experiments/STT
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

Fix parameter inference from model shape

Browse Source
This commit is contained in:
Reuben Morais 2018-09-14 12:03:02 -03:00
parent 778f5deb9d
commit 799baf1f99
1 changed files with 10 additions and 10 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

20
native_client/deepspeech.cc
View File

@ -323,14 +323,15 @@ ModelState::infer(const float* aMfcc, unsigned int n_frames, vector<float>& logi
return;
#endif // DS_NATIVE_MODEL
} else {
Tensor input(DT_FLOAT, TensorShape({BATCH_SIZE, n_steps, mfcc_feats_per_timestep}));
Tensor input(DT_FLOAT, TensorShape({BATCH_SIZE, n_steps, 2*n_context+1, MFCC_FEATURES}));
auto input_mapped = input.tensor<float, 3>();
int idx = 0;
for (int i = 0; i < n_frames; i++) {
for (int j = 0; j < mfcc_feats_per_timestep; j++, idx++) {
input_mapped(0, i, j) = aMfcc[idx];
}
auto input_mapped = input.flat<float>();
int i;
for (i = 0; i < n_frames*mfcc_feats_per_timestep; ++i) {
input_mapped(i) = aMfcc[i];
}
for (; i < n_steps*mfcc_feats_per_timestep; ++i) {
input_mapped(i) = 0;
}
Tensor input_lengths(DT_INT32, TensorShape({1}));
@ -482,9 +483,8 @@ DS_CreateModel(const char* aModelPath,
if (node.name() == "input_node") {
const auto& shape = node.attr().at("shape").shape();
model->n_steps = shape.dim(1).size();
model->mfcc_feats_per_timestep = shape.dim(2).size();
// mfcc_features_per_timestep = MFCC_FEATURES * ((2*n_context) + 1)
model->n_context = (model->mfcc_feats_per_timestep - MFCC_FEATURES) / (2 * MFCC_FEATURES);
model->n_context = (shape.dim(2).size()-1)/2;
model->mfcc_feats_per_timestep = shape.dim(2).size() * shape.dim(3).size();
} else if (node.name() == "logits_shape") {
Tensor logits_shape = Tensor(DT_INT32, TensorShape({3}));
if (!logits_shape.FromProto(node.attr().at("value").tensor())) {