gx

chenyc

2025-06-12 7b72ac13a83764a662159d4a49b7fffb90476ecb

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
/**
 * @license
 * Copyright 2023 Google LLC. All Rights Reserved.
 * Licensed under the Apache License, Version 2.0 (the "License");
 * you may not use this file except in compliance with the License.
 * You may obtain a copy of the License at
 *
 * http://www.apache.org/licenses/LICENSE-2.0
 *
 * Unless required by applicable law or agreed to in writing, software
 * distributed under the License is distributed on an "AS IS" BASIS,
 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 * See the License for the specific language governing permissions and
 * limitations under the License.
 * =============================================================================
 */
!function(e,t){"object"==typeof exports&&"undefined"!=typeof module?t(exports,require("@tensorflow/tfjs-core"),require("seedrandom")):"function"==typeof define&&define.amd?define(["exports","@tensorflow/tfjs-core","seedrandom"],t):t((e="undefined"!=typeof globalThis?globalThis:e||self).tf=e.tf||{},e.tf,e.seedrandom)}(this,(function(e,t,n){"use strict";function a(e){var t=Object.create(null);return e&&Object.keys(e).forEach((function(n){if("default"!==n){var a=Object.getOwnPropertyDescriptor(e,n);Object.defineProperty(t,n,a.get?a:{enumerable:!0,get:function(){return e[n]}})}})),t.default=e,t}var s=a(n);function r(e,n){Array.isArray(e)||(e=[e]),e.forEach((e=>{null!=e&&t.util.assert("complex64"!==e.dtype,(()=>`${n} does not support complex64 tensors in the CPU backend.`))}))}const o=t.kernel_impls.whereImpl;class i extends t.KernelBackend{nextDataId(){return i.nextDataId++}constructor(){super(),this.blockSize=48,this.firstUse=!0,this.data=new t.DataStorage(this,t.engine())}write(e,n,a){this.firstUse&&(this.firstUse=!1,t.env().get("IS_NODE")&&t.backend_util.warn("\n============================\nHi, looks like you are running TensorFlow.js in Node.js. To speed things up dramatically, install our node backend, visit https://github.com/tensorflow/tfjs-node for more details. \n============================"));const s={id:this.nextDataId()};return this.data.set(s,{values:e,dtype:a,refCount:1}),s}makeTensorInfo(e,n,a){let s;if("string"===n&&null!=a&&a.length>0&&t.util.isString(a[0])){const r=a.map((e=>t.util.encodeString(e)));s=this.write(r,e,n)}else s=this.write(a,e,n);return{dataId:s,shape:e,dtype:n}}refCount(e){if(this.data.has(e)){return this.data.get(e).refCount}return 0}incRef(e){this.data.get(e).refCount++}decRef(e){if(this.data.has(e)){this.data.get(e).refCount--}}move(e,t,n,a,s){this.data.set(e,{values:t,dtype:a,refCount:s})}numDataIds(){return this.data.numDataIds()}async read(e){return this.readSync(e)}readSync(e){const{dtype:n,complexTensorInfos:a}=this.data.get(e);if("complex64"===n){const e=this.readSync(a.real.dataId),n=this.readSync(a.imag.dataId);return t.backend_util.mergeRealAndImagArrays(e,n)}return t.util.convertBackendValuesAndArrayBuffer(this.data.get(e).values,n)}bufferSync(e){const n=this.readSync(e.dataId);if("string"===e.dtype)try{const a=n.map((e=>t.util.decodeString(e)));return t.buffer(e.shape,e.dtype,a)}catch(e){throw new Error("Failed to decode encoded string bytes into utf-8")}return t.buffer(e.shape,e.dtype,n)}makeOutput(e,n,a){return t.engine().makeTensorFromTensorInfo(this.makeTensorInfo(n,a,e),this)}disposeData(e,t=!1){if(this.data.has(e)){if(this.data.get(e).refCount--,!t&&this.data.get(e).refCount>0)return!1;const{complexTensorInfos:n}=this.data.get(e);null!=n&&(this.disposeData(n.real.dataId,!0),this.disposeData(n.imag.dataId,!0)),this.data.delete(e)}return!0}disposeIntermediateTensorInfo(e){this.disposeData(e.dataId)}async time(e){const n=t.util.now();e();return{kernelMs:t.util.now()-n}}memory(){return{unreliable:!0,reasons:["The reported memory is an upper bound. Due to automatic garbage collection, the true allocated memory may be less."]}}where(e){r([e],"where");const t=this.readSync(e.dataId);return o(e.shape,t)}dispose(){}floatPrecision(){return 32}epsilon(){return super.epsilon()}}function l(e){const t=new Float32Array(e.length);for(let n=0;n<e.length;++n)t[n]=Math.abs(e[n]);return t}i.nextDataId=0;const u={kernelName:t.Abs,backendName:"cpu",kernelFunc:e=>{const{x:n}=e.inputs,a=e.backend;r(n,"abs");let s=new Float32Array(t.util.sizeFromShape(n.shape));return s=l(a.data.get(n.dataId).values),a.makeOutput(s,n.shape,n.dtype)}};function c(e){return(n,a,s,r,o)=>{const i=t.backend_util.assertAndGetBroadcastShape(n,a),l=i.length,u=t.util.computeStrides(i),c=t.util.sizeFromShape(i),d=t.util.getTypedArrayFromDType(o,c),p=n.length,h=a.length,f=t.util.computeStrides(n),m=t.util.computeStrides(a),k=t.backend_util.getBroadcastDims(n,i),g=t.backend_util.getBroadcastDims(a,i);if(k.length+g.length===0)for(let t=0;t<d.length;++t)d[t]=e(s[t%s.length],r[t%r.length]);else for(let n=0;n<d.length;++n){const a=t.util.indexToLoc(n,l,u),o=a.slice(-p);k.forEach((e=>o[e]=0));const i=t.util.locToIndex(o,p,f),c=a.slice(-h);g.forEach((e=>c[e]=0));const b=t.util.locToIndex(c,h,m);d[n]=e(s[i],r[b])}return[d,i]}}function d(e){const{inputs:t,backend:n}=e,{real:a,imag:s}=t,r=n.data.get(a.dataId).values,o=n.data.get(s.dataId).values,i=n.makeTensorInfo(a.shape,"complex64");return n.data.get(i.dataId).complexTensorInfos={real:n.makeTensorInfo(a.shape,"float32",r),imag:n.makeTensorInfo(s.shape,"float32",o)},i}const p={kernelName:t.Complex,backendName:"cpu",kernelFunc:d};function h(e,n,a="float32"){if("complex64"===a){return d({inputs:{real:h(e,n,"float32"),imag:h(e,n,"float32")},backend:e})}const s=t.util.makeZerosTypedArray(t.util.sizeFromShape(n),a);return e.makeTensorInfo(n,a,s)}function f(e){const{inputs:t,backend:n}=e,{x:a}=t;return n.incRef(a.dataId),{dataId:a.dataId,shape:a.shape,dtype:a.dtype}}const m={kernelName:t.Identity,backendName:"cpu",kernelFunc:f};function k(e){const{inputs:t,backend:n}=e,{input:a}=t,s=n.data.get(a.dataId).complexTensorInfos.real,r=n.data.get(s.dataId).values;return n.makeTensorInfo(s.shape,s.dtype,r)}const g={kernelName:t.Real,backendName:"cpu",kernelFunc:k};function b(e,n,a,s){if("int32"===s){return[n,"int32",Int32Array.from(e)]}if("bool"===s){const s=t.util.toTypedArray([0],a),[r,o]=c(((e,t)=>e!==t?1:0))(n,[],e,s,"bool");return[o,"bool",r]}throw new Error(`Error in Cast: failed to cast ${a} to ${s}`)}function I(e){const{inputs:n,backend:a,attrs:s}=e,{x:r}=n,{dtype:o}=s;if("complex64"===o){if("complex64"===r.dtype)return f({inputs:{x:r},backend:a});const e=h(a,r.shape,r.dtype),t=I({inputs:{x:r},backend:a,attrs:{dtype:"float32"}}),n=d({inputs:{real:t,imag:e},backend:a});return a.disposeIntermediateTensorInfo(e),a.disposeIntermediateTensorInfo(t),n}if("complex64"===r.dtype){const e=k({inputs:{input:r},backend:a}),t=I({inputs:{x:e},backend:a,attrs:{dtype:o}});return a.disposeIntermediateTensorInfo(e),t}if(!t.util.hasEncodingLoss(r.dtype,o)){const e=f({inputs:{x:r},backend:a});return{dataId:e.dataId,shape:e.shape,dtype:o}}const i=a.data.get(r.dataId).values,[l,u,c]=b(i,r.shape,r.dtype,o);return a.makeTensorInfo(l,u,c)}const y={kernelName:t.Cast,backendName:"cpu",kernelFunc:I};function S(e,n,a,s){return null==a?({inputs:a,backend:o})=>{const{a:i,b:l}=a,u=o;r([i,l],e);const c=u.data.get(i.dataId).values,d=u.data.get(l.dataId).values,p="string"===i.dtype?t.backend_util.fromUint8ToStringArray(c):c,h="string"===i.dtype?t.backend_util.fromUint8ToStringArray(d):d,f=s||i.dtype,[m,k]=n(i.shape,l.shape,p,h,f);return u.makeTensorInfo(k,f,m)}:({inputs:e,backend:t})=>{const{a:r,b:o}=e,i=t;if("complex64"===r.dtype||"complex64"===o.dtype){const e=I({inputs:{x:r},backend:i,attrs:{dtype:"complex64"}}),t=i.data.get(e.dataId),n=t.complexTensorInfos.real,s=t.complexTensorInfos.imag,l=i.data.get(n.dataId).values,u=i.data.get(s.dataId).values,c=I({inputs:{x:o},backend:i,attrs:{dtype:"complex64"}}),p=i.data.get(c.dataId),h=p.complexTensorInfos.real,f=p.complexTensorInfos.imag,m=i.data.get(h.dataId).values,k=i.data.get(f.dataId).values,[g,b,y]=a(r.shape,o.shape,l,u,m,k),S=i.makeTensorInfo(y,"float32",g),T=i.makeTensorInfo(y,"float32",b),N=d({inputs:{real:S,imag:T},backend:i});return i.disposeIntermediateTensorInfo(e),i.disposeIntermediateTensorInfo(c),i.disposeIntermediateTensorInfo(S),i.disposeIntermediateTensorInfo(T),N}{const e=i.data.get(r.dataId).values,t=i.data.get(o.dataId).values,a=s||r.dtype,[l,u]=n(r.shape,o.shape,e,t,a);return i.makeTensorInfo(u,a,l)}}}function T(e){return(n,a,s,r,o,i)=>{const l=t.backend_util.assertAndGetBroadcastShape(n,a),u=t.util.sizeFromShape(l),c=l.length,d=t.util.computeStrides(l),p=t.util.getTypedArrayFromDType("float32",u),h=t.util.getTypedArrayFromDType("float32",u),f=t.backend_util.getBroadcastDims(n,l),m=t.backend_util.getBroadcastDims(a,l),k=t.backend_util.mergeRealAndImagArrays(s,r),g=t.backend_util.mergeRealAndImagArrays(o,i),b=n.length,I=t.util.computeStrides(n),y=a.length,S=t.util.computeStrides(a);if(f.length+m.length===0)for(let t=0;t<p.length;t++){const n=t%k.length,a=t%g.length,s=e(k[2*n],k[2*n+1],g[2*a],g[2*a+1]);p[t]=s.real,h[t]=s.imag}else for(let n=0;n<p.length;n++){const a=t.util.indexToLoc(n,c,d),s=a.slice(-b);f.forEach((e=>s[e]=0));const r=t.util.locToIndex(s,b,I),o=a.slice(-y);m.forEach((e=>o[e]=0));const i=t.util.locToIndex(o,y,S),l=e(k[2*r],k[2*r+1],g[2*i],g[2*i+1]);p[n]=l.real,h[n]=l.imag}return[p,h,l]}}const N=c(((e,t)=>e+t)),x=T(((e,t,n,a)=>({real:e+n,imag:t+a}))),v=S(t.Add,N,x),F={kernelName:t.Add,backendName:"cpu",kernelFunc:v};function w(e,n,a,s,r){const o=t.util.sizeFromShape(s),i=t.util.makeZerosTypedArray(r,a);for(let t=0;t<e.length;t++){const a=e[t];if(a<0)throw new Error("Input x must be non-negative!");a>=r||(i[a]+=o>0?n[t]:1)}return i}function M(e,n,a,s=!1){const r=e.shape[0],o=e.shape[1],i=t.buffer([r,a],n.dtype);for(let t=0;t<r;t++)for(let r=0;r<o;r++){const o=e.get(t,r);if(o<0)throw new Error("Input x must be non-negative!");o>=a||(s?i.set(1,t,o):n.size>0?i.set(i.get(t,o)+n.get(t,r),t,o):i.set(i.get(t,o)+1,t,o))}return i}const A=c(((e,t)=>e&t)),D=S(t.BitwiseAnd,A),E={kernelName:t.BitwiseAnd,backendName:"cpu",kernelFunc:D};function _(e){return(n,a,s)=>{const r=t.util.getArrayFromDType(a,n.length);for(let t=0;t<n.length;++t)r[t]=e(n[t],s);return r}}function z(e,t,n){return R(e,_(t),n)}function R(e,n,a){return({inputs:s,attrs:o,backend:i})=>{const{x:l}=s;r(l,e);const u=i,c=u.data.get(l.dataId).values;let d;if("string"===l.dtype){if(!Array.isArray(c))throw new Error("String tensor's value was not an instance of Array");d=t.backend_util.fromUint8ToStringArray(c)}else d=c;const p=a||l.dtype,h=n(d,p,o);return u.makeTensorInfo(l.shape,p,h)}}const W=_((e=>Math.ceil(e))),P=R(t.Ceil,W),C={kernelName:t.Ceil,backendName:"cpu",kernelFunc:P};function H(e,n,a,s){const r=t.util.getArrayFromDType(a,t.util.sizeFromShape(n));if(s&&"string"!==a){let n=0;e.forEach((e=>{const a=t.util.sizeFromShape(e.shape);r.set(e.vals,n),n+=a}))}else{let s=0;e.forEach((e=>{const o="string"===a?t.backend_util.fromUint8ToStringArray(e.vals):e.vals;let i=0;for(let t=0;t<e.shape[0];++t){const a=t*n[1]+s;for(let t=0;t<e.shape[1];++t)r[a+t]=o[i++]}s+=e.shape[1]}))}return r}const O=c(((e,t)=>e===t?1:0)),B=S(t.Equal,O,null,"bool"),$={kernelName:t.Equal,backendName:"cpu",kernelFunc:B},V=_((e=>Math.exp(e))),G=R(t.Exp,V,"float32"),L={kernelName:t.Exp,backendName:"cpu",kernelFunc:G},q=_((e=>Math.expm1(e))),U=R(t.Expm1,q),Z={kernelName:t.Expm1,backendName:"cpu",kernelFunc:U},j=_((e=>Math.floor(e))),K=R(t.Floor,j),Y={kernelName:t.Floor,backendName:"cpu",kernelFunc:K},J=c(((e,t)=>Math.floor(e/t))),Q=S(t.FloorDiv,J,null,"int32"),X={kernelName:t.FloorDiv,backendName:"cpu",kernelFunc:Q};function ee(e,n,a,s,r,o,i,l,u){const c=t.buffer([s,o],a);for(let t=0;t<s;t++){const a=[];let s=0;for(let n=0;n<r;n++){const o=e[t*r+n];s+=o*i[n],a.push(o)}if(s<0||s>=u/o)throw new Error(`Invalid indices: ${a} does not index into ${l}`);for(let e=0;e<o;e++)c.values[t*o+e]=n.get(...n.indexToLoc(s*o+e))}return c}function te(e,n,a){const s=t.buffer(a,e.dtype);for(let t=0;t<s.size;++t){const a=s.indexToLoc(t).slice(),r=a[0],o=a[2],i=n.locToIndex([r,o]);a[2]=n.values[i];const l=e.locToIndex(a);0<=l&&l<e.values.length&&(s.values[t]=e.values[l])}return s}const ne=c(((e,t)=>e>t?1:0)),ae=S(t.Greater,ne,null,"bool"),se={kernelName:t.Greater,backendName:"cpu",kernelFunc:ae},re=c(((e,t)=>e>=t?1:0)),oe=S(t.GreaterEqual,re,null,"bool"),ie={kernelName:t.GreaterEqual,backendName:"cpu",kernelFunc:oe},le=c(((e,t)=>e<t?1:0)),ue=S(t.Less,le,null,"bool"),ce={kernelName:t.Less,backendName:"cpu",kernelFunc:ue},de=c(((e,t)=>e<=t?1:0)),pe=S(t.LessEqual,de,null,"bool"),he={kernelName:t.LessEqual,backendName:"cpu",kernelFunc:pe};function fe(e,n,a){const s=(n-e)/(a-1),r=t.util.makeZerosTypedArray(a,"float32");r[0]=e;for(let e=1;e<r.length;e++)r[e]=r[e-1]+s;return r}const me=_((e=>Math.log(e))),ke=R(t.Log,me),ge={kernelName:t.Log,backendName:"cpu",kernelFunc:ke};function be(e,n,a,s){const r=t.util.getTypedArrayFromDType(s,t.util.sizeFromShape(a));for(let t=0;t<r.length;++t){const a=t*n;let s=e[a];for(let t=0;t<n;++t){const n=e[a+t];(Number.isNaN(n)||n>s)&&(s=n)}r[t]=s}return r}const Ie=c(((e,t)=>Math.max(e,t))),ye=S(t.Maximum,Ie),Se={kernelName:t.Maximum,backendName:"cpu",kernelFunc:ye},Te=c(((e,t)=>Math.min(e,t))),Ne=S(t.Minimum,Te),xe={kernelName:t.Minimum,backendName:"cpu",kernelFunc:Ne},ve=c(((e,t)=>e*t)),Fe=T(((e,t,n,a)=>({real:e*n-t*a,imag:e*a+t*n}))),we=S(t.Multiply,ve,Fe),Me={kernelName:t.Multiply,backendName:"cpu",kernelFunc:we};function Ae(e,n,a){const s=t.util.createScalarValue(-1,a);return ve([],n,s,e,a)}const De={kernelName:t.Neg,backendName:"cpu",kernelFunc:function(e){const{inputs:t,backend:n}=e,{x:a}=t;r(a,"neg");const s=n.data.get(a.dataId).values,[o,i]=Ae(s,a.shape,a.dtype);return n.makeTensorInfo(i,a.dtype,o)}},Ee=c(((e,t)=>e!==t?1:0)),_e=S(t.NotEqual,Ee,null,"bool"),ze={kernelName:t.NotEqual,backendName:"cpu",kernelFunc:_e};function Re(e,n,a,s,r){const o=n.length,i=t.util.sizeFromShape(n),l=t.util.computeStrides(n),u=t.util.computeStrides(r),c=t.util.getTypedArrayFromDType(a,t.util.sizeFromShape(r));for(let n=0;n<i;++n){const a=t.util.indexToLoc(n,o,l),r=new Array(a.length);for(let e=0;e<r.length;e++)r[e]=a[s[e]];c[t.util.locToIndex(r,o,u)]=e[n]}return c}function We(e){const{inputs:t,attrs:n,backend:a}=e,{x:s}=t,{perm:o}=n;r(s,"transpose");const i=s.shape.length,l=new Array(i);for(let e=0;e<l.length;e++)l[e]=s.shape[o[e]];const u=Re(a.data.get(s.dataId).values,s.shape,s.dtype,o,l);return{dataId:a.write(u,l,s.dtype),shape:l,dtype:s.dtype}}const Pe={kernelName:t.Transpose,backendName:"cpu",kernelFunc:We};function Ce(e,n,a,s){const[r,o]=t.backend_util.computeOutAndReduceShapes(e,s),i=t.upcastType(n,"int32"),l=t.util.makeZerosTypedArray(t.util.sizeFromShape(r),i),u=t.util.sizeFromShape(o);for(let e=0;e<l.length;++e){const t=e*u;let n=1;for(let e=0;e<u;++e)n*=a[t+e];l[e]=n}return{outVals:l,outShape:r,outDtype:i}}const He={kernelName:t.Prod,backendName:"cpu",kernelFunc:function(e){const{inputs:n,backend:a,attrs:s}=e,{x:o}=n,{axis:i,keepDims:l}=s;r(o,"prod");const u=o.shape.length,c=t.util.parseAxisParam(i,o.shape),d=t.backend_util.getAxesPermutation(c,u);let p=c,h=o;const f=[];null!=d&&(h=We({inputs:{x:o},backend:a,attrs:{perm:d}}),f.push(h),p=t.backend_util.getInnerMostAxes(p.length,u));const m=a.data.get(h.dataId).values,{outVals:k,outShape:g,outDtype:b}=Ce(h.shape,h.dtype,m,p);let I=g;return l&&(I=t.backend_util.expandShapeToKeepDim(g,c)),f.forEach((e=>a.disposeIntermediateTensorInfo(e))),a.makeTensorInfo(I,b,k)}};function Oe(e,t,n,a){const s=[];let r=0;const o=t.length-1+n.length,i=new Array(o).fill(null).map((()=>[0]));!function(e,t){for(let n=0;n<e.length;++n){const a=e[n],s=n===e.length-1?t:e[n+1].length;if(0===a.length)throw new Error("Ragged splits may not be empty");if(a[0]<0)throw new Error("Ragged splits must be non-negative");if(a[a.length-1]>s)throw new Error("Ragged splits must not point past values");for(let e=1;e<a.length;++e)if(a[e-1]>a[e])throw new Error("Ragged splits must be sorted in ascending order")}}(n,a);let l=1;for(let e=0;e<t.length-1;++e){l*=t[e];const n=t[e+1];for(let t=1;t<l+1;++t)i[e].push(t*n)}for(let a=0;a<e.length;++a){let o=e[a],l=e[a]+1;for(let e=0;e<n.length;++e){const a=n[e],s=e+t.length-1;if(s>=0){const e=i[s],t=e[e.length-1]-a[o];for(let e=o;e<l;++e)i[s].push(a[e+1]+t)}o=a[o],l=a[l]}l!==o&&(s.push([o,l]),r+=l-o)}return{outSplits:i,valueSlices:s,numValues:r}}function Be(e,t){const n=e.slice(0,t);for(;n.length<t;)n.push(1);for(let a=t;a<e.length;a++)n[t-1]*=e[a];return n}function $e(e,n,a,s,r){const o=n.slice();o[0]=r;const i=t.util.getArrayFromDType(a,t.util.sizeFromShape(o)),l=e.length;return function(e,t,n,a,s,r){const o=Be(t,2)[1],i=Be(r,2)[1];let l=0;for(const t of n)for(let n=t[0];n<t[1];++n){for(let t=0;t<a;++t)s[l*i+t]=e[n*o+t];++l}}(e,n,s,0===l?0:l/n[0],i,o),[i,o]}function Ve(e,n,a,s,r,o,i,l){if(0===e.length)throw new Error("paramsNestedSplits must be non empty");if(0===n[0].length)throw new Error("Split tensors must not be scalars");if(function(e,n,a){e.forEach(((e,s)=>{if(e<0||e>=a){const r=t.util.indexToLoc(s,n.length,t.util.computeStrides(n)).join(",");throw new Error(`indices[${r}] = ${e} is not in [0, ${a})`)}}))}(o,i,n[0][0]-1),0===s.length)throw new Error("params.rank must be nonzero");const u=s[0],{outSplits:c,valueSlices:d,numValues:p}=Oe(o,i,e,u),h=function(e){const n=[];for(let a=0;a<e.length;++a){const s=e[a].length,r=t.util.getArrayFromDType("int32",s);n.push(r),e[a].forEach(((e,t)=>r[t]=e))}return n}(c),f=$e(a,s,r,d,p);return[h,f[0],f[1]]}const Ge=2147483647;function Le(e,n,a,s,r,o,i){if(n.length>1)throw new Error("starts must be a scalar or vector");if(r.length>1)throw new Error("limits must be a scalar or vector");if(i.length>1)throw new Error("deltas must be a scalar or vector");const l=0===n.length,u=0===r.length,c=0===i.length,d=[];l||d.push(n[0]),u||d.push(r[0]),c||d.push(i[0]);for(let e=1;e<d.length;++e)if(d[e]!==d[e-1])throw new Error("starts, limits, and deltas must have the same shape");const p=0===d.length?1:d[0],h=t.util.getArrayFromDType("int32",p+1);h[0]=0;for(let t=0;t<p;++t){const n=l?e[0]:e[t],a=u?s[0]:s[t],r=c?o[0]:o[t];if(0===r)throw new Error("Requires delta != 0");let i;if(r>0&&a<n||r<0&&a>n)i=0;else if(i=Math.ceil(Math.abs((a-n)/r)),i>Ge)throw new Error("Requires ((limit - start) / delta) <= 2147483647");h[t+1]=h[t]+i}const f=h[p],m=t.util.getArrayFromDType(a,f);let k=0;for(let t=0;t<p;++t){const n=h[t+1]-h[t];let a=l?e[0]:e[t];const s=c?o[0]:o[t];for(let e=0;e<n;++e)m[k++]=a,a+=s}return[h,m]}var qe=t.backend_util.RowPartitionType;class Ue{constructor(e,n,a,s,r,o,i,l,u,c){this.shape=e,this.shapeShape=n,this.values=a,this.valuesShape=s,this.valuesDType=r,this.defaultValue=o,this.defaultValueShape=i,this.rowPartitionValues=l,this.rowPartitionValuesShapes=u,this.rowPartitionTypes=t.backend_util.getRowPartitionTypesHelper(c),this.raggedRank=t.backend_util.getRaggedRank(this.rowPartitionTypes)}getRowPartitionTypeByDimension(e){return this.rowPartitionTypes[0]===qe.FIRST_DIM_SIZE?this.rowPartitionTypes[e+1]:this.rowPartitionTypes[e]}getRowPartitionTensor(e){return this.rowPartitionTypes[0]===qe.FIRST_DIM_SIZE?this.rowPartitionValues[e+1]:this.rowPartitionValues[e]}getMaxWidth(e){const t=this.getRowPartitionTensor(e-1);switch(this.getRowPartitionTypeByDimension(e-1)){case qe.VALUE_ROWIDS:return Ue.getMaxWidthValueRowID(t);case qe.ROW_SPLITS:return Ue.getMaxWidthRowSplit(t);default:throw new Error(`Cannot handle partition type ${qe[this.getRowPartitionTypeByDimension(e-1)]}`)}}static getMaxWidthRowSplit(e){const t=e.length;if(0===t||1===t)return 0;let n=0;for(let a=0;a<t-1;++a){const t=e[a+1]-e[a];t>n&&(n=t)}return n}static getMaxWidthValueRowID(e){const t=e.length;if(0===t)return 0;let n=0,a=e[0],s=0;for(let r=1;r<t;++r){const t=e[r];t!==a&&(a=t,s=Math.max(r-n,s),n=r)}return Math.max(t-n,s)}tensorShapeFromTensor(e,t,n=!0){if(0===t.length){if(-1===e[0])return[];throw new Error("The only valid scalar shape tensor is the fully unknown shape specified as -1.")}return je(e,n)}calculateOutputSize(e){const n=this.valuesShape,a=this.defaultValueShape;t.backend_util.validateDefaultValueShape(a,n);const s=this.tensorShapeFromTensor(this.shape,this.shapeShape),r=t.backend_util.combineRaggedTensorToTensorShapes(this.raggedRank,s,n);r[0]<0&&(r[0]=e);for(let e=1;e<=this.raggedRank;++e)r[e]<0&&(r[e]=this.getMaxWidth(e));return r}calculateFirstParentOutputIndex(e,n,a){const s=Math.min(e,a),r=[];let o=0;for(let e=0;e<s;++e,o+=n)r.push(o);for(let t=s;t<e;++t)r.push(-1);return t.util.assert(r.length===e,(()=>"Final length of result must be equal to firstDimension.")),r}calculateOutputIndexRowSplit(e,t,n,a){const s=e.length,r=[];for(let o=0;o<s-1;++o){const s=e[o+1]-e[o];let i=Math.min(a,s),l=t[o];-1===l&&(i=0);for(let e=0;e<i;++e)r.push(l),l+=n;for(let e=0;e<s-i;++e)r.push(-1)}if(s>0&&r.length!==e[s-1])throw new Error("Invalid row split size.");return r}calculateOutputIndexValueRowID(e,t,n,a){const s=e.length,r=[];if(0===s)return[];let o=0,i=e[0];if(i>=t.length)throw new Error(`Got currentValueRowId=${i}, which is not less than ${t.length}`);let l=t[i];r.push(l);for(let u=1;u<s;++u){const s=e[u];if(s===i)l>=0&&(++o,o<a?l+=n:l=-1);else{if(o=0,i=s,s>=t.length)throw new Error(`Got nextValueRowId=${s} which is not less than ${t.length}`);l=t[s]}r.push(l)}if(r.length!==e.length)throw new Error("Invalid row ids.");return r}calculateOutputIndex(e,t,n,a){const s=this.getRowPartitionTensor(e),r=this.getRowPartitionTypeByDimension(e);switch(r){case qe.VALUE_ROWIDS:return this.calculateOutputIndexValueRowID(s,t,n,a);case qe.ROW_SPLITS:if(s.length-1>t.length)throw new Error(`Row partition size is greater than output size: ${s.length-1} > ${t.length}`);return this.calculateOutputIndexRowSplit(s,t,n,a);default:throw new Error(`Unsupported partition type: ${qe[r]}`)}}getFirstDimensionSize(){const e=this.rowPartitionValues[0];if(0===this.rowPartitionTypes.length)throw new Error("No row_partition_types given.");const t=this.rowPartitionTypes[0];switch(t){case qe.FIRST_DIM_SIZE:return e[0];case qe.VALUE_ROWIDS:throw new Error("Cannot handle VALUE_ROWIDS in first dimension.");case qe.ROW_SPLITS:return this.rowPartitionValuesShapes[0][0]-1;default:throw new Error(`Cannot handle type ${qe[t]}`)}}compute(){if(this.rowPartitionValues[0].length<=0)throw new Error("Invalid first partition input. Tensor requires at least one element.");const e=this.getFirstDimensionSize(),n=this.calculateOutputSize(e),a=new Array(this.raggedRank+1);a[a.length-1]=1;for(let e=a.length-2;e>=0;--e)a[e]=a[e+1]*n[e+1];const s=je(n,!1),r=t.util.getArrayFromDType(this.valuesDType,t.util.sizeFromShape(s));if(a[0]*n[0]>0){let t=this.calculateFirstParentOutputIndex(e,a[0],n[0]);for(let e=1;e<=this.raggedRank;++e){t=this.calculateOutputIndex(e-1,t,a[e],n[e])}this.setOutput(this.raggedRank,t,r,s)}return[s,r]}setOutput(e,n,a,s){if(0===a.length)return;const r=this.values,o=a;let i=s.slice();i=i.slice(e+1);const l=t.util.sizeFromShape(i),u=n.length;let c=this.defaultValue;if(c.length!==l&&1!==c.length){const e=this.defaultValueShape;t.tidy((()=>{const n=t.reshape(c,e),a=t.broadcastTo(n,i);c=a.dataSync()}))}let d=0,p=0,h=0;for(let e=0;e<=u;++e){let t=e<u?n[e]:-1;if(t!==h){if(p<h){const e=r.subarray(d*l);Ze(o.subarray(p*l),e,(h-p)*l)}if(e>=u){const e=a.length;t=Math.floor(e/l)}if(t>h)if(1===this.defaultValue.length)o.subarray(h*l,t*l).fill(this.defaultValue[0]),h=t;else for(;t>h;){Ze(o.slice(h*l),c,l),++h}t<0?(d=e+1,p=h):(d=e,p=h,h=p+1)}else++h}}}function Ze(e,t,n){for(let a=0;a<n;a++)e[a]=t[a]}function je(e,t){const n=[];for(let a of e){if(a<0){if(!t)throw new Error(`Dimension ${a} must be >= 0`);if(a<-1)throw new Error(`Dimension ${a} must be >= -1`);a=-1}n.push(a)}return n}function Ke(e,t,n,a,s,r,o,i,l,u){return new Ue(e,t,n,a,s,r,o,i,l,u).compute()}function Ye(e,n,a,s){if(e===n||e<n&&a<0||n<e&&a>1)return t.util.makeZerosTypedArray(0,s);const r=Math.abs(Math.ceil((n-e)/a)),o=t.util.makeZerosTypedArray(r,s);n<e&&1===a&&(a=-1),o[0]=e;for(let e=1;e<o.length;e++)o[e]=o[e-1]+a;return o}const Je=_((e=>1/Math.sqrt(e))),Qe=R(t.Rsqrt,Je),Xe={kernelName:t.Rsqrt,backendName:"cpu",kernelFunc:Qe};function et(e,n,a,s,r,o,i,l,u,c){const d=[s/r,r],p=e.values,h=n.values;if(0===s)return t.buffer(a,n.dtype);const f=u instanceof t.TensorBuffer?u:t.buffer(d,n.dtype);"string"==typeof u||"number"==typeof u?f.values.fill(u):"boolean"==typeof u&&f.values.fill(+u);for(let e=0;e<o;e++){const t=[];let o=0;for(let n=0;n<i;n++){const a=p[e*i+n];t.push(a),o+=a*l[n]}if(o<0||o>=s/r)throw new Error(`Invalid indices: ${t} does not index into ${a}`);for(let t=0;t<r;t++)c?f.values[o*r+t]+=h[e*r+t]:f.values[o*r+t]=0===n.rank?h[0]:h[e*r+t]}return f}const tt=_((e=>1/(1+Math.exp(-e)))),nt=z(t.Sigmoid,(e=>1/(1+Math.exp(-e)))),at={kernelName:t.Sigmoid,backendName:"cpu",kernelFunc:nt};function st(e,n,a,s,r){const o=t.slice_util.isSliceContinous(s,n,a),i=t.util.sizeFromShape(a),l=t.util.computeStrides(s);if(o){const a=t.slice_util.computeFlatOffset(n,l);return"string"===r?e.slice(a,a+i):e.subarray(a,a+i)}const u="string"===r?t.backend_util.fromUint8ToStringArray(e):e,c=t.buffer(s,r,u),d=t.buffer(a,r);for(let e=0;e<d.size;++e){const t=d.indexToLoc(e),a=t.map(((e,t)=>e+n[t]));d.set(c.get(...a),...t)}return"string"===r?t.backend_util.fromStringArrayToUint8(d.values):d.values}function rt(e){const{inputs:n,backend:a,attrs:s}=e,{x:o}=n,{begin:i,size:l}=s;r(o,"slice");const[u,c]=t.slice_util.parseSliceParams(o,i,l);t.slice_util.assertParamsValid(o,u,c);const d=st(a.data.get(o.dataId).values,u,c,o.shape,o.dtype);return a.makeTensorInfo(c,o.dtype,d)}const ot={kernelName:t.Slice,backendName:"cpu",kernelFunc:rt};function it(e,n,a,s,r,o,i){const l=n[0],u=o[0],c=new Array(u),d=new Array(l),p=n[1];if(0===u){if(0!==l)throw new Error(t.backend_util.getSparseFillEmptyRowsIndicesDenseShapeMismatch(l));return[t.util.getArrayFromDType(a,0),[0,p],t.util.getArrayFromDType(r,0),c,d]}let h=!0,f=0;const m=new Array(u).fill(0);for(let n=0;n<l;++n){const a=e[n*p];if(a<0)throw new Error(t.backend_util.getSparseFillEmptyRowsNegativeIndexErrorMessage(n,a));if(a>=u)throw new Error(t.backend_util.getSparseFillEmptyRowsOutOfRangeIndexErrorMessage(n,a,u));++m[a],h=h&&a>=f,f=a}let k=!0;for(let e=0;e<u;++e){const t=0===m[e];c[e]=t,k=k&&!t,m[e]=Math.max(m[e],1),e>0&&(m[e]+=m[e-1])}if(k&&h){const t=e,n=s;for(let e=0;e<l;++e)d[e]=e;return[t,[l,p],n,c,d]}{const n=m[u-1],o=t.util.getArrayFromDType(a,n*p),h=t.util.getArrayFromDType(r,n),f=new Array(u).fill(0);for(let t=0;t<l;++t){const n=e[t*p],a=f[n],r=(0===n?0:m[n-1])+a;f[n]++;for(let n=0;n<p;++n)o[r*p+n]=e[t*p+n];h[r]=s[t],d[t]=r}for(let e=0;e<u;++e){if(0===f[e]){const t=0===e?0:m[e-1];o[t*p+0]=e;for(let e=1;e<p;++e)o[t*p+e]=0;h[t]=i}}return[o,[n,p],h,c,d]}}function lt(e,n,a,s,r){const o=t.util.sizeFromShape(s),i=n[0],l=r.length,u=[];let c=1,d=-1;for(let e=0;e<l;++e){const n=r[e];if(-1===n){if(-1!==d)throw new Error(t.backend_util.getSparseReshapeMultipleNegativeOneOutputDimErrorMessage(d,e));d=e,u.push(1)}else{if(n<0)throw new Error(t.backend_util.getSparseReshapeNegativeOutputDimErrorMessage(e,n));c*=n,u.push(n)}}if(-1!==d){if(c<=0)throw new Error(t.backend_util.getSparseReshapeEmptyTensorZeroOutputDimErrorMessage());const e=Math.trunc(o/c);if(c*e!==o)throw new Error(t.backend_util.getSparseReshapeInputOutputMultipleErrorMessage(s,u));u[d]=e}if(t.util.sizeFromShape(u)!==o)throw new Error(t.backend_util.getSparseReshapeInputOutputMismatchErrorMessage(s,u));const p=s.length,h=[];if(p>0){h[p-1]=1;for(let e=p-2;e>=0;--e)h[e]=h[e+1]*s[e+1]}const f=[];if(l>0){f[l-1]=1;for(let e=l-2;e>=0;--e)f[e]=f[e+1]*u[e+1]}const m=t.util.getArrayFromDType(a,i*l);for(let t=0;t<i;++t){let n=0;for(let a=0;a<p;++a)n+=e[t*p+a]*h[a];for(let e=0;e<l;++e)m[t*l+e]=Math.trunc(n/f[e]),n%=f[e]}return[m,[i,l],u]}function ut(e,n,a,s,r,o=!1,i=0){const l=s.length,u=[n[0],e.length/n[0]],c=u[1],d=l>0?r[l-1]+1:0;if(d<0)throw new Error(t.backend_util.getSparseSegmentReductionNegativeSegmentIdsErrorMessage());const p=n.slice();p[0]=d;const h=p.reduce(((e,t)=>e*t),1),f=t.util.getArrayFromDType(a,h);if(0===l)return d>0&&f.fill(i),[f,p];if(d<=0)throw new Error(t.backend_util.getSparseSegmentReductionNegativeSegmentIdsErrorMessage());let m=0,k=1,g=0,b=r[m];for(;;){let n=0;if(k<l){if(n=r[k],b===n){++k;continue}if(b>=n)throw new Error(t.backend_util.getSparseSegmentReductionNonIncreasingSegmentIdsErrorMessage())}if(b<0||b>=d)throw new Error(t.backend_util.getSparseSegmentReductionSegmentIdOutOfRangeErrorMessage(b,d));b>g&&f.fill(i,g*c,b*c);for(let n=m;n<k;++n){const a=s[n];if(a<0||a>=u[0])throw new Error(t.backend_util.getSparseSegmentReductionIndicesOutOfRangeErrorMessage(n,s[n],u[0]));for(let t=0;t<c;t++)f[b*c+t]+=e[a*c+t]}if(o)for(let e=0;e<c;e++)f[b*c+e]/=k-m;if(m=k,++k,g=b+1,b=n,k>l)break}return g<d&&f.fill(i,g*c,d*c),[f,p]}const ct=_((e=>Math.sqrt(e))),dt=z(t.Sqrt,(e=>Math.sqrt(e))),pt={kernelName:t.Sqrt,backendName:"cpu",kernelFunc:dt},ht=c(((e,t)=>{const n=e-t;return n*n})),ft=S(t.SquaredDifference,ht),mt={kernelName:t.SquaredDifference,backendName:"cpu",kernelFunc:ft},kt=_(((e,t)=>{const{pattern:n,replaceGlobal:a,rewrite:s}=t;return e.replace(new RegExp(n,a?"g":""),s)})),gt=R(t.StaticRegexReplace,kt),bt={kernelName:t.StaticRegexReplace,backendName:"cpu",kernelFunc:gt};function It(e,n,a,s){const r=t.buffer(e,n.dtype);for(let e=0;e<r.size;e++){const t=r.indexToLoc(e),o=new Array(t.length);for(let e=0;e<o.length;e++)o[e]=t[e]*a[e]+s[e];r.set(n.get(...o),...t)}return r}class yt{constructor(e,n,a,s,r,o){this.separator=t.util.encodeString(e),this.nGramWidths=n,this.leftPad=t.util.encodeString(a),this.rightPad=t.util.encodeString(s),this.padWidth=r,this.preserveShort=o}getPadWidth(e){return Math.min(this.padWidth<0?e-1:this.padWidth,e-1)}getNumNGrams(e,t){const n=this.getPadWidth(t);return Math.max(0,e+2*n-t+1)}createNGrams(e,t,n,a,s,r){for(let o=0;o<s;++o){const i=this.getPadWidth(r),l=Math.max(0,i-o),u=Math.max(0,i-(s-(o+1))),c=r-(l+u),d=t+(l>0?0:o-i);let p=0;p+=l*this.leftPad.length;for(let t=0;t<c;++t)p+=e[d+t].length;p+=u*this.rightPad.length;p+=(l+u+c-1)*this.separator.length,n[a+o]=new Uint8Array(p);const h=n[a+o];let f=0;const m=e=>e.forEach((e=>h[f++]=e));for(let e=0;e<l;++e)m(this.leftPad),m(this.separator);for(let t=0;t<c-1;++t)m(e[d+t]),m(this.separator);if(c>0){m(e[d+c-1]);for(let e=0;e<u;++e)m(this.separator),m(this.rightPad)}else{for(let e=0;e<u-1;++e)m(this.rightPad),m(this.separator);m(this.rightPad)}}}compute(e,n){const a=e.length,s=n.length;if(s>0){let e=n[0];if(0!==e)throw new Error(`First split value must be 0, got ${e}`);for(let t=1;t<s;++t){let s=n[t]>=e;if(s=s&&n[t]<=a,!s)throw new Error(`Invalid split value ${n[t]}, must be in [${e}, ${a}]`);e=n[t]}if(e!==a)throw new Error(`Last split value must be data size. Expected ${a}, got ${e}`)}const r=s-1,o=t.util.getArrayFromDType("int32",s);if(0===a||0===s){const e=new Array(a);for(let e=0;e<=r;++e)o[e]=0;return[e,o]}o[0]=0;for(let e=1;e<=r;++e){const t=n[e]-n[e-1];let a=0;this.nGramWidths.forEach((e=>{a+=this.getNumNGrams(t,e)})),this.preserveShort&&t>0&&0===a&&(a=1),o[e]=o[e-1]+a}const i=new Array(o[r]);for(let t=0;t<r;++t){const a=n[t];let s=o[t];if(this.nGramWidths.forEach((r=>{const o=n[t+1]-n[t],l=this.getNumNGrams(o,r);this.createNGrams(e,a,i,s,l,r),s+=l})),this.preserveShort&&s===o[t]){const r=n[t+1]-n[t];if(0===r)continue;const o=r+2*this.padWidth,l=1;this.createNGrams(e,a,i,s,l,o)}}return[i,o]}}function St(e,t,n,a,s,r,o,i){return new yt(n,a,s,r,o,i).compute(e,t)}function Tt(e,t,n,a){if(!e.length)return;if(0===t.length){for(let t=0;t<e.length;++t)a.push(e.subarray(t,t+1));return}if(1===t.length){const s=t[0];let r=e.indexOf(s);for(;-1!==r;){const t=e.subarray(0,r);n&&0===t.length||a.push(t),r=(e=e.subarray(r+1)).indexOf(s)}return void(n&&0===e.length||a.push(e))}let s=0;for(let r=0;r<e.length+1;r++)if(r===e.length||-1!==t.indexOf(e[r])){const t=e.subarray(s,r);n&&0===t.length||a.push(t),s=r+1}}function Nt(e,n,a){const s=e.length,r=[];let o=0,i=0;const l=new Array(s);for(let t=0;t<s;++t){const s=r.length;Tt(e[t],n,a,r);const u=r.length-s;l[t]=u,o+=u,i=Math.max(i,u)}const u=t.util.getArrayFromDType("int32",2*o),c=new Array(o),d=[s,i];let p=0;for(let e=0;e<s;++e)for(let t=0;t<l[e];++t)u[2*p]=e,u[2*p+1]=t,c[p]=r[p],++p;return[u,c,d]}function xt(e,n){const a=t.util.getArrayFromDType("int32",e.length);for(let s=0;s<e.length;++s)a[s]=t.util.fingerPrint64(e[s]).modulo(n).getLowBitsUnsigned();return a}const vt=c(((e,t)=>e-t)),Ft=T(((e,t,n,a)=>({real:e-n,imag:t-a}))),wt=S(t.Sub,vt,Ft),Mt={kernelName:t.Sub,backendName:"cpu",kernelFunc:wt};function At(e,n){const a=new Array(e.rank);for(let t=0;t<a.length;t++)a[t]=e.shape[t]*n[t];const s=t.buffer(a,e.dtype);for(let t=0;t<s.values.length;++t){const n=s.indexToLoc(t),a=new Array(e.rank);for(let t=0;t<a.length;t++)a[t]=n[t]%e.shape[t];const r=e.locToIndex(a);s.values[t]=e.values[r]}return s}const Dt=(e,t)=>{const n=t.value-e.value;return 0===n?e.index-t.index:n};function Et(e,n,a=0,s=e.length-1){for(;s>a;){if(s-a>600){const t=s-a+1,r=n-a+1,o=Math.log(t),i=.5*Math.exp(2*o/3),l=.5*Math.sqrt(o*i*(t-i)/t)*Math.sign(r-t/2);Et(e,n,Math.max(a,Math.floor(n-r*i/t+l)),Math.min(s,Math.floor(n+(t-r)*i/t+l)))}const r=e[n];let o=a,i=s;for(t.util.swap(e,a,n),Dt(e[s],r)>0&&t.util.swap(e,a,s);o<i;){for(t.util.swap(e,o,i),o++,i--;Dt(e[o],r)<0;)o+=1;for(;Dt(e[i],r)>0;)i-=1}0===Dt(e[a],r)?t.util.swap(e,a,i):(i+=1,t.util.swap(e,i,s)),i<=n&&(a=i+1),n<=i&&(s=i-1)}}function _t(e,n,a,s,r){const o=n[n.length-1],[i,l]=[e.length/o,o],u=t.util.getTypedArrayFromDType(a,i*s),c=t.util.getTypedArrayFromDType("int32",i*s);for(let t=0;t<i;t++){const n=t*l,a=e.subarray(n,n+l);let o=new Array(a.length);a.forEach(((e,t)=>o[t]={value:e,index:t})),s<o.length&&(Et(o,s),o=o.slice(0,s)),r&&o.sort(Dt);const i=t*s,d=u.subarray(i,i+s),p=c.subarray(i,i+s);for(let e=0;e<s;e++)d[e]=o[e].value,p[e]=o[e].index}const d=n.slice();return d[d.length-1]=s,[t.buffer(d,a,u),t.buffer(d,"int32",c)]}function zt(e,n,a,s){const r=t.util.parseAxisParam(n,a)[0],o=[1,a[0],1];for(let e=0;e<r;e++)o[0]*=a[e];o[1]=a[r];for(let e=r+1;e<a.length;e++)o[2]*=a[e];const i=new Map,l=new Int32Array(a[r]),u=new t.TensorBuffer(o,s,e),c=[],d=1===o[0]&&1===o[2];for(let t=0;t<a[r];t++){let n;if(d)n=e[t].toString();else{const e=[];for(let n=0;n<o[0];n++)for(let a=0;a<o[2];a++)e.push(u.get(n,t,a));n=e.join(",")}const a=i.get(n);if(null!=a)l[t]=a;else{const e=i.size;i.set(n,e),l[t]=e,c.push(t)}}const p=o.slice();p[1]=i.size;const h=new t.TensorBuffer(p,s);c.forEach(((e,t)=>{for(let n=0;n<o[0];n++)for(let a=0;a<o[2];a++)h.set(u.get(n,e,a),n,t,a)}));const f=a.slice();return f[r]=p[1],{outputValues:h.values,outputShape:f,indices:l}}var Rt={__proto__:null,addImpl:N,bincountImpl:w,bincountReduceImpl:M,bitwiseAndImpl:A,castImpl:b,ceilImpl:W,concatImpl:H,equalImpl:O,expImpl:V,expm1Impl:q,floorDivImpl:J,floorImpl:j,gatherNdImpl:ee,gatherV2Impl:te,greaterEqualImpl:re,greaterImpl:ne,lessEqualImpl:de,lessImpl:le,linSpaceImpl:fe,logImpl:me,maxImpl:be,maximumImpl:Ie,minimumImpl:Te,multiplyImpl:ve,negImpl:Ae,notEqualImpl:Ee,prodImpl:Ce,raggedGatherImpl:Ve,raggedRangeImpl:Le,raggedTensorToTensorImpl:Ke,rangeImpl:Ye,rsqrtImpl:Je,scatterImpl:et,sigmoidImpl:tt,simpleAbsImpl:l,sliceImpl:st,sparseFillEmptyRowsImpl:it,sparseReshapeImpl:lt,sparseSegmentReductionImpl:ut,sqrtImpl:ct,squaredDifferenceImpl:ht,staticRegexReplaceImpl:kt,stridedSliceImpl:It,stringNGramsImpl:St,stringSplitImpl:Nt,stringToHashBucketFastImpl:xt,subImpl:vt,tileImpl:At,topKImpl:_t,transposeImpl:Re,uniqueImpl:zt};t.registerBackend("cpu",(()=>new i),1);const Wt=z(t.Elu,(e=>e>=0?e:Math.exp(e)-1)),Pt={kernelName:t.Elu,backendName:"cpu",kernelFunc:Wt};function Ct(e){const{inputs:n,backend:a,attrs:s}=e,{x:o}=n,{alpha:i}=s;r([o],"leakyRelu");const l=t.util.sizeFromShape(o.shape),u=a.data.get(o.dataId).values,c=t.util.getTypedArrayFromDType("float32",l);for(let e=0;e<u.length;e++)c[e]=u[e]<0?i*u[e]:u[e];return a.makeTensorInfo(o.shape,"float32",c)}const Ht={kernelName:t.LeakyRelu,backendName:"cpu",kernelFunc:Ct},Ot=c(((e,t)=>e<0?t*e:e));function Bt(e){const{inputs:t,backend:n}=e,{x:a,alpha:s}=t;r([a,s],"prelu");const o=n.data.get(a.dataId).values,i=n.data.get(s.dataId).values,[l,u]=Ot(a.shape,s.shape,o,i,"float32");return n.makeTensorInfo(u,"float32",l)}const $t={kernelName:t.Prelu,backendName:"cpu",kernelFunc:Bt},Vt=z(t.Relu,(e=>Math.max(0,e))),Gt={kernelName:t.Relu,backendName:"cpu",kernelFunc:Vt},Lt=z(t.Relu6,(e=>Math.min(Math.max(0,e),6))),qt={kernelName:t.Relu6,backendName:"cpu",kernelFunc:Lt};function Ut(e,t,n,a,s){if("linear"===n)return f({inputs:{x:t},backend:e});if("relu"===n)return Vt({inputs:{x:t},backend:e});if("elu"===n)return Wt({inputs:{x:t},backend:e});if("relu6"===n)return Lt({inputs:{x:t},backend:e});if("prelu"===n)return Bt({inputs:{x:t,alpha:a},backend:e});if("leakyrelu"===n)return Ct({inputs:{x:t},backend:e,attrs:{alpha:s}});if("sigmoid"===n)return nt({inputs:{x:t},backend:e});throw new Error(`Activation ${n} has not been implemented for the CPU backend.`)}function Zt(e){const{inputs:n,backend:a,attrs:s}=e,{x:r}=n,{shape:o}=s,i=t.util.sizeFromShape(r.shape),l=t.util.inferFromImplicitShape(o,i),u=t.util.sizeFromShape(l);t.util.assert(i===u,(()=>`The new shape (${l}) has ${u} elements and the old shape (${r.shape}) has ${i} elements. The new shape and old shape must have the same number of elements.`)),a.incRef(r.dataId);const c=a.data.get(r.dataId);if(null!=c.complexTensorInfos){const e=c.complexTensorInfos.real,t=c.complexTensorInfos.imag;e.shape=l,t.shape=l}return{dataId:r.dataId,shape:l,dtype:r.dtype}}const jt={kernelName:t.Reshape,backendName:"cpu",kernelFunc:Zt};function Kt(e){const{inputs:n,backend:a,attrs:s}=e,{a:o,b:i}=n,{transposeA:l,transposeB:u}=s;r([o,i],"matMul");const c=o.shape.length,d=i.shape.length,p=l?o.shape[c-2]:o.shape[c-1],h=u?i.shape[d-1]:i.shape[d-2],f=l?o.shape[c-1]:o.shape[c-2],m=u?i.shape[d-2]:i.shape[d-1],k=o.shape.slice(0,-2),g=i.shape.slice(0,-2),b=t.util.sizeFromShape(k),I=t.util.sizeFromShape(g),y=t.broadcast_util.assertAndGetBroadcastShape(o.shape.slice(0,-2),i.shape.slice(0,-2)).concat([f,m]);t.util.assert(p===h,(()=>`Error in matMul: inner shapes (${p}) and (${h}) of Tensors with shapes ${o.shape} and ${i.shape} and transposeA=${l} and transposeB=${u} must match.`));const S=u?[I,m,h]:[I,h,m],T=Zt({inputs:{x:o},backend:a,attrs:{shape:l?[b,p,f]:[b,f,p]}}),N=Zt({inputs:{x:i},backend:a,attrs:{shape:S}}),x=l?T.shape[1]:T.shape[2],v=l?T.shape[2]:T.shape[1],F=u?N.shape[1]:N.shape[2],w=Math.max(b,I),M=a.data.get(T.dataId).values,A=a.data.get(N.dataId).values,D=t.util.computeStrides(T.shape),E=t.util.computeStrides(N.shape),[_,z,R]=l?[D[0],1,D[1]]:[D[0],D[1],1],[W,P,C]=u?[1,E[1],E[0]]:[E[1],1,E[0]],H=v*F,O=t.buffer([w,v,F],T.dtype),B=O.values,$=a.blockSize;for(let e=0;e<w;e++){const t=e%b,n=e%I;for(let a=0;a<v;a+=$){const s=Math.min(a+$,v);for(let r=0;r<F;r+=$){const o=Math.min(r+$,F);for(let i=0;i<x;i+=$){const l=Math.min(i+$,x);for(let u=a;u<s;u++)for(let a=r;a<o;a++){let s=0;for(let e=i;e<l;e++){s+=M[t*_+u*z+e*R]*A[e*W+a*P+n*C]}B[e*H+(u*F+a)]+=s}}}}}return a.disposeIntermediateTensorInfo(T),a.disposeIntermediateTensorInfo(N),a.makeTensorInfo(y,O.dtype,O.values)}const Yt={kernelName:t.BatchMatMul,backendName:"cpu",kernelFunc:Kt};const Jt={kernelName:t._FusedMatMul,backendName:"cpu",kernelFunc:function(e){const{inputs:t,backend:n,attrs:a}=e,{a:s,b:r,bias:o,preluActivationWeights:i}=t,{transposeA:l,transposeB:u,activation:c,leakyreluAlpha:d}=a;let p,h,f;const m=[];p=Kt({inputs:{a:s,b:r},attrs:{transposeA:l,transposeB:u},backend:n}),o&&(h=v({inputs:{a:p,b:o},backend:n}),m.push(p),p=h),c&&(f=Ut(n,p,c,i,d),m.push(p),p=f);for(const e of m)n.disposeIntermediateTensorInfo(e);return p}},Qt=z(t.Acos,(e=>Math.acos(e))),Xt={kernelName:t.Acos,backendName:"cpu",kernelFunc:Qt},en=z(t.Acosh,(e=>Math.acosh(e))),tn={kernelName:t.Acosh,backendName:"cpu",kernelFunc:en};const nn={kernelName:t.AddN,backendName:"cpu",kernelFunc:function(e){const{inputs:n,backend:a}=e,s=n;r(n,"addN");const o=s.map((e=>a.data.get(e.dataId).values)),i=t.buffer(s[0].shape,s[0].dtype),l=i.values;for(let e=0;e<s.length;e++){const t=o[e];for(let e=0;e<l.length;e++)l[e]+=t[e]}return a.makeTensorInfo(i.shape,i.dtype,i.values)}};const an={kernelName:t.All,backendName:"cpu",kernelFunc:function(e){const{inputs:n,backend:a,attrs:s}=e,{x:o}=n,{axis:i,keepDims:l}=s;r(o,"all");const u=t.util.parseAxisParam(i,o.shape);let c=u;const d=t.backend_util.getAxesPermutation(c,o.shape.length);let p=o;null!=d&&(p=We({inputs:{x:o},backend:a,attrs:{perm:d}}),c=t.backend_util.getInnerMostAxes(c.length,o.shape.length)),t.backend_util.assertAxesAreInnerMostDims("all",c,p.shape.length);const[h,f]=t.backend_util.computeOutAndReduceShapes(p.shape,c),m=t.util.sizeFromShape(f),k=t.util.makeZerosTypedArray(t.util.sizeFromShape(h),p.dtype),g=a.data.get(p.dataId).values;for(let e=0;e<k.length;++e){const t=e*m;let n=g[t];for(let e=0;e<m;++e){const a=g[t+e];n=n&&a}k[e]=n}null!=d&&a.disposeIntermediateTensorInfo(p);const b=a.makeTensorInfo(h,p.dtype,k);if(l){const e=Zt({inputs:{x:b},backend:a,attrs:{shape:t.backend_util.expandShapeToKeepDim(h,u)}});return a.disposeIntermediateTensorInfo(b),e}return b}};const sn={kernelName:t.Any,backendName:"cpu",kernelFunc:function(e){const{inputs:n,backend:a,attrs:s}=e,{x:o}=n,{axis:i,keepDims:l}=s;r(o,"any");const u=t.util.parseAxisParam(i,o.shape);let c=u;const d=t.backend_util.getAxesPermutation(c,o.shape.length);let p=o;null!=d&&(p=We({inputs:{x:o},backend:a,attrs:{perm:d}}),c=t.backend_util.getInnerMostAxes(c.length,o.shape.length)),t.backend_util.assertAxesAreInnerMostDims("any",c,p.shape.length);const[h,f]=t.backend_util.computeOutAndReduceShapes(p.shape,c),m=t.util.sizeFromShape(f),k=t.util.makeZerosTypedArray(t.util.sizeFromShape(h),p.dtype),g=a.data.get(p.dataId).values;for(let e=0;e<k.length;++e){const t=e*m;let n=g[t];for(let e=0;e<m;++e){const a=g[t+e];n=n||a}k[e]=n}null!=d&&a.disposeIntermediateTensorInfo(p);const b=a.makeTensorInfo(h,p.dtype,k);if(l){const e=Zt({inputs:{x:b},backend:a,attrs:{shape:t.backend_util.expandShapeToKeepDim(h,u)}});return a.disposeIntermediateTensorInfo(b),e}return b}};const rn={kernelName:t.ArgMax,backendName:"cpu",kernelFunc:function(e){const{inputs:n,backend:a,attrs:s}=e,{x:o}=n,{axis:i}=s;r(o,"argMax");let l=t.util.parseAxisParam(i,o.shape);const u=t.backend_util.getAxesPermutation(l,o.shape.length);let c=o;const d=[];null!=u&&(c=We({inputs:{x:o},backend:a,attrs:{perm:u}}),d.push(c),l=t.backend_util.getInnerMostAxes(l.length,c.shape.length)),l=[l[0]],t.backend_util.assertAxesAreInnerMostDims("argMax",l,c.shape.length);const[p,h]=t.backend_util.computeOutAndReduceShapes(c.shape,l),f=t.util.sizeFromShape(p),m=t.util.makeZerosTypedArray(f,"int32"),k=t.util.sizeFromShape(h),g=a.data.get(c.dataId).values;for(let e=0;e<m.length;++e){const t=e*k;let n=g[t],a=0;for(let e=0;e<k;++e){const s=g[t+e];s>n&&(n=s,a=e)}m[e]=a}return d.forEach((e=>a.disposeIntermediateTensorInfo(e))),a.makeTensorInfo(p,"int32",m)}};const on={kernelName:t.ArgMin,backendName:"cpu",kernelFunc:function(e){const{inputs:n,backend:a,attrs:s}=e,{x:o}=n,{axis:i}=s;r(o,"argMin");let l=t.util.parseAxisParam(i,o.shape);const u=t.backend_util.getAxesPermutation(l,o.shape.length);let c=o;const d=[];null!=u&&(c=We({inputs:{x:o},backend:a,attrs:{perm:u}}),d.push(c),l=t.backend_util.getInnerMostAxes(l.length,c.shape.length)),l=[l[0]],t.backend_util.assertAxesAreInnerMostDims("argMin",l,c.shape.length);const[p,h]=t.backend_util.computeOutAndReduceShapes(c.shape,l),f=t.util.sizeFromShape(p),m=t.util.makeZerosTypedArray(f,"int32"),k=t.util.sizeFromShape(h),g=a.data.get(c.dataId).values;for(let e=0;e<m.length;++e){const t=e*k;let n=g[t],a=0;for(let e=0;e<k;++e){const s=g[t+e];s<n&&(n=s,a=e)}m[e]=a}return d.forEach((e=>a.disposeIntermediateTensorInfo(e))),a.makeTensorInfo(p,"int32",m)}},ln=z(t.Asin,(e=>Math.asin(e))),un={kernelName:t.Asin,backendName:"cpu",kernelFunc:ln},cn=z(t.Asinh,(e=>Math.asinh(e))),dn={kernelName:t.Asinh,backendName:"cpu",kernelFunc:cn},pn=z(t.Atan,(e=>Math.atan(e))),hn={kernelName:t.Atan,backendName:"cpu",kernelFunc:pn},fn=c(((e,t)=>Math.atan2(e,t))),mn=S(t.Atan2,fn),kn={kernelName:t.Atan2,backendName:"cpu",kernelFunc:mn},gn=z(t.Atanh,(e=>Math.atanh(e))),bn={kernelName:t.Atanh,backendName:"cpu",kernelFunc:gn};function In(e,n,a,s,r,o){const i=r.strideHeight,l=r.strideWidth,u=r.dilationHeight,c=r.dilationWidth,d=r.effectiveFilterHeight,p=r.effectiveFilterWidth,h=r.padInfo.top,f=r.padInfo.left,m="max"===o?Number.NEGATIVE_INFINITY:Number.POSITIVE_INFINITY,k=t.buffer(r.outShape,a),g=k.values,b=r.outShape[1]*r.outShape[2]*r.outShape[3],I=r.outShape[2]*r.outShape[3],y=r.outShape[3];for(let t=0;t<r.batchSize;++t){const n=t*b,a=t*s[0];for(let t=0;t<r.inChannels;++t)for(let k=0;k<r.outHeight;++k){const b=k*i-h,S=Math.max(0,b),T=Math.min(r.inHeight,d+b),N=n+k*I;for(let n=0;n<r.outWidth;++n){const i=n*l-f,d=Math.max(0,i),h=Math.min(r.inWidth,p+i);let k=m,b=0,I=0;for(let n=S;n<T;n+=u){const r=a+n*s[1];for(let n=d;n<h;n+=c){const a=e[r+n*s[2]+t];"max"===o&&a>k?k=a:"avg"===o&&(b+=a,I++)}if(isNaN(k))break}g[N+n*y+t]="avg"===o?b/I:k}}}return k}function yn(e,n,a,s,r=!1,o=!1){const i=t.buffer(s.outShape,"int32"),l=s.strideHeight,u=s.strideWidth,c=s.dilationHeight,d=s.dilationWidth,p=s.effectiveFilterHeight,h=s.effectiveFilterWidth,f=s.padInfo.top,m=s.padInfo.left,k=t.buffer(n,a,e);for(let e=0;e<s.batchSize;++e)for(let t=0;t<s.inChannels;++t)for(let n=0;n<s.outHeight;++n){const a=n*l-f;let g=a;for(;g<0;)g+=c;const b=Math.min(s.inHeight,p+a);for(let l=0;l<s.outWidth;++l){const p=l*u-m;let f=p;for(;f<0;)f+=d;const I=Math.min(s.inWidth,h+p);let y=Number.NEGATIVE_INFINITY,S=-1;for(let n=g;n<b;n+=c){const i=n-a;for(let a=f;a<I;a+=d){const l=a-p,u=k.get(e,n,a,t);u>y&&(y=u,S=r?o?((e*s.inHeight+n)*s.inWidth+a)*s.inChannels+t:(n*s.inWidth+a)*s.inChannels+t:i*h+l)}}i.set(S,e,n,l,t)}}return i}function Sn(e,n,a,s,r,o){const i=r.strideDepth,l=r.strideHeight,u=r.strideWidth,c=r.dilationDepth,d=r.dilationHeight,p=r.dilationWidth,h=r.effectiveFilterDepth,f=r.effectiveFilterHeight,m=r.effectiveFilterWidth,k=r.padInfo.front,g=r.padInfo.top,b=r.padInfo.left,I="max"===o?Number.NEGATIVE_INFINITY:Number.POSITIVE_INFINITY,y=t.buffer(r.outShape,a),S=y.values,T=r.outShape[1]*r.outShape[2]*r.outShape[3]*r.outShape[4],N=r.outShape[2]*r.outShape[3]*r.outShape[4],x=r.outShape[3]*r.outShape[4],v=r.outShape[4];for(let t=0;t<r.batchSize;++t){const n=t*T,a=t*s[0];for(let t=0;t<r.inChannels;++t)for(let y=0;y<r.outDepth;++y){const T=y*i-k;let F=T;for(;F<0;)F+=c;const w=Math.min(r.inDepth,h+T),M=n+y*N;for(let n=0;n<r.outHeight;++n){const i=n*l-g;let h=i;for(;h<0;)h+=d;const k=Math.min(r.inHeight,f+i),y=M+n*x;for(let n=0;n<r.outWidth;++n){const i=n*u-b;let l=i;for(;l<0;)l+=p;const f=Math.min(r.inWidth,m+i),g=y+n*v;let T=I,N=0,x=0;for(let n=F;n<w;n+=c){const r=a+n*s[1];for(let n=h;n<k;n+=d){const a=r+n*s[2];for(let n=l;n<f;n+=p){const r=e[a+n*s[3]+t];if("max"===o&&r>T?T=r:"avg"===o&&(N+=r,x++),isNaN(T))break}if(isNaN(T))break}if(isNaN(T))break}S[g+t]="avg"===o?N/Math.max(x,1):T}}}}return y}const Tn={kernelName:t.AvgPool,backendName:"cpu",kernelFunc:function(e){const{inputs:n,backend:a,attrs:s}=e,{x:o}=n;r(o,"avgPool");const{filterSize:i,strides:l,pad:u,dimRoundingMode:c}=s;t.util.assert(t.backend_util.eitherStridesOrDilationsAreOne(l,1),(()=>`Error in avgPool: Either strides or dilations must be 1. Got strides ${l} and dilations '1'`));const d=t.backend_util.computePool2DInfo(o.shape,i,l,1,u,c);let p;if(1===d.filterWidth&&1===d.filterHeight&&t.util.arraysEqual(d.inShape,d.outShape))p=f({inputs:{x:o},backend:a});else{const e=a.data.get(o.dataId).values,n=t.util.computeStrides(o.shape),s=In(e,o.shape,o.dtype,n,d,"avg");p=a.makeTensorInfo(d.outShape,o.dtype,s.values)}return p}};const Nn={kernelName:t.AvgPool3D,backendName:"cpu",kernelFunc:function(e){const{inputs:n,backend:a,attrs:s}=e,{x:o}=n,{filterSize:i,strides:l,pad:u,dimRoundingMode:c,dataFormat:d}=s;r(o,"avgPool3d");const p=t.backend_util.computePool3DInfo(o.shape,i,l,1,u,c,d),h=Sn(a.data.get(o.dataId).values,o.shape,o.dtype,t.util.computeStrides(o.shape),p,"avg");return a.makeTensorInfo(h.shape,"float32",h.values)}};const xn={kernelName:t.AvgPool3DGrad,backendName:"cpu",kernelFunc:function(e){const{inputs:n,backend:a,attrs:s}=e,{dy:o,input:i}=n,{filterSize:l,strides:u,pad:c,dimRoundingMode:d}=s;r([o,i],"avgPool3DGrad");const p=t.backend_util.computePool3DInfo(i.shape,l,u,1,c,d),h=p.strideDepth,f=p.strideHeight,m=p.strideWidth,k=p.filterDepth,g=p.filterHeight,b=p.filterWidth,I=p.dilationDepth,y=p.dilationHeight,S=p.dilationWidth,T=p.effectiveFilterDepth,N=p.effectiveFilterHeight,x=p.effectiveFilterWidth,v=T-1-p.padInfo.front,F=x-1-p.padInfo.left,w=N-1-p.padInfo.top,M=t.buffer(i.shape,"float32"),A=1/(k*g*b),D=a.bufferSync(o);for(let e=0;e<p.batchSize;++e)for(let t=0;t<p.inChannels;++t)for(let n=0;n<p.inDepth;++n)for(let a=0;a<p.inHeight;++a)for(let s=0;s<p.inWidth;++s){const r=n-v,o=a-w,i=s-F;let l=0;for(let n=0;n<T;n+=I){const a=(r+n)/h;if(!(a<0||a>=p.outDepth||Math.floor(a)!==a))for(let n=0;n<N;n+=y){const s=(o+n)/f;if(!(s<0||s>=p.outHeight||Math.floor(s)!==s))for(let n=0;n<x;n+=S){const r=(i+n)/m;if(r<0||r>=p.outWidth||Math.floor(r)!==r)continue;l+=D.get(e,a,s,r,t)}}}M.set(l*A,e,n,a,s,t)}return a.makeTensorInfo(M.shape,M.dtype,M.values)}};const vn={kernelName:t.AvgPoolGrad,backendName:"cpu",kernelFunc:function(e){const{inputs:n,backend:a,attrs:s}=e,{dy:o,input:i}=n,l=i;r([o,i],"avgPoolGrad");const{filterSize:u,strides:c,pad:d}=s,p=t.backend_util.computePool2DInfo(l.shape,u,c,1,d),h=p.strideHeight,f=p.strideWidth,m=p.filterHeight,k=p.filterWidth,g=p.dilationHeight,b=p.dilationWidth,I=p.effectiveFilterHeight,y=p.effectiveFilterWidth,S=y-1-p.padInfo.left,T=I-1-p.padInfo.top,N=t.buffer(l.shape,"float32"),x=1/(m*k),v=a.data.get(o.dataId).values,F=t.buffer(o.shape,"float32",v);for(let e=0;e<p.batchSize;++e)for(let t=0;t<p.inChannels;++t)for(let n=0;n<p.inHeight;++n)for(let a=0;a<p.inWidth;++a){const s=n-T,r=a-S;let o=0;for(let n=0;n<I;n+=g){const a=(s+n)/h;if(!(a<0||a>=p.outHeight||Math.floor(a)!==a))for(let n=0;n<y;n+=b){const s=(r+n)/f;if(s<0||s>=p.outWidth||Math.floor(s)!==s)continue;o+=F.get(e,a,s,t)}}N.set(o*x,e,n,a,t)}return a.makeTensorInfo(N.shape,N.dtype,N.values)}};const Fn={kernelName:t.FusedBatchNorm,backendName:"cpu",kernelFunc:function(e){const{inputs:n,backend:a,attrs:s}=e,{x:o,scale:i,offset:l,mean:u,variance:c}=n;t.util.assert(u.shape.length===c.shape.length,(()=>"Batch normalization gradient requires mean and variance to have equal ranks.")),t.util.assert(null==l||u.shape.length===l.shape.length,(()=>"Batch normalization gradient requires mean and offset to have equal ranks.")),t.util.assert(null==i||u.shape.length===i.shape.length,(()=>"Batch normalization gradient requires mean and scale to have equal ranks.")),r([o,u,c,i,l],"batchNorm");let{varianceEpsilon:d}=s;null==d&&(d=.001);const p=a.data.get(o.dataId).values,h=a.data.get(u.dataId).values,f=a.data.get(c.dataId).values,m=i?a.data.get(i.dataId).values:new Float32Array([1]),k=l?a.data.get(l.dataId).values:new Float32Array([0]),g=new Float32Array(p.length),b=k.length,I=m.length,y=f.length,S=h.length;let T=0,N=0,x=0,v=0;for(let e=0;e<p.length;++e)g[e]=k[T++]+(p[e]-h[N++])*m[x++]/Math.sqrt(f[v++]+d),T>=b&&(T=0),N>=S&&(N=0),x>=I&&(x=0),v>=y&&(v=0);return a.makeTensorInfo(o.shape,o.dtype,g)}};const wn={kernelName:t.BatchToSpaceND,backendName:"cpu",kernelFunc:function(e){const{inputs:n,backend:a,attrs:s}=e,{x:o}=n,{blockShape:i,crops:l}=s;r([o],"batchToSpaceND");const u=i.reduce(((e,t)=>e*t)),c=t.backend_util.getReshaped(o.shape,i,u),d=t.backend_util.getPermuted(c.length,i.length),p=t.backend_util.getReshapedPermuted(o.shape,i,u),h=t.backend_util.getSliceBeginCoords(l,i.length),f=t.backend_util.getSliceSize(p,l,i.length),m=Zt({inputs:{x:o},backend:a,attrs:{shape:c}}),k=We({inputs:{x:m},backend:a,attrs:{perm:d}}),g=Zt({inputs:{x:k},backend:a,attrs:{shape:p}}),b=rt({inputs:{x:g},backend:a,attrs:{begin:h,size:f}});return a.disposeIntermediateTensorInfo(m),a.disposeIntermediateTensorInfo(k),a.disposeIntermediateTensorInfo(g),b}};const Mn={kernelName:t.Bincount,backendName:"cpu",kernelFunc:function(e){const{inputs:t,backend:n,attrs:a}=e,{x:s,weights:r}=t,{size:o}=a,i=w(n.data.get(s.dataId).values,n.data.get(r.dataId).values,r.dtype,r.shape,o);return n.makeTensorInfo([o],r.dtype,i)}};const An={kernelName:t.BroadcastArgs,backendName:"cpu",kernelFunc:function(e){const{inputs:n,backend:a}=e,{s0:s,s1:r}=n,o=a.data.get(s.dataId).values,i=a.data.get(r.dataId).values,l=t.backend_util.assertAndGetBroadcastShape(Array.from(o),Array.from(i));return a.makeTensorInfo([l.length],"int32",Int32Array.from(l))}},Dn=z(t.ClipByValue,((e,t)=>{const n=t;return e>n.clipValueMax?n.clipValueMax:e<n.clipValueMin?n.clipValueMin:e})),En={kernelName:t.ClipByValue,backendName:"cpu",kernelFunc:Dn},_n={kernelName:t.ComplexAbs,backendName:"cpu",kernelFunc:e=>{const{x:n}=e.inputs,a=e.backend,s=new Float32Array(t.util.sizeFromShape(n.shape)),r=a.data.get(n.dataId),o=r.complexTensorInfos.real,i=r.complexTensorInfos.imag,l=a.data.get(o.dataId).values,u=a.data.get(i.dataId).values;for(let e=0;e<l.length;e++){const t=l[e],n=u[e];s[e]=Math.hypot(t,n)}return a.makeOutput(s,n.shape,"float32")}};function zn(e){const{inputs:t,backend:n}=e,{input:a}=t,s=n.data.get(a.dataId).complexTensorInfos.imag,r=n.data.get(s.dataId).values;return n.makeTensorInfo(s.shape,s.dtype,r)}const Rn={kernelName:t.Imag,backendName:"cpu",kernelFunc:zn};function Wn(e){const{inputs:n,backend:a,attrs:s}=e,{axis:r}=s,o=t.util.parseAxisParam(r,n[0].shape)[0],i=n.map((e=>e.shape));t.backend_util.assertParamsConsistent(i,o);let l=t.backend_util.computeOutShape(n.map((e=>e.shape)),o);if(0===t.util.sizeFromShape(l))return a.makeTensorInfo(l,n[0].dtype,[]);const u=n.filter((e=>t.util.sizeFromShape(e.shape)>0));if(1===u.length)return f({inputs:{x:u[0]},backend:a});if("complex64"===u[0].dtype){const e=u.map((e=>k({inputs:{input:e},backend:a}))),t=u.map((e=>zn({inputs:{input:e},backend:a}))),n=Wn({inputs:e,backend:a,attrs:{axis:o}}),s=Wn({inputs:t,backend:a,attrs:{axis:o}}),r=d({inputs:{real:n,imag:s},backend:a});return e.forEach((e=>a.disposeIntermediateTensorInfo(e))),t.forEach((e=>a.disposeIntermediateTensorInfo(e))),a.disposeIntermediateTensorInfo(n),a.disposeIntermediateTensorInfo(s),r}const c=u.map((e=>{const n=t.util.sizeFromShape(e.shape.slice(o));return Zt({inputs:{x:e},backend:a,attrs:{shape:[-1,n]}})})),p=c.map((e=>({vals:a.data.get(e.dataId).values,shape:e.shape})));l=t.backend_util.computeOutShape(c.map((e=>e.shape)),1);const h=1===c[0].shape[0],m=H(p,l,n[0].dtype,h),g=t.backend_util.computeOutShape(u.map((e=>e.shape)),o),b=a.makeTensorInfo(g,n[0].dtype,m);return c.forEach((e=>a.disposeIntermediateTensorInfo(e))),b}const Pn={kernelName:t.Concat,backendName:"cpu",kernelFunc:Wn};function Cn(e){const{inputs:n,backend:a,attrs:s}=e,{x:o,filter:i}=n,{strides:l,pad:u,dataFormat:c,dilations:d,dimRoundingMode:p}=s;r([o,i],"conv2d");const h=t.backend_util.convertConv2DDataFormat(c),f=t.backend_util.computeConv2DInfo(o.shape,i.shape,l,d,u,p,!1,h),m=f.filterHeight,k=f.filterWidth,g=f.dilationHeight,b=f.dilationWidth,I=f.padInfo.left,y=f.padInfo.top,S="channelsLast"===f.dataFormat,T=new t.TensorBuffer(f.outShape,o.dtype),N=t.util.computeStrides(o.shape),x=t.util.computeStrides(i.shape),v=N[0],F=S?N[1]:N[2],w=S?N[2]:1,M=S?1:N[1],A=T.strides[0],D=S?T.strides[1]:T.strides[2],E=S?T.strides[2]:1,_=S?1:T.strides[1],z=a.data.get(o.dataId).values,R=a.data.get(i.dataId).values,W=T.values;for(let e=0;e<f.batchSize;++e){const t=e*v,n=e*A;for(let e=0;e<f.outHeight;++e){const a=n+e*D,s=e*f.strideHeight-y;for(let e=0;e<m;++e){const n=s+e*g;if(n<0||n>=f.inHeight)continue;const r=e*x[0],o=t+n*F;for(let e=0;e<f.outWidth;++e){const t=a+e*E,n=e*f.strideWidth-I;for(let e=0;e<k;++e){const a=n+e*b;if(a<0||a>=f.inWidth)continue;const s=o+a*w;let i=r+e*x[1];for(let e=0;e<f.inChannels;++e){const n=z[s+e*M];for(let e=0;e<f.outChannels;++e)W[t+e*_]+=n*R[i+e];i+=f.outChannels}}}}}}return a.makeTensorInfo(T.shape,T.dtype,W)}const Hn={kernelName:t.Conv2D,backendName:"cpu",kernelFunc:Cn};const On={kernelName:t.Conv2DBackpropFilter,backendName:"cpu",kernelFunc:function(e){const{inputs:n,backend:a,attrs:s}=e,{x:o,dy:i}=n,{strides:l,pad:u,dataFormat:c,dimRoundingMode:d,filterShape:p}=s;r([o,i],"conv2dBackpropFilter");const h=t.backend_util.convertConv2DDataFormat(c),f=t.backend_util.computeConv2DInfo(o.shape,p,l,1,u,d,!1,h),{strideHeight:m,strideWidth:k,filterHeight:g,filterWidth:b}=f,I="channelsLast"===f.dataFormat,y=new t.TensorBuffer(f.filterShape,"float32"),S=f.padInfo.left,T=f.padInfo.top,N=a.data.get(o.dataId).values,x=a.data.get(i.dataId).values,v=new t.TensorBuffer(o.shape,o.dtype,N),F=new t.TensorBuffer(i.shape,i.dtype,x);for(let e=0;e<g;++e){const t=Math.max(0,Math.ceil((T-e)/m)),n=Math.min(f.outHeight,(f.inHeight+T-e)/m);for(let a=0;a<b;++a){const s=Math.max(0,Math.ceil((S-a)/k)),r=Math.min(f.outWidth,(f.inWidth+S-a)/k);for(let o=0;o<f.inChannels;++o)for(let i=0;i<f.outChannels;++i){let l=0;for(let u=0;u<f.batchSize;++u)for(let c=t;c<n;++c){const t=e+c*m-T;for(let e=s;e<r;++e){const n=a+e*k-S;l+=I?v.get(u,t,n,o)*F.get(u,c,e,i):v.get(u,o,t,n)*F.get(u,i,c,e)}}y.set(l,e,a,o,i)}}}return a.makeTensorInfo(y.shape,y.dtype,y.values)}};const Bn={kernelName:t.Conv2DBackpropInput,backendName:"cpu",kernelFunc:function(e){const{inputs:n,backend:a,attrs:s}=e,{dy:o,filter:i}=n,{inputShape:l,strides:u,pad:c,dataFormat:d,dimRoundingMode:p}=s;r([o,i],"conv2dBackpropInput");const h=t.util.computeStrides(i.shape),f=t.util.computeStrides(o.shape);let m=t.backend_util.convertConv2DDataFormat(d);const k=t.backend_util.computeConv2DInfo(l,i.shape,u,1,c,p,!1,m),g=new t.TensorBuffer(k.inShape,"float32"),b=g.values,I=a.data.get(o.dataId).values,y=a.data.get(i.dataId).values,[S,T,N]=h,{batchSize:x,filterHeight:v,filterWidth:F,inChannels:w,inHeight:M,inWidth:A,outChannels:D,outHeight:E,outWidth:_,strideHeight:z,strideWidth:R}=k;m=k.dataFormat;const W=v-1-k.padInfo.top,P=F-1-k.padInfo.left,C="channelsLast"===m,H=g.strides[0],O=C?g.strides[1]:g.strides[2],B=C?g.strides[2]:1,$=C?1:g.strides[1],V=f[0],G=C?f[1]:f[2],L=C?f[2]:1,q=C?1:f[1];for(let e=0;e<x;++e)for(let t=0;t<w;++t)for(let n=0;n<M;++n){const a=n-W,s=Math.max(0,Math.ceil(a/z)),r=Math.min(E,(v+a)/z);for(let o=0;o<A;++o){const i=o-P,l=Math.max(0,Math.ceil(i/R)),u=Math.min(_,(F+i)/R);let c=0;for(let n=s;n<r;++n){const s=n*z-a;for(let a=l;a<u;++a){const r=V*e+G*n+L*a,o=S*(v-1-s)+T*(F-1-(a*R-i))+N*t;for(let e=0;e<D;++e){c+=I[r+q*e]*y[o+e]}}}b[H*e+O*n+B*o+$*t]=c}}return a.makeTensorInfo(g.shape,g.dtype,g.values)}};const $n={kernelName:t.Conv3D,backendName:"cpu",kernelFunc:function(e){const{inputs:n,backend:a,attrs:s}=e,{x:o,filter:i}=n,{strides:l,pad:u,dilations:c}=s;r([o,i],"conv3d");const d=t.backend_util.computeConv3DInfo(o.shape,i.shape,l,c,u),{filterDepth:p,filterHeight:h,filterWidth:f,dilationDepth:m,dilationHeight:k,dilationWidth:g,padInfo:b}=d,I=b.front,y=b.left,S=b.top,T=new t.TensorBuffer(d.outShape,o.dtype),N=a.data.get(o.dataId).values,x=a.data.get(i.dataId).values,v=T.values,F=t.util.computeStrides(o.shape),w=t.util.computeStrides(i.shape);for(let e=0;e<d.batchSize;++e){const t=e*F[0],n=e*T.strides[0];for(let e=0;e<d.outDepth;++e){const a=n+e*T.strides[1],s=e*d.strideDepth-I;for(let e=0;e<p;++e){const n=s+e*m;if(n<0||n>=d.inDepth)continue;const r=e*w[0],o=t+n*F[1];for(let e=0;e<d.outHeight;++e){const t=a+e*T.strides[2],n=e*d.strideHeight-S;for(let e=0;e<h;++e){const a=n+e*k;if(a<0||a>=d.inHeight)continue;const s=r+e*w[1],i=o+a*F[2];for(let e=0;e<d.outWidth;++e){const n=t+e*d.outChannels,a=e*d.strideWidth-y;for(let e=0;e<f;++e){const t=a+e*g;if(t<0||t>=d.inWidth)continue;const r=s+e*w[2],o=i+t*d.inChannels;let l=r;for(let e=0;e<d.inChannels;++e){const t=N[o+e];for(let e=0;e<d.outChannels;++e)v[n+e]+=t*x[l+e];l+=d.outChannels}}}}}}}}return a.makeTensorInfo(T.shape,T.dtype,T.values)}};const Vn={kernelName:t.Conv3DBackpropFilterV2,backendName:"cpu",kernelFunc:function(e){const{inputs:n,backend:a,attrs:s}=e,{x:o,dy:i}=n,{strides:l,pad:u,filterShape:c}=s;r([o,i],"conv3dBackpropFilterV2");const d=t.util.computeStrides(o.shape),p=t.util.computeStrides(i.shape),h=t.backend_util.computeConv3DInfo(o.shape,c,l,1,u),f=h.strideDepth,m=h.strideHeight,k=h.strideWidth,g=h.filterDepth,b=h.filterHeight,I=h.filterWidth,y=new t.TensorBuffer(h.filterShape,"float32"),S=y.values,[T,N,x,v]=y.strides,F=a.data.get(i.dataId).values,[w,M,A,D]=p,E=a.data.get(o.dataId).values,[_,z,R,W]=d,P=h.padInfo.front,C=h.padInfo.left,H=h.padInfo.top;for(let e=0;e<g;++e){const t=Math.max(0,Math.ceil((P-e)/f)),n=Math.min(h.outDepth,(h.inDepth+P-e)/f),a=e*T;for(let s=0;s<b;++s){const r=Math.max(0,Math.ceil((H-s)/m)),o=Math.min(h.outHeight,(h.inHeight+H-s)/m),i=s*N+a;for(let a=0;a<I;++a){const l=Math.max(0,Math.ceil((C-a)/k)),u=Math.min(h.outWidth,(h.inWidth+C-a)/k),c=a*x+i;for(let i=0;i<h.inChannels;++i){const d=i*v+c;for(let c=0;c<h.outChannels;++c){let p=0;for(let d=0;d<h.batchSize;++d){const h=d*_,g=d*w;for(let d=t;d<n;++d){const t=(e+d*f-P)*z+h,n=d*M+g;for(let e=r;e<o;++e){const r=(s+e*m-H)*R+t,o=e*A+n;for(let e=l;e<u;++e){const t=e*D+o;p+=E[(a+e*k-C)*W+r+i]*F[t+c]}}}}S[d+c]=p}}}}}return a.makeTensorInfo(y.shape,y.dtype,y.values)}};const Gn={kernelName:t.Conv3DBackpropInputV2,backendName:"cpu",kernelFunc:function(e){const{inputs:n,backend:a,attrs:s}=e,{dy:o,filter:i}=n,{pad:l,strides:u,inputShape:c}=s;r([o],"conv3dBackpropInputV2");const d=t.util.computeStrides(o.shape),p=t.util.computeStrides(i.shape),h=t.backend_util.computeConv3DInfo(c,i.shape,u,1,l),f=new t.TensorBuffer(h.inShape,"float32"),m=f.values,[k,g,b,I]=f.strides,y=a.data.get(o.dataId).values,[S,T,N,x]=d,v=a.data.get(i.dataId).values,[F,w,M,A]=p,{batchSize:D,filterDepth:E,filterHeight:_,filterWidth:z,inChannels:R,inDepth:W,inHeight:P,inWidth:C,outChannels:H,outDepth:O,outHeight:B,outWidth:$,strideDepth:V,strideHeight:G,strideWidth:L}=h,q=E-1-h.padInfo.front,U=_-1-h.padInfo.top,Z=z-1-h.padInfo.left;for(let e=0;e<D;++e)for(let t=0;t<R;++t)for(let n=0;n<W;++n){const a=n-q,s=Math.max(0,Math.ceil(a/V)),r=Math.min(O,(E+a)/V);for(let o=0;o<P;++o){const i=o-U,l=Math.max(0,Math.ceil(i/G)),u=Math.min(B,(_+i)/G);for(let c=0;c<C;++c){const d=c-Z,p=Math.max(0,Math.ceil(d/L)),h=Math.min($,(z+d)/L);let f=0;for(let n=s;n<r;++n){const s=n*V-a;for(let a=l;a<u;++a){const r=a*G-i;for(let o=p;o<h;++o){const i=S*e+T*n+N*a+x*o,l=F*(E-1-s)+w*(_-1-r)+M*(z-1-(o*L-d))+A*t;for(let e=0;e<H;++e){f+=y[i+e]*v[l+e]}}}}m[k*e+g*n+b*o+I*c+t]=f}}}return a.makeTensorInfo(f.shape,f.dtype,f.values)}},Ln=z(t.Cos,(e=>Math.cos(e))),qn={kernelName:t.Cos,backendName:"cpu",kernelFunc:Ln},Un=z(t.Cosh,(e=>Math.cosh(e))),Zn={kernelName:t.Cosh,backendName:"cpu",kernelFunc:Un};const jn={kernelName:t.CropAndResize,backendName:"cpu",kernelFunc:function(e){const{inputs:n,backend:a,attrs:s}=e,{image:r,boxes:o,boxInd:i}=n,{cropSize:l,method:u,extrapolationValue:c}=s,[d,p,h,f]=r.shape,m=o.shape[0],[k,g]=l,b=t.buffer([m,k,g,f],"float32"),I=a.data.get(o.dataId).values,y=a.data.get(i.dataId).values,S=a.data.get(r.dataId).values,T=t.util.computeStrides(r.shape),N=t.util.computeStrides(b.shape);for(let e=0;e<m;e++){const t=4*e,n=I[t],a=I[t+1],s=I[t+2],r=I[t+3],o=y[e];if(o>=d)continue;const i=k>1?(s-n)*(p-1)/(k-1):0,l=g>1?(r-a)*(h-1)/(g-1):0;for(let t=0;t<k;t++){const d=k>1?n*(p-1)+t*i:.5*(n+s)*(p-1);if(d<0||d>p-1)for(let n=0;n<g;n++)for(let a=0;a<f;a++){const s=a+n*N[2]+t*N[1]+e*N[0];b.values[s]=c}else if("bilinear"===u){const n=Math.floor(d),s=Math.ceil(d),i=d-n;for(let u=0;u<g;u++){const d=g>1?a*(h-1)+u*l:.5*(a+r)*(h-1);if(d<0||d>h-1){for(let n=0;n<f;n++){const a=n+u*N[2]+t*N[1]+e*N[0];b.values[a]=c}continue}const p=Math.floor(d),m=Math.ceil(d),k=d-p;for(let a=0;a<f;a++){let r=a+p*T[2]+n*T[1]+o*T[0];const l=S[r];r=a+m*T[2]+n*T[1]+o*T[0];const c=S[r];r=a+p*T[2]+s*T[1]+o*T[0];const d=S[r];r=a+m*T[2]+s*T[1]+o*T[0];const h=l+(c-l)*k,f=d+(S[r]-d)*k;r=a+u*N[2]+t*N[1]+e*N[0],b.values[r]=h+(f-h)*i}}}else for(let n=0;n<g;++n){const s=g>1?a*(h-1)+n*l:.5*(a+r)*(h-1);if(s<0||s>h-1){for(let a=0;a<f;a++){const s=a+n*N[2]+t*N[1]+e*N[0];b.values[s]=c}continue}const i=Math.round(s),u=Math.round(d);for(let a=0;a<f;a++){const s=a+i*T[2]+u*T[1]+o*T[0],r=a+n*N[2]+t*N[1]+e*N[0];b.values[r]=S[s]}}}}return a.makeTensorInfo(b.shape,b.dtype,b.values)}};const Kn={kernelName:t.Cumprod,backendName:"cpu",kernelFunc:function(e){const{inputs:n,backend:a,attrs:s}=e,{x:o}=n,{axis:i,exclusive:l,reverse:u}=s;r(o,"cumprod");const c=t.backend_util.getAxesPermutation([i],o.shape.length);let d=o;null!=c&&(d=We({inputs:{x:o},backend:a,attrs:{perm:c}}));const p=t.backend_util.getInnerMostAxes(1,o.shape.length)[0];if(p!==d.shape.length-1)throw new Error(`backend.cumprod in CPU expects an inner-most axis=${d.shape.length-1} but got axis=${p}`);const h=t.upcastType(d.dtype,"int32"),f=t.util.makeOnesTypedArray(t.util.sizeFromShape(d.shape),h),m=a.data.get(d.dataId).values,k=d.shape[d.shape.length-1],g=u?(e,t)=>e+k-t-1:(e,t)=>e+t;for(let e=0;e<m.length;e+=k)for(let t=0;t<k;t++){const n=g(e,t);if(0===t)f[n]=l?1:m[n];else{const a=g(e,t-1);f[n]=l?m[a]*f[a]:m[n]*f[a]}}const b=a.makeTensorInfo(d.shape,h,f);if(null!=c){const e=We({inputs:{x:b},backend:a,attrs:{perm:t.backend_util.getUndoAxesPermutation(c)}});return a.disposeIntermediateTensorInfo(b),a.disposeIntermediateTensorInfo(d),e}return b}};const Yn={kernelName:t.Cumsum,backendName:"cpu",kernelFunc:function(e){const{inputs:n,backend:a,attrs:s}=e,{x:o}=n,{axis:i,exclusive:l,reverse:u}=s;r(o,"cumsum");const c=t.backend_util.getAxesPermutation([i],o.shape.length);let d=o;null!=c&&(d=We({inputs:{x:o},backend:a,attrs:{perm:c}}));const p=t.backend_util.getInnerMostAxes(1,o.shape.length)[0];if(p!==d.shape.length-1)throw new Error(`backend.cumsum in CPU expects an inner-most axis=${d.shape.length-1} but got axis=${p}`);const h=t.upcastType(d.dtype,"int32"),f=t.util.makeZerosTypedArray(t.util.sizeFromShape(d.shape),h),m=a.data.get(d.dataId).values,k=d.shape[d.shape.length-1],g=u?(e,t)=>e+k-t-1:(e,t)=>e+t;for(let e=0;e<m.length;e+=k)for(let t=0;t<k;t++){const n=g(e,t);if(0===t)f[n]=l?0:m[n];else{const a=g(e,t-1);f[n]=l?m[a]+f[a]:m[n]+f[a]}}const b=a.makeTensorInfo(d.shape,h,f);if(null!=c){const e=We({inputs:{x:b},backend:a,attrs:{perm:t.backend_util.getUndoAxesPermutation(c)}});return a.disposeIntermediateTensorInfo(b),a.disposeIntermediateTensorInfo(d),e}return b}};const Jn={kernelName:t.DenseBincount,backendName:"cpu",kernelFunc:function(e){const{inputs:t,backend:n,attrs:a}=e,{x:s,weights:r}=t,{size:o,binaryOutput:i}=a;if(1===s.shape.length){const e=w(n.data.get(s.dataId).values,n.data.get(r.dataId).values,r.dtype,r.shape,o);return n.makeTensorInfo([o],r.dtype,e)}if(2===s.shape.length){const e=M(n.bufferSync(s),n.bufferSync(r),o,i);return n.makeTensorInfo(e.shape,r.dtype,e.values)}throw new Error(`Error in denseBincount: input must be at most rank 2, but got rank${s.shape.length}.`)}};const Qn={kernelName:t.DepthToSpace,backendName:"cpu",kernelFunc:function(e){const{inputs:n,backend:a,attrs:s}=e,{x:r}=n,{blockSize:o,dataFormat:i}=s;t.util.assert("NHWC"===i,(()=>`Only NHWC dataFormat supported on CPU for depthToSpace. Got ${i}`));const l=r.shape[0],u=r.shape[1],c=r.shape[2],d=r.shape[3],p=u*o,h=c*o,f=d/(o*o),m=a.data.get(r.dataId).values,k=new Float32Array(l*p*h*f);let g=0;for(let e=0;e<l;++e)for(let t=0;t<p;++t){const n=Math.floor(t/o),a=t%o;for(let t=0;t<h;++t){const s=Math.floor(t/o),r=(a*o+t%o)*f;for(let t=0;t<f;++t){const a=t+r+d*(s+c*(n+u*e));k[g++]=m[a]}}}return a.makeTensorInfo([l,p,h,f],r.dtype,k)}};function Xn(e){const{inputs:n,backend:a,attrs:s}=e,{x:o,filter:i}=n,{strides:l,pad:u,dilations:c,dimRoundingMode:d}=s;r([o,i],"depthwiseConv2DNative");const p=t.util.computeStrides(o.shape),h=t.util.computeStrides(i.shape);let f=c;null==f&&(f=[1,1]),t.util.assert(t.backend_util.eitherStridesOrDilationsAreOne(l,f),(()=>`Error in depthwiseConv2d: Either strides or dilations must be 1. Got strides ${l} and dilations '${f}'`));const m=t.backend_util.computeConv2DInfo(o.shape,i.shape,l,f,u,d,!0),{filterHeight:k,filterWidth:g,dilationHeight:b,dilationWidth:I,padInfo:y}=m,S=y.left,T=y.top,N=m.outChannels/m.inChannels,x=new t.TensorBuffer(m.outShape,o.dtype),v=a.data.get(o.dataId).values,F=a.data.get(i.dataId).values,w=x.values;for(let e=0;e<m.batchSize;++e){const t=e*p[0],n=e*x.strides[0];for(let e=0;e<m.outHeight;++e){const a=n+e*x.strides[1],s=e*m.strideHeight-T;for(let e=0;e<k;++e){const n=s+e*b;if(n<0||n>=m.inHeight)continue;const r=e*h[0],o=t+n*p[1];for(let e=0;e<m.outWidth;++e){const t=a+e*x.strides[2],n=e*m.strideWidth-S;for(let e=0;e<g;++e){const a=n+e*I;if(a<0||a>=m.inWidth)continue;const s=r+e*h[1],i=o+a*m.inChannels;let l=t,u=s;for(let e=0;e<m.inChannels;++e){const t=v[i+e];for(let e=0;e<N;++e)w[l+e]+=t*F[u+e];l+=N,u+=N}}}}}}return a.makeTensorInfo(x.shape,x.dtype,x.values)}const ea={kernelName:t.DepthwiseConv2dNative,backendName:"cpu",kernelFunc:Xn};const ta={kernelName:t.DepthwiseConv2dNativeBackpropFilter,backendName:"cpu",kernelFunc:function(e){const{inputs:n,backend:a,attrs:s}=e,{x:o,dy:i}=n,{strides:l,dilations:u,pad:c,dimRoundingMode:d,filterShape:p}=s;r([o,i],"depthwiseConv2dNativeBackpropFilter");const h=t.backend_util.computeConv2DInfo(o.shape,p,l,u,c,d,!0),{strideHeight:f,strideWidth:m,filterHeight:k,filterWidth:g}=h,b=new t.TensorBuffer(h.filterShape,"float32"),I=h.padInfo.left,y=h.padInfo.top,S=h.outChannels/h.inChannels,T=a.data.get(o.dataId).values,N=new t.TensorBuffer(o.shape,o.dtype,T),x=a.data.get(i.dataId).values,v=new t.TensorBuffer(i.shape,i.dtype,x);for(let e=0;e<k;++e){const t=Math.max(0,Math.ceil((y-e)/f)),n=Math.min(h.outHeight,(h.inHeight+y-e)/f);for(let a=0;a<g;++a){const s=Math.max(0,Math.ceil((I-a)/m)),r=Math.min(h.outWidth,(h.inWidth+I-a)/m);for(let o=0;o<h.outChannels;++o){const i=Math.trunc(o/S),l=o%S;let u=0;for(let l=0;l<h.batchSize;++l)for(let c=t;c<n;++c){const t=e+c*f-y;for(let e=s;e<r;++e){const n=a+e*m-I;u+=N.get(l,t,n,i)*v.get(l,c,e,o)}}b.set(u,e,a,i,l)}}}return a.makeTensorInfo(b.shape,b.dtype,b.values)}};const na={kernelName:t.DepthwiseConv2dNativeBackpropInput,backendName:"cpu",kernelFunc:function(e){const{inputs:n,backend:a,attrs:s}=e,{dy:o,filter:i}=n,{strides:l,dilations:u,pad:c,dimRoundingMode:d,inputShape:p}=s;r([o,i],"depthwiseConv2DNativeBackpropInput");const h=t.util.computeStrides(o.shape),f=t.util.computeStrides(i.shape),m=t.backend_util.computeConv2DInfo(p,i.shape,l,u,c,d,!0),k=new t.TensorBuffer(m.inShape,"float32"),g=k.values,[b,I,y]=k.strides,S=a.data.get(o.dataId).values,[T,N,x]=h,v=a.data.get(i.dataId).values,[F,w,M]=f,{batchSize:A,filterHeight:D,filterWidth:E,inChannels:_,inHeight:z,inWidth:R,outChannels:W,outHeight:P,outWidth:C,strideHeight:H,strideWidth:O}=m,B=D-1-m.padInfo.top,$=E-1-m.padInfo.left,V=W/_;for(let e=0;e<A;++e)for(let t=0;t<_;++t)for(let n=0;n<z;++n){const a=n-B,s=Math.max(0,Math.ceil(a/H)),r=Math.min(P,(D+a)/H);for(let o=0;o<R;++o){const i=o-$,l=Math.max(0,Math.ceil(i/O)),u=Math.min(C,(E+i)/O);let c=0;for(let n=s;n<r;++n){const s=n*H-a;for(let a=l;a<u;++a){const r=T*e+N*n+x*a,o=F*(D-1-s)+w*(E-1-(a*O-i))+M*t;for(let e=0;e<V;++e){c+=S[r+(t*V+e)]*v[o+e]}}}g[b*e+I*n+y*o+t]=c}}return a.makeTensorInfo(k.shape,k.dtype,k.values)}};const aa={kernelName:t.Diag,backendName:"cpu",kernelFunc:function(e){const{inputs:n,backend:a}=e,{x:s}=n,r=t.util.sizeFromShape(s.shape),o=a.data.get(s.dataId).values,i=t.buffer([r,r],s.dtype),l=i.values;for(let e=0;e<o.length;e++)l[e*r+e]=o[e];const u=[...s.shape,...s.shape];return a.makeTensorInfo(u,i.dtype,i.values)}},sa={kernelName:t.Dilation2D,backendName:"cpu",kernelFunc:({inputs:e,backend:n,attrs:a})=>{const{x:s,filter:r}=e,{strides:o,pad:i,dilations:l}=a,u=n,c=u.data.get(s.dataId).values,d=s.shape.length,p=u.data.get(r.dataId).values,h=r.shape.length,{batchSize:f,inHeight:m,inWidth:k,inChannels:g,outHeight:b,outWidth:I,padInfo:y,strideHeight:S,strideWidth:T,filterHeight:N,filterWidth:x,dilationHeight:v,dilationWidth:F,outShape:w}=t.backend_util.computeDilation2DInfo(s.shape,r.shape,o,i,"NHWC",l),M=t.util.sizeFromShape(w),A=w.length,D=t.util.getArrayFromDType(s.dtype,M);for(let e=0;e<f;++e)for(let n=0;n<b;++n){const a=n*S-y.top;for(let o=0;o<I;++o){const i=o*T-y.left;for(let l=0;l<g;++l){let u=Number.MIN_SAFE_INTEGER;for(let n=0;n<N;++n){const o=a+n*v;if(o>=0&&o<m)for(let a=0;a<x;++a){const f=i+a*F;if(f>=0&&f<k){const i=t.util.locToIndex([e,o,f,l],d,t.util.computeStrides(s.shape)),m=t.util.locToIndex([n,a,l],h,t.util.computeStrides(r.shape)),k=c[i]+p[m];k>u&&(u=k)}}}D[t.util.locToIndex([e,n,o,l],A,t.util.computeStrides(w))]=u}}}return{dataId:u.write(t.util.toTypedArray(D,s.dtype),w,s.dtype),shape:w,dtype:s.dtype}}},ra={kernelName:t.Dilation2DBackpropFilter,backendName:"cpu",kernelFunc:({inputs:e,backend:n,attrs:a})=>{const{x:s,filter:r,dy:o}=e,{strides:i,pad:l,dilations:u}=a,c=n,d=t.util.toNestedArray(s.shape,c.data.get(s.dataId).values),p=t.util.toNestedArray(r.shape,c.data.get(r.dataId).values),{batchSize:h,inHeight:f,inWidth:m,inChannels:k,outHeight:g,outWidth:b,padInfo:I,strideHeight:y,strideWidth:S,filterHeight:T,filterWidth:N,dilationHeight:x,dilationWidth:v,outShape:F}=t.backend_util.computeDilation2DInfo(s.shape,r.shape,i,l,"NHWC",u);t.util.assert(o.rank===F.length,(()=>`Error in ${t.Dilation2DBackpropFilter}, dy must have the same rank as output ${F.length}, but got ${o.rank}`));const w=t.util.toNestedArray(F,c.data.get(o.dataId).values),M=t.util.makeZerosNestedTypedArray(r.shape,r.dtype);for(let e=0;e<h;++e)for(let t=0;t<g;++t){const n=t*y-I.top;for(let a=0;a<b;++a){const s=a*S-I.left;for(let r=0;r<k;++r){let o=Number.MIN_SAFE_INTEGER,i=0,l=0;for(let t=0;t<T;++t){const a=n+t*x;if(a>=0&&a<f)for(let n=0;n<N;++n){const u=s+n*v;if(u>=0&&u<m){const s=d[e][a][u][r]+p[t][n][r];s>o&&(o=s,i=t,l=n)}}}M[i][l][r]+=w[e][t][a][r]}}}return{dataId:c.write(t.util.toTypedArray(M,s.dtype),r.shape,r.dtype),shape:r.shape,dtype:r.dtype}}},oa={kernelName:t.Dilation2DBackpropInput,backendName:"cpu",kernelFunc:({inputs:e,backend:n,attrs:a})=>{const{x:s,filter:r,dy:o}=e,{strides:i,pad:l,dilations:u}=a,c=n,d=t.util.toNestedArray(s.shape,c.data.get(s.dataId).values),p=t.util.toNestedArray(r.shape,c.data.get(r.dataId).values),{batchSize:h,inHeight:f,inWidth:m,inChannels:k,outHeight:g,outWidth:b,padInfo:I,strideHeight:y,strideWidth:S,filterHeight:T,filterWidth:N,dilationHeight:x,dilationWidth:v,outShape:F}=t.backend_util.computeDilation2DInfo(s.shape,r.shape,i,l,"NHWC",u);t.util.assert(o.rank===F.length,(()=>`Error in ${t.Dilation2DBackpropInput}, dy must have the same rank as output ${F.length}, but got ${o.rank}`));const w=t.util.toNestedArray(F,c.data.get(o.dataId).values),M=t.util.makeZerosNestedTypedArray(s.shape,s.dtype);for(let e=0;e<h;++e)for(let t=0;t<g;++t){const n=t*y-I.top;for(let a=0;a<b;++a){const s=a*S-I.left;for(let r=0;r<k;++r){let o=Number.MIN_SAFE_INTEGER,i=n<0?0:n,l=s<0?0:s;for(let t=0;t<T;++t){const a=n+t*x;if(a>=0&&a<f)for(let n=0;n<N;++n){const u=s+n*v;if(u>=0&&u<m){const s=d[e][a][u][r]+p[t][n][r];s>o&&(o=s,i=a,l=u)}}}M[e][i][l][r]+=w[e][t][a][r]}}}return{dataId:c.write(t.util.toTypedArray(M,s.dtype),s.shape,s.dtype),shape:s.shape,dtype:s.dtype}}};const ia={kernelName:t.Draw,backendName:"cpu",kernelFunc:function(e){const{inputs:t,backend:n,attrs:a}=e,{image:s}=t,{canvas:r,options:o}=a,{contextOptions:i,imageOptions:l}=o||{},u=(null==l?void 0:l.alpha)||1,c=(null==i?void 0:i.contextType)||"2d";if("2d"!==c)throw new Error(`Context type ${i.contextType} is not supported by the CPU backend.`);const d=r.getContext(c,(null==i?void 0:i.contextAttributes)||{});if(null==d)throw new Error(`Could not get the context with ${c} type.`);const[p,h]=s.shape.slice(0,2),f=2===s.shape.length?1:s.shape[2],m=n.data.get(s.dataId).values,k="float32"===s.dtype?255:1,g=new Uint8ClampedArray(h*p*4);for(let e=0;e<p*h;++e){const t=[0,0,0,255*u];for(let n=0;n<f;n++){const a=m[e*f+n];if("float32"===s.dtype){if(a<0||a>1)throw new Error(`Tensor values for a float32 Tensor must be in the range [0 - 1] but encountered ${a}.`)}else if("int32"===s.dtype&&(a<0||a>255))throw new Error(`Tensor values for a int32 Tensor must be in the range [0 - 255] but encountered ${a}.`);1===f?(t[0]=a*k,t[1]=a*k,t[2]=a*k):t[n]=a*k}const n=4*e;g[n+0]=Math.round(t[0]),g[n+1]=Math.round(t[1]),g[n+2]=Math.round(t[2]),g[n+3]=Math.round(t[3])}r.width=h,r.height=p;const b=new ImageData(g,h,p);return d.putImageData(b,0,0),s}};function la(e){const{inputs:n,backend:a,attrs:s}=e,{x:o}=n,{axis:i,keepDims:l}=s;let u;r(o,"sum"),u="bool"===o.dtype?I({inputs:{x:o},backend:a,attrs:{dtype:"int32"}}):f({inputs:{x:o},backend:a});const c=u.shape.length,d=t.util.parseAxisParam(i,u.shape),p=t.backend_util.getAxesPermutation(d,c);let m=d,k=u;null!=p&&(k=We({inputs:{x:u},backend:a,attrs:{perm:p}}),m=t.backend_util.getInnerMostAxes(m.length,c)),t.backend_util.assertAxesAreInnerMostDims("sum",m,k.shape.length);const[g,b]=t.backend_util.computeOutAndReduceShapes(k.shape,m);let y=h(a,g,t.backend_util.upcastType(k.dtype,"int32"));const S=t.util.sizeFromShape(b),T=a.data.get(y.dataId).values,N=a.data.get(k.dataId).values;for(let e=0;e<T.length;++e){const t=e*S;let n=0;for(let e=0;e<S;++e)n+=N[t+e];T[e]=n}if(l){const e=y;y=Zt({inputs:{x:y},backend:a,attrs:{shape:t.backend_util.expandShapeToKeepDim(y.shape,d)}}),a.disposeIntermediateTensorInfo(e)}return a.disposeIntermediateTensorInfo(u),null!=p&&a.disposeIntermediateTensorInfo(k),y}const ua={kernelName:t.Sum,backendName:"cpu",kernelFunc:la};const ca={kernelName:t.Einsum,backendName:"cpu",kernelFunc:function(e){const{inputs:n,backend:a,attrs:s}=e,{equation:r}=s,o=n,{allDims:i,summedDims:l,idDims:u}=t.backend_util.decodeEinsumEquation(r,o.length);t.backend_util.checkEinsumDimSizes(i.length,u,o);const{path:c,steps:d}=t.backend_util.getEinsumComputePath(l,u),p=d.length;let h=null,f=i.length;const m=[];for(let e=0;e<p;++e){for(const n of d[e]){const{permutationIndices:e,expandDims:s}=t.backend_util.getEinsumPermutation(f,u[n]);let r;t.backend_util.isIdentityPermutation(e)?r=o[n]:(r=We({inputs:{x:o[n]},backend:a,attrs:{perm:e}}),m.push(r));const i=r.shape.slice();for(let e=0;e<s.length;++e)i.splice(s[e],0,1);t.util.arraysEqual(r.shape,i)||(r=Zt({inputs:{x:r},backend:a,attrs:{shape:i}}),m.push(r)),null===h?h=r:(h=we({inputs:{a:r,b:h},backend:a}),m.push(h))}e<p-1&&(c[e]>=0&&(h=la({inputs:{x:h},backend:a,attrs:{axis:c[e]-(i.length-f),keepDims:!1}}),m.push(h)),f--)}for(const e of m)e!==h&&a.disposeIntermediateTensorInfo(e);return h}};const da={kernelName:t.EluGrad,backendName:"cpu",kernelFunc:function(e){const{inputs:n,backend:a}=e,{dy:s,y:o}=n;r([s,o],"eluGrad");const i=new Float32Array(t.util.sizeFromShape(o.shape)),l=a.data.get(o.dataId).values,u=a.data.get(s.dataId).values;for(let e=0;e<l.length;++e){const t=l[e];i[e]=t>=0?u[e]:u[e]*(t+1)}return a.makeTensorInfo(o.shape,"float32",i)}},pa=t.backend_util.ERF_P,ha=t.backend_util.ERF_A1,fa=t.backend_util.ERF_A2,ma=t.backend_util.ERF_A3,ka=t.backend_util.ERF_A4,ga=t.backend_util.ERF_A5,ba=z(t.Erf,(e=>{const t=Math.sign(e),n=Math.abs(e),a=1/(1+pa*n);return t*(1-((((ga*a+ka)*a+ma)*a+fa)*a+ha)*a*Math.exp(-n*n))})),Ia={kernelName:t.Erf,backendName:"cpu",kernelFunc:ba};function ya(e){const{inputs:n,backend:a,attrs:s}=e,{input:r}=n,{dim:o}=s,i=r.shape.length,l=r.shape.slice();let u=o;return o<0&&(t.util.assert(-(i+1)<=o,(()=>`Axis must be in the interval [${-(i+1)}, ${i}]`)),u=i+o+1),l.splice(u,0,1),Zt({inputs:{x:r},backend:a,attrs:{shape:l}})}const Sa={kernelName:t.ExpandDims,backendName:"cpu",kernelFunc:ya},Ta=c(((e,t)=>e/t)),Na=S(t.RealDiv,Ta),xa={kernelName:t.RealDiv,backendName:"cpu",kernelFunc:Na};function va(e,n,a){const s=e.shape,r=s[0],o=s[1],i=a.data.get(e.dataId),l=i.complexTensorInfos.real,u=i.complexTensorInfos.imag,c=[r,o],p=t.util.sizeFromShape(c),h=t.util.getTypedArrayFromDType("float32",p),f=t.util.getTypedArrayFromDType("float32",p);for(let e=0;e<r;e++){const s=rt({inputs:{x:l},backend:a,attrs:{begin:[e,0],size:[1,o]}}),r=rt({inputs:{x:u},backend:a,attrs:{begin:[e,0],size:[1,o]}}),i=d({inputs:{real:s,imag:r},backend:a}),{real:c,imag:p}=Fa(i,n,a),m=t.backend_util.mergeRealAndImagArrays(c,p);for(let n=0;n<o;n++){const a=t.backend_util.getComplexWithIndex(m,n);h[e*o+n]=a.real,f[e*o+n]=a.imag}a.disposeIntermediateTensorInfo(s),a.disposeIntermediateTensorInfo(r),a.disposeIntermediateTensorInfo(i)}const m=a.makeTensorInfo(c,"float32",h),k=a.makeTensorInfo(c,"float32",f),g=d({inputs:{real:m,imag:k},backend:a});return a.disposeIntermediateTensorInfo(m),a.disposeIntermediateTensorInfo(k),g}function Fa(e,n,a){const s=t.util.sizeFromShape(e.shape),r=a.data.get(e.dataId),o=a.data.get(r.complexTensorInfos.real.dataId).values,i=a.data.get(r.complexTensorInfos.imag.dataId).values;if(0==((l=s)&l-1)){const r=wa(o,i,s,n,a),l=[e.shape[0],e.shape[1]];if(n){const e=a.makeTensorInfo(l,"float32",r.real),n=a.makeTensorInfo(l,"float32",r.imag),o=a.makeTensorInfo([],"float32",t.util.createScalarValue(s,"float32")),i=f({inputs:{x:o},backend:a}),u=xa.kernelFunc({inputs:{a:e,b:o},backend:a}),c=xa.kernelFunc({inputs:{a:n,b:i},backend:a}),d=a.data.get(u.dataId).values,p=a.data.get(c.dataId).values;return a.disposeIntermediateTensorInfo(e),a.disposeIntermediateTensorInfo(n),a.disposeIntermediateTensorInfo(o),a.disposeIntermediateTensorInfo(i),a.disposeIntermediateTensorInfo(u),a.disposeIntermediateTensorInfo(c),{real:d,imag:p}}return r}{const e=function(e,n,a){const s=new Float32Array(2*n);for(let r=0;r<n;r++){let o=0,i=0;for(let s=0;s<n;s++){const l=t.backend_util.exponent(r*s,n,a),u=t.backend_util.getComplexWithIndex(e,s);o+=u.real*l.real-u.imag*l.imag,i+=u.real*l.imag+u.imag*l.real}a&&(o/=n,i/=n),t.backend_util.assignToTypedArray(s,o,i,r)}return s}(t.backend_util.mergeRealAndImagArrays(o,i),s,n);return t.backend_util.splitRealAndImagArrays(e)}var l}function wa(e,n,a,s,r){if(1===a)return{real:e,imag:n};const o=t.backend_util.mergeRealAndImagArrays(e,n),i=a/2,l=t.backend_util.complexWithEvenIndex(o),u=l.real,c=l.imag,p=[u.length],h=r.makeTensorInfo(p,"float32",u),f=r.makeTensorInfo(p,"float32",c),m=d({inputs:{real:h,imag:f},backend:r}),g=t.backend_util.complexWithOddIndex(o),b=g.real,I=g.imag,y=[b.length],S=r.makeTensorInfo(y,"float32",b),T=r.makeTensorInfo(y,"float32",I),N=d({inputs:{real:S,imag:T},backend:r}),x=wa(u,c,i,s,r),F=x.real,w=x.imag,M=[F.length],A=r.makeTensorInfo(M,"float32",F),D=r.makeTensorInfo(M,"float32",w),E=d({inputs:{real:A,imag:D},backend:r}),_=wa(b,I,i,s,r),z=_.real,R=_.imag,W=[z.length],P=r.makeTensorInfo(W,"float32",z),C=r.makeTensorInfo(W,"float32",R),H=d({inputs:{real:P,imag:C},backend:r}),O=t.backend_util.exponents(a,s),B=[O.real.length],$=r.makeTensorInfo(B,"float32",O.real),V=r.makeTensorInfo(B,"float32",O.imag),G=d({inputs:{real:$,imag:V},backend:r}),L=we({inputs:{a:G,b:H},backend:r}),q=v({inputs:{a:E,b:L},backend:r}),U=wt({inputs:{a:E,b:L},backend:r}),Z=k({inputs:{input:q},backend:r}),j=k({inputs:{input:U},backend:r}),K=zn({inputs:{input:q},backend:r}),Y=zn({inputs:{input:U},backend:r}),J=Wn({inputs:[Z,j],backend:r,attrs:{axis:0}}),Q=Wn({inputs:[K,Y],backend:r,attrs:{axis:0}}),X=r.data.get(J.dataId).values,ee=r.data.get(Q.dataId).values;return r.disposeIntermediateTensorInfo(h),r.disposeIntermediateTensorInfo(f),r.disposeIntermediateTensorInfo(m),r.disposeIntermediateTensorInfo(S),r.disposeIntermediateTensorInfo(T),r.disposeIntermediateTensorInfo(N),r.disposeIntermediateTensorInfo(A),r.disposeIntermediateTensorInfo(D),r.disposeIntermediateTensorInfo(E),r.disposeIntermediateTensorInfo(P),r.disposeIntermediateTensorInfo(C),r.disposeIntermediateTensorInfo(H),r.disposeIntermediateTensorInfo($),r.disposeIntermediateTensorInfo(V),r.disposeIntermediateTensorInfo(G),r.disposeIntermediateTensorInfo(L),r.disposeIntermediateTensorInfo(q),r.disposeIntermediateTensorInfo(U),r.disposeIntermediateTensorInfo(Z),r.disposeIntermediateTensorInfo(K),r.disposeIntermediateTensorInfo(j),r.disposeIntermediateTensorInfo(Y),r.disposeIntermediateTensorInfo(J),r.disposeIntermediateTensorInfo(Q),{real:X,imag:ee}}const Ma={kernelName:t.FFT,backendName:"cpu",kernelFunc:function(e){const{inputs:n,backend:a}=e,{input:s}=n,r=t.util.sizeFromShape(s.shape),o=s.shape[s.shape.length-1],i=Zt({inputs:{x:s},backend:a,attrs:{shape:[r/o,o]}}),l=va(i,!1,a),u=Zt({inputs:{x:l},backend:a,attrs:{shape:s.shape}});return a.disposeIntermediateTensorInfo(i),a.disposeIntermediateTensorInfo(l),u}};function Aa(e){const{backend:n,attrs:a}=e,{shape:s,value:r,dtype:o}=a,i=o||t.util.inferDtype(r),l=t.util.getArrayFromDType(i,t.util.sizeFromShape(s));return function(e,t,n){e.fill(t)}(l,r),n.makeTensorInfo(s,i,l)}const Da={kernelName:t.Fill,backendName:"cpu",kernelFunc:Aa};const Ea={kernelName:t.FlipLeftRight,backendName:"cpu",kernelFunc:({inputs:e,attrs:n,backend:a})=>{const{image:s}=e,r=a,o=t.util.getTypedArrayFromDType(s.dtype,t.util.sizeFromShape(s.shape)),[i,l,u,c]=s.shape,d=r.data.get(s.dataId).values;for(let e=0;e<i;e++){const t=e*u*l*c;for(let e=0;e<l;e++){const n=e*(u*c);for(let e=0;e<u;e++){const a=e*c;for(let s=0;s<c;s++){const r=Math.round(u-e-1),i=t+n+a+s;let l=d[i];if(r>=0&&r<u){l=d[t+n+r*c+s]}o[i]=l}}}}return{dataId:r.write(o,s.shape,s.dtype),shape:s.shape,dtype:s.dtype}}};const _a={kernelName:t.FusedConv2D,backendName:"cpu",kernelFunc:function(e){const{inputs:t,backend:n,attrs:a}=e,{x:s,filter:r,bias:o,preluActivationWeights:i}=t,{strides:l,pad:u,dataFormat:c,dilations:d,dimRoundingMode:p,activation:h,leakyreluAlpha:f}=a;let m=Cn({inputs:{x:s,filter:r},backend:n,attrs:{strides:l,pad:u,dataFormat:c,dilations:d,dimRoundingMode:p}});if(o){const e=m;if("NCHW"===c&&1===o.shape.length&&1!==o.shape[0]){const e=Zt({inputs:{x:o},backend:n,attrs:{shape:[o.shape[0],1,1]}});m=v({inputs:{a:m,b:e},backend:n}),n.disposeIntermediateTensorInfo(e)}else m=v({inputs:{a:m,b:o},backend:n});n.disposeIntermediateTensorInfo(e)}if(h){const e=m;if("NCHW"===c&&"prelu"===h&&1===i.shape.length&&1!==i.shape[0]){const e=Zt({inputs:{x:i},backend:n,attrs:{shape:[i.shape[0],1,1]}});m=Ut(n,m,h,e,f),n.disposeIntermediateTensorInfo(e)}else m=Ut(n,m,h,i,f);n.disposeIntermediateTensorInfo(e)}return m}};const za={kernelName:t.FusedDepthwiseConv2D,backendName:"cpu",kernelFunc:function(e){const{inputs:t,backend:n,attrs:a}=e,{x:s,filter:r,bias:o,preluActivationWeights:i}=t,{strides:l,pad:u,dataFormat:c,dilations:d,dimRoundingMode:p,activation:h,leakyreluAlpha:f}=a;let m=Xn({inputs:{x:s,filter:r},backend:n,attrs:{strides:l,pad:u,dataFormat:c,dilations:d,dimRoundingMode:p}});if(o){const e=m;m=v({inputs:{a:m,b:o},backend:n}),n.disposeIntermediateTensorInfo(e)}if(h){const e=m;m=Ut(n,m,h,i,f),n.disposeIntermediateTensorInfo(e)}return m}};const Ra={kernelName:t.GatherNd,backendName:"cpu",kernelFunc:function(e){const{inputs:n,backend:a}=e,{params:s,indices:r}=n,o=t.util.sizeFromShape(s.shape),i=r.shape,l=i[i.length-1],[u,c,d,p]=t.backend_util.prepareAndValidate(s,r);if(0===c)return a.makeTensorInfo(u,s.dtype,[]);const h=ee(a.data.get(r.dataId).values,a.bufferSync(s),s.dtype,c,l,d,p,s.shape,o);return a.makeTensorInfo(u,s.dtype,h.values)}};const Wa={kernelName:t.GatherV2,backendName:"cpu",kernelFunc:function(e){const{inputs:n,backend:a,attrs:s}=e,{x:o,indices:i}=n,{axis:l,batchDims:u}=s;r([o,i],"gatherV2");const c=t.util.parseAxisParam(l,o.shape)[0],d=a.data.get(i.dataId).values,p=o.shape[c];for(let e=0;e<d.length;++e){const n=d[e];t.util.assert(n<=p-1&&n>=0,(()=>`GatherV2: the index value ${n} is not in [0, ${p-1}]`))}let h=u;null==u&&(h=0);const f=t.util.sizeFromShape(i.shape),m=t.backend_util.segment_util.collectGatherOpShapeInfo(o,i,c,h),k=Zt({inputs:{x:o},backend:a,attrs:{shape:[m.batchSize,m.outerSize,m.dimSize,m.sliceSize]}}),g=Zt({inputs:{x:i},backend:a,attrs:{shape:[m.batchSize,f/m.batchSize]}}),b=[m.batchSize,m.outerSize,f/m.batchSize,m.sliceSize],I=a.bufferSync(g),y=te(a.bufferSync(k),I,b);return a.disposeIntermediateTensorInfo(k),a.disposeIntermediateTensorInfo(g),a.makeTensorInfo(m.outputShape,y.dtype,y.values)}};const Pa={kernelName:t.IFFT,backendName:"cpu",kernelFunc:function(e){const{inputs:n,backend:a}=e,{input:s}=n,r=t.util.sizeFromShape(s.shape),o=s.shape[s.shape.length-1],i=Zt({inputs:{x:s},backend:a,attrs:{shape:[r/o,o]}}),l=va(i,!0,a),u=Zt({inputs:{x:l},backend:a,attrs:{shape:s.shape}});return a.disposeIntermediateTensorInfo(i),a.disposeIntermediateTensorInfo(l),u}},Ca=z(t.IsFinite,(e=>Number.isFinite(e)?1:0),"bool"),Ha={kernelName:t.IsFinite,backendName:"cpu",kernelFunc:Ca},Oa=z(t.IsInf,(e=>Math.abs(e)===1/0?1:0),"bool"),Ba={kernelName:t.IsInf,backendName:"cpu",kernelFunc:Oa},$a=z(t.IsNan,(e=>Number.isNaN(e)?1:0),"bool"),Va={kernelName:t.IsNan,backendName:"cpu",kernelFunc:$a};const Ga={kernelName:t.LinSpace,backendName:"cpu",kernelFunc:function(e){const{backend:t,attrs:n}=e,{start:a,stop:s,num:r}=n,o=fe(a,s,r);return t.makeTensorInfo([o.length],"float32",o)}},La=z(t.Log1p,(e=>Math.log1p(e))),qa={kernelName:t.Log1p,backendName:"cpu",kernelFunc:La},Ua=c(((e,t)=>e&&t)),Za=S(t.LogicalAnd,Ua,null,"bool"),ja={kernelName:t.LogicalAnd,backendName:"cpu",kernelFunc:Za},Ka=z(t.LogicalNot,(e=>e?0:1),"bool"),Ya={kernelName:t.LogicalNot,backendName:"cpu",kernelFunc:Ka},Ja=c(((e,t)=>e||t)),Qa=S(t.LogicalOr,Ja,null,"bool"),Xa={kernelName:t.LogicalOr,backendName:"cpu",kernelFunc:Qa};const es={kernelName:t.LRN,backendName:"cpu",kernelFunc:function(e){const{inputs:n,backend:a,attrs:s}=e,{x:o}=n,{depthRadius:i,bias:l,alpha:u,beta:c}=s;r(o,"LRN");const d=o.shape[3],p=d-1,h=a.data.get(o.dataId).values,f=t.util.sizeFromShape(o.shape),m=new Float32Array(f);function k(e){const t=e%d;let n=e-t+Math.max(0,t-i);const a=e-t+Math.min(t+i,p);let s=0;for(;n<=a;n++){const e=h[n];s+=e*e}return s}for(let e=0;e<f;e++){const t=k(e),n=h[e]*Math.pow(l+u*t,-c);m[e]=n}return a.makeTensorInfo(o.shape,o.dtype,m)}};const ts={kernelName:t.LRNGrad,backendName:"cpu",kernelFunc:function(e){const{inputs:n,backend:a,attrs:s}=e,{x:o,y:i,dy:l}=n,{depthRadius:u,bias:c,alpha:d,beta:p}=s;r(l,"LRNGrad");const h=t.util.sizeFromShape(l.shape),f=l.shape[3],m=a.data.get(l.dataId).values,k=a.data.get(o.dataId).values,g=a.data.get(i.dataId).values,b=new Float32Array(h),I=h;for(let e=0;e<I;e++){const t=e%f,n=e-t+Math.max(0,t-u),a=e-t+Math.min(f,t+u+1);let s=0;for(let e=n;e<a;e++)s+=Math.pow(k[e],2);s=d*s+c;for(let t=n;t<a;t++){let n=-2*d*p*k[t]*g[e]/s;e===t&&(n+=Math.pow(s,-p)),n*=m[e],b[t]+=n}}return a.makeTensorInfo(l.shape,o.dtype,b)}};function ns(e){const{inputs:n,backend:a,attrs:s}=e,{x:o}=n,{reductionIndices:i,keepDims:l}=s,u=a;let c=o.shape;const d=c.length,p=t.util.parseAxisParam(i,c);let h=p;const f=t.backend_util.getAxesPermutation(h,d);let m=u.data.get(o.dataId).values;if(null!=f){const e=new Array(d);for(let t=0;t<e.length;t++)e[t]=c[f[t]];m=Re(m,c,o.dtype,f,e),h=t.backend_util.getInnerMostAxes(h.length,d),c=e}r(o,"max"),t.backend_util.assertAxesAreInnerMostDims("max",h,d);const[k,g]=t.backend_util.computeOutAndReduceShapes(c,h),b=be(m,t.util.sizeFromShape(g),k,o.dtype),I=u.write(b,k,o.dtype);let y=k;if(l){y=t.backend_util.expandShapeToKeepDim(k,p)}return{dataId:I,shape:y,dtype:o.dtype}}const as={kernelName:t.Max,backendName:"cpu",kernelFunc:ns};const ss={kernelName:t.MaxPool,backendName:"cpu",kernelFunc:function(e){const{inputs:n,backend:a,attrs:s}=e,{x:o}=n;r(o,"maxPool");const{filterSize:i,strides:l,pad:u,dimRoundingMode:c}=s;t.util.assert(t.backend_util.eitherStridesOrDilationsAreOne(l,1),(()=>`Error in maxPool: Either strides or dilations must be 1. Got strides ${l} and dilations '1'`));const d=t.backend_util.computePool2DInfo(o.shape,i,l,1,u,c);let p;if(1===d.filterWidth&&1===d.filterHeight&&t.util.arraysEqual(d.inShape,d.outShape))p=f({inputs:{x:o},backend:a});else{const e=a.data.get(o.dataId).values,n=t.util.computeStrides(o.shape),s=In(e,o.shape,o.dtype,n,d,"max");p=a.makeTensorInfo(d.outShape,o.dtype,s.values)}return p}};const rs={kernelName:t.MaxPool3D,backendName:"cpu",kernelFunc:function(e){const{inputs:n,backend:a,attrs:s}=e,{x:o}=n,{filterSize:i,strides:l,pad:u,dimRoundingMode:c,dataFormat:d}=s;r(o,"maxPool3d");const p=t.backend_util.computePool3DInfo(o.shape,i,l,1,u,c,d),h=Sn(a.data.get(o.dataId).values,o.shape,o.dtype,t.util.computeStrides(o.shape),p,"max");return a.makeTensorInfo(h.shape,"float32",h.values)}};const os={kernelName:t.MaxPool3DGrad,backendName:"cpu",kernelFunc:function(e){const{inputs:n,backend:a,attrs:s}=e,{dy:o,input:i}=n,{filterSize:l,strides:u,pad:c,dimRoundingMode:d}=s;r([o,i],"maxPool3DGrad");const p=t.backend_util.computePool3DInfo(i.shape,l,u,1,c,d),h=function(e,n){const a=t.buffer(n.outShape,"int32"),s=n.strideDepth,r=n.strideHeight,o=n.strideWidth,i=n.dilationDepth,l=n.dilationHeight,u=n.dilationWidth,c=n.effectiveFilterDepth,d=n.effectiveFilterHeight,p=n.effectiveFilterWidth,h=n.padInfo.front,f=n.padInfo.top,m=n.padInfo.left;for(let t=0;t<n.batchSize;++t)for(let k=0;k<n.inChannels;++k)for(let g=0;g<n.outDepth;++g){const b=g*s-h;let I=b;for(;I<0;)I+=i;const y=Math.min(n.inDepth,c+b);for(let s=0;s<n.outHeight;++s){const c=s*r-f;let h=c;for(;h<0;)h+=l;const S=Math.min(n.inHeight,d+c);for(let r=0;r<n.outWidth;++r){const f=r*o-m;let T=f;for(;T<0;)T+=u;const N=Math.min(n.inWidth,p+f);let x=Number.NEGATIVE_INFINITY,v=-1;for(let n=I;n<y;n+=i){const a=n-b;for(let s=h;s<S;s+=l){const r=s-c;for(let o=T;o<N;o+=u){const i=o-f,l=e.get(t,n,s,o,k);l>=x&&(x=l,v=a*d*p+r*d+i)}}}a.set(v,t,g,s,r,k)}}}return a}(a.bufferSync(i),p),f=p.strideDepth,m=p.strideHeight,k=p.strideWidth,g=p.dilationDepth,b=p.dilationHeight,I=p.dilationWidth,y=p.effectiveFilterDepth,S=p.effectiveFilterHeight,T=p.effectiveFilterWidth,N=y-1-p.padInfo.front,x=T-1-p.padInfo.left,v=S-1-p.padInfo.top,F=t.buffer(i.shape,"float32"),w=a.bufferSync(o);for(let e=0;e<p.batchSize;++e)for(let t=0;t<p.inChannels;++t)for(let n=0;n<p.inDepth;++n)for(let a=0;a<p.inHeight;++a)for(let s=0;s<p.inWidth;++s){const r=n-N,o=a-v,i=s-x;let l=0;for(let n=0;n<y;n+=g){const a=(r+n)/f;if(!(a<0||a>=p.outDepth||Math.floor(a)!==a))for(let s=0;s<S;s+=b){const r=(o+s)/m;if(!(r<0||r>=p.outHeight||Math.floor(r)!==r))for(let o=0;o<T;o+=I){const u=(i+o)/k;if(u<0||u>=p.outWidth||Math.floor(u)!==u)continue;const c=y*S*T-1-h.get(e,a,r,u,t)===n*S*T+s*T+o?1:0;if(0===c)continue;l+=w.get(e,a,r,u,t)*c}}}F.set(l,e,n,a,s,t)}return a.makeTensorInfo(F.shape,F.dtype,F.values)}};const is={kernelName:t.MaxPoolGrad,backendName:"cpu",kernelFunc:function(e){const{inputs:n,backend:a,attrs:s}=e,{dy:o,input:i,output:l}=n,u=i;r([i,l],"maxPoolGrad");const{filterSize:c,strides:d,pad:p,dimRoundingMode:h}=s,f=t.backend_util.computePool2DInfo(u.shape,c,d,1,p,h),m=a.data.get(u.dataId).values,k=t.buffer(f.outShape,u.dtype,yn(m,u.shape,u.dtype,f).values),g=f.strideHeight,b=f.strideWidth,I=f.dilationHeight,y=f.dilationWidth,S=f.effectiveFilterHeight,T=f.effectiveFilterWidth,N=T-1-f.padInfo.left,x=S-1-f.padInfo.top,v=t.buffer(u.shape,"float32"),F=a.data.get(o.dataId).values,w=t.buffer(o.shape,"float32",F);for(let e=0;e<f.batchSize;++e)for(let t=0;t<f.inChannels;++t)for(let n=0;n<f.inHeight;++n)for(let a=0;a<f.inWidth;++a){const s=n-x,r=a-N;let o=0;for(let n=0;n<S;n+=I){const a=(s+n)/g;if(!(a<0||a>=f.outHeight||Math.floor(a)!==a))for(let s=0;s<T;s+=y){const i=(r+s)/b;if(i<0||i>=f.outWidth||Math.floor(i)!==i)continue;const l=S*T-1-k.get(e,a,i,t)===n*T+s?1:0;if(0===l)continue;o+=w.get(e,a,i,t)*l}}v.set(o,e,n,a,t)}return a.makeTensorInfo(v.shape,v.dtype,v.values)}};const ls={kernelName:t.MaxPoolWithArgmax,backendName:"cpu",kernelFunc:({inputs:e,attrs:n,backend:a})=>{const{x:s}=e,{filterSize:o,strides:i,pad:l,includeBatchInIndex:u}=n,c=a;r(s,"MaxPoolWithArgmax");const d=c.data.get(s.dataId).values,p=t.backend_util.computePool2DInfo(s.shape,o,i,[1,1],l),[h,f]=function(e,n,a,s,r){const o=In(e,0,a,t.util.computeStrides(n),r,"max"),i=yn(e,n,a,r,!0,s);return[o.values,i.values]}(d,s.shape,s.dtype,u,p),m=c.write(h,p.outShape,s.dtype),k=c.write(f,p.outShape,s.dtype);return[{dataId:m,shape:p.outShape,dtype:s.dtype},{dataId:k,shape:p.outShape,dtype:"int32"}]}};const us={kernelName:t.Mean,backendName:"cpu",kernelFunc:function(e){const{inputs:n,backend:a,attrs:s}=e,{x:r}=n,{axis:o,keepDims:i}=s,l=t.util.parseAxisParam(o,r.shape),u=t.backend_util.computeOutAndReduceShapes(r.shape,l)[1],c=t.util.sizeFromShape(u),d=[],p=a.makeTensorInfo([],"float32",new Float32Array([c]));d.push(p);const h=I({inputs:{x:r},backend:a,attrs:{dtype:"float32"}});d.push(h);const f=Na({inputs:{a:h,b:p},backend:a});d.push(f);const m=la({inputs:{x:f},backend:a,attrs:{axis:o,keepDims:i}});return d.forEach((e=>a.disposeIntermediateTensorInfo(e))),m}};const cs={kernelName:t.Min,backendName:"cpu",kernelFunc:function(e){const{inputs:n,backend:a,attrs:s}=e,{x:o}=n,{axis:i,keepDims:l}=s;r(o,"min");const u=t.util.parseAxisParam(i,o.shape);let c=u;const d=t.backend_util.getAxesPermutation(c,o.shape.length);let p=o;null!=d&&(p=We({inputs:{x:o},backend:a,attrs:{perm:d}}),c=t.backend_util.getInnerMostAxes(c.length,o.shape.length)),t.backend_util.assertAxesAreInnerMostDims("min",c,p.shape.length);const[h,f]=t.backend_util.computeOutAndReduceShapes(p.shape,c),m=t.util.sizeFromShape(f),k=t.util.makeZerosTypedArray(t.util.sizeFromShape(h),p.dtype),g=a.data.get(p.dataId).values;for(let e=0;e<k.length;++e){const t=e*m;let n=g[t];for(let e=0;e<m;++e){const a=g[t+e];(Number.isNaN(a)||a<n)&&(n=a)}k[e]=n}null!=d&&a.disposeIntermediateTensorInfo(p);const b=a.makeTensorInfo(h,p.dtype,k);if(l){const e=Zt({inputs:{x:b},backend:a,attrs:{shape:t.backend_util.expandShapeToKeepDim(h,u)}});return a.disposeIntermediateTensorInfo(b),e}return b}};const ds={kernelName:t.MirrorPad,backendName:"cpu",kernelFunc:function(e){const{inputs:n,backend:a,attrs:s}=e,{x:o}=n,{paddings:i,mode:l}=s;r(o,"mirrorPad");const u=i.map(((e,t)=>e[0]+o.shape[t]+e[1])),c=i.map((e=>e[0])),d=i.map(((e,t)=>e[0]+o.shape[t])),p="reflect"===l?0:1,h=a.data.get(o.dataId).values,f=o.shape.length,m=t.util.computeStrides(o.shape),k=t.util.sizeFromShape(u),g=u.length,b=t.util.computeStrides(u),I=t.util.getTypedArrayFromDType(o.dtype,k);for(let e=0;e<k;e++){let n=t.util.indexToLoc(e,g,b);for(let e=0;e<g;e++)n[e]<c[e]?n[e]=2*c[e]-n[e]-p:n[e]>=d[e]&&(n[e]=2*(d[e]-1)-n[e]+p);n=n.map(((e,t)=>e-c[t]));const a=t.util.locToIndex(n,f,m);I[e]=h[a]}return{dataId:a.write(I,u,o.dtype),shape:u,dtype:o.dtype}}},ps=c(((e,t)=>{const n=e%t;return e<0&&t<0||e>=0&&t>=0?n:(n+t)%t})),hs=S(t.Mod,ps),fs={kernelName:t.Mod,backendName:"cpu",kernelFunc:hs};function ms(e){const{inputs:n,backend:a,attrs:s}=e,{logits:r}=n,{dim:o}=s,i=r.shape.length;let l=o;if(-1===l&&(l=i-1),l!==i-1)throw Error(`Softmax along a non-last dimension is not yet supported. Logits was rank ${i} and dim was ${l}`);const u=t.util.parseAxisParam([l],r.shape),c=ns({inputs:{x:r},backend:a,attrs:{reductionIndices:u,keepDims:!1}}),d=t.backend_util.expandShapeToKeepDim(c.shape,u),p=Zt({inputs:{x:c},backend:a,attrs:{shape:d}}),h=wt({inputs:{a:r,b:p},backend:a}),f=G({inputs:{x:h},backend:a}),m=la({inputs:{x:f},backend:a,attrs:{axis:u,keepDims:!1}}),k=Zt({inputs:{x:m},backend:a,attrs:{shape:d}}),g=Na({inputs:{a:f,b:k},backend:a});return a.disposeIntermediateTensorInfo(c),a.disposeIntermediateTensorInfo(p),a.disposeIntermediateTensorInfo(h),a.disposeIntermediateTensorInfo(f),a.disposeIntermediateTensorInfo(m),a.disposeIntermediateTensorInfo(k),g}const ks={kernelName:t.Softmax,backendName:"cpu",kernelFunc:ms};const gs={kernelName:t.Multinomial,backendName:"cpu",kernelFunc:function(e){const{inputs:n,backend:a,attrs:o}=e,{logits:i}=n,{numSamples:l,seed:u,normalized:c}=o;r(i,"multinomial");const d=c?i:ms({inputs:{logits:i},backend:a,attrs:{dim:-1}}),p=d.shape[0],h=d.shape[1],f=a.data.get(d.dataId).values,m=[p,l],k=t.util.makeZerosTypedArray(t.util.sizeFromShape(m),"int32");for(let e=0;e<p;++e){const t=e*h,n=new Float32Array(h-1);n[0]=f[t];for(let e=1;e<n.length;++e)n[e]=n[e-1]+f[t+e];const a=s.alea(u.toString()),r=e*l;for(let e=0;e<l;++e){const t=a();k[r+e]=n.length;for(let a=0;a<n.length;a++)if(t<n[a]){k[r+e]=a;break}}}return c||a.disposeIntermediateTensorInfo(d),a.makeTensorInfo(m,"int32",k)}},bs=t.kernel_impls.nonMaxSuppressionV3Impl;const Is={kernelName:t.NonMaxSuppressionV3,backendName:"cpu",kernelFunc:function(e){const{inputs:t,backend:n,attrs:a}=e,{boxes:s,scores:o}=t,{maxOutputSize:i,iouThreshold:l,scoreThreshold:u}=a;r(s,"NonMaxSuppression");const c=n.data.get(s.dataId).values,d=n.data.get(o.dataId).values,{selectedIndices:p}=bs(c,d,i,l,u);return n.makeTensorInfo([p.length],"int32",new Int32Array(p))}},ys=t.kernel_impls.nonMaxSuppressionV4Impl;const Ss={kernelName:t.NonMaxSuppressionV4,backendName:"cpu",kernelFunc:function(e){const{inputs:t,backend:n,attrs:a}=e,{boxes:s,scores:o}=t,{maxOutputSize:i,iouThreshold:l,scoreThreshold:u,padToMaxOutputSize:c}=a;r(s,"NonMaxSuppressionPadded");const d=n.data.get(s.dataId).values,p=n.data.get(o.dataId).values,{selectedIndices:h,validOutputs:f}=ys(d,p,i,l,u,c);return[n.makeTensorInfo([h.length],"int32",new Int32Array(h)),n.makeTensorInfo([],"int32",new Int32Array([f]))]}},Ts=t.kernel_impls.nonMaxSuppressionV5Impl;const Ns={kernelName:t.NonMaxSuppressionV5,backendName:"cpu",kernelFunc:function(e){const{inputs:t,backend:n,attrs:a}=e,{boxes:s,scores:o}=t,{maxOutputSize:i,iouThreshold:l,scoreThreshold:u,softNmsSigma:c}=a;r(s,"NonMaxSuppressionWithScore");const d=n.data.get(s.dataId).values,p=n.data.get(o.dataId).values,h=i,f=l,m=u,k=c,{selectedIndices:g,selectedScores:b}=Ts(d,p,h,f,m,k);return[n.makeTensorInfo([g.length],"int32",new Int32Array(g)),n.makeTensorInfo([b.length],"float32",new Float32Array(b))]}};const xs={kernelName:t.OneHot,backendName:"cpu",kernelFunc:function(e){const{inputs:n,backend:a,attrs:s}=e,{indices:o}=n,{dtype:i,depth:l,onValue:u,offValue:c}=s;r(o,"oneHot");const d=t.util.sizeFromShape(o.shape),p=new Float32Array(d*l);p.fill(c);const h=a.data.get(o.dataId).values;for(let e=0;e<d;++e)h[e]>=0&&h[e]<l&&(p[e*l+h[e]]=u);return a.makeTensorInfo([...o.shape,l],i,p)}};function vs(e){const{inputs:t,backend:n}=e,{x:a}=t;if("string"===a.dtype)throw new Error("zerosLike is not supported for string tensors");if("complex64"===a.dtype){const e=k({inputs:{input:a},backend:n}),t=vs({inputs:{x:e},backend:n}),s=zn({inputs:{input:a},backend:n}),r=vs({inputs:{x:s},backend:n}),o=d({inputs:{real:t,imag:r},backend:n});return n.disposeIntermediateTensorInfo(e),n.disposeIntermediateTensorInfo(t),n.disposeIntermediateTensorInfo(s),n.disposeIntermediateTensorInfo(r),o}return Aa({backend:n,attrs:{shape:a.shape,value:0,dtype:a.dtype}})}const Fs={kernelName:t.ZerosLike,backendName:"cpu",kernelFunc:vs};const ws={kernelName:t.OnesLike,backendName:"cpu",kernelFunc:function e(t){const{inputs:n,backend:a}=t,{x:s}=n;if("string"===s.dtype)throw new Error("onesLike is not supported for string tensors");if("complex64"===s.dtype){const t=k({inputs:{input:s},backend:a}),n=e({inputs:{x:t},backend:a}),r=zn({inputs:{input:s},backend:a}),o=vs({inputs:{x:r},backend:a}),i=d({inputs:{real:n,imag:o},backend:a});return a.disposeIntermediateTensorInfo(t),a.disposeIntermediateTensorInfo(n),a.disposeIntermediateTensorInfo(r),a.disposeIntermediateTensorInfo(o),i}return Aa({backend:a,attrs:{shape:s.shape,value:1,dtype:s.dtype}})}};function Ms(e){const{inputs:n,backend:a,attrs:s}=e,{axis:r}=s;if(1===n.length)return ya({inputs:{input:n[0]},backend:a,attrs:{dim:r}});const o=n[0].shape,i=n[0].dtype;n.forEach((e=>{t.util.assertShapesMatch(o,e.shape,"All tensors passed to stack must have matching shapes"),t.util.assert(i===e.dtype,(()=>"All tensors passed to stack must have matching dtypes"))}));const l=[],u=Wn({inputs:n.map((e=>{const t=ya({inputs:{input:e},backend:a,attrs:{dim:r}});return l.push(t),t})),backend:a,attrs:{axis:r}});return l.forEach((e=>a.disposeIntermediateTensorInfo(e))),u}const As={kernelName:t.Pack,backendName:"cpu",kernelFunc:Ms};const Ds={kernelName:t.PadV2,backendName:"cpu",kernelFunc:function(e){const{inputs:n,backend:a,attrs:s}=e,{x:o}=n,{paddings:i,constantValue:l}=s;r(o,"pad");const u=i.map(((e,t)=>e[0]+o.shape[t]+e[1])),c=i.map((e=>e[0])),d=a.data.get(o.dataId).values,p=t.util.sizeFromShape(o.shape),h=o.shape.length,f=t.util.computeStrides(o.shape),m=t.util.sizeFromShape(u),k=u.length,g=t.util.computeStrides(u),b=t.util.getTypedArrayFromDType(o.dtype,m);0!==l&&b.fill(l);for(let e=0;e<p;e++){const n=t.util.indexToLoc(e,h,f).map(((e,t)=>e+c[t]));b[t.util.locToIndex(n,k,g)]=d[e]}return{dataId:a.write(b,u,o.dtype),shape:u,dtype:o.dtype}}},Es=c(((e,t)=>Math.pow(e,t))),_s=S(t.Pow,Es),zs={kernelName:t.Pow,backendName:"cpu",kernelFunc:_s};const Rs={kernelName:t.RaggedGather,backendName:"cpu",kernelFunc:function(e){const{inputs:t,backend:n,attrs:a}=e,{paramsNestedSplits:s,paramsDenseValues:r,indices:o}=t,i=s.map((e=>n.data.get(e.dataId).values)),l=s.map((e=>e.shape)),u=n.data.get(r.dataId).values,c=n.data.get(o.dataId).values,[d,p,h]=Ve(i,l,u,r.shape,r.dtype,c,o.shape),f=d.map((e=>n.makeTensorInfo([e.length],"int32",e))),m=n.makeTensorInfo(h,r.dtype,p);return f.concat([m])}};const Ws={kernelName:t.RaggedRange,backendName:"cpu",kernelFunc:function(e){const{inputs:t,backend:n}=e,{starts:a,limits:s,deltas:r}=t,o=n.data.get(a.dataId).values,i=n.data.get(s.dataId).values,l=n.data.get(r.dataId).values,[u,c]=Le(o,a.shape,a.dtype,i,s.shape,l,r.shape);return[n.makeTensorInfo([u.length],"int32",u),n.makeTensorInfo([c.length],a.dtype,c)]}};const Ps={kernelName:t.RaggedTensorToTensor,backendName:"cpu",kernelFunc:function(e){const{inputs:t,backend:n,attrs:a}=e,{shape:s,values:r,defaultValue:o,rowPartitionTensors:i}=t,{rowPartitionTypes:l}=a,u=n.data.get(s.dataId).values,c=n.data.get(r.dataId).values,d=n.data.get(o.dataId).values,p=i.map((e=>n.data.get(e.dataId).values)),h=i.map((e=>e.shape)),[f,m]=Ke(u,s.shape,c,r.shape,r.dtype,d,o.shape,p,h,l);return n.makeTensorInfo(f,r.dtype,m)}};const Cs={kernelName:t.Range,backendName:"cpu",kernelFunc:function(e){const{backend:t,attrs:n}=e,{start:a,stop:s,dtype:r,step:o}=n,i=Ye(a,s,o,r);return t.makeTensorInfo([i.length],r,i)}},Hs=z(t.Reciprocal,(e=>1/e)),Os={kernelName:t.Reciprocal,backendName:"cpu",kernelFunc:Hs};const Bs={kernelName:t.ResizeBilinear,backendName:"cpu",kernelFunc:function(e){const{inputs:n,backend:a,attrs:s}=e,{images:o}=n,{alignCorners:i,halfPixelCenters:l,size:u}=s;r(o,"resizeBilinear");const c=t.util.computeStrides(o.shape),[d,p]=u,[h,f,m,k]=o.shape,g=a.data.get(o.dataId).values,b=new Float32Array(t.util.sizeFromShape([h,d,p,k])),I=[i&&d>1?f-1:f,i&&p>1?m-1:m],y=[i&&d>1?d-1:d,i&&p>1?p-1:p];let S=0;const T=I[0]/y[0],N=I[1]/y[1];for(let e=0;e<h;e++)for(let t=0;t<d;t++){let n;n=l?T*(t+.5)-.5:T*t;const a=Math.max(0,Math.floor(n)),s=n-a,r=Math.min(f-1,Math.ceil(n)),o=e*c[0]+a*c[1],i=e*c[0]+r*c[1];for(let e=0;e<p;e++){let t;t=l?N*(e+.5)-.5:N*e;const n=Math.max(0,Math.floor(t)),a=t-n,r=Math.min(m-1,Math.ceil(t)),u=o+n*c[2],d=i+n*c[2],p=o+r*c[2],h=i+r*c[2];for(let e=0;e<k;e++){const t=g[u+e],n=g[d+e],r=t+(g[p+e]-t)*a,o=r+(n+(g[h+e]-n)*a-r)*s;b[S++]=o}}}return a.makeTensorInfo([h,d,p,k],"float32",b)}};const $s={kernelName:t.ResizeBilinearGrad,backendName:"cpu",kernelFunc:function(e){const{inputs:n,backend:a,attrs:s}=e,{images:o,dy:i}=n,{alignCorners:l}=s;r([i,o],"resizeBilinearGrad");const u=t.util.computeStrides(o.shape),[c,d,p,h]=o.shape,[,f,m]=i.shape,k=new Float32Array(c*d*p*h),g=[l&&f>1?d-1:d,l&&m>1?p-1:p],b=[l&&f>1?f-1:f,l&&m>1?m-1:m],I=g[0]/b[0],y=g[1]/b[1],S=a.data.get(i.dataId).values;let T=0;for(let e=0;e<c;e++){const t=e*u[0];for(let e=0;e<f;e++){const n=e*I,a=Math.floor(n),s=Math.min(Math.ceil(n),d-1),r=t+a*u[1],o=t+s*u[1],i=n-a,l=1-i;for(let e=0;e<m;e++){const t=e*y,n=Math.floor(t),a=Math.min(Math.ceil(t),p-1),s=t-n,c=1-s,d=r+n*u[2],f=r+a*u[2],m=o+n*u[2],g=o+a*u[2],b=l*c,I=l*s,N=i*c,x=i*s;for(let e=0;e<h;e++){const t=S[T++];k[d+e]+=t*b,k[f+e]+=t*I,k[m+e]+=t*N,k[g+e]+=t*x}}}}return a.makeTensorInfo([c,p,d,h],"float32",k)}};const Vs={kernelName:t.ResizeNearestNeighbor,backendName:"cpu",kernelFunc:function(e){const{inputs:n,backend:a,attrs:s}=e,{images:o}=n,{alignCorners:i,halfPixelCenters:l,size:u}=s;r(o,"resizeNearestNeighbor");const c=t.util.computeStrides(o.shape),[d,p]=u,[h,f,m,k]=o.shape,g=a.data.get(o.dataId).values,b=new Float32Array(h*d*p*k),I=[i&&d>1?f-1:f,i&&p>1?m-1:m],y=[i&&d>1?d-1:d,i&&p>1?p-1:p],S=I[0]/y[0],T=I[1]/y[1];let N=0;for(let e=0;e<h;e++){const t=e*c[0];for(let e=0;e<d;e++){const n=l?S*(e+.5):S*e;let a=Math.min(f-1,i?Math.round(n):Math.floor(n));l&&(a=Math.max(0,a));const s=t+a*c[1];for(let e=0;e<p;e++){const t=l?T*(e+.5):T*e;let n=Math.min(m-1,i?Math.round(t):Math.floor(t));l&&(n=Math.max(0,n));const a=s+n*c[2];for(let e=0;e<k;e++){const t=g[a+e];b[N++]=t}}}}return a.makeTensorInfo([h,d,p,k],o.dtype,b)}};const Gs={kernelName:t.ResizeNearestNeighborGrad,backendName:"cpu",kernelFunc:function(e){const{inputs:n,backend:a,attrs:s}=e,{images:o,dy:i}=n,{alignCorners:l}=s;r([i,o],"resizeNearestNeighborGrad");const u=t.util.computeStrides(o.shape),c=t.util.computeStrides(i.shape),[d,p,h,f]=o.shape,[,m,k]=i.shape,g=new Float32Array(d*p*h*f),b=a.data.get(i.dataId).values,I=[l&&m>1?p-1:p,l&&k>1?h-1:h],y=[l&&m>1?m-1:m,l&&k>1?k-1:k],S=I[0]/y[0],T=I[1]/y[1],N=1/S,x=1/T,v=2*Math.ceil(N)+2,F=2*Math.ceil(x)+2;for(let e=0;e<d;e++){const t=e*u[0];for(let e=0;e<p;e++){const n=t+e*u[1],a=Math.floor(e*N),s=Math.floor(a-v/2);for(let a=0;a<h;a++){const r=n+a*u[2],o=Math.floor(a*x),i=Math.floor(o-F/2);for(let n=0;n<f;n++){let o=0;for(let r=0;r<v;r++){const u=r+s;if(u<0||u>=m)continue;const d=t+u*c[1],f=u*S;if(e===Math.min(p-1,l?Math.round(f):Math.floor(f)))for(let e=0;e<F;e++){const t=e+i;if(t<0||t>=k)continue;const s=d+t*c[2],r=t*T;a===Math.min(h-1,l?Math.round(r):Math.floor(r))&&(o+=b[s+n])}}g[r+n]=o}}}}return a.makeTensorInfo(o.shape,o.dtype,g)}};const Ls={kernelName:t.Reverse,backendName:"cpu",kernelFunc:function(e){const{inputs:n,backend:a,attrs:s}=e,{x:o}=n,{dims:i}=s;r(o,"reverse");const l=o.shape.length,u=t.util.parseAxisParam(i,o.shape);if(0===l)return f({inputs:{x:o},backend:a});const c=new t.TensorBuffer(o.shape,o.dtype),d=a.bufferSync(o);for(let e=0;e<c.size;e++){const t=c.indexToLoc(e),n=t.slice();u.forEach((e=>n[e]=o.shape[e]-1-n[e])),c.set(d.get(...n),...t)}return a.makeTensorInfo(c.shape,c.dtype,c.values)}},qs={kernelName:t.RotateWithOffset,backendName:"cpu",kernelFunc:({inputs:e,attrs:n,backend:a})=>{const{image:s}=e,{radians:r,fillValue:o,center:i}=n,l=a,u=t.util.getTypedArrayFromDType(s.dtype,t.util.sizeFromShape(s.shape)),[c,d,p,h]=s.shape,[f,m]=t.backend_util.getImageCenter(i,d,p),k=Math.sin(r),g=Math.cos(r),b=l.data.get(s.dataId).values;for(let e=0;e<c;e++){const t=e*p*d*h;for(let e=0;e<d;e++){const n=e*(p*h);for(let a=0;a<p;a++){const s=a*h;for(let r=0;r<h;r++){const i=[c,e,a,r],l=i[2],I=i[1];let y=(l-f)*g-(I-m)*k,S=(l-f)*k+(I-m)*g;y=Math.round(y+f),S=Math.round(S+m);let T=o;if("number"!=typeof o&&(T=3===r?255:o[r]),y>=0&&y<p&&S>=0&&S<d){T=b[t+S*(p*h)+y*h+r]}u[t+n+s+r]=T}}}}return{dataId:l.write(u,s.shape,s.dtype),shape:s.shape,dtype:s.dtype}}},Us=z(t.Round,(e=>{const t=Math.floor(e);return e-t<.5?Math.floor(e):e-t>.5?Math.ceil(e):t%2==0?t:t+1})),Zs={kernelName:t.Round,backendName:"cpu",kernelFunc:Us};const js={kernelName:t.ScatterNd,backendName:"cpu",kernelFunc:function(e){const{inputs:n,backend:a,attrs:s}=e,{indices:r,updates:o}=n,{shape:i}=s,{sliceRank:l,numUpdates:u,sliceSize:c,strides:d,outputSize:p}=t.backend_util.calculateShapes(o,r,i),h=et(a.bufferSync(r),a.bufferSync(o),i,p,c,u,l,d,0,!0);return a.makeTensorInfo(i,h.dtype,h.values)}};function Ks(e,t){let n=0,a=e.length,s=0;for(;n<a;)s=Math.floor((n+a)/2),e[s]<t?n=s+1:a=s;return a}function Ys(e,t){let n=0,a=e.length,s=0;for(;n<a;)s=Math.floor((n+a)/2),e[s]<=t?n=s+1:a=s;return a}const Js={kernelName:t.SearchSorted,backendName:"cpu",kernelFunc:function(e){const{inputs:n,backend:a,attrs:s}=e,{sortedSequence:r,values:o}=n,{side:i}=s,l=function(e,n,a,s,r,o){const i=t.util.getArrayFromDType("int32",a*r);for(let t=0;t<a;++t){const a=e.slice(t*s,(t+1)*s),l=t*r;for(let e=0;e<r;++e)i[l+e]="left"===o?Ks(a,n[e+l]):Ys(a,n[e+l])}return i}(a.data.get(r.dataId).values,a.data.get(o.dataId).values,r.shape[0],r.shape[1],o.shape[1],i);return a.makeTensorInfo(o.shape,"int32",l)}};const Qs={kernelName:t.Select,backendName:"cpu",kernelFunc:function(e){const{inputs:n,backend:a}=e,{condition:s,t:o,e:i}=n;r([s,o,i],"select");const l=s.shape.length,u=a.data.get(s.dataId).values,c=a.data.get(o.dataId).values,d=a.data.get(i.dataId).values,p=t.upcastType(o.dtype,i.dtype),h=t.util.makeZerosTypedArray(t.util.sizeFromShape(o.shape),p);let f=0;const m=0===l||l>1||1===o.shape.length?1:t.util.sizeFromShape(o.shape.slice(1));for(let e=0;e<u.length;e++)for(let t=0;t<m;t++)1===u[e]?h[f++]=c[e]:h[f++]=d[e];return a.makeTensorInfo(o.shape,p,h)}},Xs=t.backend_util.SELU_SCALEALPHA,er=t.backend_util.SELU_SCALE,tr=z(t.Selu,(e=>e>=0?er*e:Xs*(Math.exp(e)-1))),nr={kernelName:t.Selu,backendName:"cpu",kernelFunc:tr},ar=z(t.Sign,(e=>e<0?-1:e>0?1:0)),sr={kernelName:t.Sign,backendName:"cpu",kernelFunc:ar},rr=z(t.Sin,(e=>Math.sin(e))),or={kernelName:t.Sin,backendName:"cpu",kernelFunc:rr},ir=z(t.Sinh,(e=>Math.sinh(e))),lr={kernelName:t.Sinh,backendName:"cpu",kernelFunc:ir},ur=Math.log(1.1920928955078125e-7)+2,cr=z(t.Softplus,(e=>{const t=e>-ur,n=e<ur,a=Math.exp(e);let s;return s=n?a:t?e:Math.log(1+a),s})),dr={kernelName:t.Softplus,backendName:"cpu",kernelFunc:cr};const pr={kernelName:t.SpaceToBatchND,backendName:"cpu",kernelFunc:function(e){const{inputs:n,backend:a,attrs:s}=e,{x:o}=n,{blockShape:i,paddings:l}=s;r([o],"spaceToBatchND");const u=t.util.sizeFromShape(i),c=[[0,0]];c.push(...l);for(let e=1+i.length;e<o.shape.length;++e)c.push([0,0]);const d=Ds.kernelFunc({inputs:{x:o},backend:a,attrs:{paddings:c,constantValue:0}}),p=t.backend_util.getReshaped(d.shape,i,u,!1),h=t.backend_util.getPermuted(p.length,i.length,!1),f=t.backend_util.getReshapedPermuted(d.shape,i,u,!1),m=Zt({inputs:{x:d},backend:a,attrs:{shape:p}}),k=We({inputs:{x:m},backend:a,attrs:{perm:h}}),g=Zt({inputs:{x:k},backend:a,attrs:{shape:f}});return a.disposeIntermediateTensorInfo(d),a.disposeIntermediateTensorInfo(m),a.disposeIntermediateTensorInfo(k),g}};const hr={kernelName:t.SparseFillEmptyRows,backendName:"cpu",kernelFunc:function(e){const{inputs:t,backend:n}=e,{indices:a,values:s,denseShape:r,defaultValue:o}=t;if(1!==r.shape.length)throw new Error(`Dense shape must be a vector, saw:\n        ${r.shape}`);if(2!==a.shape.length)throw new Error(`Indices must be a matrix, saw:\n        ${a.shape}`);if(1!==s.shape.length)throw new Error(`Values must be a vector, saw:\n        ${s.shape}`);if(0!==o.shape.length)throw new Error(`Default value must be a scalar, saw:\n        ${o.shape}`);const i=n.data.get(a.dataId).values,l=n.data.get(s.dataId).values,u=n.data.get(r.dataId).values,c=n.data.get(o.dataId).values[0],[d,p,h,f,m]=it(i,a.shape,a.dtype,l,s.dtype,u,c);return[n.makeTensorInfo(p,a.dtype,d),n.makeTensorInfo([p[0]],s.dtype,h),n.makeTensorInfo([f.length],"bool",new Uint8Array(f.map((e=>Number(e))))),n.makeTensorInfo([m.length],a.dtype,new Int32Array(m))]}};const fr={kernelName:t.SparseReshape,backendName:"cpu",kernelFunc:function(e){const{inputs:t,backend:n}=e,{inputIndices:a,inputShape:s,newShape:r}=t;if(2!==a.shape.length)throw new Error(`Input indices should be a matrix but received shape\n        ${a.shape}`);if(1!==s.shape.length)throw new Error(`Input shape should be a vector but received shape\n        ${s.shape}`);if(1!==r.shape.length)throw new Error(`Target shape should be a vector but received shape ${r.shape}`);const o=Array.from(n.data.get(s.dataId).values),i=n.data.get(a.dataId).values,l=Array.from(n.data.get(r.dataId).values),[u,c,d]=lt(i,a.shape,a.dtype,o,l);return[n.makeTensorInfo(c,a.dtype,u),n.makeTensorInfo([d.length],r.dtype,new Int32Array(d))]}};const mr={kernelName:t.SparseSegmentMean,backendName:"cpu",kernelFunc:function(e){const{inputs:t,backend:n}=e,{data:a,indices:s,segmentIds:r}=t;if(a.shape.length<1)throw new Error("Data should be at least 1 dimensional but received scalar");if(1!==s.shape.length)throw new Error(`Indices should be a vector but received shape\n          ${s.shape}`);if(1!==r.shape.length)throw new Error(`Segment ids should be a vector but received shape\n          ${r.shape}`);if(s.shape[0]!==r.shape[0])throw new Error("segmentIds and indices should have same size.");const o=n.data.get(a.dataId).values,i=n.data.get(s.dataId).values,l=n.data.get(r.dataId).values,[u,c]=ut(o,a.shape,a.dtype,i,l,!0);return n.makeTensorInfo(c,a.dtype,u)}};const kr={kernelName:t.SparseSegmentSum,backendName:"cpu",kernelFunc:function(e){const{inputs:t,backend:n}=e,{data:a,indices:s,segmentIds:r}=t;if(a.shape.length<1)throw new Error("Data should be at least 1 dimensional but received scalar");if(1!==s.shape.length)throw new Error(`Indices should be a vector but received shape\n         ${s.shape}`);if(1!==r.shape.length)throw new Error(`Segment ids should be a vector but received shape\n         ${r.shape}`);if(s.shape[0]!==r.shape[0])throw new Error("segmentIds and indices should have same size.");const o=n.data.get(a.dataId).values,i=n.data.get(s.dataId).values,l=n.data.get(r.dataId).values,[u,c]=ut(o,a.shape,a.dtype,i,l);return n.makeTensorInfo(c,a.dtype,u)}};const gr={kernelName:t.SparseToDense,backendName:"cpu",kernelFunc:function(e){const{inputs:n,backend:a,attrs:s}=e,{sparseIndices:r,sparseValues:o,defaultValue:i}=n,{outputShape:l}=s,{sliceRank:u,numUpdates:c,sliceSize:d,strides:p,outputSize:h}=t.backend_util.calculateShapes(o,r,l),f=!1,m=a.bufferSync(r);let k;switch(o.dtype){case"bool":k=et(m,a.bufferSync(o),l,h,d,c,u,p,Boolean(a.data.get(i.dataId).values[0]),f);break;case"float32":k=et(m,a.bufferSync(o),l,h,d,c,u,p,a.data.get(i.dataId).values[0],f);break;case"int32":k=et(m,a.bufferSync(o),l,h,d,c,u,p,a.data.get(i.dataId).values[0],f);break;case"string":k=et(m,a.bufferSync(o),l,h,d,c,u,p,t.util.decodeString(a.data.get(i.dataId).values[0]),f);break;default:throw new Error(`Unsupported type ${o.dtype}`)}return a.makeTensorInfo(l,k.dtype,k.values)}};const br={kernelName:t.SplitV,backendName:"cpu",kernelFunc:function(e){const{inputs:n,backend:a,attrs:s}=e,{x:r}=n,{numOrSizeSplits:o,axis:i}=s,l=t.util.parseAxisParam(i,r.shape)[0],u=t.backend_util.prepareSplitSize(r,o,l),c=new Array(r.shape.length).fill(0),d=r.shape.slice();return u.map((e=>{const t=[...d];t[l]=e;const n=rt({inputs:{x:r},backend:a,attrs:{begin:c,size:t}});return c[l]+=e,n}))}},Ir={kernelName:t.Square,backendName:"cpu",kernelFunc:({inputs:e,backend:t})=>{const{x:n}=e,a=t;r(n,"square");const s=a.data.get(n.dataId).values,o=new Float32Array(s.length);for(let e=0;e<s.length;++e){const t=s[e];o[e]=t*t}return{dataId:a.write(o,n.shape,n.dtype),shape:n.shape,dtype:n.dtype}}},yr=z(t.Step,((e,t)=>{const n=t;return isNaN(e)?NaN:e>0?1:n.alpha})),Sr={kernelName:t.Step,backendName:"cpu",kernelFunc:yr};const Tr={kernelName:t.StridedSlice,backendName:"cpu",kernelFunc:function(e){const{inputs:n,backend:a,attrs:s}=e,{x:o}=n,{begin:i,end:l,strides:u,beginMask:c,endMask:d,ellipsisMask:p,newAxisMask:h,shrinkAxisMask:f}=s;r(o,"stridedSlice");const{finalShapeSparse:m,finalShape:k,isIdentity:g,sliceDim0:b,isSimpleSlice:I,begin:y,end:S,strides:T}=t.slice_util.sliceInfo(o.shape,i,l,u,c,d,p,h,f);let N;if(g)N=Zt({inputs:{x:o},backend:a,attrs:{shape:k}});else if(b||I){t.util.assert(o.shape.length>=1,(()=>`Input must have rank at least 1, got: ${o.shape.length}`));const e=t.slice_util.computeOutShape(y,S,T),n=rt({inputs:{x:o},backend:a,attrs:{begin:y,size:e}});N=Zt({inputs:{x:n},backend:a,attrs:{shape:k}}),a.disposeIntermediateTensorInfo(n)}else{const e=It(m,a.bufferSync(o),T,y);N=a.makeTensorInfo(k,e.dtype,e.values)}return N}};const Nr={kernelName:t.StringNGrams,backendName:"cpu",kernelFunc:function(e){const{inputs:t,backend:n,attrs:a}=e,{separator:s,nGramWidths:r,leftPad:o,rightPad:i,padWidth:l,preserveShortSequences:u}=a,{data:c,dataSplits:d}=t,p=n.data.get(c.dataId).values,h=n.data.get(d.dataId).values,[f,m]=St(p,h,s,r,o,i,l,u);return[n.makeTensorInfo([f.length],"string",f),n.makeTensorInfo(d.shape,"int32",m)]}};const xr={kernelName:t.StringSplit,backendName:"cpu",kernelFunc:function(e){const{inputs:t,backend:n,attrs:a}=e,{skipEmpty:s}=a,{input:r,delimiter:o}=t;if("string"!==r.dtype)throw new Error("Input must be of datatype string");if(1!==r.shape.length)throw new Error(`Input must be a vector, got shape: ${r.shape}`);if(0!==o.shape.length)throw new Error(`Delimiter must be a scalar, got shape: ${o.shape}`);const i=n.data.get(r.dataId).values,l=n.data.get(o.dataId).values[0],[u,c,d]=Nt(i,l,s),p=c.length;return[n.makeTensorInfo([p,2],"int32",u),n.makeTensorInfo([p],"string",c),n.makeTensorInfo([2],"int32",new Int32Array(d))]}};const vr={kernelName:t.StringToHashBucketFast,backendName:"cpu",kernelFunc:function(e){const{inputs:t,backend:n,attrs:a}=e,{numBuckets:s}=a,{input:r}=t;if("string"!==r.dtype)throw new Error("Input must be of datatype string");if(s<=0)throw new Error("Number of buckets must be at least 1");const o=xt(n.data.get(r.dataId).values,s);return n.makeTensorInfo(r.shape,"int32",o)}},Fr=z(t.Tan,(e=>Math.tan(e))),wr={kernelName:t.Tan,backendName:"cpu",kernelFunc:Fr},Mr=z(t.Tanh,(e=>Math.tanh(e))),Ar={kernelName:t.Tanh,backendName:"cpu",kernelFunc:Mr};const Dr={kernelName:t.TensorScatterUpdate,backendName:"cpu",kernelFunc:function(e){const{inputs:n,backend:a}=e,{tensor:s,indices:r,updates:o}=n,{sliceRank:i,numUpdates:l,sliceSize:u,strides:c,outputSize:d}=t.backend_util.calculateShapes(o,r,s.shape),p=a.bufferSync(r),h=a.bufferSync(o),f=a.bufferSync(s),m=et(p,h,s.shape,d,u,l,i,c,f,!1);return a.makeTensorInfo(s.shape,m.dtype,m.values)}};const Er={kernelName:t.Tile,backendName:"cpu",kernelFunc:function(e){const{inputs:t,backend:n,attrs:a}=e,{x:s}=t,{reps:o}=a;r(s,"tile");const i=At(n.bufferSync(s),o);return n.makeTensorInfo(i.shape,i.dtype,i.values)}};const _r={kernelName:t.TopK,backendName:"cpu",kernelFunc:function(e){const{inputs:t,backend:n,attrs:a}=e,{x:s}=t,{k:o,sorted:i}=a;r(s,"topk");const l=n.data.get(s.dataId).values,[u,c]=_t(l,s.shape,s.dtype,o,i);return[n.makeTensorInfo(u.shape,u.dtype,u.values),n.makeTensorInfo(c.shape,c.dtype,c.values)]}};const zr={kernelName:t.Transform,backendName:"cpu",kernelFunc:function(e){const{inputs:n,attrs:a,backend:s}=e,{image:r,transforms:o}=n,{interpolation:i,fillMode:l,fillValue:u,outputShape:c}=a,[d,p,h,f]=r.shape,[m,k]=null!=c?c:[p,h],g=[d,m,k,f],b=t.util.computeStrides(r.shape),I=b[0],y=b[1],S=b[2],T=t.util.computeStrides(g),N=T[0],x=T[1],v=T[2],F=t.util.getTypedArrayFromDType(r.dtype,t.util.sizeFromShape(g));F.fill(u);const w=s.data.get(r.dataId).values,M=s.data.get(o.dataId).values;for(let e=0;e<d;++e){const t=1===o.shape[0]?M:M.subarray(8*e,8*e+8);for(let n=0;n<m;++n)for(let a=0;a<k;++a)for(let s=0;s<f;++s){let r;const o=t[6]*a+t[7]*n+1;if(0===o)continue;const c=(t[0]*a+t[1]*n+t[2])/o,d=(t[3]*a+t[4]*n+t[5])/o,f=Rr(c,h,l),m=Rr(d,p,l);switch(i){case"nearest":r=Pr(w,p,h,I,y,S,e,m,f,s,u);break;case"bilinear":r=Cr(w,p,h,I,y,S,e,m,f,s,u);break;default:throw new Error(`Error in Transform: Expect 'nearest' or 'bilinear', but got ${i}`)}F[e*N+n*x+a*v+s]=r}return s.makeTensorInfo(g,r.dtype,F)}return{dataId:s.write(F,g,r.dtype),shape:r.shape,dtype:r.dtype}}};function Rr(e,n,a){switch(a){case"reflect":return function(e,n){let a=e;if(a<0)if(n<=1)a=0;else{const e=2*n;a<e&&(a=e*Math.trunc(-a/e)+a),a=a<-n?a+e:-a-1}else if(a>n-1)if(n<=1)a=0;else{const e=2*n;a-=e*Math.trunc(a/e),a>=n&&(a=e-a-1)}return t.util.clamp(0,a,n-1)}(e,n);case"wrap":return function(e,n){let a=e;if(a<0)if(n<=1)a=0;else{const e=n-1;a+=n*(Math.trunc(-a/e)+1)}else if(a>n-1)if(n<=1)a=0;else{const e=n-1;a-=n*Math.trunc(a/e)}return t.util.clamp(0,a,n-1)}(e,n);case"nearest":return function(e,n){return t.util.clamp(0,e,n-1)}(e,n);default:return function(e,t){return e}(e)}}function Wr(e,t,n,a,s,r,o,i,l,u,c){return 0<=i&&i<t&&0<=l&&l<n?e[o*a+i*s+l*r+u]:c}function Pr(e,t,n,a,s,r,o,i,l,u,c){return Wr(e,t,n,a,s,r,o,Math.round(i),Math.round(l),u,c)}function Cr(e,t,n,a,s,r,o,i,l,u,c){const d=Math.floor(i),p=Math.floor(l),h=d+1,f=p+1;return(h-i)*((f-l)*Wr(e,t,n,a,s,r,o,d,p,u,c)+(l-p)*Wr(e,t,n,a,s,r,o,d,f,u,c))+(i-d)*((f-l)*Wr(e,t,n,a,s,r,o,h,p,u,c)+(l-p)*Wr(e,t,n,a,s,r,o,h,f,u,c))}const Hr={kernelName:t.Unique,backendName:"cpu",kernelFunc:function(e){const{inputs:t,attrs:n,backend:a}=e,{axis:s}=n,{x:o}=t;r(o,"unique");const i=a.data.get(o.dataId).values,{outputValues:l,outputShape:u,indices:c}=zt(i,s,o.shape,o.dtype);return[a.makeTensorInfo(u,o.dtype,l),a.makeTensorInfo([c.length],"int32",c)]}};const Or={kernelName:t.Unpack,backendName:"cpu",kernelFunc:function(e){const{inputs:t,backend:n,attrs:a}=e,{value:s}=t;let{axis:r}=a;r<0&&(r+=s.shape.length);const o=s.shape.length,i=s.shape[r],l=new Array(o-1);let u=0;for(let e=0;e<o;e++)e!==r&&(l[u++]=s.shape[e]);const c=new Array(o).fill(0),d=s.shape.slice();d[r]=1;const p=new Array(i);for(let e=0;e<p.length;e++){c[r]=e;const t=rt({inputs:{x:s},backend:n,attrs:{begin:c,size:d}});p[e]=Zt({inputs:{x:t},backend:n,attrs:{shape:l}}),n.disposeIntermediateTensorInfo(t)}return p}};const Br={kernelName:t.UnsortedSegmentSum,backendName:"cpu",kernelFunc:function(e){const{inputs:n,backend:a,attrs:s}=e,{x:o,segmentIds:i}=n,{numSegments:l}=s;r(o,"unsortedSegmentSum");const u=[],c=[],d=o.shape.length-i.shape.length;let p=i;for(let e=0;e<d;++e){const t=ya({inputs:{input:p},backend:a,attrs:{dim:e+1}});p=t,c.push(t)}for(let e=0;e<l;++e){const n=t.util.createScalarValue(e,"int32"),s=a.makeTensorInfo([],"int32",n),r=B({inputs:{a:s,b:p},backend:a}),i=I({inputs:{x:r},backend:a,attrs:{dtype:"float32"}}),l=we({inputs:{a:i,b:o},backend:a}),d=la({inputs:{x:l},backend:a,attrs:{axis:0,keepDims:!1}});u.push(d),c.push(s),c.push(r),c.push(i),c.push(l),c.push(d)}const h=Ms({inputs:u,backend:a,attrs:{axis:0}});return c.forEach((e=>a.disposeIntermediateTensorInfo(e))),h}},$r=[Jt,u,Xt,tn,F,nn,an,sn,rn,on,un,dn,hn,kn,bn,Tn,Nn,xn,vn,Yt,Fn,wn,Mn,E,An,y,C,En,p,_n,Pn,Hn,On,Bn,$n,Vn,Gn,qn,Zn,jn,Kn,Yn,Jn,Qn,ea,ta,na,aa,sa,ra,oa,ia,ca,Pt,da,$,Ia,L,Sa,Z,Ma,Da,Ea,Y,X,_a,za,Ra,Wa,se,ie,m,Pa,Rn,Ha,Ba,Va,Ht,ce,he,Ga,ge,qa,ja,Ya,Xa,es,ts,as,Se,ss,rs,os,is,ls,us,cs,xe,ds,fs,gs,Me,De,Is,Ss,Ns,ze,xs,ws,As,Ds,zs,$t,He,Rs,Ws,Ps,Cs,g,xa,Os,Gt,qt,jt,Bs,$s,Vs,Gs,Ls,qs,Zs,Xe,js,Js,Qs,nr,at,sr,or,lr,ot,ks,dr,pr,hr,fr,mr,kr,gr,br,pt,Ir,mt,bt,Sr,Tr,Nr,xr,vr,Mt,ua,wr,Ar,Dr,Er,_r,zr,Pe,Hr,Or,Br,Fs];for(const e of $r)t.registerKernel(e);e.MathBackendCPU=i,e.shared=Rt,e.version_cpu="4.15.0"}));
//# sourceMappingURL=tf-backend-cpu.es2017.min.js.map