[GNA] Added ConcatAlignFilter support for Accuracy Aware mode (openvi…

…notoolkit#7834) * Added ConcatAlignFilter support for Accuracy Aware mode * fails after reviewing and added tests * moved the test to gna/scale_factors_tests; added comments to test
tadamowicz · Aug 24, 2023 · 8c106fd · 8c106fd
1 parent bc2f7c5
commit 8c106fd
Show file tree

Hide file tree

Showing 3 changed files with 21 additions and 4 deletions.
diff --git a/inference-engine/src/gna_plugin/frontend/layer_quantizer.hpp b/inference-engine/src/gna_plugin/frontend/layer_quantizer.hpp
@@ -653,11 +653,18 @@ class DataQuantizer<Desc, InferenceEngine::WeightableLayer *> : public DataQuant
 
     void operator()(InferenceEngine::WeightableLayer *wl, const FakeQuant&) const {
         auto quantData = InferenceEngine::getInjectedData<QuantizedLayerParams>(*wl);
-        IE_ASSERT(quantData->_weights_quant.IsStatsSet());
-        if (quantData->_weights_quant.GetLevels() <= std::numeric_limits<uint8_t>::max()) {
-            quantizeWeightsBiases<FakeQuantI8>(FakeQuantI8(), wl, Quant<FakeQuantI8>());
+        if (quantData->_weights_quant.IsStatsSet()) {
+            if (quantData->_weights_quant.GetLevels() <= std::numeric_limits<uint8_t>::max()) {
+                quantizeWeightsBiases<FakeQuantI8>(FakeQuantI8(), wl, Quant<FakeQuantI8>());
+            } else {
+                quantizeWeightsBiases<FakeQuantI16>(FakeQuantI16(), wl, Quant<FakeQuantI16>());
+            }
         } else {
-            quantizeWeightsBiases<FakeQuantI16>(FakeQuantI16(), wl, Quant<FakeQuantI16>());
+            if (std::is_same<typename Desc::OptionalType, FakeQuant>()) {
+                THROW_GNA_EXCEPTION << "Infinite recursion. The type Desc::OptionalType is equal FakeQuant.";
+            }
+
+            (*this)(wl, typename Desc::OptionalType());
         }
     }
 };

diff --git a/inference-engine/src/gna_plugin/frontend/scale_factor_calc.hpp b/inference-engine/src/gna_plugin/frontend/scale_factor_calc.hpp
@@ -1354,6 +1354,12 @@ class ScaleFactorCalculator {
                 } else {
                     return frontend::FakeQuantI16().getWeightsPrecision().size();
                 }
+            } else {
+                if (!info.isSynthetic()) {
+                    gnawarn() << "The layer (" << ptr->name << ") has not quantization statistics\n";
+                }
+
+                return GetOptionalWeightsBytesSize();
             }
         }
 

diff --git a/inference-engine/src/gna_plugin/layers/gna_layer_info.hpp b/inference-engine/src/gna_plugin/layers/gna_layer_info.hpp
@@ -344,6 +344,10 @@ class LayerInfo {
         return isConcatAlignFilter() || isSyntheticScaleShift() || isCropAffined();
     }
 
+    bool isSynthetic() const noexcept {
+        return isConcatAlignFilter() || isSyntheticScaleShift() || isConvolutionFilter() || isAffineFilter();
+    }
+
     size_t paddingSize() const {
         static InferenceEngine::details::caseless_set<std::string> layersWithPossiblePadding = {"FullyConnected",
                                                                         "InnerProduct",