[LPT] NNCF GroupConvolution 5D on weights support

openvinotoolkit · Mar 20, 2023 · 401b90d · 401b90d
1 parent 4ffecce
commit 401b90d
Show file tree

Hide file tree

Showing 10 changed files with 1,173 additions and 602 deletions.
diff --git a/src/common/low_precision_transformations/src/convolution.cpp b/src/common/low_precision_transformations/src/convolution.cpp
@@ -237,8 +237,15 @@ bool ConvolutionTransformation::transform(TransformationContext &context, ngraph
             Shape newScaleShape = newScalePShape.to_shape();
 
             if (!newScaleShape.empty()) {
-                // that's all we need: [C, 1, 1, 1] => [C, 1, 1]
-                newScaleShape.pop_back();
+                const auto input_shape = convolution->get_input_partial_shape(0);
+                const auto diff = newScaleShape.size() - input_shape.size();
+                OPENVINO_ASSERT(
+                    newScaleShape.empty() || ((0 <= diff) && (diff <= 2ull)),
+                    "unexpected shape size on weights");
+
+                for (size_t i = 0; i <= diff; ++i) {
+                    newScaleShape.pop_back();
+                }
             }
 
             if (reshapeFromWeights != nullptr) {
@@ -282,7 +289,12 @@ bool ConvolutionTransformation::transform(TransformationContext &context, ngraph
 
                 const size_t weightsRankValue = weightsPShape.rank().get_length();
                 Shape zeroPointShape(weightsRankValue, 1ul);
+                // output channel or group
                 zeroPointShape[0] = static_cast<size_t>(weightsPShape[0].get_length());
+                if ((reshapeFromWeights == nullptr) && (weightsRankValue == 5ull)) {
+                    // output channel
+                    zeroPointShape[1] = static_cast<size_t>(weightsPShape[1].get_length());
+                }
 
                 auto zeroPointConstant = fold<opset1::Broadcast>(
                     subtractFromWeights->input_value(1),

diff --git a/src/common/low_precision_transformations/src/pull_reshape_through_dequantization.cpp b/src/common/low_precision_transformations/src/pull_reshape_through_dequantization.cpp
@@ -114,11 +114,6 @@ ngraph::pass::low_precision::PullReshapeThroughDequantization::PullReshapeThroug
         const auto& opsMap = m.get_pattern_value_map();
         auto reshape = opsMap.at(reshapeWrapper).get_node_shared_ptr();
 
-        auto child = reshape->get_output_target_inputs(0).begin()->get_node();
-        if (ov::is_type<opset1::GroupConvolution>(child)) {
-            return false;
-        }
-
         while (reshape != nullptr) {
             const auto parent = reshape->get_input_node_shared_ptr(0);
             if (ov::is_type<opset1::Multiply>(parent) || ov::is_type<opset1::Subtract>(parent)) {

diff --git a/src/common/low_precision_transformations/src/weightable_layer_transformation.cpp b/src/common/low_precision_transformations/src/weightable_layer_transformation.cpp
@@ -230,16 +230,16 @@ bool WeightableLayerTransformation::isQuantizedStatic(const std::shared_ptr<cons
     FakeQuantizeDequantization dequantizationOnWeights;
     if (reshapeIsRequired) {
         const auto reshape = layer->get_input_node_shared_ptr(1);
-        if (!ov::is_type<opset1::Reshape>(reshape)) {
-            return false;
-        }
+        std::shared_ptr<Node> parent = ov::is_type<opset1::Reshape>(reshape) ?
+            reshape->get_input_node_shared_ptr(0) :
+            reshape;
 
-        if (ov::is_type<opset1::FakeQuantize>(reshape->get_input_node_shared_ptr(0))) {
-            const std::shared_ptr<opset1::FakeQuantize> fq = ov::as_type_ptr<opset1::FakeQuantize>(reshape->get_input_node_shared_ptr(0));
+        const auto fq = ov::as_type_ptr<opset1::FakeQuantize>(parent);
+        if (fq != nullptr) {
             return NetworkHelper::isQuantizeSupported(fq);
         }
 
-        dequantizationOnWeights = NetworkHelper::getDequantization(reshape, defaultPrecisions, 0);
+        dequantizationOnWeights = NetworkHelper::getDequantization(parent, defaultPrecisions, 0, true);
     } else if (ov::is_type<opset1::FakeQuantize>(layer->get_input_node_shared_ptr(1))) {
         const std::shared_ptr<opset1::FakeQuantize> fq = ov::as_type_ptr<opset1::FakeQuantize>(layer->get_input_node_shared_ptr(1));
         return NetworkHelper::isQuantizeSupported(fq);