A set of common optimization passes. More...
Data Structures | |
class | ngraph::pass::BidirectionalLSTMSequenceDecomposition |
Decompose LSTMSequence to forward and reverse LSTMSequence. More... | |
class | ngraph::pass::BidirectionalGRUSequenceDecomposition |
Decompose GRUSequence to forward and reverse GRUSequence. More... | |
class | ngraph::pass::BidirectionalRNNSequenceDecomposition |
Decompose RNNSequence to forward and reverse RNNSequence. More... | |
class | ngraph::pass::FakeQuantizeMulFusion |
This transformation looks for a FQ + Mul pair in the graph and moves the Mul operation above the FQ node. The last two inputs of FQ are multiplied by the value that was originally below the FQ node. More... | |
class | ngraph::pass::ConvertBatchToSpace |
ConvertBatchToSpace transformation decomposes BatchToSpace layer to Reshape->Transpose->Reshape->Crop. More... | |
class | ngraph::pass::ConvertExtractImagePatchesToReorgYolo |
ConvertExtractImagePatchesToReorgYolo transformation replaces ExtractImagePatches with a ReorgYolo op. More... | |
class | ngraph::pass::ConvertLSTMSequenceMatcher |
Converts LSTMSequence to legacy LSTMSequenceIE. SequenceIE op doesn't use seq_length input and num_direction (direction) attribute. We squeeze num_direction dimension for all corresponding inputs and unsqueeze them after the SequenceIE op. More... | |
class | ngraph::pass::ConvertGRUSequenceMatcher |
Converts GRUSequence to legacy GRUSequenceIE. SequenceIE op doesn't use seq_length input and num_direction (direction) attribute. We squeeze num_direction dimension for all corresponding inputs and unsqueeze them after the SequenceIE op. More... | |
class | ngraph::pass::ConvertRNNSequenceMatcher |
Converts RNNSequence to legacy RNNSequenceIE. SequenceIE op doesn't use seq_length input and num_direction (direction) attribute. We squeeze num_direction dimension for all corresponding inputs and unsqueeze them after the SequenceIE op. More... | |
class | ngraph::pass::ConvertPadToGroupConvolution |
ConvertPadToGroupConvolution transformation replaces Pad operation with GroupConvolution but has some restrictions on Pad parameters: More... | |
class | ngraph::pass::ConvertPrecision |
ConvertPrecision transformation convert precision for entire ngraph::Function List of supported precision conversion: FROM -> TO u8 -> i32 u16 -> i32 u32 -> i32 u64 -> i32 i64 -> i32 f16 -> f32 bool -> u8 bool -> i32. More... | |
class | ngraph::pass::ConvertQuantizeDequantize |
ConvertQuantizeDequantize transformation replaces following graph: FakeQuantize->Convert->Convert->Subtract->Multiply with a single FakeQuantize. Restrictions: More... | |
class | ngraph::pass::ConvertScatterElementsToScatter |
ConvertScatterElementsToScatter convert opset3::ScatterElementsUpdate to opset3::ScatterUpdate. More... | |
class | ngraph::pass::ConvertSpaceToBatch |
ConvertSpaceToBatch transformation decomposes SpaceToBatch layer to Pad->Reshape->Transpose->Reshape. More... | |
class | ngraph::pass::DepthToSpaceFusion |
DepthToSpaceFusion transformation detects Reshape-Transpose-Reshape pattern and tries to fuse it into a single DepthToSpace layer. More... | |
class | ngraph::pass::GRUCellDecomposition |
GRUCellDecomposition transformation decomposes GRUCell layer with inputs X, H, W, R, B to Add, Split, MatMul, Multiply and Subtract ops according to the formula: (.) - Denotes element-wise multiplication. More... | |
class | ngraph::pass::HSwishDecomposition |
HSwishDecomposition transformation into sub-graph x * (min(Relu(x + 3), 6) * const(1/6). More... | |
class | ngraph::pass::HSwishFusion |
HSwishFusion transformation replaces various sub-graphs with a HSwish op. More... | |
class | ngraph::pass::HSwishFusionWithReluDiv |
HSwishFusion transformation replaces a sub-graph (x * (min(Relu(x + 3), 6))) / 6 with a HSwish op. More... | |
class | ngraph::pass::HSwishFusionWithReluMul |
HSwishFusion transformation replaces a sub-graph (x * (min(Relu(x + 3), 6)) * const(1/6) with a HSwish op. More... | |
class | ngraph::pass::HSwishFusionWithoutRelu |
HSwishFusion transformation replaces a sub-graph x * (min(max(x + 3, 0), 6) / 6) with a HSwish op. More... | |
class | ngraph::pass::HSwishFusionWithClamp |
HSwishFusion transformation replaces a sub-graph x * (Clamp(x + 3, 0, 6) * const(1/6)) with a HSwish op. More... | |
class | ngraph::pass::InitNodeInfo |
InitNodeInfo transformation helps to set runtime info attributes in a single place. More... | |
class | ngraph::pass::LSTMCellDecomposition |
LSTMCellDecomposition transformation decomposes LSTMCell layer with inputs X, H, C, W, R, B to Add, Split, MatMul, Multiply ops according to the formula: (.) - Denotes element-wise multiplication. More... | |
class | ngraph::pass::MishFusion |
MishFusion transformation replaces group of operations: x * tanh(log(exp(x) + 1)) to Mish op. More... | |
class | ngraph::pass::NormalizeL2Fusion |
NormalizeL2Fusion transformation replaces various sub-graphs with a NormalizeL2 op. More... | |
class | ngraph::pass::NormalizeL2FusionWithMax |
NormalizeL2FusionWithMax transformation replaces a sub-graph x/(max(sqrt(sum(x[j0, ..., jN]**2), eps)) with a NormalizeL2 op. More... | |
class | ngraph::pass::NormalizeL2FusionWithAdd |
NormalizeL2FusionWithAdd transformation replaces a sub-graph x/(add(sqrt(sum(x[j0, ..., jN]**2), eps)) with a NormalizeL2 op. More... | |
class | ngraph::pass::UselessStridedSliceEraser |
UselessStridedSliceEraser transformation removes StridedSlice operations with equal input and output shapes. More... | |
class | ngraph::pass::SharedStridedSliceEraser |
SharedStridedSliceEraser transformation replaces group of StridedSlice operations with first StridedSlice in this group. All SrtideSlices in this group must be equal and consume the same output port. More... | |
class | ngraph::pass::GroupedStridedSliceOptimizer |
GroupedStridedSliceOptimizer transformation replaces group of StridedSlice operations with VariadicSplit. All StridedSlice operations must slice data with the same axis and stride = 1. More... | |
class | ngraph::pass::StridedSliceOptimization |
StridedSliceOptimization transformation executes all transformations related to StridedSlice optimizations. More... | |
class | ngraph::pass::ReduceL1Decomposition |
Decomposes ReduceL1 into ReduceSum(abs(x)). More... | |
class | ngraph::pass::ReduceL2Decomposition |
Decomposes ReduceL2 into sqrt(ReduceSum(x * x)). More... | |
class | ngraph::pass::RNNCellDecomposition |
RNNCellDecomposition transformation decomposes RNNCell layer with inputs X, H, W, R, B to Add, MatMul ops according to the formula: More... | |
class | ngraph::pass::SoftPlusDecomposition |
SoftPlusDecomposition transformation replaces SoftPlus op to group of operations: log(exp(x) + 1). More... | |
class | ngraph::pass::SoftPlusFusion |
SoftPlusFusion transformation replaces group of operations: log(exp(x) + 1) to SoftPlus op. More... | |
class | ngraph::pass::SoftPlusToMishFusion |
SoftPlusToMishFusion transformation replaces group of operations: x * tanh(softplus(x)) to Mish op. More... | |
class | ngraph::pass::SwishFusion |
SwishFusion transformation replaces various sub-graphs with a Swish op. More... | |
class | ngraph::pass::SwishFusionWithSigmoid |
SwishFusionWithSigmoid replaces a sub-graphs x * Sigmoid(x) with a Swish op. More... | |
class | ngraph::pass::SwishFusionWithSigmoidWithBeta |
SwishFusionWithSigmoid replaces a sub-graphs x * Sigmoid(x * beta) with a Swish op. More... | |
class | ngraph::pass::SwishFusionWithBeta |
SwishFusionWithSigmoid replaces a sub-graphs x / (1.0 + exp(-x * beta)) with a Swish op. More... | |
class | ngraph::pass::SwishFusionWithoutBeta |
SwishFusionWithSigmoid replaces a sub-graphs x / (1.0 + exp(-x)) with a Swish op. More... | |
class | ngraph::pass::ApplyTransformationsToTIBody |
Applies transformations from given pass::Manager for each TensorIterator body. More... | |
class | ngraph::pass::UnrollTensorIterator |
Unrolls the body of the TensorIterator layer. Multiple body copies, the number of which is determined by the number of iterations of the TensorIterator layer, are created and connected to each other and to the external network. If the number of TensorIterator iterations is greater than 1, then additional Concat and Split layers are added to the network. More... | |
A set of common optimization passes.