From d213072178ceba23f075139f62b9e436d2e0029d Mon Sep 17 00:00:00 2001
From: =?utf8?q?=D0=9F=D0=B0=D0=B2=D0=B5=D0=BB=20=D0=98=D0=BB=D1=8C=D1=8E?=
=?utf8?q?=D1=82=D1=87=D0=B5=D0=BD=D0=BA=D0=BE/AI=20Tools=20Lab=20/SRR/Eng?=
=?utf8?q?ineer/=EC=82=BC=EC=84=B1=EC=A0=84=EC=9E=90?=
Date: Thu, 10 Jan 2019 16:04:34 +0300
Subject: [PATCH] [nnc] Support Clip and Reshape operations on caffe2 importer
(#2751)
* Support GivenTensorInt64Fill const tensors
* Convert Clip operation in CappedRelu
* Support Reshape operation
Signed-off-by: Pavel Iliutchenko
---
contrib/nnc/include/core/modelIR/DataType.h | 3 +-
contrib/nnc/include/core/modelIR/Scalar.h | 2 +
.../nnc/passes/caffe2_frontend/caffe2_importer.cpp | 109 ++++++++++-------
.../nnc/passes/caffe2_frontend/caffe2_importer.h | 2 +-
.../passes/caffe2_frontend/caffe2_op_creator.cpp | 130 ++++++++++++++++++---
.../nnc/passes/caffe2_frontend/caffe2_op_creator.h | 9 +-
.../nnc/passes/caffe2_frontend/caffe2_op_types.h | 5 +-
7 files changed, 193 insertions(+), 67 deletions(-)
diff --git a/contrib/nnc/include/core/modelIR/DataType.h b/contrib/nnc/include/core/modelIR/DataType.h
index 22d6f90..d7c79c4 100644
--- a/contrib/nnc/include/core/modelIR/DataType.h
+++ b/contrib/nnc/include/core/modelIR/DataType.h
@@ -23,7 +23,8 @@ namespace mir {
enum class DTYPE {
UNKNOWN,
FLOAT32,
- INT32
+ INT32,
+ INT64
};
} // namespace mir
diff --git a/contrib/nnc/include/core/modelIR/Scalar.h b/contrib/nnc/include/core/modelIR/Scalar.h
index dba1e08..01f1444 100644
--- a/contrib/nnc/include/core/modelIR/Scalar.h
+++ b/contrib/nnc/include/core/modelIR/Scalar.h
@@ -62,6 +62,8 @@ public:
case DTYPE::FLOAT32:
case DTYPE::INT32:
return 4;
+ case DTYPE::INT64:
+ return 8;
}
}
/**
diff --git a/contrib/nnc/passes/caffe2_frontend/caffe2_importer.cpp b/contrib/nnc/passes/caffe2_frontend/caffe2_importer.cpp
index e2a5fa8..5a0e613 100644
--- a/contrib/nnc/passes/caffe2_frontend/caffe2_importer.cpp
+++ b/contrib/nnc/passes/caffe2_frontend/caffe2_importer.cpp
@@ -122,10 +122,13 @@ void Caffe2Importer::collectUnsupportedOp(const OperatorDef& op) {
case SupportedCaffe2OpType::constantFill:
case SupportedCaffe2OpType::dropout:
case SupportedCaffe2OpType::givenTensorFill:
+ case SupportedCaffe2OpType::givenTensorInt64Fill:
case SupportedCaffe2OpType::relu:
case SupportedCaffe2OpType::sigmoid:
case SupportedCaffe2OpType::softmax:
case SupportedCaffe2OpType::sum:
+ case SupportedCaffe2OpType::clip:
+ case SupportedCaffe2OpType::reshape:
_opCreator->commonCheck(op, _problemsOpSet);
break;
default:
@@ -139,7 +142,8 @@ void Caffe2Importer::preloadAllTensors() {
// All tensor values are stored in 'GivenTensorFill' and 'ConstantFill' operators, so skip rest
auto opType = _operatorTypes.at(op.type());
if ((opType == SupportedCaffe2OpType::givenTensorFill
- || opType == SupportedCaffe2OpType::constantFill)
+ || opType == SupportedCaffe2OpType::constantFill
+ || opType == SupportedCaffe2OpType::givenTensorInt64Fill)
&& hasArgument(op.arg(), "values")) {
_MIRTensors.insert(std::make_pair(op.output(0), createTensor(op)));
}
@@ -165,6 +169,7 @@ void Caffe2Importer::createMIRNodesFromOp(const OperatorDef& op) {
switch (opType) {
case SupportedCaffe2OpType::constantFill:
case SupportedCaffe2OpType::givenTensorFill:
+ case SupportedCaffe2OpType::givenTensorInt64Fill:
return;
case SupportedCaffe2OpType::add:
outputs = _opCreator->convertAdd(inputs, op, _MIRTensors);
@@ -205,6 +210,12 @@ void Caffe2Importer::createMIRNodesFromOp(const OperatorDef& op) {
case SupportedCaffe2OpType::sum:
outputs = _opCreator->convertSum(inputs);
break;
+ case SupportedCaffe2OpType::clip:
+ outputs = _opCreator->convertClip(inputs, op);
+ break;
+ case SupportedCaffe2OpType::reshape:
+ outputs = _opCreator->convertReshape(inputs, op, _MIRTensors);
+ break;
default:
assert(false && "All unsupported types should have been found before this pass.");
}
@@ -225,14 +236,36 @@ mir::TensorVariant Caffe2Importer::createTensor(const OperatorDef& op) {
const auto& values = findArgumentByName(op.arg(), "values");
// Create untyped tensor. Note, tensor contents will be *copied* here.
- auto element_type = mir::DTYPE::FLOAT32;
- size_t element_size = sizeof(float);
- size_t data_size = values.floats().size() * element_size;
- std::shared_ptr data(new char[data_size], std::default_delete());
- memcpy(data.get(), values.floats().data(), data_size);
+ mir::DTYPE element_type(mir::DTYPE::UNKNOWN);
+ const auto opType(_operatorTypes.at(op.type()));
+ size_t element_size = 0;
+ size_t data_size = 0;
+ char* data_ptr = nullptr;
+ // if values on floats
+ if (values.floats().size() > 0) {
+ element_type = mir::DTYPE::FLOAT32;
+ element_size = sizeof(float);
+ data_size = values.floats().size() * element_size;
+ data_ptr = new char[data_size];
+ memcpy(data_ptr, values.floats().data(), data_size);
+ }
+ // if values on ints
+ if (values.ints().size() > 0) {
+ if (opType == SupportedCaffe2OpType::givenTensorInt64Fill) {
+ element_size = sizeof(int64_t);
+ element_type = mir::DTYPE::INT64;
+ } else {
+ element_size = sizeof(int32_t);
+ element_type = mir::DTYPE::INT32;
+ }
+ data_size = values.ints().size() * element_size;
+ data_ptr = new char[data_size];
+ memcpy(data_ptr, values.ints().data(), data_size);
+ }
+ std::shared_ptr data(data_ptr, std::default_delete());
Shape tensor_shape = ShapeHelper::createShape(
- shape.ints(), static_cast(shape.ints().size()));
+ shape.ints(), static_cast(shape.ints().size()));
return mir::TensorVariant(tensor_shape, data, element_type, element_size);
}
@@ -242,30 +275,13 @@ std::vector Caffe2Importer::getInputMIROps(const OperatorDef&
// so choose caffe2 inputs, which are 'real' inputs
std::vector inputs;
SupportedCaffe2OpType opType = _operatorTypes.at(op.type());
- switch (opType) {
- case SupportedCaffe2OpType::givenTensorFill:
- case SupportedCaffe2OpType::constantFill:
- break;
- case SupportedCaffe2OpType::add:
- case SupportedCaffe2OpType::averagePool:
- case SupportedCaffe2OpType::conv:
- case SupportedCaffe2OpType::dropout:
- case SupportedCaffe2OpType::FC:
- case SupportedCaffe2OpType::maxPool:
- case SupportedCaffe2OpType::mul:
- case SupportedCaffe2OpType::relu:
- case SupportedCaffe2OpType::sigmoid:
- case SupportedCaffe2OpType::softmax:
- case SupportedCaffe2OpType::spatialBN:
- inputs.push_back(_blobNameToIODescriptor[op.input(0)]);
- break;
- case SupportedCaffe2OpType::sum:
- case SupportedCaffe2OpType::concat:
- for (auto& i : op.input())
+ if (opType != SupportedCaffe2OpType::givenTensorFill &&
+ opType != SupportedCaffe2OpType::constantFill &&
+ opType != SupportedCaffe2OpType::givenTensorInt64Fill)
+ {
+ for (auto& i : op.input())
+ if (_blobNameToIODescriptor.find(i) != _blobNameToIODescriptor.end())
inputs.push_back(_blobNameToIODescriptor[i]);
- break;
- default:
- assert(false && "All unsupported types should have been found before this pass.");
}
return inputs;
@@ -280,21 +296,24 @@ void Caffe2Importer::setGraphOutputs() {
}
const std::map Caffe2Importer::_operatorTypes = {
- {"Add", SupportedCaffe2OpType::add},
- {"AveragePool", SupportedCaffe2OpType::averagePool},
- {"Conv", SupportedCaffe2OpType::conv},
- {"Concat", SupportedCaffe2OpType::concat},
- {"ConstantFill", SupportedCaffe2OpType::constantFill},
- {"Dropout", SupportedCaffe2OpType::dropout},
- {"FC", SupportedCaffe2OpType::FC},
- {"GivenTensorFill", SupportedCaffe2OpType::givenTensorFill},
- {"MaxPool", SupportedCaffe2OpType::maxPool},
- {"Mul", SupportedCaffe2OpType::mul},
- {"Relu", SupportedCaffe2OpType::relu},
- {"Sigmoid", SupportedCaffe2OpType::sigmoid},
- {"Softmax", SupportedCaffe2OpType::softmax},
- {"SpatialBN", SupportedCaffe2OpType::spatialBN},
- {"Sum", SupportedCaffe2OpType::sum}
+{"Add", SupportedCaffe2OpType::add},
+{"AveragePool", SupportedCaffe2OpType::averagePool},
+{"Conv", SupportedCaffe2OpType::conv},
+{"Concat", SupportedCaffe2OpType::concat},
+{"ConstantFill", SupportedCaffe2OpType::constantFill},
+{"Dropout", SupportedCaffe2OpType::dropout},
+{"FC", SupportedCaffe2OpType::FC},
+{"GivenTensorFill", SupportedCaffe2OpType::givenTensorFill},
+{"MaxPool", SupportedCaffe2OpType::maxPool},
+{"Mul", SupportedCaffe2OpType::mul},
+{"Relu", SupportedCaffe2OpType::relu},
+{"Sigmoid", SupportedCaffe2OpType::sigmoid},
+{"Softmax", SupportedCaffe2OpType::softmax},
+{"SpatialBN", SupportedCaffe2OpType::spatialBN},
+{"Sum", SupportedCaffe2OpType::sum},
+{"Clip", SupportedCaffe2OpType::clip},
+{"Reshape", SupportedCaffe2OpType::reshape},
+{"GivenTensorInt64Fill", SupportedCaffe2OpType::givenTensorInt64Fill},
};
} // namespace nnc
diff --git a/contrib/nnc/passes/caffe2_frontend/caffe2_importer.h b/contrib/nnc/passes/caffe2_frontend/caffe2_importer.h
index c01de54..8720442 100644
--- a/contrib/nnc/passes/caffe2_frontend/caffe2_importer.h
+++ b/contrib/nnc/passes/caffe2_frontend/caffe2_importer.h
@@ -66,7 +66,7 @@ private:
// This map maps caffe2 operators names to MIR operators
// that correspond to previous caffe2 operators
- std::map _blobNameToIODescriptor;
+ std::unordered_map _blobNameToIODescriptor;
mir::Operation* _lastMIROp = nullptr;
std::map _MIRTensors;
diff --git a/contrib/nnc/passes/caffe2_frontend/caffe2_op_creator.cpp b/contrib/nnc/passes/caffe2_frontend/caffe2_op_creator.cpp
index 478ec6b..ee92ff9 100644
--- a/contrib/nnc/passes/caffe2_frontend/caffe2_op_creator.cpp
+++ b/contrib/nnc/passes/caffe2_frontend/caffe2_op_creator.cpp
@@ -61,8 +61,30 @@ using nnc::mir::transposeTensor;
static std::pair, std::vector>
getPadding(const ::caffe2::OperatorDef& op) {
+
+ if (hasArgument(op.arg(), "pads")) {
+ // pads order: t l b r
+ auto pads_arg = findArgumentByName(op.arg(), "pads");
+
+ std::vector paddings;
+ for (const auto& pad : pads_arg.ints())
+ paddings.push_back(static_cast(pad));
+
+ assert(paddings.size() == 4);
+
+ int32_t pad_t = paddings[0];
+ int32_t pad_l = paddings[1];
+ int32_t pad_b = paddings[2];
+ int32_t pad_r = paddings[3];
+
+ std::vector padding_before{pad_t, pad_l};
+ std::vector padding_after{pad_b, pad_r};
+ return {padding_before, padding_after};
+ }
+
bool has_custom_pad = hasArgument(op.arg(), "pad_l") || hasArgument(op.arg(), "pad_r")
|| hasArgument(op.arg(), "pad_t") || hasArgument(op.arg(), "pad_b");
+
if (has_custom_pad) {
int32_t pad_l = getSingleArgument(op, "pad_l", 0);
int32_t pad_t = getSingleArgument(op, "pad_t", 0);
@@ -78,13 +100,35 @@ getPadding(const ::caffe2::OperatorDef& op) {
return {{pad, pad}, {pad, pad}};
};
+static std::vector
+getStrides(const ::caffe2::OperatorDef& op) {
+ std::vector strides;
+
+ if (hasArgument(op.arg(), "stride")) {
+ int stride = getSingleArgument(op, "stride", 1);
+ strides = {stride, stride};
+ }
+
+ if (hasArgument(op.arg(), "strides")) {
+ // strides order: h w
+ auto strides_arg = findArgumentByName(op.arg(), "strides");
+ for (const auto& s : strides_arg.ints())
+ strides.push_back(s);
+ }
+
+ assert(!strides.empty() && "Strides not found");
+
+ return strides;
+}
+
static Shape getWindowShape(const ::caffe2::OperatorDef& op,
const std::vector& inputs) {
int is_global_pooling = getSingleArgument(op, "global_pooling", 0);
- bool has_custom_kernel_size = hasArgument(op.arg(), "kernel_h")
- || hasArgument(op.arg(), "kernel_w");
+ bool has_custom_kernel_size = hasArgument(op.arg(), "kernel_h") ||
+ hasArgument(op.arg(), "kernel_w");
+ bool has_custom_kernels_size = hasArgument(op.arg(), "kernels");
- int kernel_h, kernel_w;
+ int kernel_h(0), kernel_w(0);
if (is_global_pooling) {
auto& input_shape = inputs[0].op->getOutputShape(inputs[0].index);
assert(input_shape.rank() == 4 && "getWindowShape() inputs must be of rank 4");
@@ -95,10 +139,21 @@ static Shape getWindowShape(const ::caffe2::OperatorDef& op,
kernel_h = getSingleArgument(op, "kernel_h", 0);
kernel_w = getSingleArgument(op, "kernel_w", 0);
} else {
- kernel_h = kernel_w = getSingleArgument(op, "kernel", 0);
+ if (has_custom_kernels_size) {
+ // kernels order: h w
+ std::vector kernels;
+ auto kernels_arg = findArgumentByName(op.arg(), "kernels");
+ for (const auto& ker : kernels_arg.ints())
+ kernels.push_back(static_cast(ker));
+ assert(kernels.size() == 2);
+ kernel_h = kernels[0];
+ kernel_w = kernels[1];
+ } else {
+ kernel_h = kernel_w = getSingleArgument(op, "kernel", 0);
+ }
}
}
- return Shape({kernel_h, kernel_w});
+ return Shape{kernel_h, kernel_w};
}
mir::IODescriptor Caffe2OpCreator::convertCaffeToMIR(const mir::IODescriptor& arg) {
@@ -200,14 +255,20 @@ Caffe2OpCreator::convertAdd(const std::vector& inputs,
const ::caffe2::OperatorDef& op,
const MIRTensors& mir_tensors) {
const auto& input_shape = inputs[0].op->getOutputShape(inputs[0].index);
- const auto& addend = mir_tensors.at(op.input(1));
- assert(addend.getShape().rank() == 1 && "Only 1-rank addend is supported");
- assert(addend.getShape().numElements() == input_shape.dim(1)
- && "Only addend size equal to number of input channels is supported");
- // TODO: replace with elementwise op, when broadcating will be added in elementwise op
- auto add = createOp(convertCaffeToMIR(inputs[0]), addend);
+ std::vector add_input;
+ for (const auto& i : inputs)
+ add_input.push_back(convertCaffeToMIR(i.op->getOutput(0)));
+
+ // check mir tensors contain operand
+ if (mir_tensors.find(op.input(1)) != mir_tensors.end()) {
+ auto next_input = createOp(mir_tensors.at(op.input(1)));
+ add_input.push_back(next_input[0].getOutput(0));
+ }
+
+ auto add = createOp(add_input, ops::ElementwiseOp::OpType::add);
+
return {convertMIRToCaffe(add->getOutput(0))};
}
@@ -216,8 +277,7 @@ Caffe2OpCreator::convertAveragePool(const std::vector& inputs,
const OperatorDef& op) {
Shape window_shape = getWindowShape(op, inputs);
- int stride = getSingleArgument(op, "stride", 1);
- Shape strides = Shape({stride, stride});
+ Shape strides(getStrides(op));
ops::PoolOp::PoolingType pool_type = ops::PoolOp::PoolingType::AVG;
ops::PoolOp::BorderType border_type = ops::PoolOp::BorderType::EMPTY;
@@ -235,8 +295,8 @@ Caffe2OpCreator::convertAveragePool(const std::vector& inputs,
std::vector Caffe2OpCreator::convertConv(const std::vector& inputs,
const ::caffe2::OperatorDef& op,
const MIRTensors& mir_tensors) {
- int stride = getSingleArgument(op, "stride", 1);
- Shape stride_shape = Shape({stride, stride});
+ // dilation order: h w (not used)
+ Shape stride_shape(getStrides(op));
std::vector pad_before, pad_after;
std::tie(pad_before, pad_after) = getPadding(op);
@@ -306,9 +366,7 @@ Caffe2OpCreator::convertFullyConnected(const std::vector& inputs,
std::vector Caffe2OpCreator::convertMaxPool(const std::vector& inputs,
const OperatorDef& op) {
Shape window_shape = getWindowShape(op, inputs);
-
- int stride = getSingleArgument(op, "stride", 1);
- Shape strides = Shape({stride, stride});
+ Shape strides(getStrides(op));
ops::PoolOp::PoolingType pool_type = ops::PoolOp::PoolingType::MAX;
ops::PoolOp::BorderType border_type = ops::PoolOp::BorderType::EMPTY;
@@ -397,6 +455,42 @@ std::vector Caffe2OpCreator::convertSum(const std::vectorgetOutput(0)};
}
+std::vector
+Caffe2OpCreator::convertClip(const std::vector& inputs,
+ const ::caffe2::OperatorDef& op) {
+
+ float max = getSingleArgument(op, "max", float(0));
+ float min = getSingleArgument(op, "min", float(0));
+
+ assert(max > 0.0 && min == 0.0 && "Support only if clip is CappedRelu");
+ auto cap_relu = createOp(inputs[0], max);
+
+ return {cap_relu->getOutput(0)};
+}
+
+
+std::vector
+Caffe2OpCreator::convertReshape(const std::vector& inputs,
+ const ::caffe2::OperatorDef& op,
+ const MIRTensors& mir_tensors) {
+ // Check new shape input
+ assert(mir_tensors.find(op.input(1)) != mir_tensors.end());
+ const auto& shape_tensor = mir_tensors.at(op.input(1));
+
+ Tensor out_shape_tensor(shape_tensor);
+
+ ShapeRange range(out_shape_tensor.getShape());
+ std::vector shape_vec;
+ for (const auto& index: range) {
+ shape_vec.push_back(static_cast(out_shape_tensor.at(index)));
+ }
+ Shape out_shape(shape_vec);
+
+ auto reshape = createOp(inputs[0], out_shape);
+
+ return {reshape->getOutput(0)};
+}
+
std::vector
Caffe2OpCreator::createInput(const std::string& input_name, const mir::Shape& input_shape) {
// TODO For now we only support convolutional networks with one element per batch.
diff --git a/contrib/nnc/passes/caffe2_frontend/caffe2_op_creator.h b/contrib/nnc/passes/caffe2_frontend/caffe2_op_creator.h
index 5bbbf9d..ec2bbc5 100644
--- a/contrib/nnc/passes/caffe2_frontend/caffe2_op_creator.h
+++ b/contrib/nnc/passes/caffe2_frontend/caffe2_op_creator.h
@@ -18,7 +18,7 @@
#define NNCC_CAFFE2_OP_CREATOR_H
#include
-#include