// matrix/matrix-functions.cc // Copyright 2009-2011 Microsoft Corporation; Go Vivace Inc.; Jan Silovsky // Yanmin Qian; Saarland University; Johns Hopkins University (Author: Daniel Povey) // See ../../COPYING for clarification regarding multiple authors // // 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 // // THIS CODE IS PROVIDED *AS IS* BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY // KIND, EITHER EXPRESS OR IMPLIED, INCLUDING WITHOUT LIMITATION ANY IMPLIED // WARRANTIES OR CONDITIONS OF TITLE, FITNESS FOR A PARTICULAR PURPOSE, // MERCHANTABLITY OR NON-INFRINGEMENT. // See the Apache 2 License for the specific language governing permissions and // limitations under the License. // // (*) incorporates, with permission, FFT code from his book // "Signal Processing with Lapped Transforms", Artech, 1992. #include "matrix/matrix-functions.h" #include "matrix/sp-matrix.h" namespace kaldi { template void ComplexFt (const VectorBase &in, VectorBase *out, bool forward) { int exp_sign = (forward ? -1 : 1); KALDI_ASSERT(out != NULL); KALDI_ASSERT(in.Dim() == out->Dim()); KALDI_ASSERT(in.Dim() % 2 == 0); int twoN = in.Dim(), N = twoN / 2; const Real *data_in = in.Data(); Real *data_out = out->Data(); Real exp1N_re, exp1N_im; // forward -> exp(-2pi / N), backward -> exp(2pi / N). Real fraction = exp_sign * M_2PI / static_cast(N); // forward -> -2pi/N, backward->-2pi/N ComplexImExp(fraction, &exp1N_re, &exp1N_im); Real expm_re = 1.0, expm_im = 0.0; // forward -> exp(-2pi m / N). for (int two_m = 0; two_m < twoN; two_m+=2) { // For each output component. Real expmn_re = 1.0, expmn_im = 0.0; // forward -> exp(-2pi m n / N). Real sum_re = 0.0, sum_im = 0.0; // complex output for index m (the sum expression) for (int two_n = 0; two_n < twoN; two_n+=2) { ComplexAddProduct(data_in[two_n], data_in[two_n+1], expmn_re, expmn_im, &sum_re, &sum_im); ComplexMul(expm_re, expm_im, &expmn_re, &expmn_im); } data_out[two_m] = sum_re; data_out[two_m + 1] = sum_im; if (two_m % 10 == 0) { // occasionally renew "expm" from scratch to avoid // loss of precision. int nextm = 1 + two_m/2; Real fraction_mult = fraction * nextm; ComplexImExp(fraction_mult, &expm_re, &expm_im); } else { ComplexMul(exp1N_re, exp1N_im, &expm_re, &expm_im); } } } template void ComplexFt (const VectorBase &in, VectorBase *out, bool forward); template void ComplexFt (const VectorBase &in, VectorBase *out, bool forward); #define KALDI_COMPLEXFFT_BLOCKSIZE 8192 // This #define affects how we recurse in ComplexFftRecursive. // We assume that memory-caching happens on a scale at // least as small as this. //! ComplexFftRecursive is a recursive function that computes the //! complex FFT of size N. The "nffts" arguments specifies how many //! separate FFTs to compute in parallel (we assume the data for //! each one is consecutive in memory). The "forward argument" //! specifies whether to do the FFT (true) or IFFT (false), although //! note that we do not include the factor of 1/N (the user should //! do this if required. The iterators factor_begin and factor_end //! point to the beginning and end (i.e. one past the last element) //! of an array of small factors of N (typically prime factors). //! See the comments below this code for the detailed equations //! of the recursion. template void ComplexFftRecursive (Real *data, int nffts, int N, const int *factor_begin, const int *factor_end, bool forward, Vector *tmp_vec) { if (factor_begin == factor_end) { KALDI_ASSERT(N == 1); return; } { // an optimization: compute in smaller blocks. // this block of code could be removed and it would still work. MatrixIndexT size_perblock = N * 2 * sizeof(Real); if (nffts > 1 && size_perblock*nffts > KALDI_COMPLEXFFT_BLOCKSIZE) { // can break it up... // Break up into multiple blocks. This is an optimization. We make // no progress on the FFT when we do this. int block_skip = KALDI_COMPLEXFFT_BLOCKSIZE / size_perblock; // n blocks per call if (block_skip == 0) block_skip = 1; if (block_skip < nffts) { int blocks_left = nffts; while (blocks_left > 0) { int skip_now = std::min(blocks_left, block_skip); ComplexFftRecursive(data, skip_now, N, factor_begin, factor_end, forward, tmp_vec); blocks_left -= skip_now; data += skip_now * N*2; } return; } // else do the actual algorithm. } // else do the actual algorithm. } int P = *factor_begin; KALDI_ASSERT(P > 1); int Q = N / P; if (P > 1 && Q > 1) { // Do the rearrangement. C.f. eq. (8) below. Transform // (a) to (b). Real *data_thisblock = data; if (tmp_vec->Dim() < (MatrixIndexT)N) tmp_vec->Resize(N); Real *data_tmp = tmp_vec->Data(); for (int thisfft = 0; thisfft < nffts; thisfft++, data_thisblock+=N*2) { for (int offset = 0; offset < 2; offset++) { // 0 == real, 1 == im. for (int p = 0; p < P; p++) { for (int q = 0; q < Q; q++) { int aidx = q*P + p, bidx = p*Q + q; data_tmp[bidx] = data_thisblock[2*aidx+offset]; } } for (int n = 0;n < P*Q;n++) data_thisblock[2*n+offset] = data_tmp[n]; } } } { // Recurse. ComplexFftRecursive(data, nffts*P, Q, factor_begin+1, factor_end, forward, tmp_vec); } int exp_sign = (forward ? -1 : 1); Real rootN_re, rootN_im; // Nth root of unity. ComplexImExp(static_cast(exp_sign * M_2PI / N), &rootN_re, &rootN_im); Real rootP_re, rootP_im; // Pth root of unity. ComplexImExp(static_cast(exp_sign * M_2PI / P), &rootP_re, &rootP_im); { // Do the multiplication // could avoid a bunch of complex multiplies by moving the loop over data_thisblock // inside. if (tmp_vec->Dim() < (MatrixIndexT)(P*2)) tmp_vec->Resize(P*2); Real *temp_a = tmp_vec->Data(); Real *data_thisblock = data, *data_end = data+(N*2*nffts); for (; data_thisblock != data_end; data_thisblock += N*2) { // for each separate fft. Real qd_re = 1.0, qd_im = 0.0; // 1^(q'/N) for (int qd = 0; qd < Q; qd++) { Real pdQ_qd_re = qd_re, pdQ_qd_im = qd_im; // 1^((p'Q+q') / N) == 1^((p'/P) + (q'/N)) // Initialize to q'/N, corresponding to p' == 0. for (int pd = 0; pd < P; pd++) { // pd == p' { // This is the p = 0 case of the loop below [an optimization]. temp_a[pd*2] = data_thisblock[qd*2]; temp_a[pd*2 + 1] = data_thisblock[qd*2 + 1]; } { // This is the p = 1 case of the loop below [an optimization] // **** MOST OF THE TIME (>60% I think) gets spent here. *** ComplexAddProduct(pdQ_qd_re, pdQ_qd_im, data_thisblock[(qd+Q)*2], data_thisblock[(qd+Q)*2 + 1], &(temp_a[pd*2]), &(temp_a[pd*2 + 1])); } if (P > 2) { Real p_pdQ_qd_re = pdQ_qd_re, p_pdQ_qd_im = pdQ_qd_im; // 1^(p(p'Q+q')/N) for (int p = 2; p < P; p++) { ComplexMul(pdQ_qd_re, pdQ_qd_im, &p_pdQ_qd_re, &p_pdQ_qd_im); // p_pdQ_qd *= pdQ_qd. int data_idx = p*Q + qd; ComplexAddProduct(p_pdQ_qd_re, p_pdQ_qd_im, data_thisblock[data_idx*2], data_thisblock[data_idx*2 + 1], &(temp_a[pd*2]), &(temp_a[pd*2 + 1])); } } if (pd != P-1) ComplexMul(rootP_re, rootP_im, &pdQ_qd_re, &pdQ_qd_im); // pdQ_qd *= (rootP == 1^{1/P}) // (using 1/P == Q/N) } for (int pd = 0; pd < P; pd++) { data_thisblock[(pd*Q + qd)*2] = temp_a[pd*2]; data_thisblock[(pd*Q + qd)*2 + 1] = temp_a[pd*2 + 1]; } ComplexMul(rootN_re, rootN_im, &qd_re, &qd_im); // qd *= rootN. } } } } /* Equations for ComplexFftRecursive. We consider here one of the "nffts" separate ffts; it's just a question of doing them all in parallel. We also write all equations in terms of complex math (the conversion to real arithmetic is not hard, and anyway takes place inside function calls). Let the input (i.e. "data" at start) be a_n, n = 0..N-1, and the output (Fourier transform) be d_k, k = 0..N-1. We use these letters because there will be two intermediate variables b and c. We want to compute: d_k = \sum_n a_n 1^(kn/N) (1) where we use 1^x as shorthand for exp(-2pi x) for the forward algorithm and exp(2pi x) for the backward one. We factorize N = P Q (P small, Q usually large). With p = 0..P-1 and q = 0..Q-1, and also p'=0..P-1 and q'=0..P-1, we let: k == p'Q + q' (2) n == qP + p (3) That is, we let p, q, p', q' range over these indices and observe that this way we can cover all n, k. Expanding (1) using (2) and (3), we can write: d_k = \sum_{p, q} a_n 1^((p'Q+q')(qP+p)/N) = \sum_{p, q} a_n 1^(p'pQ/N) 1^(q'qP/N) 1^(q'p/N) (4) using 1^(PQ/N) = 1 to get rid of the terms with PQ in them. Rearranging (4), d_k = \sum_p 1^(p'pQ/N) 1^(q'p/N) \sum_q 1^(q'qP/N) a_n (5) The point here is to separate the index q. Now we can expand out the remaining instances of k and n using (2) and (3): d_(p'Q+q') = \sum_p 1^(p'pQ/N) 1^(q'p/N) \sum_q 1^(q'qP/N) a_(qP+p) (6) The expression \sum_q varies with the indices p and q'. Let us define C_{p, q'} = \sum_q 1^(q'qP/N) a_(qP+p) (7) Here, C_{p, q'}, viewed as a sequence in q', is just the DFT of the points a_(qP+p) for q = 1..Q-1. These points are not consecutive in memory though, they jump by P each time. Let us define b as a rearranged version of a, so that b_(pQ+q) = a_(qP+p) (8) How to do this rearrangement in place? In We can rearrange (7) to be written in terms of the b's, using (8), so that C_{p, q'} = \sum_q 1^(q'q (P/N)) b_(pQ+q) (9) Here, the sequence of C_{p, q'} over q'=0..Q-1, is just the DFT of the sequence of b_(pQ) .. b_(p(Q+1)-1). Let's arrange the C_{p, q'} in a single array in memory in the same way as the b's, i.e. we define c_(pQ+q') == C_{p, q'}. (10) Note that we could have written (10) with q in place of q', as there is only one index of type q present, but q' is just a more natural variable name to use since we use q' elsewhere to subscript c and C. Rewriting (9), we have: c_(pQ+q') = \sum_q 1^(q'q (P/N)) b_(pQ+q) (11) which is the DFT computed by the recursive call to this function [after computing the b's by rearranging the a's]. From the c's we want to compute the d's. Taking (6), substituting in the sum (7), and using (10) to write it as an array, we have: d_(p'Q+q') = \sum_p 1^(p'pQ/N) 1^(q'p/N) c_(pQ+q') (12) This sum is independent for different values of q'. Note that d overwrites c in memory. We compute this in a direct way, using a little array of size P to store the computed d values for one value of q' (we reuse the array for each value of q'). So the overall picture is this: We get a call to compute DFT on size N. - If N == 1 we return (nothing to do). - We factor N = P Q (typically, P is small). - Using (8), we rearrange the data in memory so that we have b not a in memory (this is the block "do the rearrangement"). The pseudocode for this is as follows. For simplicity we use a temporary array. for p = 0..P-1 for q = 0..Q-1 bidx = pQ + q aidx = qP + p tmp[bidx] = data[aidx]. end end data <-- tmp else endif The reason this accomplishes (8) is that we want pQ+q and qP+p to be swapped over for each p, q, and the "if m > n" is a convenient way of ensuring that this swapping happens only once (otherwise it would happen twice, since pQ+q and qP+p both range over the entire set of numbers 0..N-1). - We do the DFT on the smaller block size to compute c from b (this eq eq. (11)). Note that this is actually multiple DFTs, one for each value of p, but this goes to the "nffts" argument of the function call, which we have ignored up to now. -We compute eq. (12) via a loop, as follows allocate temporary array e of size P. For q' = 0..Q-1: for p' = 0..P-1: set sum to zero [this will go in e[p']] for p = p..P-1: sum += 1^(p'pQ/N) 1^(q'p/N) c_(pQ+q') end e[p'] = sum end for p' = 0..P-1: d_(p'Q+q') = e[p'] end end delete temporary array e */ // This is the outer-layer calling code for ComplexFftRecursive. // It factorizes the dimension and then calls the FFT routine. template void ComplexFft(VectorBase *v, bool forward, Vector *tmp_in) { KALDI_ASSERT(v != NULL); if (v->Dim()<=1) return; KALDI_ASSERT(v->Dim() % 2 == 0); // complex input. int N = v->Dim() / 2; std::vector factors; Factorize(N, &factors); int *factor_beg = NULL; if (factors.size() > 0) factor_beg = &(factors[0]); Vector tmp; // allocated in ComplexFftRecursive. ComplexFftRecursive(v->Data(), 1, N, factor_beg, factor_beg+factors.size(), forward, (tmp_in?tmp_in:&tmp)); } //! Inefficient version of Fourier transform, for testing purposes. template void RealFftInefficient (VectorBase *v, bool forward) { KALDI_ASSERT(v != NULL); MatrixIndexT N = v->Dim(); KALDI_ASSERT(N%2 == 0); if (N == 0) return; Vector vtmp(N*2); // store as complex. if (forward) { for (MatrixIndexT i = 0; i < N; i++) vtmp(i*2) = (*v)(i); ComplexFft(&vtmp, forward); // this is already tested so we can use this. v->CopyFromVec( vtmp.Range(0, N) ); (*v)(1) = vtmp(N); // Copy the N/2'th fourier component, which is real, // to the imaginary part of the 1st complex output. } else { // reverse the transformation above to get the complex spectrum. vtmp(0) = (*v)(0); // copy F_0 which is real vtmp(N) = (*v)(1); // copy F_{N/2} which is real for (MatrixIndexT i = 1; i < N/2; i++) { // Copy i'th to i'th fourier component vtmp(2*i) = (*v)(2*i); vtmp(2*i+1) = (*v)(2*i+1); // Copy i'th to N-i'th, conjugated. vtmp(2*(N-i)) = (*v)(2*i); vtmp(2*(N-i)+1) = -(*v)(2*i+1); } ComplexFft(&vtmp, forward); // actually backward since forward == false // Copy back real part. Complex part should be zero. for (MatrixIndexT i = 0; i < N; i++) (*v)(i) = vtmp(i*2); } } template void RealFftInefficient (VectorBase *v, bool forward); template void RealFftInefficient (VectorBase *v, bool forward); template void ComplexFft(VectorBase *v, bool forward, Vector *tmp_in); template void ComplexFft(VectorBase *v, bool forward, Vector *tmp_in); // See the long comment below for the math behind this. template void RealFft (VectorBase *v, bool forward) { KALDI_ASSERT(v != NULL); MatrixIndexT N = v->Dim(), N2 = N/2; KALDI_ASSERT(N%2 == 0); if (N == 0) return; if (forward) ComplexFft(v, true); Real *data = v->Data(); Real rootN_re, rootN_im; // exp(-2pi/N), forward; exp(2pi/N), backward int forward_sign = forward ? -1 : 1; ComplexImExp(static_cast(M_2PI/N *forward_sign), &rootN_re, &rootN_im); Real kN_re = -forward_sign, kN_im = 0.0; // exp(-2pik/N), forward; exp(-2pik/N), backward // kN starts out as 1.0 for forward algorithm but -1.0 for backward. for (MatrixIndexT k = 1; 2*k <= N2; k++) { ComplexMul(rootN_re, rootN_im, &kN_re, &kN_im); Real Ck_re, Ck_im, Dk_re, Dk_im; // C_k = 1/2 (B_k + B_{N/2 - k}^*) : Ck_re = 0.5 * (data[2*k] + data[N - 2*k]); Ck_im = 0.5 * (data[2*k + 1] - data[N - 2*k + 1]); // re(D_k)= 1/2 (im(B_k) + im(B_{N/2-k})): Dk_re = 0.5 * (data[2*k + 1] + data[N - 2*k + 1]); // im(D_k) = -1/2 (re(B_k) - re(B_{N/2-k})) Dk_im =-0.5 * (data[2*k] - data[N - 2*k]); // A_k = C_k + 1^(k/N) D_k: data[2*k] = Ck_re; // A_k <-- C_k data[2*k+1] = Ck_im; // now A_k += D_k 1^(k/N) ComplexAddProduct(Dk_re, Dk_im, kN_re, kN_im, &(data[2*k]), &(data[2*k+1])); MatrixIndexT kdash = N2 - k; if (kdash != k) { // Next we handle the index k' = N/2 - k. This is necessary // to do now, to avoid invalidating data that we will later need. // The quantities C_{k'} and D_{k'} are just the conjugates of C_k // and D_k, so the equations are simple modifications of the above, // replacing Ck_im and Dk_im with their negatives. data[2*kdash] = Ck_re; // A_k' <-- C_k' data[2*kdash+1] = -Ck_im; // now A_k' += D_k' 1^(k'/N) // We use 1^(k'/N) = 1^((N/2 - k) / N) = 1^(1/2) 1^(-k/N) = -1 * (1^(k/N))^* // so it's the same as 1^(k/N) but with the real part negated. ComplexAddProduct(Dk_re, -Dk_im, -kN_re, kN_im, &(data[2*kdash]), &(data[2*kdash+1])); } } { // Now handle k = 0. // In simple terms: after the complex fft, data[0] becomes the sum of real // parts input[0], input[2]... and data[1] becomes the sum of imaginary // pats input[1], input[3]... // "zeroth" [A_0] is just the sum of input[0]+input[1]+input[2].. // and "n2th" [A_{N/2}] is input[0]-input[1]+input[2]... . Real zeroth = data[0] + data[1], n2th = data[0] - data[1]; data[0] = zeroth; data[1] = n2th; if (!forward) { data[0] /= 2; data[1] /= 2; } } if (!forward) { ComplexFft(v, false); v->Scale(2.0); // This is so we get a factor of N increase, rather than N/2 which we would // otherwise get from [ComplexFft, forward] + [ComplexFft, backward] in dimension N/2. // It's for consistency with our normal FFT convensions. } } template void RealFft (VectorBase *v, bool forward); template void RealFft (VectorBase *v, bool forward); /* Notes for real FFTs. We are using the same convention as above, 1^x to mean exp(-2\pi x) for the forward transform. Actually, in a slight abuse of notation, we use this meaning for 1^x in both the forward and backward cases because it's more convenient in this section. Suppose we have real data a[0...N-1], with N even, and want to compute its Fourier transform. We can make do with the first N/2 points of the transform, since the remaining ones are complex conjugates of the first. We want to compute: for k = 0...N/2-1, A_k = \sum_{n = 0}^{N-1} a_n 1^(kn/N) (1) We treat a[0..N-1] as a complex sequence of length N/2, i.e. a sequence b[0..N/2 - 1]. Viewed as sequences of length N/2, we have: b = c + i d, where c = a_0, a_2 ... and d = a_1, a_3 ... We can recover the length-N/2 Fourier transforms of c and d by doing FT on b and then doing the equations below. Derivation is marked by (*) in a comment below (search for it). Let B, C, D be the FTs. We have C_k = 1/2 (B_k + B_{N/2 - k}^*) (z0) D_k =-1/2i (B_k - B_{N/2 - k}^*) (z1) so: re(D_k)= 1/2 (im(B_k) + im(B_{N/2-k})) (z2) im(D_k) = -1/2 (re(B_k) - re(B_{N/2-k})) (z3) To recover the FT A from C and D, we write, rearranging (1): A_k = \sum_{n = 0, 2, ..., N-2} a_n 1^(kn/N) +\sum_{n = 1, 3, ..., N-1} a_n 1^(kn/N) = \sum_{n = 0, 1, ..., N/2-1} a_n 1^(2kn/N) + a_{n+1} 1^(2kn/N) 1^(k/N) = \sum_{n = 0, 1, ..., N/2-1} c_n 1^(2kn/N) + d_n 1^(2kn/N) 1^(k/N) A_k = C_k + 1^(k/N) D_k (a0) This equation is valid for k = 0...N/2-1, which is the range of the sequences B_k and C_k. We don't use is for k = 0, which is a special case considered below. For 1 < k < N/2, it's convenient to consider the pair k, k', where k' = N/2 - k. Remember that C_k' = C_k^ *and D_k' = D_k^* [where * is conjugation]. Also, 1^(N/2 / N) = -1. So we have: A_k' = C_k^* - 1^(k/N) D_k^* (a0b) We do (a0) and (a0b) together. By symmetry this gives us the Fourier components for N/2+1, ... N, if we want them. However, it doesn't give us the value for exactly k = N/2. For k = 0 and k = N/2, it is easiest to argue directly about the meaning of the A_k, B_k and C_k in terms of sums of points. A_0 and A_{N/2} are both real, with A_0=\sum_n a_n, and A_1 an alternating sum A_1 = a_0 - a_1 + a_2 ... It's easy to show that A_0 = B_0 + C_0 (a1) A_{N/2} = B_0 - C_0. (a2) Since B_0 and C_0 are both real, B_0 is the real coefficient of D_0 and C_0 is the imaginary coefficient. *REVERSING THE PROCESS* Next we want to reverse this process. We just need to work out C_k and D_k from the sequence A_k. Then we do the inverse complex fft and we get back where we started. For 0 and N/2, working from (a1) and (a2) above, we can see that: B_0 = 1/2 (A_0 + A_{N/2}) (y0) C_0 = 1/2 (A_0 + A_{N/2}) (y1) and we use D_0 = B_0 + i C_0 to get the 1st complex coefficient of D. This is exactly the same as the forward process except with an extra factor of 1/2. Consider equations (a0) and (a0b). We want to work out C_k and D_k from A_k and A_k'. Remember k' = N/2 - k. Write down A_k = C_k + 1^(k/N) D_k (copying a0) A_k'^* = C_k - 1^(k/N) D_k (conjugate of a0b) So C_k = 0.5 (A_k + A_k'^*) (p0) D_k = 1^(-k/N) . 0.5 (A_k - A_k'^*) (p1) Next, we want to compute B_k and B_k' from C_k and D_k. C.f. (z0)..(z3), and remember that k' = N/2-k. We can see that B_k = C_k + i D_k (p2) B_k' = C_k - i D_k (p3) We would like to make the equations (p0) ... (p3) look like the forward equations (z0), (z1), (a0) and (a0b) so we can reuse the code. Define E_k = -i 1^(k/N) D_k. Then write down (p0)..(p3). We have C_k = 0.5 (A_k + A_k'^*) (p0') E_k = -0.5 i (A_k - A_k'^*) (p1') B_k = C_k - 1^(-k/N) E_k (p2') B_k' = C_k + 1^(-k/N) E_k (p3') So these are exactly the same as (z0), (z1), (a0), (a0b) except replacing 1^(k/N) with -1^(-k/N) . Remember that we defined 1^x above to be exp(-2pi x/N), so the signs here might be opposite to what you see in the code. MODIFICATION: we need to take care of a factor of two. The complex FFT we implemented does not divide by N in the reverse case. So upon inversion we get larger by N/2. However, this is not consistent with normal FFT conventions where you get a factor of N. For this reason we multiply by two after the process described above. */ /* (*) [this token is referred to in a comment above]. Notes for separating 2 real transforms from one complex one. Note that the letters here (A, B, C and N) are all distinct from the same letters used in the place where this comment is used. Suppose we have two sequences a_n and b_n, n = 0..N-1. We combine them into a complex number, c_n = a_n + i b_n. Then we take the fourier transform to get C_k = \sum_{n = 0}^{N-1} c_n 1^(n/N) . Then we use symmetry. Define A_k and B_k as the DFTs of a and b. We use A_k = A_{N-k}^*, and B_k = B_{N-k}^*, since a and b are real. Using C_k = A_k + i B_k, C_{N-k} = A_k^* + i B_k^* = A_k^* - (i B_k)^* So: A_k = 1/2 (C_k + C_{N-k}^*) i B_k = 1/2 (C_k - C_{N-k}^*) -> B_k =-1/2i (C_k - C_{N-k}^*) -> re(B_k) = 1/2 (im(C_k) + im(C_{N-k})) im(B_k) =-1/2 (re(C_k) - re(C_{N-k})) */ template void ComputeDctMatrix(Matrix *M) { //KALDI_ASSERT(M->NumRows() == M->NumCols()); MatrixIndexT K = M->NumRows(); MatrixIndexT N = M->NumCols(); KALDI_ASSERT(K > 0); KALDI_ASSERT(N > 0); Real normalizer = std::sqrt(1.0 / static_cast(N)); // normalizer for // X_0. for (MatrixIndexT j = 0; j < N; j++) (*M)(0, j) = normalizer; normalizer = std::sqrt(2.0 / static_cast(N)); // normalizer for other // elements. for (MatrixIndexT k = 1; k < K; k++) for (MatrixIndexT n = 0; n < N; n++) (*M)(k, n) = normalizer * std::cos( static_cast(M_PI)/N * (n + 0.5) * k ); } template void ComputeDctMatrix(Matrix *M); template void ComputeDctMatrix(Matrix *M); template void ComputePca(const MatrixBase &X, MatrixBase *U, MatrixBase *A, bool print_eigs, bool exact) { // Note that some of these matrices may be transposed w.r.t. the // way it's most natural to describe them in math... it's the rows // of X and U that correspond to the (data-points, basis elements). MatrixIndexT N = X.NumRows(), D = X.NumCols(); // N = #points, D = feature dim. KALDI_ASSERT(U != NULL && U->NumCols() == D); MatrixIndexT G = U->NumRows(); // # of retained basis elements. KALDI_ASSERT(A == NULL || (A->NumRows() == N && A->NumCols() == G)); KALDI_ASSERT(G <= N && G <= D); if (D < N) { // Do conventional PCA. SpMatrix Msp(D); // Matrix of outer products. Msp.AddMat2(1.0, X, kTrans, 0.0); // M <-- X^T X Matrix Utmp; Vector l; if (exact) { Utmp.Resize(D, D); l.Resize(D); //Matrix M(Msp); //M.DestructiveSvd(&l, &Utmp, NULL); Msp.Eig(&l, &Utmp); } else { Utmp.Resize(D, G); l.Resize(G); Msp.TopEigs(&l, &Utmp); } SortSvd(&l, &Utmp); for (MatrixIndexT g = 0; g < G; g++) U->Row(g).CopyColFromMat(Utmp, g); if (print_eigs) KALDI_LOG << (exact ? "" : "Retained ") << "PCA eigenvalues are " << l; if (A != NULL) A->AddMatMat(1.0, X, kNoTrans, *U, kTrans, 0.0); } else { // Do inner-product PCA. SpMatrix Nsp(N); // Matrix of inner products. Nsp.AddMat2(1.0, X, kNoTrans, 0.0); // M <-- X X^T Matrix Vtmp; Vector l; if (exact) { Vtmp.Resize(N, N); l.Resize(N); Matrix Nmat(Nsp); Nmat.DestructiveSvd(&l, &Vtmp, NULL); } else { Vtmp.Resize(N, G); l.Resize(G); Nsp.TopEigs(&l, &Vtmp); } MatrixIndexT num_zeroed = 0; for (MatrixIndexT g = 0; g < G; g++) { if (l(g) < 0.0) { KALDI_WARN << "In PCA, setting element " << l(g) << " to zero."; l(g) = 0.0; num_zeroed++; } } SortSvd(&l, &Vtmp); // Make sure zero elements are last, this // is necessary for Orthogonalize() to work properly later. Vtmp.Transpose(); // So eigenvalues are the rows. for (MatrixIndexT g = 0; g < G; g++) { Real sqrtlg = sqrt(l(g)); if (l(g) != 0.0) { U->Row(g).AddMatVec(1.0 / sqrtlg, X, kTrans, Vtmp.Row(g), 0.0); } else { U->Row(g).SetZero(); (*U)(g, g) = 1.0; // arbitrary direction. Will later orthogonalize. } if (A != NULL) for (MatrixIndexT n = 0; n < N; n++) (*A)(n, g) = sqrtlg * Vtmp(g, n); } // Now orthogonalize. This is mainly useful in // case there were zero eigenvalues, but we do it // for all of them. U->OrthogonalizeRows(); if (print_eigs) KALDI_LOG << "(inner-product) PCA eigenvalues are " << l; } } template void ComputePca(const MatrixBase &X, MatrixBase *U, MatrixBase *A, bool print_eigs, bool exact); template void ComputePca(const MatrixBase &X, MatrixBase *U, MatrixBase *A, bool print_eigs, bool exact); // Added by Dan, Feb. 13 2012. // This function does: *plus += max(0, a b^T), // *minus += max(0, -(a b^T)). template void AddOuterProductPlusMinus(Real alpha, const VectorBase &a, const VectorBase &b, MatrixBase *plus, MatrixBase *minus) { KALDI_ASSERT(a.Dim() == plus->NumRows() && b.Dim() == plus->NumCols() && a.Dim() == minus->NumRows() && b.Dim() == minus->NumCols()); int32 nrows = a.Dim(), ncols = b.Dim(), pskip = plus->Stride() - ncols, mskip = minus->Stride() - ncols; const Real *adata = a.Data(), *bdata = b.Data(); Real *plusdata = plus->Data(), *minusdata = minus->Data(); for (int32 i = 0; i < nrows; i++) { const Real *btmp = bdata; Real multiple = alpha * *adata; if (multiple > 0.0) { for (int32 j = 0; j < ncols; j++, plusdata++, minusdata++, btmp++) { if (*btmp > 0.0) *plusdata += multiple * *btmp; else *minusdata -= multiple * *btmp; } } else { for (int32 j = 0; j < ncols; j++, plusdata++, minusdata++, btmp++) { if (*btmp < 0.0) *plusdata += multiple * *btmp; else *minusdata -= multiple * *btmp; } } plusdata += pskip; minusdata += mskip; adata++; } } // Instantiate template template void AddOuterProductPlusMinus(float alpha, const VectorBase &a, const VectorBase &b, MatrixBase *plus, MatrixBase *minus); template void AddOuterProductPlusMinus(double alpha, const VectorBase &a, const VectorBase &b, MatrixBase *plus, MatrixBase *minus); } // end namespace kaldi