holm/whisper.cpp
1
1
Fork 0
forked from extern/whisper.cpp
Code Issues Pull Requests Packages Projects Releases Wiki Activity

diarization : try to cluster embedings from last encoder layer

Browse Source
This commit is contained in:
Georgi Gerganov 2023-02-19 10:33:03 +02:00
parent d5d7769fa7
commit d11f35920e
No known key found for this signature in database
GPG Key ID: 449E073F9DC10735
2 changed files with 108 additions and 37 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

23
ggml.c
View File

@ -8652,16 +8652,16 @@ void ggml_svd_reduce_dims(
} }
// normalize U // normalize U
for (int i = 0; i < n; ++i) { //for (int i = 0; i < n; ++i) {
double sum = 0.0; // double sum = 0.0;
for (int j = 0; j < m; ++j) { // for (int j = 0; j < m; ++j) {
sum += U[i * m + j] * U[i * m + j]; // sum += U[i * m + j] * U[i * m + j];
} // }
sum = sqrt(sum); // sum = sqrt(sum);
for (int j = 0; j < m; ++j) { // for (int j = 0; j < m; ++j) {
U[i * m + j] /= sum*sqrt((double) m); // U[i * m + j] /= sum*sqrt((double) m);
} // }
} //}
// print U // print U
//printf("U:\n"); //printf("U:\n");
@ -8675,9 +8675,10 @@ void ggml_svd_reduce_dims(
//printf("\n"); //printf("\n");
printf("n = %d, m = %d, nd = %d\n", n, m, nd);
// project A0 onto U // project A0 onto U
for (int i = 0; i < n; ++i) { for (int i = 0; i < n; ++i) {
for (int j = 0; j < n; ++j) { for (int j = 0; j < nd; ++j) {
A[i * nd + j] = 0.0f; A[i * nd + j] = 0.0f;
for (int k = 0; k < m; ++k) { for (int k = 0; k < m; ++k) {
A[i * nd + j] += A0[i * m + k] * U[j * m + k]; A[i * nd + j] += A0[i * m + k] * U[j * m + k];

110
whisper.cpp
View File

@ -603,6 +603,8 @@ struct whisper_context {
// [EXPERIMENTAL] speed-up techniques // [EXPERIMENTAL] speed-up techniques
int32_t exp_n_audio_ctx; // 0 - use default int32_t exp_n_audio_ctx; // 0 - use default
std::vector<float> audio_embd;
void use_buf(struct ggml_context * ctx, int i) { void use_buf(struct ggml_context * ctx, int i) {
#if defined(WHISPER_USE_SCRATCH) #if defined(WHISPER_USE_SCRATCH)
size_t last_size = 0; size_t last_size = 0;
@ -1723,17 +1725,35 @@ static bool whisper_encode(
} }
// cur // cur
//{
// printf("ne0 = %d\n", cur->ne[0]);
// printf("ne1 = %d\n", cur->ne[1]);
// for (int i = 0; i < 10; ++i) {
// printf("%8.4f ", ((float *)(cur->data))[i]);
// }
// printf("... ");
// for (int i = cur->ne[0] - 10; i < cur->ne[0]; ++i) {
// printf("%8.4f ", ((float *)(cur->data))[i]);
// }
// printf("\n");
//}
{ {
//printf("ne0 = %d\n", cur->ne[0]); //const int i0 = std::min(mel_offset, mel_inp.n_len);
//printf("ne1 = %d\n", cur->ne[1]); //const int i1 = std::min(mel_offset + 2*n_ctx, mel_inp.n_len);
//for (int i = 0; i < 10; ++i) { const int i0 = 0;
// printf("%8.4f ", ((float *)(cur->data))[i]); const int i1 = cur->ne[1];
//}
//printf("... "); //printf("i0 = %d, i1 = %d, (i1 - i0) = %d, embd size = %d\n", i0, i1, i1 - i0, cur->ne[0]);
//for (int i = cur->ne[0] - 10; i < cur->ne[0]; ++i) {
// printf("%8.4f ", ((float *)(cur->data))[i]); wctx.audio_embd.clear();
//} wctx.audio_embd.resize(cur->ne[0], 0.0f);
//printf("\n"); for (int j = 0; j < cur->ne[0]; ++j) {
for (int i = i0; i < i1; ++i) {
wctx.audio_embd[j] += ((float *)(cur->data))[(i - i0)*cur->ne[0] + j];
}
wctx.audio_embd[j] /= (i1 - i0);
}
} }
// pre-compute cross-attention memory // pre-compute cross-attention memory
@ -4838,6 +4858,28 @@ void whisper_full_cluster_segments(struct whisper_context * ctx) {
const int n_state = ctx->model.hparams.n_audio_state; const int n_state = ctx->model.hparams.n_audio_state;
const int n_layer = ctx->model.hparams.n_audio_layer; const int n_layer = ctx->model.hparams.n_audio_layer;
#if 1
// use the last layer of the encoder
{
std::vector<float> embd(n_segments*n_state);
for (int i = 0; i < n_segments; ++i) {
const auto & segment_i = ctx->result_all[i];
printf("%s: segment %3d: t0 = %7d, t1 = %7d, text = %s\n", __func__, i, (int) segment_i.t0, (int) segment_i.t1, segment_i.text.c_str());
ctx->mel.n_len = segment_i.t1;
whisper_encode(*ctx, segment_i.t0, 7, true);
for (int j = 0; j < n_state; ++j) {
embd[i*n_state + j] = ctx->audio_embd[j];
}
}
const int n_features = std::min(4, n_segments);
ggml_svd_reduce_dims(n_state, n_segments, embd.data(), n_features);
#else
// use cross kv cache of various layers
for (int il = 0; il < n_layer; ++il) { for (int il = 0; il < n_layer; ++il) {
std::vector<float> embd(n_segments*n_ctx*n_state); std::vector<float> embd(n_segments*n_ctx*n_state);
@ -4856,9 +4898,10 @@ void whisper_full_cluster_segments(struct whisper_context * ctx) {
} }
} }
const int n_features = 64; const int n_features = std::min(4, n_segments);
ggml_svd_reduce_dims(n_ctx*n_state, n_segments, embd.data(), n_features); ggml_svd_reduce_dims(n_ctx*n_state, n_segments, embd.data(), n_features);
#endif
std::vector<std::vector<float>> features(n_segments); std::vector<std::vector<float>> features(n_segments);
@ -4927,32 +4970,59 @@ void whisper_full_cluster_segments(struct whisper_context * ctx) {
for (int l = 0; l < n_clusters; ++l) { for (int l = 0; l < n_clusters; ++l) {
//sum += std::pow(whisper_distance(features[j], centroids[k])/whisper_distance(features[j], centroids[l]), 2.0/(2.0 - 1.0)); //sum += std::pow(whisper_distance(features[j], centroids[k])/whisper_distance(features[j], centroids[l]), 2.0/(2.0 - 1.0));
// use the euclidean distance
double d0 = 0.0; double d0 = 0.0;
double d1 = 0.0;
// use the euclidean distance
{
for (int m = 0; m < n_features; ++m) { for (int m = 0; m < n_features; ++m) {
d0 += std::pow(features[j][m] - centroids[k][m], 2.0); d0 += std::pow(features[j][m] - centroids[k][m], 2.0);
} }
d0 = std::sqrt(d0); d0 = std::sqrt(d0);
double d1 = 0.0;
for (int m = 0; m < n_features; ++m) { for (int m = 0; m < n_features; ++m) {
d1 += std::pow(features[j][m] - centroids[l][m], 2.0); d1 += std::pow(features[j][m] - centroids[l][m], 2.0);
} }
d1 = std::sqrt(d1); d1 = std::sqrt(d1);
if (d1 == 0.0) {
sum += 1.0;
} else {
sum += std::pow(d0/d1, 2.0/(1.10 - 1.0));
}
} }
membership[j][k] = 1.0/sum; // use the cosine distance
//{
// double dot = 0.0;
// double norm0 = 0.0;
// double norm1 = 0.0;
// for (int m = 0; m < n_features; ++m) {
// dot += features[j][m]*centroids[k][m];
// norm0 += std::pow(features[j][m], 2.0);
// norm1 += std::pow(centroids[k][m], 2.0);
// }
// d0 = 1.0 - dot/(std::sqrt(norm0)*std::sqrt(norm1));
// dot = 0.0;
// norm0 = 0.0;
// norm1 = 0.0;
// for (int m = 0; m < n_features; ++m) {
// dot += features[j][m]*centroids[l][m];
// norm0 += std::pow(features[j][m], 2.0);
// norm1 += std::pow(centroids[l][m], 2.0);
// }
// d1 = 1.0 - dot/(std::sqrt(norm0)*std::sqrt(norm1));
//}
sum += std::pow(d0/d1, 2.0/(1.15 - 1.0));
}
membership[j][k] = sum == 0.0 ? 0.0 : 1.0/sum;
} }
} }
// print the membership // print the membership
if (i == niter - 1) { if (i == niter - 1) {
//{
for (int i = 0; i < n_segments; ++i) { for (int i = 0; i < n_segments; ++i) {
printf("%s: membership %3d: ", __func__, i); printf("%s: membership %3d: ", __func__, i);
for (int j = 0; j < n_clusters; ++j) { for (int j = 0; j < n_clusters; ++j) {