mirror/opencv
1
0
Fork 0
mirror of https://github.com/opencv/opencv.git synced 2025-08-06 06:26:29 +08:00
Code Issues Actions Packages Projects Releases Wiki Activity

Merge pull request #23409 from dkurt:dnn_tflite_quant

Browse Source 
Import and inference INT8 quantized TFLite model #23409

### Pull Request Readiness Checklist

* Support quantized TFLite models
* Enable fused activations (FP32, INT8)

**Merge with extra**: https://github.com/opencv/opencv_extra/pull/1048

![res](https://user-images.githubusercontent.com/25801568/231433201-566b4bd6-ccff-462c-9e74-adbdcdf3648b.png)

on the image, green boxes are from TFLite and red boxes from OpenCV

See details at https://github.com/opencv/opencv/wiki/How_to_contribute#making-a-good-pull-request

- [x] I agree to contribute to the project under Apache 2 License.
- [x] To the best of my knowledge, the proposed patch is not based on a code under GPL or another license that is incompatible with OpenCV
- [x] The PR is proposed to the proper branch
- [x] There is a reference to the original bug report and related work
- [x] There is accuracy test, performance test and test data in opencv_extra repository, if applicable
      Patch to opencv_extra has the same branch name.
- [x] The feature is well documented and sample code can be built with the project CMake
This commit is contained in:
Dmitry Kurtaev 2023-04-24 13:44:10 +03:00 committed by GitHub
parent e4e774d42b
commit aa57833ad5
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
2 changed files with 455 additions and 122 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

523
modules/dnn/src/tflite/tflite_importer.cpp
View File

@ -61,8 +61,17 @@ private:
void parseConcat(const Operator& op, const std::string& opcode, LayerParams& layerParams);
void parseResize(const Operator& op, const std::string& opcode, LayerParams& layerParams);
void parseDeconvolution(const Operator& op, const std::string& opcode, LayerParams& layerParams);
void parseQuantize(const Operator& op, const std::string& opcode, LayerParams& layerParams);
void parseDequantize(const Operator& op, const std::string& opcode, LayerParams& layerParams);
void parseDetectionPostProcess(const Operator& op, const std::string& opcode, LayerParams& layerParams);
void parseActivation(const Operator& op, const std::string& opcode, LayerParams& layerParams);
int addPermuteLayer(const std::vector<int>& order, const std::string& permName, const std::pair<int, int>& inpId);
void parseFusedActivation(const Operator& op, ActivationFunctionType activ);
void parseActivation(const Operator& op, const std::string& opcode, LayerParams& layerParams, bool isFused);
void addLayer(LayerParams& layerParams, const Operator& op);
int addPermuteLayer(const std::vector<int>& order, const std::string& permName, const std::pair<int, int>& inpId, int dtype);
inline bool isInt8(const Operator& op);
inline void getQuantParams(const Operator& op, float& inpScale, int& inpZero, float& outScale, int& outZero);
};
Mat TFLiteImporter::parseTensor(const Tensor& tensor)
@ -75,7 +84,9 @@ Mat TFLiteImporter::parseTensor(const Tensor& tensor)
const Buffer* buffer = model->buffers()->Get(bufferIdx);
CV_Assert(buffer);
const auto buffer_data = buffer->data();
CV_Assert(buffer_data);
if (!buffer_data)
return Mat();
const void* data = buffer_data->data();
int dtype = -1;
@ -89,6 +100,9 @@ Mat TFLiteImporter::parseTensor(const Tensor& tensor)
case TensorType_FLOAT16:
dtype = CV_16S;
break;
case TensorType_INT8:
dtype = CV_8S;
break;
default:
CV_Error(Error::StsNotImplemented, format("Parse tensor with type %s", EnumNameTensorType(tensor.type())));
}
@ -207,37 +221,103 @@ void TFLiteImporter::populateNet()
if (type == "DEQUANTIZE") {
// Convert from FP16 to FP32
Mat data = allTensors[op_inputs->Get(0)];
if (!data.empty()) {
// Dequantize a buffer
Mat dataFP32;
convertFp16(data, dataFP32);
allTensors[op_outputs->Get(0)] = dataFP32;
continue;
}
}
DispatchMap::const_iterator iter = dispatch.find(type);
if (iter == dispatch.end())
CV_Error(Error::StsNotImplemented, "Unsupported operator type " + type);
CALL_MEMBER_FN(*this, iter->second)(*op, type, layerParams);
}
catch (const cv::Exception& e)
{
CV_LOG_ERROR(NULL, "DNN/TFLite: Problem during import of operator "
<< cv::format("[%s]:(%s)", type.c_str(), layerParams.name.c_str())
<< " (" << op_idx << "/" << all_operators_size << "). Exception: " << e.what());
if (DNN_DIAGNOSTICS_RUN)
{
continue;
}
throw;
}
}
}
TFLiteImporter::DispatchMap TFLiteImporter::buildDispatchMap()
{
static DispatchMap dispatch;
if (!dispatch.empty())
return dispatch;
dispatch["CONV_2D"] = &TFLiteImporter::parseConvolution;
dispatch["DEPTHWISE_CONV_2D"] = &TFLiteImporter::parseDWConvolution;
dispatch["ADD"] = dispatch["MUL"] = &TFLiteImporter::parseEltwise;
dispatch["RELU"] = dispatch["PRELU"] = dispatch["HARD_SWISH"] =
dispatch["LOGISTIC"] = &TFLiteImporter::parseActivation;
dispatch["MAX_POOL_2D"] = dispatch["AVERAGE_POOL_2D"] = &TFLiteImporter::parsePooling;
dispatch["MaxPoolingWithArgmax2D"] = &TFLiteImporter::parsePoolingWithArgmax;
dispatch["MaxUnpooling2D"] = &TFLiteImporter::parseUnpooling;
dispatch["PAD"] = &TFLiteImporter::parsePadding;
dispatch["RESHAPE"] = &TFLiteImporter::parseReshape;
dispatch["CONCATENATION"] = &TFLiteImporter::parseConcat;
dispatch["RESIZE_BILINEAR"] = dispatch["RESIZE_NEAREST_NEIGHBOR"] = &TFLiteImporter::parseResize;
dispatch["Convolution2DTransposeBias"] = &TFLiteImporter::parseDeconvolution;
dispatch["QUANTIZE"] = &TFLiteImporter::parseQuantize;
dispatch["DEQUANTIZE"] = &TFLiteImporter::parseDequantize;
dispatch["TFLite_Detection_PostProcess"] = &TFLiteImporter::parseDetectionPostProcess;
return dispatch;
}
void TFLiteImporter::addLayer(LayerParams& layerParams, const Operator& op) {
const auto op_inputs = op.inputs();
const auto op_outputs = op.outputs();
// Collect input blobs
std::vector<int> layerInputs;
std::vector<DataLayout> inpLayouts;
if (layerParams.blobs.empty()) {
for (int idx : *op_inputs) {
if (layerIds.find(idx) != layerIds.end()) {
layerInputs.push_back(idx);
inpLayouts.push_back(layouts[idx]);
continue; // Output from a different layer
}
Mat blob = allTensors[idx];
layerParams.blobs.push_back(blob.u ? blob : blob.clone()); // some tensors are owned by OpenCV
}
}
int layerId = dstNet.addLayer(layerParams.name, layerParams.type, layerParams);
int dtype = CV_32F;
if (isInt8(op)) {
dtype = CV_8S;
if (layerParams.type != "Quantize")
layerParams.type += "Int8";
if (!layerParams.has("zeropoints")) {
float inpScale, outScale;
int inpZero, outZero;
getQuantParams(op, inpScale, inpZero, outScale, outZero);
layerParams.set("input_scale", inpScale);
layerParams.set("input_zeropoint", inpZero);
layerParams.set("scales", outScale);
layerParams.set("zeropoints", outZero);
}
}
int layerId = dstNet.addLayer(layerParams.name, layerParams.type, dtype, layerParams);
// Connect layer to inputs
int i = 0;
for (int idx : layerInputs) {
std::vector<DataLayout> inpLayouts;
for (int idx : *op_inputs) {
if (layerIds.find(idx) == layerIds.end()) {
continue; // Const input
}
inpLayouts.push_back(layouts[idx]);
auto it = layerIds.find(idx);
CV_Assert(it != layerIds.end());
dstNet.connect(it->second.first, it->second.second, layerId, i++);
@ -265,49 +345,12 @@ void TFLiteImporter::populateNet()
for (int idx : *op_outputs) {
layerIds[idx] = std::make_pair(layerId, i++);
}
}
catch (const cv::Exception& e)
{
CV_LOG_ERROR(NULL, "DNN/TFLite: Problem during import of operator "
<< cv::format("[%s]:(%s)", type.c_str(), layerParams.name.c_str())
<< " (" << op_idx << "/" << all_operators_size << "). Exception: " << e.what());
if (DNN_DIAGNOSTICS_RUN)
{
continue;
}
throw;
}
}
}
TFLiteImporter::DispatchMap TFLiteImporter::buildDispatchMap()
{
static DispatchMap dispatch;
if (!dispatch.empty())
return dispatch;
dispatch["CONV_2D"] = &TFLiteImporter::parseConvolution;
dispatch["DEPTHWISE_CONV_2D"] = &TFLiteImporter::parseDWConvolution;
dispatch["RELU"] = dispatch["ADD"] = dispatch["MUL"] = dispatch["PRELU"] =
dispatch["HARD_SWISH"] = dispatch["LOGISTIC"] = &TFLiteImporter::parseEltwise;
dispatch["MAX_POOL_2D"] = dispatch["AVERAGE_POOL_2D"] = &TFLiteImporter::parsePooling;
dispatch["MaxPoolingWithArgmax2D"] = &TFLiteImporter::parsePoolingWithArgmax;
dispatch["MaxUnpooling2D"] = &TFLiteImporter::parseUnpooling;
dispatch["PAD"] = &TFLiteImporter::parsePadding;
dispatch["RESHAPE"] = &TFLiteImporter::parseReshape;
dispatch["CONCATENATION"] = &TFLiteImporter::parseConcat;
dispatch["RESIZE_BILINEAR"] = &TFLiteImporter::parseResize;
dispatch["Convolution2DTransposeBias"] = &TFLiteImporter::parseDeconvolution;
return dispatch;
}
void TFLiteImporter::parseConvolution(const Operator& op, const std::string& opcode, LayerParams& layerParams) {
layerParams.type = "Convolution";
auto options = reinterpret_cast<const Conv2DOptions*>(op.builtin_options());
if (options->fused_activation_function() != ActivationFunctionType_NONE) {
CV_Error(Error::StsNotImplemented, "Convolution with fused activation");
}
layerParams.set("pad_mode", EnumNamePadding(options->padding()));
layerParams.set("stride_w", options->stride_w());
layerParams.set("stride_h", options->stride_h());
@ -320,42 +363,57 @@ void TFLiteImporter::parseConvolution(const Operator& op, const std::string& opc
int oc = filter.size[0];
int kh = filter.size[1];
int kw = filter.size[2];
int ic = filter.size[3];
layerParams.set("kernel_w", kw);
layerParams.set("kernel_h", kh);
layerParams.set("num_output", oc);
bool isInt8 = filter.depth() == CV_8S;
// Fill convolutions blobs here because of two reasons:
// 1. Kernel transposition
// 2. Extra blob with kernel scales in case of INT8 mode
bool hasBias = op.inputs()->size() > 2;
layerParams.blobs.resize(1 + (int)hasBias + (int)isInt8);
if (hasBias) {
Mat bias = allTensors[op.inputs()->Get(2)];
layerParams.blobs[1] = bias.u ? bias : bias.clone();
}
// Reorder filter data from OHWI to OIHW and change shape correspondingly.
filter = allTensors[filterIdx] = filter.reshape(1, {oc, ic, kh, kw});
transposeND(filter, {0, 3, 1, 2}, layerParams.blobs[0]);
CV_CheckTypeEQ(filter.type(), CV_32F, "");
Mat filterCopy = filter.clone();
float* data = filterCopy.ptr<float>();
float* dstData = filter.ptr<float>();
if (isInt8) {
float inpScale, outScale;
int inpZero, outZero;
getQuantParams(op, inpScale, inpZero, outScale, outZero);
int total = oc * ic * kh * kw;
for (int i_oc = 0; i_oc < oc; i_oc++) {
for (int i_ic = 0; i_ic < ic; i_ic++) {
for (int i_h = 0; i_h < kh; i_h++) {
for (int i_w = 0; i_w < kw; i_w++) {
int dst_i = kw * (kh * (ic * i_oc + i_ic) + i_h) + i_w;
int src_i = ic * (kw * (kh * i_oc + i_h) + i_w) + i_ic;
CV_CheckLT(dst_i, total, "");
CV_CheckLT(src_i, total, "");
dstData[dst_i] = data[src_i];
}
layerParams.blobs[2] = Mat(oc, 1, CV_32F);
auto filterScales = modelTensors->Get(filterIdx)->quantization()->scale();
if (filterScales->size() == 1) {
layerParams.blobs[2].setTo(inpScale * filterScales->Get(0) / outScale);
} else {
for (size_t i = 0; i < filterScales->size(); ++i) {
layerParams.blobs[2].at<float>(i) = inpScale * filterScales->Get(i) / outScale;
}
}
if (hasBias) {
Mat bias = layerParams.blobs[1].reshape(1, oc);
Mat weights_2d = layerParams.blobs[0].reshape(1, oc);
for (int i = 0; i < oc; i++)
{
bias.at<int>(i) -= inpZero * (cv::sum(weights_2d.row(i))[0]);
}
}
}
addLayer(layerParams, op);
parseFusedActivation(op, options->fused_activation_function());
}
void TFLiteImporter::parseDWConvolution(const Operator& op, const std::string& opcode, LayerParams& layerParams) {
layerParams.type = "Convolution";
auto options = reinterpret_cast<const DepthwiseConv2DOptions*>(op.builtin_options());
if (options->fused_activation_function() != ActivationFunctionType_NONE) {
CV_Error(Error::StsNotImplemented, "Depthwise convolution with fused activation");
}
layerParams.set("pad_mode", EnumNamePadding(options->padding()));
layerParams.set("stride_w", options->stride_w());
layerParams.set("stride_h", options->stride_h());
@ -372,13 +430,51 @@ void TFLiteImporter::parseDWConvolution(const Operator& op, const std::string& o
layerParams.set("num_output", oc);
layerParams.set("group", oc);
filter = allTensors[filterIdx] = filter.reshape(1, {oc, 1, kh, kw});
cv::transpose(filter.reshape(1, kh * kw).clone(), filter.reshape(1, oc));
bool isInt8 = filter.depth() == CV_8S;
// Fill convolutions blobs here because of two reasons:
// 1. Kernel transposition
// 2. Extra blob with kernel scales in case of INT8 mode
bool hasBias = op.inputs()->size() > 2;
layerParams.blobs.resize(1 + (int)hasBias + (int)isInt8);
if (hasBias) {
Mat bias = allTensors[op.inputs()->Get(2)];
layerParams.blobs[1] = bias.u ? bias : bias.clone();
}
transposeND(filter, {3, 0, 1, 2}, layerParams.blobs[0]);
if (isInt8) {
float inpScale, outScale;
int inpZero, outZero;
getQuantParams(op, inpScale, inpZero, outScale, outZero);
layerParams.blobs[2] = Mat(oc, 1, CV_32F);
auto filterScales = modelTensors->Get(filterIdx)->quantization()->scale();
if (filterScales->size() == 1) {
layerParams.blobs[2].setTo(inpScale * filterScales->Get(0) / outScale);
} else {
for (size_t i = 0; i < filterScales->size(); ++i) {
layerParams.blobs[2].at<float>(i) = inpScale * filterScales->Get(i) / outScale;
}
}
if (hasBias) {
Mat bias = layerParams.blobs[1].reshape(1, oc);
Mat weights_2d = layerParams.blobs[0].reshape(1, oc);
for (int i = 0; i < oc; i++)
{
bias.at<int>(i) -= inpZero * (cv::sum(weights_2d.row(i))[0]);
}
}
}
addLayer(layerParams, op);
parseFusedActivation(op, options->fused_activation_function());
}
void TFLiteImporter::parsePadding(const Operator& op, const std::string& opcode, LayerParams& layerParams) {
layerParams.type = "Padding";
Mat paddings = allTensors[op.inputs()->Get(1)];
Mat paddings = allTensors[op.inputs()->Get(1)].clone();
CV_CheckTypeEQ(paddings.type(), CV_32S, "");
// N H W C
@ -393,43 +489,60 @@ void TFLiteImporter::parsePadding(const Operator& op, const std::string& opcode,
// 0 1 2 3 4 5 6 7
layerParams.set("paddings", DictValue::arrayInt<int32_t*>((int32_t*)paddings.data, paddings.total()));
addLayer(layerParams, op);
}
void TFLiteImporter::parseEltwise(const Operator& op, const std::string& opcode, LayerParams& layerParams) {
if (opcode == "PRELU") {
layerParams.type = "PReLU";
} else if (opcode == "RELU") {
layerParams.type = "ReLU";
} else if (opcode == "ADD") {
ActivationFunctionType activ = ActivationFunctionType_NONE;
layerParams.type = "Eltwise";
if (opcode == "ADD") {
auto options = reinterpret_cast<const AddOptions*>(op.builtin_options());
if (options->fused_activation_function() != ActivationFunctionType_NONE) {
CV_Error(Error::StsNotImplemented, "Add with fused activation");
}
layerParams.type = "Eltwise";
activ = options->fused_activation_function();
layerParams.set("operation", "sum");
} else if (opcode == "MUL") {
}
else if (opcode == "MUL") {
auto options = reinterpret_cast<const MulOptions*>(op.builtin_options());
if (options->fused_activation_function() != ActivationFunctionType_NONE) {
CV_Error(Error::StsNotImplemented, "Mul with fused activation");
}
layerParams.type = "Eltwise";
activ = options->fused_activation_function();
layerParams.set("operation", "prod");
} else if (opcode == "HARD_SWISH") {
layerParams.type = "HardSwish";
} else if (opcode == "LOGISTIC") {
layerParams.type = "Sigmoid";
} else {
CV_Error(Error::StsNotImplemented, "Unknown eltwise operator opcode: " + opcode);
CV_Error(Error::StsNotImplemented, "Unknown opcode for Eltwise layer: " + opcode);
}
if (isInt8(op)) {
const Tensor* out = modelTensors->Get(op.outputs()->Get(0));
float outScale = out->quantization()->scale()->Get(0);
int outZero = out->quantization()->zero_point()->Get(0);
const size_t numInps = op.inputs()->size();
std::vector<float> inputScales(numInps);
std::vector<int> inputZeros(numInps);
std::vector<float> coeffs(numInps);
float offset = outZero;
for (int i = 0; i < numInps; ++i) {
const Tensor* inp = modelTensors->Get(op.inputs()->Get(i));
float inpScale = inp->quantization()->scale()->Get(0);
int inpZero = inp->quantization()->zero_point()->Get(0);
inputScales[i] = inpScale;
inputZeros[i] = inpZero;
coeffs[i] = inpScale / outScale;
offset -= coeffs[i] * inpZero;
}
layerParams.set("input_scales", DictValue::arrayReal(inputScales.data(), numInps));
layerParams.set("input_zeropoints", DictValue::arrayInt(inputZeros.data(), numInps));
layerParams.set("coeff", DictValue::arrayReal(coeffs.data(), numInps));
layerParams.set("offset", offset);
layerParams.set("scales", outScale);
layerParams.set("zeropoints", outZero);
}
addLayer(layerParams, op);
parseFusedActivation(op, activ);
}
void TFLiteImporter::parsePooling(const Operator& op, const std::string& opcode, LayerParams& layerParams) {
layerParams.type = "Pooling";
auto options = reinterpret_cast<const Pool2DOptions*>(op.builtin_options());
if (options->fused_activation_function() != ActivationFunctionType_NONE) {
CV_Error(Error::StsNotImplemented, "Pooling with fused activation");
}
layerParams.set("pad_mode", EnumNamePadding(options->padding()));
layerParams.set("stride_w", options->stride_w());
layerParams.set("stride_h", options->stride_h());
@ -441,6 +554,8 @@ void TFLiteImporter::parsePooling(const Operator& op, const std::string& opcode,
layerParams.set("pool", "ave");
else
CV_Error(Error::StsNotImplemented, "Pool type selection for " + opcode);
addLayer(layerParams, op);
parseFusedActivation(op, options->fused_activation_function());
}
void TFLiteImporter::parsePoolingWithArgmax(const Operator& op, const std::string& opcode, LayerParams& layerParams) {
@ -458,6 +573,7 @@ void TFLiteImporter::parsePoolingWithArgmax(const Operator& op, const std::strin
layerParams.set("kernel_w", params->filter_width);
layerParams.set("kernel_h", params->filter_height);
layerParams.set("pool", "max");
addLayer(layerParams, op);
}
void TFLiteImporter::parseUnpooling(const Operator& op, const std::string& opcode, LayerParams& layerParams) {
@ -474,6 +590,7 @@ void TFLiteImporter::parseUnpooling(const Operator& op, const std::string& opcod
layerParams.set("pool_k_h", params->filter_height);
layerParams.set("pool_pad_w", 0);
layerParams.set("pool_pad_h", 0);
addLayer(layerParams, op);
}
void TFLiteImporter::parseReshape(const Operator& op, const std::string& opcode, LayerParams& layerParams) {
@ -481,24 +598,30 @@ void TFLiteImporter::parseReshape(const Operator& op, const std::string& opcode,
if (inpLayout == DNN_LAYOUT_NHWC) {
// Permute to NCHW
int permId = addPermuteLayer({0, 2, 3, 1}, layerParams.name + "/permute", layerIds[op.inputs()->Get(0)]); // NCHW -> NHWC
std::vector<int> order = {0, 2, 3, 1};
const std::string name = layerParams.name + "/permute";
auto inpId = layerIds[op.inputs()->Get(0)];
int permId = addPermuteLayer(order, name, inpId, isInt8(op) ? CV_8S : CV_32F); // NCHW -> NHWC
layerIds[op.inputs()->Get(0)] = std::make_pair(permId, 0);
layouts[op.outputs()->Get(0)] = DNN_LAYOUT_NCHW;
}
layerParams.type = "Reshape";
auto options = reinterpret_cast<const ReshapeOptions*>(op.builtin_options());
std::vector<int> shape(options->new_shape()->begin(), options->new_shape()->end());
// std::swap(shape[1], shape[2]);
std::vector<int> shape;
if (op.inputs()->size() > 1) {
shape = allTensors[op.inputs()->Get(1)];
} else {
auto options = op.builtin_options_as_ReshapeOptions();
CV_Assert(options);
shape.assign(options->new_shape()->begin(), options->new_shape()->end());
}
layerParams.set("dim", DictValue::arrayInt<int*>(shape.data(), shape.size()));
addLayer(layerParams, op);
}
void TFLiteImporter::parseConcat(const Operator& op, const std::string& opcode, LayerParams& layerParams) {
layerParams.type = "Concat";
auto options = reinterpret_cast<const ConcatenationOptions*>(op.builtin_options());
if (options->fused_activation_function() != ActivationFunctionType_NONE) {
CV_Error(Error::StsNotImplemented, "Concat with fused activation");
}
int axis = options->axis();
DataLayout inpLayout = layouts[op.inputs()->Get(0)];
@ -509,28 +632,36 @@ void TFLiteImporter::parseConcat(const Operator& op, const std::string& opcode,
axis = remap[axis];
}
layerParams.set("axis", axis);
addLayer(layerParams, op);
parseFusedActivation(op, options->fused_activation_function());
}
void TFLiteImporter::parseResize(const Operator& op, const std::string& opcode, LayerParams& layerParams) {
layerParams.type = "Resize";
auto options = reinterpret_cast<const ResizeBilinearOptions*>(op.builtin_options());
if (opcode == "RESIZE_BILINEAR") {
auto options = op.builtin_options_as_ResizeBilinearOptions();
layerParams.set("interpolation", "bilinear");
layerParams.set("align_corners", options->align_corners());
layerParams.set("half_pixel_centers", options->half_pixel_centers());
} else if (opcode == "RESIZE_NEAREST_NEIGHBOR") {
auto options = op.builtin_options_as_ResizeNearestNeighborOptions();
layerParams.set("interpolation", "nearest");
layerParams.set("align_corners", options->align_corners());
layerParams.set("half_pixel_centers", options->half_pixel_centers());
}
Mat shape = allTensors[op.inputs()->Get(1)].reshape(1, 1);
layerParams.set("height", shape.at<int>(0, 0));
layerParams.set("width", shape.at<int>(0, 1));
addLayer(layerParams, op);
}
int TFLiteImporter::addPermuteLayer(const std::vector<int>& order, const std::string& permName,
const std::pair<int, int>& inpId)
const std::pair<int, int>& inpId, int dtype)
{
LayerParams permLP;
permLP.set("order", DictValue::arrayInt<const int*>(order.data(), order.size()));
int permId = dstNet.addLayer(permName, "Permute", permLP);
int permId = dstNet.addLayer(permName, "Permute", dtype, permLP);
dstNet.connect(inpId.first, inpId.second, permId, 0);
return permId;
}
@ -593,6 +724,192 @@ void TFLiteImporter::parseDeconvolution(const Operator& op, const std::string& o
}
}
}
addLayer(layerParams, op);
}
void TFLiteImporter::parseQuantize(const Operator& op, const std::string& opcode, LayerParams& layerParams) {
layerParams.type = "Quantize";
layerParams.set("scales", 1);
layerParams.set("zeropoints", -128);
addLayer(layerParams, op);
}
void TFLiteImporter::parseDequantize(const Operator& op, const std::string& opcode, LayerParams& layerParams) {
layerParams.type = "Dequantize";
float inpScale, outScale;
int inpZero, outZero;
getQuantParams(op, inpScale, inpZero, outScale, outZero);
layerParams.set("scales", inpScale);
layerParams.set("zeropoints", inpZero);
addLayer(layerParams, op);
}
void TFLiteImporter::parseDetectionPostProcess(const Operator& op, const std::string& opcode, LayerParams& layerParams) {
// Parse parameters;
std::vector<std::string> keys(1, "");
const uint8_t* data = op.custom_options()->Data();
int offset = 0;
// Read zero delimited keys
while (data[offset] != 10 && offset < op.custom_options()->size()) {
if (data[offset]) {
keys.back() += data[offset];
} else {
keys.emplace_back("");
}
offset += 1;
}
keys.pop_back();
std::sort(keys.begin(), keys.end());
// TODO: Replace empirical offset to something more reliable.
offset += 25;
std::map<std::string, uint32_t> parameters;
for (int i = 0; i < keys.size(); ++i) {
parameters[keys[i]] = *reinterpret_cast<const uint32_t*>(data + offset + i * 4);
}
layerParams.type = "DetectionOutput";
layerParams.set("num_classes", parameters["num_classes"]);
layerParams.set("share_location", true);
layerParams.set("background_label_id", parameters["num_classes"] + 1);
layerParams.set("nms_threshold", *(float*)&parameters["nms_iou_threshold"]);
layerParams.set("confidence_threshold", *(float*)&parameters["nms_score_threshold"]);
layerParams.set("top_k", parameters["max_detections"]);
layerParams.set("keep_top_k", parameters["max_detections"]);
layerParams.set("code_type", "CENTER_SIZE");
layerParams.set("variance_encoded_in_target", true);
layerParams.set("loc_pred_transposed", true);
// Replace third input from tensor to Const layer with the priors
Mat priors = allTensors[op.inputs()->Get(2)].clone();
// Change priors data from (ycenter, xcenter, h, w) to (xmin, ymin, xmax, ymax)
priors = priors.reshape(1, priors.total() / 4);
Mat tmp = priors.col(0).clone();
priors.col(0) = priors.col(1) - 0.5 * priors.col(3);
priors.col(1) = tmp - 0.5 * priors.col(2);
tmp = priors.col(2).clone();
priors.col(2) = priors.col(0) + priors.col(3);
priors.col(3) = priors.col(1) + tmp;
LayerParams priorsLP;
priorsLP.name = layerParams.name + "/priors";
priorsLP.type = "Const";
priorsLP.blobs.resize(1, priors.reshape(1, {1, 1, (int)priors.total()}));
int priorsId = dstNet.addLayer(priorsLP.name, priorsLP.type, priorsLP);
layerIds[op.inputs()->Get(2)] = std::make_pair(priorsId, 0);
addLayer(layerParams, op);
}
void TFLiteImporter::parseFusedActivation(const Operator& op, ActivationFunctionType activ) {
LayerParams activParams;
activParams.name = modelTensors->Get(op.outputs()->Get(0))->name()->str() + "/activ";
parseActivation(op, EnumNameActivationFunctionType(activ), activParams, true);
}
void TFLiteImporter::parseActivation(const Operator& op, const std::string& opcode, LayerParams& activParams) {
parseActivation(op, opcode, activParams, false);
}
void TFLiteImporter::parseActivation(const Operator& op, const std::string& opcode, LayerParams& activParams, bool isFused) {
if (opcode == "NONE")
return;
else if (opcode == "RELU6")
activParams.type = "ReLU6";
else if (opcode == "PRELU")
activParams.type = "PReLU";
else if (opcode == "RELU")
activParams.type = "ReLU";
else if (opcode == "HARD_SWISH")
activParams.type = "HardSwish";
else if (opcode == "LOGISTIC")
activParams.type = "Sigmoid";
else
CV_Error(Error::StsNotImplemented, "Unsupported activation " + opcode);
if (isInt8(op)) {
float inpScale, outScale;
int inpZero, outZero;
getQuantParams(op, inpScale, inpZero, outScale, outZero);
if (isFused) {
activParams.type += "Int8";
activParams.set("input_scale", outScale);
activParams.set("input_zeropoint", outZero);
activParams.set("scales", outScale);
activParams.set("zeropoints", outZero);
}
Mat lookUpTable(1, 256, CV_8S);
int8_t* table = lookUpTable.ptr<int8_t>();
for (int i = -128; i < 128; i++) {
float x, y = i;
if (isFused)
x = outScale * (i - outZero);
else
x = inpScale * (i - inpZero);
if (opcode == "RELU6")
y = std::min(std::max(x, 0.f), 6.f);
else if (opcode == "LOGISTIC")
y = 1.0f / (1.0f + std::exp(-x));
else
CV_Error(Error::StsNotImplemented, "Lookup table for " + opcode);
int quantized = outZero + cvRound(y / outScale);
table[i + 128] = saturate_cast<int8_t>(quantized);
}
activParams.blobs.resize(1, lookUpTable);
}
if (isFused) {
int dtype = isInt8(op) ? CV_8S : CV_32F;
int layerId = dstNet.addLayerToPrev(activParams.name, activParams.type, dtype, activParams);
// Override layer ids mapping
int i = 0;
for (int idx : *op.outputs()) {
layerIds[idx] = std::make_pair(layerId, i++);
}
} else {
addLayer(activParams, op);
}
}
bool TFLiteImporter::isInt8(const Operator& op) {
const Tensor* out = modelTensors->Get(op.outputs()->Get(0));
return out->type() == TensorType_INT8;
}
void TFLiteImporter::getQuantParams(const Operator& op, float& inpScale, int& inpZero, float& outScale, int& outZero) {
const Tensor* inp = modelTensors->Get(op.inputs()->Get(0));
const Tensor* out = modelTensors->Get(op.outputs()->Get(0));
inpScale = outScale = inpZero = outZero = 0;
if (inp->quantization()) {
if (inp->quantization()->scale()) {
CV_Assert(inp->quantization()->scale()->size() == 1);
inpScale = inp->quantization()->scale()->Get(0);
}
if (inp->quantization()->zero_point()) {
CV_Assert(inp->quantization()->zero_point()->size() == 1);
inpZero = inp->quantization()->zero_point()->Get(0);
}
}
if (out->quantization()) {
if (out->quantization()->scale()) {
CV_Assert(out->quantization()->scale()->size() == 1);
outScale = out->quantization()->scale()->Get(0);
}
if (out->quantization()->zero_point()) {
CV_Assert(out->quantization()->zero_point()->size() == 1);
outZero = out->quantization()->zero_point()->Get(0);
}
}
}
Net readNetFromTFLite(const String &modelPath) {

16
modules/dnn/test/test_tflite_importer.cpp
View File

@ -130,6 +130,22 @@ TEST(Test_TFLite, max_unpooling)
}
}
TEST(Test_TFLite, EfficientDet_int8) {
Net net = readNet(findDataFile("dnn/tflite/coco_efficientdet_lite0_v1_1.0_quant_2021_09_06.tflite", false));
Mat img = imread(findDataFile("dnn/dog416.png"));
Mat blob = blobFromImage(img, 1.0, Size(320, 320));
net.setInput(blob);
Mat out = net.forward();
Mat_<float> ref({3, 7}, {
0, 7, 0.62890625, 0.6014542579650879, 0.13300055265426636, 0.8977657556533813, 0.292389452457428,
0, 17, 0.56640625, 0.15983937680721283, 0.35905322432518005, 0.5155506730079651, 0.9409466981887817,
0, 1, 0.5, 0.14357104897499084, 0.2240825891494751, 0.7183101177215576, 0.9140362739562988
});
normAssertDetections(ref, out, "", 0.5, 0.05, 0.1);
}
}} // namespace
#endif // OPENCV_TEST_DNN_TFLITE