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

Update documentation for FC op in hexagon delegate.

Browse Source 
FC now supports relu activation function

PiperOrigin-RevId: 316966229
Change-Id: If5a42ecca8aa9b6e94474e75614153f50ca8ae3b
This commit is contained in:
Karim Nosir 2020-06-17 14:32:39 -07:00 committed by TensorFlower Gardener
parent b920ae9262
commit 69e36409c0
2 changed files with 1 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
tensorflow/lite/delegates/hexagon/README.md
View File

@ -74,7 +74,7 @@ are verified in `IsNodeSupportedByHexagon`:
- depth_multiplier == 1 - depth_multiplier == 1
- dilation only supported when stride == 1 - dilation only supported when stride == 1
- Otherwise, stride height/width <= 3 - Otherwise, stride height/width <= 3
* FullyConnected (without any activation) * FullyConnected
* Hardswish * Hardswish
* L2Normalization (without any activation) * L2Normalization (without any activation)
* Logistic (aka Sigmoid) * Logistic (aka Sigmoid)

1
tensorflow/lite/delegates/hexagon/builders/matmul_builder.cc
View File

@ -151,7 +151,6 @@ TfLiteStatus AddFullyConnectedHelper(const TfLiteIntArray* inputs,
// Data (8-bit), Weights (const, 8-bit) => MatMul => MatMul out (int32) // Data (8-bit), Weights (const, 8-bit) => MatMul => MatMul out (int32)
// MatMul out (int32), Bias (int32) => QuantizedBiasAdd => BiasAdd out (int32) // MatMul out (int32), Bias (int32) => QuantizedBiasAdd => BiasAdd out (int32)
// BiasAdd out (int32) => Requantize_32to8 => Output (8-bit) // BiasAdd out (int32) => Requantize_32to8 => Output (8-bit)
// TODO(b/129276536): Add activation support.
TfLiteStatus MatMulWithConstWeightsOpBuilder::PopulateSubGraph( TfLiteStatus MatMulWithConstWeightsOpBuilder::PopulateSubGraph(
const TfLiteIntArray* inputs, const TfLiteIntArray* outputs, const TfLiteIntArray* inputs, const TfLiteIntArray* outputs,
TfLiteContext* context) { TfLiteContext* context) {