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

Quantization propagation shouldn't change the graph if the value has been quantized

Browse Source 
Since the same pass will be used for post-training quantization and partial
quantization, we should allow the quantization pass to handle this cases.

PiperOrigin-RevId: 275334735
Change-Id: I66a01cc07454118f3cf9079bb37e07cc8e6f3704
This commit is contained in:
Feng Liu 2019-10-17 14:06:33 -07:00 committed by TensorFlower Gardener
parent b700b798ba
commit b78675a3bd
2 changed files with 13 additions and 1 deletions
tensorflow/compiler/mlir/lite
quantization
quantization_driver.cc
tests
prepare-quantize.mlir

5
tensorflow/compiler/mlir/lite/quantization/quantization_driver.cc
View File

@ -688,7 +688,10 @@ bool QuantizationDriver::PropagateParams() {
auto key = std::make_pair(8, is_signed_);
auto &restricted_outputs = spec->restricted_output_params[key];
for (int i = 0, e = restricted_outputs.size(); i != e; ++i) {
changed |= SetResultParams(op, i, restricted_outputs[i]);
// The restrict can be nullptr if the result has been quantized.
if (auto params = restricted_outputs[i]) {
changed |= SetResultParams(op, i, params);
}
}
for (auto &it : spec->biases_params) {

9
tensorflow/compiler/mlir/lite/tests/prepare-quantize.mlir
View File

@ -292,6 +292,15 @@ func @QuantizeLogistic(tensor<1x6x6x16x!quant.uniform<u8:f32, 7.812500e-03:128>>
// CHECK: return %3 : tensor<1x6x6x16xf32>
}
// CHECK-LABEL: NotRescaleLogistic
func @NotRescaleLogistic(%arg0: tensor<1x6x6x16x!quant.uniform<u8:f32, 7.812500e-03:128>>) -> tensor<1x6x6x16x!quant.uniform<u8:f32, 3.906250e-03>> {
%0 = "tfl.logistic"(%arg0) : (tensor<1x6x6x16x!quant.uniform<u8:f32, 7.812500e-03:128>>) -> tensor<1x6x6x16x!quant.uniform<u8:f32, 3.906250e-03>>
return %0 : tensor<1x6x6x16x!quant.uniform<u8:f32, 3.906250e-03>>
// CHECK: %[[log:.*]] = "tfl.logistic"(%arg0)
// CHECK: return %[[log]]
}
// CHECK-LABEL: NotQuantizeConcatConstantOperand
func @NotQuantizeConcatConstantOperand(%arg0: tensor<1x2xf32>) -> tensor<2x2xf32> {
%0 = constant dense<1.0> : tensor<1x2xf32>