Deployment

We provide deployment tools under tools/deployment directory.

Convert to ONNX (experimental)

We provide a script to convert the model to ONNX format. The converted model could be visualized by tools like Netron. Besides, we also support comparing the output results between PyTorch and ONNX model.

python tools/deployment/pytorch2onnx.py
    ${MODEL_CONFIG_PATH} \
    ${MODEL_CKPT_PATH} \
    ${MODEL_TYPE} \
    ${IMAGE_PATH} \
    --output-file ${OUTPUT_FILE} \
    --device-id ${DEVICE_ID} \
    --opset-version ${OPSET_VERSION} \
    --verify \
    --verbose \
    --show \
    --dynamic-export

Description of arguments:

ARGS	Type	Description
`model_config`	str	The path to a model config file.
`model_ckpt`	str	The path to a model checkpoint file.
`model_type`	'recog', 'det'	The model type of the config file.
`image_path`	str	The path to input image file.
`--output-file`	str	The path to output ONNX model. Defaults to `tmp.onnx`.
`--device-id`	int	Which GPU to use. Defaults to 0.
`--opset-version`	int	ONNX opset version. Defaults to 11.
`--verify`	bool	Determines whether to verify the correctness of an exported model. Defaults to `False`.
`--verbose`	bool	Determines whether to print the architecture of the exported model. Defaults to `False`.
`--show`	bool	Determines whether to visualize outputs of ONNXRuntime and PyTorch. Defaults to `False`.
`--dynamic-export`	bool	Determines whether to export ONNX model with dynamic input and output shapes. Defaults to `False`.

:::{note} This tool is still experimental. For now, some customized operators are not supported, and we only support a subset of detection and recognition algorithms. :::

List of supported models exportable to ONNX

The table below lists the models that are guaranteed to be exportable to ONNX and runnable in ONNX Runtime.

Model	Config	Dynamic Shape	Batch Inference	Note
DBNet	dbnet_r18_fpnc_1200e_icdar2015.py	Y	N
PSENet	psenet_r50_fpnf_600e_ctw1500.py	Y	Y
PSENet	psenet_r50_fpnf_600e_icdar2015.py	Y	Y
PANet	panet_r18_fpem_ffm_600e_ctw1500.py	Y	Y
PANet	panet_r18_fpem_ffm_600e_icdar2015.py	Y	Y
CRNN	crnn_academic_dataset.py	Y	Y	CRNN only accepts input with height 32

:::{note}

All models above are tested with PyTorch==1.8.1 and onnxruntime-gpu == 1.8.1
If you meet any problem with the listed models above, please create an issue and it would be taken care of soon.
Because this feature is experimental and may change fast, please always try with the latest mmcv and mmocr. :::

Convert ONNX to TensorRT (experimental)

We also provide a script to convert ONNX model to TensorRT format. Besides, we support comparing the output results between ONNX and TensorRT model.

python tools/deployment/onnx2tensorrt.py
    ${MODEL_CONFIG_PATH} \
    ${MODEL_TYPE} \
    ${IMAGE_PATH} \
    ${ONNX_FILE} \
    --trt-file ${OUT_TENSORRT} \
    --max-shape INT INT INT INT \
    --min-shape INT INT INT INT \
    --workspace-size INT \
    --fp16 \
    --verify \
    --show \
    --verbose

Description of arguments:

ARGS	Type	Description
`model_config`	str	The path to a model config file.
`model_type`	'recog', 'det'	The model type of the config file.
`image_path`	str	The path to input image file.
`onnx_file`	str	The path to input ONNX file.
`--trt-file`	str	The path of output TensorRT model. Defaults to `tmp.trt`.
`--max-shape`	int * 4	Maximum shape of model input.
`--min-shape`	int * 4	Minimum shape of model input.
`--workspace-size`	int	Max workspace size in GiB. Defaults to 1.
`--fp16`	bool	Determines whether to export TensorRT with fp16 mode. Defaults to `False`.
`--verify`	bool	Determines whether to verify the correctness of an exported model. Defaults to `False`.
`--show`	bool	Determines whether to show the output of ONNX and TensorRT. Defaults to `False`.
`--verbose`	bool	Determines whether to verbose logging messages while creating TensorRT engine. Defaults to `False`.

:::{note} This tool is still experimental. For now, some customized operators are not supported, and we only support a subset of detection and recognition algorithms. :::

List of supported models exportable to TensorRT

The table below lists the models that are guaranteed to be exportable to TensorRT engine and runnable in TensorRT.

Model	Config	Dynamic Shape	Batch Inference	Note
DBNet	dbnet_r18_fpnc_1200e_icdar2015.py	Y	N
PSENet	psenet_r50_fpnf_600e_ctw1500.py	Y	Y
PSENet	psenet_r50_fpnf_600e_icdar2015.py	Y	Y
PANet	panet_r18_fpem_ffm_600e_ctw1500.py	Y	Y
PANet	panet_r18_fpem_ffm_600e_icdar2015.py	Y	Y
CRNN	crnn_academic_dataset.py	Y	Y	CRNN only accepts input with height 32

:::{note}

All models above are tested with PyTorch==1.8.1, onnxruntime-gpu==1.8.1 and tensorrt==7.2.1.6
If you meet any problem with the listed models above, please create an issue and it would be taken care of soon.
Because this feature is experimental and may change fast, please always try with the latest mmcv and mmocr. :::

Evaluate ONNX and TensorRT Models (experimental)

We provide methods to evaluate TensorRT and ONNX models in tools/deployment/deploy_test.py.

Prerequisite

To evaluate ONNX and TensorRT models, ONNX, ONNXRuntime and TensorRT should be installed first. Install mmcv-full with ONNXRuntime custom ops and TensorRT plugins follow ONNXRuntime in mmcv and TensorRT plugin in mmcv.

Usage

python tools/deploy_test.py \
    ${CONFIG_FILE} \
    ${MODEL_PATH} \
    ${MODEL_TYPE} \
    ${BACKEND} \
    --eval ${METRICS} \
    --device ${DEVICE}

Description of all arguments

ARGS	Type	Description
`model_config`	str	The path to a model config file.
`model_file`	str	The path to a TensorRT or an ONNX model file.
`model_type`	'recog', 'det'	Detection or recognition model to deploy.
`backend`	'TensorRT', 'ONNXRuntime'	The backend for testing.
`--eval`	'acc', 'hmean-iou'	The evaluation metrics. 'acc' for recognition models, 'hmean-iou' for detection models.
`--device`	str	Device for evaluation. Defaults to `cuda:0`.

Results and Models

Model	Config	Dataset	Metric	PyTorch	ONNX Runtime	TensorRT FP32	TensorRT FP16
DBNet	dbnet_r18_fpnc_1200e_icdar2015.py	icdar2015	Recall	0.731	0.731	0.678	0.679
			Precision	0.871	0.871	0.844	0.842
			Hmean	0.795	0.795	0.752	0.752
DBNet*	dbnet_r18_fpnc_1200e_icdar2015.py	icdar2015	Recall	0.720	0.720	0.720	0.718
			Precision	0.868	0.868	0.868	0.868
			Hmean	0.787	0.787	0.787	0.786
PSENet	psenet_r50_fpnf_600e_icdar2015.py	icdar2015	Recall	0.753	0.753	0.753	0.752
			Precision	0.867	0.867	0.867	0.867
			Hmean	0.806	0.806	0.806	0.805
PANet	panet_r18_fpem_ffm_600e_icdar2015.py	icdar2015	Recall	0.740	0.740	0.687	N/A
			Precision	0.860	0.860	0.815	N/A
			Hmean	0.796	0.796	0.746	N/A
PANet*	panet_r18_fpem_ffm_600e_icdar2015.py	icdar2015	Recall	0.736	0.736	0.736	N/A
			Precision	0.857	0.857	0.857	N/A
			Hmean	0.792	0.792	0.792	N/A
CRNN	crnn_academic_dataset.py	IIIT5K	Acc	0.806	0.806	0.806	0.806

:::{note}

TensorRT upsampling operation is a little different from PyTorch. For DBNet and PANet, we suggest replacing upsampling operations with the nearest mode to operations with bilinear mode. Here for PANet, here and here for DBNet. As is shown in the above table, networks with tag * mean the upsampling mode is changed.
Note that changing upsampling mode reduces less performance compared with using the nearest mode. However, the weights of networks are trained through the nearest mode. To pursue the best performance, using bilinear mode for both training and TensorRT deployment is recommended.
All ONNX and TensorRT models are evaluated with dynamic shapes on the datasets, and images are preprocessed according to the original config file.
This tool is still experimental, and we only support a subset of detection and recognition algorithms for now. :::

C++ Inference example with OpenCV

The example below is tested with Visual Studio 2019 as console application, CPU inference only.

Prerequisites

Project should use OpenCV (tested with version 4.5.4), ONNX Runtime NuGet package (version 1.9.0).
Download DBNet_r18 detector and SATRN_small recognizer models from our Model Zoo, and export them with the following python commands (you may change the paths accordingly):

python3.9 ../mmocr/tools/deployment/pytorch2onnx.py --verify --output-file detector.onnx ../mmocr/configs/textdet/dbnet/dbnet_r18_fpnc_1200e_icdar2015.py ./dbnet_r18_fpnc_sbn_1200e_icdar2015_20210329-ba3ab597.pth --dynamic-export det ./sample_big_image_eg_1920x1080.png

python3.9 ../mmocr/tools/deployment/pytorch2onnx.py --opset 14 --verify --output-file recognizer.onnx ../mmocr/configs/textrecog/satrn/satrn_small.py ./satrn_small_20211009-2cf13355.pth recog ./sample_small_image_eg_200x50.png

:::{note}

Be aware, while exported detector.onnx file is relatively small (about 50 Mb), recognizer.onnx is pretty big (more than 600 Mb).
DBNet_r18 can use ONNX opset 11, SATRN_small can be exported with opset 14. :::

:::{warning} Be sure, that verifications of both models are successful - look through the export messages. :::

Example

Example usage of exported models with C++ is in the code below (don't forget to change paths to *.onnx files). It's applicable to these two models only, other models have another preprocessing and postprocessing logics.

#include <iostream>

#include <opencv2/core/core.hpp>
#include <opencv2/highgui.hpp>
#include <opencv2/imgproc.hpp>
#include <opencv2/dnn.hpp>

#include <onnxruntime_cxx_api.h>
#pragma comment(lib, "onnxruntime.lib")

// DB_r18
class Detector {
public:
    Detector(const std::string& model_path) {
        session = Ort::Session{ env, std::wstring(model_path.begin(), model_path.end()).c_str(), Ort::SessionOptions{nullptr} };
    }

    std::vector<cv::Rect> inference(const cv::Mat& original, float threshold = 0.3f) {

        cv::Size original_size = original.size();

        const char* input_names[] = { "input" };
        const char* output_names[] = { "output" };

        std::array<int64_t, 4> input_shape{ 1, 3, height, width };

        cv::Mat image = cv::Mat::zeros(cv::Size(width, height), original.type());
        cv::resize(original, image, cv::Size(width, height), 0, 0, cv::INTER_AREA);

        image.convertTo(image, CV_32FC3);

        cv::cvtColor(image, image, cv::COLOR_BGR2RGB);
        image = (image - cv::Scalar(123.675f, 116.28f, 103.53f)) / cv::Scalar(58.395f, 57.12f, 57.375f);

        cv::Mat blob = cv::dnn::blobFromImage(image);

        auto memory_info = Ort::MemoryInfo::CreateCpu(OrtDeviceAllocator, OrtMemTypeDefault);
        Ort::Value input_tensor = Ort::Value::CreateTensor<float>(memory_info, (float*)blob.data, blob.total(), input_shape.data(), input_shape.size());

        std::vector<Ort::Value> output_tensor = session.Run(Ort::RunOptions{ nullptr }, input_names, &input_tensor, 1, output_names, 1);

        int sizes[] = { 1, 3, height, width };
        cv::Mat output(4, sizes, CV_32F, output_tensor.front().GetTensorMutableData<float>());

        std::vector<cv::Mat> images;
        cv::dnn::imagesFromBlob(output, images);

        std::vector<cv::Rect> areas = get_detected(images[0], threshold);
        std::vector<cv::Rect> results;

        float x_ratio = original_size.width / (float)width;
        float y_ratio = original_size.height / (float)height;

        for (int index = 0; index < areas.size(); ++index) {
            cv::Rect box = areas[index];

            box.x = int(box.x * x_ratio);
            box.width = int(box.width * x_ratio);
            box.y = int(box.y * y_ratio);
            box.height = int(box.height * y_ratio);

            results.push_back(box);
        }

        return results;
    }

private:
    Ort::Env env;
    Ort::Session session{ nullptr };

    const int width = 1312, height = 736;

    cv::Rect expand_box(const cv::Rect& original, int addition = 5) {
        cv::Rect box(original);
        box.x = std::max(0, box.x - addition);
        box.y = std::max(0, box.y - addition);
        box.width = (box.x + box.width + addition * 2 > width) ? (width - box.x) : (box.width + addition * 2);
        box.height = (box.y + box.height + addition * 2) > height ? (height - box.y) : (box.height + addition * 2);
        return box;
    }

    std::vector<cv::Rect> get_detected(const cv::Mat& output, float threshold) {
        cv::Mat text_mask = cv::Mat::zeros(height, width, CV_32F);
        std::vector<cv::Mat> maps;
        cv::split(output, maps);
        cv::Mat proba_map = maps[0];

        cv::threshold(proba_map, text_mask, threshold, 1.0f, cv::THRESH_BINARY);
        cv::multiply(text_mask, 255, text_mask);
        text_mask.convertTo(text_mask, CV_8U);

        std::vector<std::vector<cv::Point>> contours;
        cv::findContours(text_mask, contours, cv::RETR_EXTERNAL, cv::CHAIN_APPROX_SIMPLE);
        std::vector<cv::Rect> boxes;

        for (int index = 0; index < contours.size(); ++index) {
            cv::Rect box = expand_box(cv::boundingRect(contours[index]));
            boxes.push_back(box);
        }

        return boxes;
    }
};

// SATRN_small
class Recognizer {
public:
    Recognizer(const std::string& model_path) {
        session = Ort::Session{ env, std::wstring(model_path.begin(), model_path.end()).c_str(), Ort::SessionOptions{nullptr} };
    }

    std::string inference(const cv::Mat& original) {
        const char* input_names[] = { "input" };
        const char* output_names[] = { "output" };

        std::array<int64_t, 4> input_shape{ 1, 3, height, width };

        cv::Mat image;
        cv::resize(original, image, cv::Size(width, height), 0, 0, cv::INTER_AREA);
        image.convertTo(image, CV_32FC3);

        cv::cvtColor(image, image, cv::COLOR_BGR2RGB);
        image = (image / 255.0f - cv::Scalar(0.485f, 0.456f, 0.406f)) / cv::Scalar(0.229f, 0.224f, 0.225f);

        cv::Mat blob = cv::dnn::blobFromImage(image);

        auto memory_info = Ort::MemoryInfo::CreateCpu(OrtDeviceAllocator, OrtMemTypeDefault);
        Ort::Value input_tensor = Ort::Value::CreateTensor<float>(memory_info, (float*)blob.data, blob.total(), input_shape.data(), input_shape.size());

        std::vector<Ort::Value> output_tensor = session.Run(Ort::RunOptions{ nullptr }, input_names, &input_tensor, 1, output_names, 1);

        int sequence_length = 25;
        std::string dictionary = "0123456789abcdefghijklmnopqrstuvwxyzABCDEFGHIJKLMNOPQRSTUVWXYZ!\"#$%&'()*+,-./:;<=>?@[\\]_`~";
        int characters = dictionary.length() + 2; // EOS + UNK

        std::vector<int> max_indices;
        for (int outer = 0; outer < sequence_length; ++outer) {
            int character_index = -1;
            float character_value = 0;
            for (int inner = 0; inner < characters; ++inner) {
                int counter = outer * characters + inner;
                float value = output_tensor[0].GetTensorMutableData<float>()[counter];
                if (value > character_value) {
                    character_value = value;
                    character_index = inner;
                }
            }
            max_indices.push_back(character_index);
        }

        std::string recognized;

        for (int index = 0; index < max_indices.size(); ++index) {
            if (max_indices[index] == dictionary.length()) {
                continue; // unk
            }
            if (max_indices[index] == dictionary.length() + 1) {
                break; // eos
            }
            recognized += dictionary[max_indices[index]];
        }

        return recognized;
    }

private:
    Ort::Env env;
    Ort::Session session{ nullptr };

    const int height = 32;
    const int width = 100;
};

int main(int argc, const char* argv[]) {
    if (argc < 2) {
        std::cout << "Usage: this_executable.exe c:/path/to/image.png" << std::endl;
        return 0;
    }

    std::chrono::steady_clock::time_point begin = std::chrono::steady_clock::now();
    std::cout << "Loading models..." << std::endl;

    Detector detector("d:/path/to/detector.onnx");
    Recognizer recognizer("d:/path/to/recognizer.onnx");

    std::chrono::steady_clock::time_point end = std::chrono::steady_clock::now();
    std::cout << "Loading models done in " << std::chrono::duration_cast<std::chrono::milliseconds>(end - begin).count() << " ms" << std::endl;

    cv::Mat image = cv::imread(argv[1], cv::IMREAD_COLOR);

    begin = std::chrono::steady_clock::now();
    std::vector<cv::Rect> detections = detector.inference(image);
    for (int index = 0; index < detections.size(); ++index) {
        cv::Mat roi = image(detections[index]);
        std::string text = recognizer.inference(roi);
        cv::rectangle(image, detections[index], cv::Scalar(255, 255, 255), 2);
        cv::putText(image, text, cv::Point(detections[index].x, detections[index].y - 10), cv::FONT_HERSHEY_COMPLEX, 0.4, cv::Scalar(255, 255, 255));
    }

    end = std::chrono::steady_clock::now();
    std::cout << "Inference time (with drawing): " << std::chrono::duration_cast<std::chrono::milliseconds>(end - begin).count() << " ms" << std::endl;

    cv::imshow("Results", image);
    cv::waitKey(0);

    return 0;
}

The output should look something like this.

Loading models...
Loading models done in 5715 ms
Inference time (with drawing): 3349 ms

And the sample result should look something like this.