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

Switch legacy quantize mode on by default for MLIR quantizer

Browse Source 
Also fix code to make legacy pass applied together with other patterns. It ensures when numeric_verify is set, float ops can be duplicated before quantization to happen.

PiperOrigin-RevId: 352948582
Change-Id: I6550697ee6a4508bc6518ec08cccc854b34d2321
This commit is contained in:
Taehee Jeong 2021-01-20 22:44:09 -08:00 committed by TensorFlower Gardener
parent 9c530d1204
commit 310b42a801
2 changed files with 10 additions and 5 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

7
tensorflow/compiler/mlir/lite/quantization/lite/quantize_model.h
View File

@ -29,6 +29,11 @@ namespace lite {
// The `input_type`, `output_type` and `inference_type` can be
// float32/qint8/int8/int16.
// Return partially quantized model if `fully_quantize` is false.
// When `verify_numeric` is true, the model will have it's original float ops
// and NumericVerify ops to compare output values from the quantized and float
// ops. When `legacy_float_scale` is true, the quantizer will use float scale
// instead of double, and call TOCO's quantization routines to maintain
// bit-exactness of the values with the TOCO quantizer.
TfLiteStatus QuantizeModel(
const tflite::ModelT& input_model, const tflite::TensorType& input_type,
const tflite::TensorType& output_type,
@ -37,7 +42,7 @@ TfLiteStatus QuantizeModel(
bool disable_per_channel, bool fully_quantize,
flatbuffers::FlatBufferBuilder* builder,
tflite::ErrorReporter* error_reporter, bool verify_numeric = false,
bool legacy_float_scale = false);
bool legacy_float_scale = true);
} // namespace lite
} // namespace mlir

8
tensorflow/compiler/mlir/lite/transforms/quantize.cc
View File

@ -88,8 +88,10 @@ struct TFLFullQuantization
};
struct LegacyQuantizePass : public OpRewritePattern<QuantizeOp> {
// This pattern should be applied before existing quantize pattern in
// `quantize_patterns.td`, so the benefit is set to some value larger than 1.
explicit LegacyQuantizePass(MLIRContext* context)
: OpRewritePattern<QuantizeOp>(context) {}
: OpRewritePattern<QuantizeOp>(context, /*benefit=*/10) {}
LogicalResult matchAndRewrite(QuantizeOp op,
PatternRewriter& rewriter) const override {
DenseFPElementsAttr attr;
@ -127,9 +129,7 @@ void QuantizePass::runOnFunction() {
auto func = getFunction();
auto* ctx = func.getContext();
if (legacy_float_scale) {
OwningRewritePatternList legacy_patterns;
legacy_patterns.insert<LegacyQuantizePass>(ctx);
applyPatternsAndFoldGreedily(func, std::move(legacy_patterns));
patterns.insert<LegacyQuantizePass>(ctx);
}
TFL::populateWithGenerated(ctx, patterns);
patterns.insert<TFLFullQuantization>(