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

Use the calculated activation range in depthwise_conv

Browse Source 
PiperOrigin-RevId: 309897016
Change-Id: I861ad3e3fd25a3c06ae57cfe51a4c0b9ad537639
This commit is contained in:
Feng Liu 2020-05-05 00:42:42 -07:00 committed by TensorFlower Gardener
parent d0a4066ffa
commit 7c39403bc7
1 changed files with 2 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

5
tensorflow/lite/kernels/depthwise_conv.cc
View File

@ -379,9 +379,8 @@ TfLiteStatus EvalQuantizedPerChannel(TfLiteContext* context, TfLiteNode* node,
op_params.input_offset = -input->params.zero_point;
op_params.weights_offset = 0;
op_params.output_offset = output->params.zero_point;
// TODO(b/130439627): Use calculated value for clamping.
op_params.quantized_activation_min = std::numeric_limits<int8_t>::min();
op_params.quantized_activation_max = std::numeric_limits<int8_t>::max();
op_params.quantized_activation_min = data->output_activation_min;
op_params.quantized_activation_max = data->output_activation_max;
TF_LITE_ENSURE_STATUS(ComputeDepthMultiplier(context, input, filter,
&op_params.depth_multiplier));