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

Add comment to shared range method in quantization. That method is needed because of restrictions in TFLite's LSTM kernel. It was not very clear in the original comments.

Browse Source 
PiperOrigin-RevId: 283399107
Change-Id: I51dbcdb49daeaf950d550405d4c31f918cf52ebe
This commit is contained in:
Jian Li 2019-12-02 13:00:12 -08:00 committed by TensorFlower Gardener
parent 654dc1e68c
commit 143e80f663
1 changed files with 5 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

5
tensorflow/lite/tools/optimize/quantize_model.cc
View File

@ -711,6 +711,11 @@ TfLiteStatus QuantizeIntemediateTensors(ModelT* model,
// Quantize tensros that have shared range. For example, in LSTM, the output // Quantize tensros that have shared range. For example, in LSTM, the output
// tensor and input state tensor should share the same range because they are // tensor and input state tensor should share the same range because they are
// using the same scale and zero point. // using the same scale and zero point.
// We have to model this explicitely because the output is modeled as an extra
// tensor in LSTM. In calibrator, state tensors are logged both before and after
// the inferece so the range is fully captured. But output, although it is
// identical to activation, is not a state tensor the input value (range) of the
// very first inference is not captured.
TfLiteStatus QuantizeSharedRange(ModelT* model, ErrorReporter* error_reporter) { TfLiteStatus QuantizeSharedRange(ModelT* model, ErrorReporter* error_reporter) {
for (size_t subgraph_idx = 0; subgraph_idx < model->subgraphs.size(); for (size_t subgraph_idx = 0; subgraph_idx < model->subgraphs.size();
subgraph_idx++) { subgraph_idx++) {