape should be a vector but received shape ${n.shape}`);const o=a.dataIdMap.get(d.dataId).id,p=a.dataIdMap.get(r.dataId).id,l=a.dataIdMap.get(n.dataId).id,s=d.shape[0],e=i.util.sizeFromShape(n.shape),m=a.makeOutput([s,e],d.dtype),f=a.dataIdMap.get(m.dataId).id,g=a.makeOutput([e],n.dtype),h=a.dataIdMap.get(g.dataId).id,b=a.makeOutput([3],"int32"),y=a.dataIdMap.get(b.dataId).id;Ju(o,p,l,s,f,h,y);const k=a.readSync(b.dataId);let I;switch(k[0]){case 0:I=i.backend_util.getSparseReshapeMultipleNegativeOneOutputDimErrorMessage(k[1],k[2]);break;case 1:I=i.backend_util.getSparseReshapeNegativeOutputDimErrorMessage(k[1],k[2]);break;case 2:I=i.backend_util.getSparseReshapeEmptyTensorZeroOutputDimErrorMessage();break;case 3:{const M=Array.from(a.readSync(r.dataId)),D=Array.from(a.readSync(g.dataId));I=i.backend_util.getSparseReshapeInputOutputMultipleErrorMessage(M,D);break}case 4:{const M=Array.from(a.readSync(r.dataId)),D=Array.from(a.readSync(g.dataId));I=i.backend_util.getSparseReshapeInputOutputMismatchErrorMessage(M,D);break}default:I=""}if(a.disposeData(b.dataId),I)throw a.disposeData(m.dataId),a.disposeData(g.dataId),new Error(I);return[m,g]}var uc={kernelName:i.SparseReshape,backendName:"wasm",setupFunc:ac,kernelFunc:rc},Qu;function Yu(t){Qu=t.wasm.cwrap("SparseSegmentReduction",null,["number","number","number","number","number","number","number","number","number"])}function Zu(t,a){const{backend:u,inputs:d}=t,{data:r,indices:n,segmentIds:o}=d,p=n.shape[0],l=u.readSync(o.dataId,p-1,p)[0],s=p>0?l+1:0;if(s<0)throw new Error(i.backend_util.getSparseSegmentReductionNegativeSegmentIdsErrorMessage());const e=r.shape.slice();e[0]=s;const m=u.dataIdMap.get(r.dataId).id,f=u.dataIdMap.get(n.dataId).id,g=u.dataIdMap.get(o.dataId).id,h=u.makeOutput(e,r.dtype),b=u.dataIdMap.get(h.dataId).id,y=u.makeOutput([4],"int32"),k=u.dataIdMap.get(y.dataId).id;Qu(m,G[r.dtype],r.shape[0],f,g,b,k,a,0);const I=u.readSync(y.dataId);let M;switch(I[0]){case 0:M=i.backend_util.getSparseSegmentReductionNegativeSegmentIdsErrorMessage();break;case 1:M=i.backend_util.getSparseSegmentReductionNonIncreasingSegmentIdsErrorMessage();break;case 2:M=i.backend_util.getSparseSegmentReductionSegmentIdOutOfRangeErrorMessage(I[1],I[2]);break;case 3:M=i.backend_util.getSparseSegmentReductionIndicesOutOfRangeErrorMessage(I[1],I[2],I[3]);break;default:M=""}if(u.disposeData(y.dataId),M)throw u.disposeData(h.dataId),new Error(M);return h}function ic(t){return Zu(t,!0)}var sc={kernelName:i.SparseSegmentMean,backendName:"wasm",setupFunc:Yu,kernelFunc:ic};function dc(t){return Zu(t,!1)}var oc={kernelName:i.SparseSegmentSum,backendName:"wasm",setupFunc:Yu,kernelFunc:dc},ju;function pc(t){ju=t.wasm.cwrap(i.SparseToDense,null,["number","number","number","number","number","number","number","number","array","number","number"])}function lc(t){const{backend:a,inputs:u,attrs:d}=t,{sparseIndices:r,sparseValues:n,defaultValue:o}=u,{outputShape:p}=d,l=a.makeOutput(p,o.dtype);if(i.util.sizeFromShape(p)===0)return l;const{sliceRank:s,numUpdates:e,sliceSize:m,strides:f,outputSize:g}=i.backend_util.calculateShapes(n,r,p),h=a.dataIdMap.get(r.dataId).id,b=a.dataIdMap.get(n.dataId).id,y=a.dataIdMap.get(o.dataId).id,k=new Uint8Array(new Int32Array(f).buffer),I=a.dataIdMap.get(l.dataId).id;return ju(h,b,n.shape.length,y,G[o.dtype],s,e,m,k,g,I),l}var cc={kernelName:i.SparseToDense,backendName:"wasm",setupFunc:pc,kernelFunc:lc};function mc(t){const{inputs:a,attrs:u,backend:d}=t,{x:r}=a,{numOrSizeSplits:n,axis:o}=u,p=i.util.parseAxisParam(o,r.shape)[0],l=i.backend_util.prepareSplitSize(r,n,p),s=new Array(r.shape.length).fill(0),e=r.shape.slice();return l.map(m=>{const f=[...e];f[p]=m;const g=De({inputs:{x:r},attrs:{begin:s,size:f},backend:d});return s[p]+=m,g})}var hc={kernelName:i.SplitV,backendName:"wasm",kernelFunc:mc},gc=W(i.Sqrt),fc=W(i.Square),bc=!0,vc=Y(i.SquaredDifference,bc),ei;function yc(t){ei=t.wasm.cwrap(i.Step,null,["number","number","number","number"])}function Ic(t){const{backend:a,inputs:u,attrs:d}=t,{alpha:r}=d,{x:n}=u,o=a.dataIdMap.get(n.dataId).id,p=a.makeOutput(n.shape,n.dtype),l=a.dataIdMap.get(p.dataId).id;return ei(o,r,G[n.dtype],l),p}var kc={kernelName:i.Step,backendName:"wasm",setupFunc:yc,kernelFunc:Ic},ni;function Sc(t){ni=t.wasm.cwrap(i.StridedSlice,null,["number","array","number","array","array","array","array","array","number","number"])}function Mc(t){const{backend:a,inputs:u,attrs:d}=t,{x:r}=u,{begin:n,end:o,strides:p,beginMask:l,endMask:s,ellipsisMask:e,newAxisMask:m,shrinkAxisMask:f}=d,{finalShapeSparse:g,finalShape:h,isIdentity:b,sliceDim0:y,isSimpleSlice:k,begin:I,end:M,strides:D}=i.slice_util.sliceInfo(r.shape,n,o,p,l,s,e,m,f);let F;if(b)F=ue({inputs:{x:r},backend:a,attrs:{shape:h}});else if(y||k){i.util.assert(r.shape.length>=1,()=>`Input must have rank at least 1, got: ${r.shape.length}`);const E=i.slice_util.computeOutShape(I,M,D),O=De({inputs:{x:r},backend:a,attrs:{begin:I,size:E}});F=ue({inputs:{x:O},backend:a,attrs:{shape:h}}),a.disposeData(O.dataId)}else{const E=a.makeOutput(g,"float32"),O=a.dataIdMap.get(r.dataId).id,B=new Uint8Array(new Int32Array(i.util.computeStrides(r.shape)).buffer),K=new Uint8Array(new Int32Array(I).buffer),X=new Uint8Array(new Int32Array(M).buffer),U=new Uint8Array(new Int32Array(D).buffer),q=new Uint8Array(new Int32Array(g).buffer),Z=new Uint8Array(new Int32Array(i.util.computeStrides(g)).buffer),ie=a.dataIdMap.get(E.dataId).id;ni(O,B,r.shape.length,K,X,U,q,Z,g.length,ie),F=ue({inputs:{x:E},backend:a,attrs:{shape:h}}),a.disposeData(E.dataId)}return F}var Ac={kernelName:i.StridedSlice,backendName:"wasm",setupFunc:Sc,kernelFunc:Mc};function xc(t){const{backend:a,inputs:u,attrs:d}=t,{data:r,dataSplits:n}=u,{separator:o,nGramWidths:p,leftPad:l,rightPad:s,padWidth:e,preserveShortSequences:m}=d,f=a.readSync(r.dataId),g=a.readSync(n.dataId),[h,b]=Is(f,g,o,p,l,s,e,m),y=a.makeOutput([h.length],"string"),k=a.dataIdMap.get(y.dataId);k.stringBytes=h;const I=a.makeOutput(n.shape,"int32");return a.typedArrayFromHeap(I).set(b),[y,I]}var Dc={kernelName:i.StringNGrams,backendName:"wasm",kernelFunc:xc};function wc(t){const{backend:a,inputs:u,attrs:d}=t,{input:r,delimiter:n}=u,{skipEmpty:o}=d,p=a.readSync(r.dataId),l=a.readSync(n.dataId),[s,e,m]=Ss(p,l[0],o),f=e.length,g=a.makeOutput([f,2],"int32");a.typedArrayFromHeap(g).set(s);const h=a.makeOutput([f],"string"),b=a.dataIdMap.get(h.dataId);b.stringBytes=e;const y=a.makeOutput([2],"int32");return a.typedArrayFromHeap(y).set(m),[g,h,y]}var Fc={kernelName:i.StringSplit,backendName:"wasm",kernelFunc:wc};function Nc(t){const{backend:a,inputs:u,attrs:d}=t,{input:r}=u,{numBuckets:n}=d,o=a.readSync(r.dataId),p=Ms(o,n),l=a.makeOutput(r.shape,"int32");return a.typedArrayFromHeap(l).set(p),l}var Cc={kernelName:i.StringToHashBucketFast,backendName:"wasm",kernelFunc:Nc},Rc=!0,_c=Y(i.Sub,Rc),ti;function Pc(t){ti=t.wasm.cwrap(i.Sum,null,["number","number","number","number"])}function Ec(t){const{backend:a,inputs:u,attrs:d}=t,{axis:r,keepDims:n}=d,{x:o}=u,p=a.dataIdMap.get(o.dataId).id;let l=p,s=o;const{transposed:e,axes:m,originalAxes:f,inputWasTransposed:g}=ke(o,r,a);let h=m;if(g){const M=a.dataIdMap.get(e.dataId).id;M!==p&&(s=e,l=M,h=i.backend_util.getInnerMostAxes(h.length,s.shape.length))}i.backend_util.assertAxesAreInnerMostDims("sum",h,s.shape.length);const[b,y]=i.backend_util.computeOutAndReduceShapes(s.shape,h),k=i.util.sizeFromShape(y),I=a.makeOutput(b,s.dtype);if(i.util.sizeFromShape(s.shape)!==0){const M=a.dataIdMap.get(I.dataId).id;ti(l,k,G[I.dtype],M)}return g&&a.disposeData(e.dataId),n&&(I.shape=i.backend_util.expandShapeToKeepDim(I.shape,f)),I}var Bc={kernelName:i.Sum,backendName:"wasm",setupFunc:Pc,kernelFunc:Ec},Tc=W(i.Tan),Oc=W(i.Tanh),ai;function Wc(t){ai=t.wasm.cwrap(i.TensorScatterUpdate,null,["number","number","number","number","number","number","array","number","number","number"])}function zc(t){const{backend:a,inputs:u,attrs:d}=t,{tensor:r,indices:n,updates:o}=u,{}=d,p=a.makeOutput(r.shape,r.dtype);if(i.util.sizeFromShape(r.shape)===0)return p;const{sliceRank:l,numUpdates:s,sliceSize:e,strides:m,outputSize:f}=i.scatter_util.calculateShapes(o,n,r.shape),g=a.dataIdMap.get(n.dataId).id,h=a.dataIdMap.get(o.dataId).id,b=a.dataIdMap.get(r.dataId).id,y=new Uint8Array(new Int32Array(m).buffer),k=a.dataIdMap.get(p.dataId).id;return ai(g,h,G[o.dtype],l,s,e,y,f,k,b),p}var Hc={kernelName:i.TensorScatterUpdate,backendName:"wasm",setupFunc:Wc,kernelFunc:zc},ri;function Lc(t){ri=t.wasm.cwrap(i.Tile,null,["number","array","number","array","number","number"])}function Gc(t){const{inputs:a,backend:u,attrs:d}=t,{x:r}=a,n=u.dataIdMap.get(r.dataId).id,{reps:o}=d,p=new Array(r.shape.length);for(let f=0;f<p.length;f++)p[f]=r.shape[f]*o[f];const l=new Uint8Array(new Int32Array(r.shape).buffer),s=new Uint8Array(new Int32Array(p).buffer),e=u.makeOutput(p,r.dtype),m=u.dataIdMap.get(e.dataId).id;return ri(n,l,r.shape.length,s,p.length,G[e.dtype],m),e}var $c={kernelName:i.Tile,backendName:"wasm",setupFunc:Lc,kernelFunc:Gc},ui;function Vc(t){ui=t.wasm.cwrap(i.TopK,null,["number","array","number","number","number","bool","number","number"])}var Uc=({inputs:t,backend:a,attrs:u})=>{const{x:d}=t,{k:r,sorted:n}=u,o=a.dataIdMap.get(d.dataId).id,p=new Uint8Array(new Int32Array(d.shape).buffer),l=d.shape.slice();l[l.length-1]=r;const s=a.makeOutput(l,d.dtype),e=a.dataIdMap.get(s.dataId).id,m=a.makeOutput(l,"int32"),f=a.dataIdMap.get(m.dataId).id;return ui(o,p,d.shape.length,G[d.dtype],r,n,e,f),[s,m]},qc={kernelName:i.TopK,backendName:"wasm",setupFunc:Vc,kernelFunc:Uc},ii;function Kc(t){ii=t.wasm.cwrap(i.Transform,null,["number","number","bool","number","number","number","number","number","number","array","number","array","number","number","number","number","number"])}function Xc(t){const{backend:a,inputs:u,attrs:d}=t,{image:r,transforms:n}=u,{interpolation:o,fillMode:p,fillValue:l,outputShape:s}=d,[e,m,f,g]=r.shape,[h,b]=s??[m,f],y=[e,h,b,g],k=new Uint8Array(new Int32Array(i.util.computeStrides(r.shape)).buffer),I=new Uint8Array(new Int32Array(i.util.computeStrides(y)).buffer),M=a.makeOutput(y,r.dtype),D=a.dataIdMap.get(M.dataId).id,F=a.dataIdMap.get(r.dataId).id,E=a.dataIdMap.get(n.dataId).id,O=o==="nearest"?1:2;let B;switch(p){case"constant":B=1;break;case"reflect":B=2;break;case"wrap":B=3;break;case"nearest":B=4;break;default:B=1}return ii(F,E,n.shape[0]>1,e,h,b,g,f,m,k,r.shape.length-1,I,y.length-1,O,B,l,D),M}var Jc={kernelName:i.Transform,backendName:"wasm",setupFunc:Kc,kernelFunc:Xc};function Qc(t){const{inputs:a,attrs:u,backend:d}=t,{axis:r}=u,{x:n}=a,{outputValues:o,outputShape:p,indices:l}=As(d.readSync(n.dataId),r,n.shape,n.dtype);return[d.makeOutput(p,n.dtype,void 0,o),d.makeOutput([l.length],"int32",void 0,l)]}var Yc={kernelName:i.Unique,backendName:"wasm",kernelFunc:Qc};function Zc(t){const{inputs:a,backend:u,attrs:d}=t,{value:r}=a;let{axis:n}=d;n<0&&(n+=r.shape.length);const o=r.shape[n],p=r.shape.length,l=new Array(p-1);let s=0;for(let g=0;g<p;g++)g!==n&&(l[s++]=r.shape[g]);const e=new Array(o),m=new Array(p).fill(0),f=r.shape.slice();f[n]=1;for(let g=0;g<e.length;g++)m[n]=g,e[g]=De({inputs:{x:r},attrs:{begin:m,size:f},backend:u});return e.map(({dataId:g,dtype:h})=>({dataId:g,dtype:h,shape:l}))}var jc={kernelName:i.Unpack,backendName:"wasm",kernelFunc:Zc};function em(t){const{inputs:{x:a},backend:u}=t,d=u.makeOutput(a.shape,a.dtype);return u.typedArrayFromHeap(d).fill(0),d}var nm={kernelName:i.ZerosLike,backendName:"wasm",kernelFunc:em},tm=[Fi,Ni,Ci,Ri,Pi,Ti,Vi,Ki,Xi,Ji,Qi,Yi,Zi,ji,es,as,cs,is,os,fs,Cs,Ps,Bs,Os,Ws,zs,Gs,$s,qs,Js,Zs,nd,rd,ud,id,od,cd,gd,vd,kd,Ad,wd,Cd,Pd,Td,Od,Hd,Gd,$d,Vd,Ud,qd,Kd,Qd,Yd,jd,to,uo,oo,co,go,bo,yo,Oi,Io,ko,So,xo,wo,No,_o,Eo,Po,To,Oo,zo,Lo,Vo,Ko,Qo,Zo,np,rp,sp,pp,mp,fp,yp,kp,Ap,Np,Cp,_p,Pp,Tp,zp,Gp,Vp,Kp,Jp,Yp,Bu,nl,rl,sl,ol,ll,cl,ml,hl,ms,bl,Il,Ml,Dl,Nl,_l,Pl,El,Ol,Hl,$l,Vl,Kl,Xl,Jl,Ql,Fs,Dp,Yl,jl,tc,uc,sc,oc,cc,hc,gc,fc,vc,kc,Ac,Dc,Fc,Cc,_c,Bc,Tc,Oc,Hc,$c,qc,Jc,Li,Yc,jc,nm];for(const t of tm)(0,i.registerKernel)(t);var Jn=(0,i.env)();Jn.registerFlag("WASM_HAS_SIMD_SUPPORT",async()=>{try{return WebAssembly.validate(new Uint8Array([0,97,115,109,1,0,0,0,1,4,1,96,0,0,3,2,1,0,10,9,1,7,0,65,0,253,15,26,11]))}catch{return!1}}),Jn.registerFlag("WASM_HAS_MULTITHREAD_SUPPORT",async()=>{if(Jn.get("IS_NODE"))return!1;try{return new MessageChannel().port1.postMessage(new SharedArrayBuffer(1)),WebAssembly.validate(new Uint8Array([0,97,115,109,1,0,0,0,1,4,1,96,0,0,3,2,1,0,5,4,1,3,1,1,10,11,1,9,0,65,0,254,16,2,0,26,11]))}catch{return!1}});var we=dn(((t,a)=>{a.exports={}})),am=dn(((t,a)=>{var u=(()=>{var d=typeof document<"u"&&document.currentScript?document.currentScript.src:void 0;return typeof __filename<"u"&&(d=d||__filename),(function(r){r=r||{};function n(){return $.buffer!=de&&he($.buffer),fe}function o(){return $.buffer!=de&&he($.buffer),We}function p(){return $.buffer!=de&&he($.buffer),Me}function l(){return $.buffer!=de&&he($.buffer),ln}function s(){return $.buffer!=de&&he($.buffer),cn}var e=typeof r<"u"?r:{},m,f;e.ready=new Promise(function(c,v){m=c,f=v});var g;typeof process<"u"&&process.listeners&&(g={uncaughtException:process.listeners("uncaughtException"),unhandledRejection:process.listeners("unhandledRejection")});var h=Object.assign({},e),b=[],y=(c,v)=>{throw v},k=typeof window=="object",I=typeof importScripts=="function",M=typeof process=="object"&&typeof process.versions=="object"&&typeof process.versions.node=="string",D=e.ENVIRONMENT_IS_PTHREAD||!1,F="";function E(c){return e.locateFile?e.locateFile(c,F):F+c}var O,B,K;function X(c){c instanceof _e||Q("exiting due to exception: "+c)}if(M){var U=we(),q=we();I?F=q.dirname(F)+"/":F=__dirname+"/",O=(v,A)=>(v=He(v)?new URL(v):q.normalize(v),U.readFileSync(v,A?void 0:"utf8")),K=v=>{var A=O(v,!0);return A.buffer||(A=new Uint8Array(A)),A},B=(v,A,_)=>{v=He(v)?new URL(v):q.normalize(v),U.readFile(v,function(R,w){R?_(R):A(w.buffer)})},process.argv.length>1&&process.argv[1].replace(/\\/g,"/"),b=process.argv.slice(2),process.on("uncaughtException",function(v){if(!(v instanceof _e))throw v}),process.on("unhandledRejection",function(v){throw v}),y=(v,A)=>{if(ze())throw process.exitCode=v,A;X(A),process.exit(v)},e.inspect=function(){return"[Emscripten Module object]"};let c;try{c=we()}catch(v){throw console.error('The "worker_threads" module is not supported in this node.js build - perhaps a newer version is needed?'),v}global.Worker=c.Worker}else(k||I)&&(I?F=self.location.href:typeof document<"u"&&document.currentScript&&(F=document.currentScript.src),typeof d<"u"&&d&&(F=d),F.indexOf("blob:")!==0?F=F.substr(0,F.replace(/[?#].*/,"").lastIndexOf("/")+1):F="",M||(O=c=>{var v=new XMLHttpRequest;return v.open("GET",c,!1),v.send(null),v.responseText},I&&(K=c=>{var v=new XMLHttpRequest;return v.open("GET",c,!1),v.responseType="arraybuffer",v.send(null),new Uint8Array(v.response)}),B=(c,v,A)=>{var _=new XMLHttpRequest;_.open("GET",c,!0),_.responseType="arraybuffer",_.onload=()=>{if(_.status==200||_.status==0&&_.response){v(_.response);return}A()},_.onerror=A,_.send(null)}));M&&typeof performance>"u"&&(global.performance=we().performance);var Z=console.log.bind(console),ie=console.warn.bind(console);M&&(Z=c=>U.writeSync(1,c+`

package/dist/backend-webgpu.js

@@ -1,4 +1,4 @@
-/*! @license Includes TensorFlow.js WebGPU backend 4.22.0 (Apache-2.0). See https://github.com/kubohiroya/turbowarp-teachable-machine/blob/v4.0.1/THIRD_PARTY_NOTICES.md. */
+/*! @license Includes TensorFlow.js WebGPU backend 4.22.0 (Apache-2.0). See https://github.com/kubohiroya/turbowarp-teachable-machine/blob/v4.0.2/THIRD_PARTY_NOTICES.md. */
 (function(u){var gi=Object.create,Se=Object.defineProperty,xi=Object.getOwnPropertyDescriptor,vi=Object.getOwnPropertyNames,Ci=Object.getPrototypeOf,je=Object.prototype.hasOwnProperty,_e=(e,i)=>{let t={};for(var s in e)Se(t,s,{get:e[s],enumerable:!0});return i||Se(t,Symbol.toStringTag,{value:"Module"}),t},bi=(e,i,t,s)=>{if(i&&typeof i=="object"||typeof i=="function")for(var a=vi(i),r=0,n=a.length,o;r<n;r++)o=a[r],!je.call(e,o)&&o!==t&&Se(e,o,{get:(l=>i[l]).bind(null,o),enumerable:!(s=xi(i,o))||s.enumerable});return e},yi=(e,i,t)=>(t=e!=null?gi(Ci(e)):{},bi(i||!e||!e.__esModule||!je.call(e,"default")?Se(t,"default",{value:e,enumerable:!0}):t,e));u=yi(u);var K=(0,u.env)();K.registerFlag("WEBGPU_DEFERRED_SUBMIT_BATCH_SIZE",()=>15),K.registerFlag("WEBGPU_CPU_FORWARD",()=>!0),K.registerFlag("WEBGPU_MATMUL_PROGRAM_TYPE",()=>-1),K.registerFlag("WEBGPU_USE_NAIVE_CONV2D_TRANSPOSE",()=>!0),K.registerFlag("WEBGPU_USE_LOW_POWER_GPU",()=>!1),K.registerFlag("WEBGPU_CPU_HANDOFF_SIZE_THRESHOLD",()=>1e3),K.registerFlag("WEBGPU_USE_PROFILE_TOOL",()=>!1),K.registerFlag("WEBGPU_IMPORT_EXTERNAL_TEXTURE",()=>!0),K.registerFlag("WEBGPU_USE_NAIVE_CONV2D_DEBUG",()=>!1),K.registerFlag("WEBGPU_THRESHOLD_TO_INCREASE_WORKGROUPS_FOR_MATMUL",()=>-1),K.registerFlag("WEBGPU_CONV_SEPARATE_IM2COL_SHADER",()=>!1),K.registerFlag("WEBGPU_PRINT_SHADER",()=>""),K.registerFlag("WEBGPU_ENGINE_COMPILE_ONLY",()=>!1);var Si=class{constructor(e){e&&(this.vendor=e.vendor,this.architecture=e.architecture,this.intelGPUGeneration=this.getIntelGPUGeneration())}getIntelGPUGeneration(){if(this.isIntel()){if(this.architecture.startsWith("gen"))return Number(this.architecture.match(/\d+/));if(this.architecture.startsWith("xe"))return 12}return 0}isIntel(){return this.vendor==="intel"}},Ii=class{constructor(e){this.device=e,this.numUsedBuffers=0,this.numFreeBuffers=0,this.freeBuffers=new Map,this.usedBuffers=new Map,this.numBytesUsed=0,this.numBytesAllocated=0}acquireBuffer(e,i,t=!1,s=!0){let a;const r=et(e,i);return s?(this.freeBuffers.has(r)||this.freeBuffers.set(r,[]),this.freeBuffers.get(r).length>0?(a=this.freeBuffers.get(r).pop(),this.numFreeBuffers--):(a=this.device.createBuffer({size:e,usage:i,mappedAtCreation:t}),this.numBytesAllocated+=e)):(a=this.device.createBuffer({size:e,usage:i,mappedAtCreation:t}),this.numBytesAllocated+=e),this.usedBuffers.has(r)||this.usedBuffers.set(r,[]),this.usedBuffers.get(r).push(a),this.numUsedBuffers++,this.numBytesUsed+=e,a}releaseBuffer(e,i=!0){if(this.freeBuffers.size===0)return;const t=e.size,s=e.usage,a=et(t,s),r=this.usedBuffers.get(a),n=r.indexOf(e);if(n<0)throw new Error("Cannot find the buffer in buffer manager");r[n]=r[r.length-1],r.pop(),this.numUsedBuffers--,this.numBytesUsed-=t,i?(this.freeBuffers.get(a).push(e),this.numFreeBuffers++):(e.destroy(),this.numBytesAllocated-=t)}getNumUsedBuffers(){return this.numUsedBuffers}getNumFreeBuffers(){return this.numFreeBuffers}dispose(){this.freeBuffers.forEach((e,i)=>{e.forEach(t=>{t.destroy()})}),this.usedBuffers.forEach((e,i)=>{e.forEach(t=>{t.destroy()})}),this.freeBuffers=new Map,this.usedBuffers=new Map,this.numUsedBuffers=0,this.numFreeBuffers=0,this.numBytesUsed=0,this.numBytesAllocated=0}};function et(e,i){return`${e}_${i}`}var ki=class{constructor(e){this.device=e,this.numUsedTextures=0,this.numFreeTextures=0,this.freeTextures=new Map,this.usedTextures=new Map,this.numBytesUsed=0,this.numBytesAllocated=0}acquireTexture(e,i,t,s){const a=it(t),r=e*i*a,n=tt(e,i,t,s);if(this.freeTextures.has(n)||this.freeTextures.set(n,[]),this.usedTextures.has(n)||this.usedTextures.set(n,[]),this.numBytesUsed+=r,this.numUsedTextures++,this.freeTextures.get(n).length>0){this.numFreeTextures--;const l=this.freeTextures.get(n).shift();return this.usedTextures.get(n).push(l),l}this.numBytesAllocated+=r;const o=this.device.createTexture({size:[e,i],format:t,usage:s});return this.usedTextures.get(n).push(o),o}releaseTexture(e){if(this.freeTextures.size===0)return;const i=e.width,t=e.height,s=e.format,a=e.usage,r=tt(i,t,s,a);this.freeTextures.has(r)||this.freeTextures.set(r,[]),this.freeTextures.get(r).push(e),this.numFreeTextures++,this.numUsedTextures--;const n=this.usedTextures.get(r),o=n.indexOf(e);if(o<0)throw new Error("Cannot release a texture that was never provided by this texture manager");n.splice(o,1);const l=it(s),d=i*t*l;this.numBytesUsed-=d}getNumUsedTextures(){return this.numUsedTextures}getNumFreeTextures(){return this.numFreeTextures}dispose(){this.freeTextures.forEach((e,i)=>{e.forEach(t=>{t.destroy()})}),this.usedTextures.forEach((e,i)=>{e.forEach(t=>{t.destroy()})}),this.freeTextures=new Map,this.usedTextures=new Map,this.numUsedTextures=0,this.numFreeTextures=0,this.numBytesUsed=0,this.numBytesAllocated=0}};function tt(e,i,t,s){return`${e}_${i}_${t}_${s}`}function it(e){if(e==="rgba8unorm")return 16;throw new Error(`${e} is not supported!`)}function wi(e,i){if(Math.max(...e)>5)throw new Error("Cannot symbolically compute strides for rank > 6 tensor.");const t=e.length,s="xyzwuv",a=e.map(n=>`${i}.${s[n]}`),r=new Array(t-1);r[t-2]=a[t-1];for(let n=t-3;n>=0;--n)r[n]=`(${r[n+1]} * ${a[n+1]})`;return r}var ee=(e,i,t)=>t==="int32"?`atomicAdd(${e}, bitcast<i32>(${i}));`:`
           {
             var oldValue = 0;

package/dist/composition.js

@@ -51,7 +51,7 @@ function requireComputeMode(value) {
 //#region src/config.ts
 var extensionConfig = {
 	id: "kubohiroyatm",
-	slug: "tm",
+	slug: "teachable-machine",
 	name: "Teachable Machine",
 	description: "Use Teachable Machine pose, image, or audio models for recognition in TurboWarp.",
 	author: "Hiroya Kubo",
@@ -1199,7 +1199,7 @@ function confidenceMultiplier(value) {
 }
 var package_default = {
 	name: "@kubohiroya/turbowarp-teachable-machine",
-	version: "4.0.1",
+	version: "4.0.2",
 	description: "A TurboWarp extension for recognition using Teachable Machine pose, image, and audio models.",
 	author: "Hiroya Kubo <hiroya@cuc.ac.jp>",
 	license: "MPL-2.0",
@@ -1223,8 +1223,10 @@ var package_default = {
 		"./dist/backend-wasm.js": "./dist/backend-wasm.js",
 		"./dist/backend-webgpu.js": "./dist/backend-webgpu.js",
 		"./runtime": "./dist/runtime.js",
-		"./dist/tm.js": "./dist/tm.js",
-		"./package.json": "./package.json"
+		"./dist/teachable-machine.js": "./dist/teachable-machine.js",
+		"./package.json": "./package.json",
+		"./dist/teachable-machine-webgpu.js": "./dist/teachable-machine-webgpu.js",
+		"./dist/teachable-machine-wasm.js": "./dist/teachable-machine-wasm.js"
 	},
 	type: "module",
 	files: [
@@ -1237,8 +1239,8 @@ var package_default = {
 		"schemas/",
 		"docs/fixed-input-editors.ja.md"
 	],
-	unpkg: "./dist/tm.js",
-	jsdelivr: "./dist/tm.js",
+	unpkg: "./dist/teachable-machine.js",
+	jsdelivr: "./dist/teachable-machine.js",
 	publishConfig: {
 		"access": "public",
 		"registry": "https://registry.npmjs.org/"
@@ -1248,7 +1250,7 @@ var package_default = {
 	scripts: {
 		"dev": "vite build --watch",
 		"typecheck": "tsc --noEmit",
-		"build": "pnpm build:extension && pnpm build:composition && pnpm build:runtime && pnpm build:backends && pnpm build:posenet && pnpm build:notices && pnpm build:types",
+		"build": "pnpm build:extension && pnpm build:standalone && pnpm build:composition && pnpm build:runtime && pnpm build:backends && pnpm build:posenet && pnpm build:notices && pnpm build:types",
 		"build:extension": "vite build",
 		"build:composition": "vite build --config vite.composition.config.ts",
 		"build:runtime": "vite build --config vite.runtime.config.ts",
@@ -1268,7 +1270,8 @@ var package_default = {
 		"pack:check": "npm pack --dry-run --ignore-scripts",
 		"release:check": "npm publish --dry-run --access public --tag latest",
 		"check": "npm run typecheck && npm run test && npm run build && npm run types:composition:check && npm run types:posenet:check && npm run docs:check && npm run docs:pages:check && npm run release:consistency:check && npm run check:dist && npm run archive:check && npm run pack:check",
-		"prepack": "npm run build && npm run release:consistency:check"
+		"prepack": "npm run build && npm run release:consistency:check",
+		"build:standalone": "vite build --config vite.standalone.config.ts --mode webgpu && vite build --config vite.standalone.config.ts --mode wasm"
 	},
 	devDependencies: {
 		"@kubohiroya/turbowarp-camera-source": "^0.21.1",