309 lines
12 KiB
C++
309 lines
12 KiB
C++
/*******************************************************************************
|
|
* Copyright (c) 2015-2018 Skymind, Inc.
|
|
*
|
|
* This program and the accompanying materials are made available under the
|
|
* terms of the Apache License, Version 2.0 which is available at
|
|
* https://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.
|
|
*
|
|
* SPDX-License-Identifier: Apache-2.0
|
|
******************************************************************************/
|
|
|
|
//
|
|
// @author Yurii Shyrma (iuriish@yahoo.com), created on 15.02.2018, Alex Black
|
|
//
|
|
|
|
// implementation of gated Recurrent Unit cell
|
|
// (cf. http://arxiv.org/abs/1406.1078).
|
|
// Kyunghyun Cho, Bart van Merrienboer, Caglar Gulcehre, Dzmitry Bahdanau, Fethi Bougares, Holger Schwenk, Yoshua Bengio
|
|
// "Learning Phrase Representations using RNN Encoder-Decoder for Statistical Machine Translation"
|
|
|
|
|
|
#include<ops/declarable/helpers/gru.h>
|
|
#include <ops/declarable/CustomOperations.h>
|
|
#include<ops/declarable/helpers/transforms.h>
|
|
#include <MmulHelper.h>
|
|
|
|
namespace nd4j {
|
|
namespace ops {
|
|
namespace helpers {
|
|
|
|
|
|
//////////////////////////////////////////////////////////////////////////
|
|
static FORCEINLINE NDArray sigmoid(const NDArray& arr) {
|
|
return (const_cast<NDArray&>(arr)).transform(transform::Sigmoid);
|
|
}
|
|
|
|
static FORCEINLINE void sigmoidInplace(const NDArray& arr) {
|
|
(const_cast<NDArray&>(arr)).applyTransform(transform::Sigmoid);
|
|
}
|
|
|
|
//////////////////////////////////////////////////////////////////////////
|
|
static FORCEINLINE NDArray tanh(const NDArray& arr) {
|
|
return (const_cast<NDArray&>(arr)).transform(transform::Tanh);
|
|
}
|
|
|
|
static FORCEINLINE void tanhInplace(const NDArray& arr) {
|
|
(const_cast<NDArray&>(arr)).applyTransform(transform::Tanh);
|
|
}
|
|
|
|
//////////////////////////////////////////////////////////////////////////
|
|
void gruCell(nd4j::LaunchContext * context, const NDArray* x, const NDArray* hLast, const NDArray* Wru, const NDArray* Wc,
|
|
const NDArray* bru, const NDArray* bc,
|
|
NDArray* r, NDArray* u, NDArray* c, NDArray* h) {
|
|
|
|
//Inputs:
|
|
// x input [bS x inSize]
|
|
// hLast previous cell output [bS x numUnits], that is at previous time step t-1
|
|
// Wru RU weights - [bS, 2*numUnits] - reset and update gates
|
|
// Wc C weights - [bS, numUnits] - cell gate
|
|
// bru r and u biases, [2*numUnits] - reset and update gates
|
|
// bc c biases, [numUnits] - cell gate
|
|
|
|
//Outputs:
|
|
// r Reset gate output [bS, numUnits]
|
|
// u Update gate output [bS, numUnits]
|
|
// c Cell gate output [bS, numUnits]
|
|
// h current cell output [bS, numUnits]
|
|
|
|
const int nIn = x->sizeAt(1);
|
|
const int nU = hLast->sizeAt(1); // number of units
|
|
|
|
//Concat inputs: [x, yt-1]: concat([bs,nIn],[bs,nOut]) -> [bs, (nIn+nOut)]
|
|
nd4j::ops::concat concatOp;
|
|
std::vector<NDArray*> inputs;
|
|
std::vector<double> targs;
|
|
std::vector<Nd4jLong> iargs({1}); //Axis = 1
|
|
std::vector<bool> bargs;
|
|
inputs.emplace_back(const_cast<NDArray*>(x));
|
|
inputs.emplace_back(const_cast<NDArray*>(hLast));
|
|
|
|
auto result = concatOp.execute(inputs, targs, iargs, bargs);
|
|
auto concatOut = result->at(0);
|
|
|
|
//mmul/z for reset and update gates: (x * weight_ux + hLast * weight_xr + b_u)
|
|
auto m = mmul(*concatOut, *Wru); //mmul: [bs, (nIn+numUnits)]* [(inSize+numUnits), 2*numUnits] = [bs, 4*numUnits]
|
|
m += (*bru);
|
|
|
|
sigmoidInplace(m); //sigmoid(rz) and sigmoid(uz)
|
|
auto mr = m({0,0, 0, nU});
|
|
auto mu = m({0,0, nU, 2*nU});
|
|
|
|
r->assign(&mr);
|
|
u->assign(&mu);
|
|
|
|
//Concatenated inputs: [x, yt-1 .* r]
|
|
auto yr = (*concatOut)({0,0, nIn, nIn+nU});
|
|
yr *= (*r);
|
|
|
|
//c = tanh(x * weight_cx + (hLast .* r) * weight_cr + b_c)
|
|
MmulHelper::mmul(concatOut, const_cast<NDArray*>(Wc), c, 1.0, 0.0); //c = 1.0 * concatOut * Wc + 0.0 * c
|
|
*c += *bc;
|
|
tanhInplace(*c);
|
|
|
|
//Output: h = (1-u).*c + u .* hPrev
|
|
//auto hResult = (*u) * (*hLast) + (1.0f - *u) * (*c); const_cast<NDArray*>(h)->assign(&hResult);
|
|
u->applyPairwiseTransform(pairwise::Multiply, hLast, h, nullptr); //h = u * hLast
|
|
auto temp = (1.0f - *u);
|
|
temp *= (*c);
|
|
(*h) += temp;
|
|
|
|
delete result;
|
|
}
|
|
|
|
//////////////////////////////////////////////////////////////////////////
|
|
void gruTimeLoop(nd4j::LaunchContext * context, const NDArray* x, const NDArray* h0, const NDArray* Wx, const NDArray* Wh, const NDArray* b, NDArray* h) {
|
|
|
|
// x input [time, bS, iS]
|
|
// h0 initial cell output (at time step = 0) [bS, nU]
|
|
// Wx input-to-hidden weights, [iS, 3*nU]
|
|
// Wh hidden-to-hidden weights, [nU, 3*nU]
|
|
// b biases, [3*nU]
|
|
|
|
// h is cell outputs at each time step [time, bS, nU]
|
|
|
|
const int time = x->sizeAt(0);
|
|
|
|
NDArray ht_1(*h0);
|
|
|
|
// loop through time steps
|
|
for (int t = 0; t < time; ++t) {
|
|
|
|
auto xt = (*x)({t,t+1, 0,0, 0,0});
|
|
auto ht = (*h)({t,t+1, 0,0, 0,0});
|
|
|
|
//helpers::gruCell(&xt, &ht_1, Wx, Wh, b, &ht);
|
|
//ht_1.assign(ht);
|
|
}
|
|
}
|
|
|
|
//////////////////////////////////////////////////////////////////////////
|
|
void gruCellBP(nd4j::LaunchContext * context, const NDArray* x, const NDArray* h0, const NDArray* Wx, const NDArray* Wh, const NDArray* b, const NDArray* dLdh, const NDArray* dLdWx0,
|
|
const NDArray* dLdWh0, const NDArray* dLdb0, NDArray* dLdx, NDArray* dLdh0, NDArray* dLdWx, NDArray* dLdWh, NDArray* dLdb) {
|
|
|
|
// x input [bS, iS]
|
|
// h0 previous cell output [bS, nU], that is at previous time step t-1
|
|
// Wx input-to-hidden weights, [iS, 3*nU]
|
|
// Wh hidden-to-hidden weights, [nU, 3*nU]
|
|
// b biases, [3*nU]
|
|
// dLdh gradient wrt output, [bS,nU], that is epsilon_next
|
|
// dLdWx0 gradient wrt Wx at previous time step, [iS, 3*nU]
|
|
// dLdWh0 gradient wrt Wh at previous time step, [nU, 3*nU]
|
|
// dLdb0 gradient wrt b at previous time step, [3*nU]
|
|
|
|
// dLdx gradient wrt x, [bS, iS], that is epsilon
|
|
// dLdh0 gradient wrt h0, [bS, nU]
|
|
// dLdWx gradient wrt Wx, [iS, 3*nU]
|
|
// dLdWh gradient wrt Wh, [nU, 3*nU]
|
|
// dLdb gradient wrt b at previous time step, [3*nU]
|
|
|
|
// h is current cell output [bS, nU], that is at current time step t
|
|
|
|
const int nU = h0->sizeAt(1);
|
|
|
|
// ***** feed forward step ***** //
|
|
// gates = sigmoid(x*Wx + h0*Wh + b)
|
|
auto gates = sigmoid(mmul(*x, (*Wx)({0,0, 0,2*nU})) + mmul(*h0, (*Wh)({0,0, 0,2*nU})) + (*b)({0,2*nU})); // [bS, 2*nU] + [bS, 2*nU] + [1, 2*nU] = [bS, 2*nU]
|
|
// reset gate
|
|
auto r = gates({0,0, 0, nU}); // [bS, nU]
|
|
// update gate
|
|
auto u = gates({0,0, nU, 2*nU}); // [bS, nU]
|
|
// ◦ means element-wise product or so called Hadamard product
|
|
// n = tanh(x*Wx + (r◦h0)*Wh + b)
|
|
auto n = tanh(mmul(*x, (*Wx)({0,0, 2*nU,3*nU})) + mmul((*h0)*r, (*Wh)({0,0, 2*nU,3*nU})) + (*b)({2*nU,3*nU})); // [bS, nU]
|
|
|
|
// ***** back prop step ***** //
|
|
auto Wxr = (*Wx)({0,0, 0, nU});
|
|
auto Wxu = (*Wx)({0,0, nU, 2*nU});
|
|
auto Wxn = (*Wx)({0,0, 2*nU,3*nU});
|
|
auto Whr = (*Wh)({0,0, 0, nU});
|
|
auto Whu = (*Wh)({0,0, nU, 2*nU});
|
|
auto Whn = (*Wh)({0,0, 2*nU,3*nU});
|
|
auto WxrT = Wxr.transp();
|
|
auto WxuT = Wxu.transp();
|
|
auto WxnT = Wxn.transp();
|
|
auto WhrT = Whr.transp();
|
|
auto WhuT = Whu.transp();
|
|
auto WhnT = Whn.transp();
|
|
auto xT = x->transp();
|
|
auto h0T = h0->transp();
|
|
|
|
auto dLdWxr = (*dLdWx)({0,0, 0, nU});
|
|
auto dLdWxu = (*dLdWx)({0,0, nU, 2*nU});
|
|
auto dLdWxn = (*dLdWx)({0,0, 2*nU,3*nU});
|
|
|
|
auto dLdWhr = (*dLdWh)({0,0, 0, nU});
|
|
auto dLdWhu = (*dLdWh)({0,0, nU, 2*nU});
|
|
auto dLdWhn = (*dLdWh)({0,0, 2*nU,3*nU});
|
|
|
|
auto dLdbr = (*dLdb)({0, nU});
|
|
auto dLdbu = (*dLdb)({nU, 2*nU});
|
|
auto dLdbn = (*dLdb)({2*nU,3*nU});
|
|
|
|
auto dhdu = *h0 - n; // [bS, nU]
|
|
auto dhdn = 1.f - u; // [bS, nU]
|
|
auto dSigdu = u * (1.f - u); // [bS, nU]
|
|
auto dSigdr = r * (1.f - r); // [bS, nU]
|
|
auto dActdn = 1.f - n * n; // [bS, nU]
|
|
auto dndr = mmul(dActdn * (*h0), WhnT);
|
|
auto drdh0 = mmul(dSigdr, WhrT);
|
|
|
|
auto dLdn = (*dLdh) * dhdn;
|
|
auto dLdu = (*dLdh) * dhdu;
|
|
auto dLdr = dLdn * dndr;
|
|
|
|
dLdx->assign( mmul(dLdu * dSigdu, WxuT) + mmul(dLdr * dSigdr, WxrT) + mmul(dLdn * dActdn, WxnT) ); // [bS,iS]
|
|
dLdh0->assign( mmul(dLdu * dSigdu, WhuT) + mmul(dLdn * dActdn * (r + drdh0), WhnT) + (*dLdh)*u ); // [bS,nU]
|
|
|
|
dLdWxr.assign( mmul(xT, dSigdr * dLdr) ); // [iS,nU]
|
|
dLdWhr.assign( mmul(h0T, dSigdr * dLdr) ); // [nU,nU]
|
|
|
|
dLdWxu.assign( mmul(xT, dSigdu * dLdu) ); // [iS,nU]
|
|
dLdWhu.assign( mmul(h0T, dSigdu * dLdu) ); // [nU,nU]
|
|
|
|
dLdWxn.assign( mmul(xT, dActdn * dLdn) ); // [iS,nU]
|
|
dLdWhn.assign( mmul((r*(*h0)).transp(), dActdn * dLdn) ); // [nU,nU]
|
|
|
|
dLdbr.assign( (dSigdr * dLdr).reduceAlongDims(reduce::Sum, {0})); // [nU]
|
|
dLdbu.assign( (dSigdu * dLdu).reduceAlongDims(reduce::Sum, {0})); // [nU]
|
|
dLdbn.assign( (dActdn * dLdn).reduceAlongDims(reduce::Sum, {0})); // [nU]
|
|
|
|
if(dLdWx0 != nullptr)
|
|
*dLdWx += *dLdWx0;
|
|
|
|
if(dLdWh0 != nullptr)
|
|
*dLdWh += *dLdWh0;
|
|
|
|
if(dLdb0 != nullptr)
|
|
*dLdb += *dLdb0;
|
|
|
|
}
|
|
|
|
// //////////////////////////////////////////////////////////////////////////
|
|
// FIXME - gruTimeLoopBP is not correct
|
|
// template <typename T>
|
|
// void gruTimeLoopBP(const std::vector<NDArray<T>*>& inArrs, const std::vector<NDArray<T>*>& outArrs) {
|
|
|
|
// NDArray<T>* x = inArrs[0]; // input [time, bS, iS]
|
|
// NDArray<T>* hi = inArrs[1]; // previous/initial cell output [bS, nU], that is at previous time step t-1
|
|
// NDArray<T>* Wx = inArrs[2]; // input-to-hidden weights, [iS, 3*nU]
|
|
// NDArray<T>* Wh = inArrs[3]; // hidden-to-hidden weights, [nU, 3*nU]
|
|
// NDArray<T>* b = inArrs[4]; // biases, [3*nU]
|
|
// NDArray<T>* dLdh = inArrs[5]; // gradient wrt output, [time, bS, nU], that is epsilon_next
|
|
|
|
// NDArray<T>* dLdx = outArrs[0]; // gradient wrt x, [time, bS, iS], that is epsilon
|
|
// NDArray<T>* dLdhi = outArrs[1]; // gradient wrt hi, [bS, nU]
|
|
// NDArray<T>* dLdWx = outArrs[2]; // gradient wrt Wx, [iS, 3*nU]
|
|
// NDArray<T>* dLdWh = outArrs[3]; // gradient wrt Wh, [nU, 3*nU]
|
|
// NDArray<T>* dLdb = outArrs[4]; // gradient wrt b, [3*nU]
|
|
|
|
// const Nd4jLong time = x->sizeAt(0);
|
|
// const Nd4jLong bS = x->sizeAt(1);
|
|
// const Nd4jLong iS = x->sizeAt(2);
|
|
// const Nd4jLong nU = hi->sizeAt(1);
|
|
|
|
// NDArray<T> h(hi->ordering(), {time, bS, nU}); // feed forward output
|
|
|
|
// // first step, time = 0, feed forward
|
|
// NDArray<T> x0 = (*x)({{0,1}, {}, {}});
|
|
// NDArray<T> h0 = h({{0,1}, {}, {}});
|
|
// helpers::gruCell<T>({&x0, hi, Wx, Wh, b}, &h0);
|
|
|
|
// // first step, time = 0, back prop
|
|
// NDArray<T> dLdx0 = (*dLdx)({{0,1}, {}, {}});
|
|
// NDArray<T> dLdh0 = (*dLdh)({{0,1}, {}, {}});
|
|
// helpers::gruCellBP<T>({&x0, hi, Wx, Wh, b, &dLdh0, nullptr, nullptr, nullptr}, {&dLdx0, dLdhi, dLdWx, dLdWh, dLdb});
|
|
|
|
// // loop through the rest time steps
|
|
// for (Nd4jLong t = time-1; t > 0; --t) {
|
|
// for (Nd4jLong t = 1; t < time; ++t) {
|
|
|
|
// NDArray<T> xt = (*x)({{t,t+1}, {}, {}});
|
|
// NDArray<T> ht = h({{t,t+1}, {}, {}});
|
|
// NDArray<T> ht_1 = h({{t-1,t}, {}, {}});
|
|
// NDArray<T> dLdxt = (*dLdx)({{t,t+1}, {}, {}});
|
|
// NDArray<T> dLdht = (*dLdh)({{t,t+1}, {}, {}});
|
|
|
|
// NDArray<T> dLdWxt_1 = dLdWx;
|
|
// NDArray<T> dLdWht_1 = dLdWh;
|
|
// NDArray<T> dLdbt_1 = dLdb;
|
|
|
|
// // feed forward, calculation of ht
|
|
// helpers::gruCell<T>({&xt, &ht_1, Wx, Wh, b}, &ht);
|
|
|
|
// // back prop
|
|
// helpers::gruCellBP<T>({&xt, &ht_1, Wx, Wh, b, &dLdht, &dLdWxt_1, &dLdWht_1, &dLdbt_1}, {&dLdxt, nullptr, dLdWx, dLdWh, dLdb});
|
|
// }
|
|
// }
|
|
|
|
|
|
}
|
|
}
|
|
}
|
|
|