Skip to content

GitLab

Commit 19eb7eb8 authored by Leif's avatar Leif
Browse files

Merge remote-tracking branch 'origin/dygraph' into dy1

parents 0afe6c32 03b7daa5
Hide whitespace changes
Inline Side-by-side
Showing with 2296 additions and 139 deletions
deploy/lite/db_post_process.cc 0 → 100644
View file @ 19eb7eb8
// Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#include "db_post_process.h" // NOLINT
#include <algorithm>
#include <utility>
void GetContourArea(std::vector<std::vector<float>> box, float unclip_ratio,
float &distance) {
int pts_num = 4;
float area = 0.0f;
float dist = 0.0f;
for (int i = 0; i < pts_num; i++) {
area += box[i][0] * box[(i + 1) % pts_num][1] -
box[i][1] * box[(i + 1) % pts_num][0];
dist += sqrtf((box[i][0] - box[(i + 1) % pts_num][0]) *
(box[i][0] - box[(i + 1) % pts_num][0]) +
(box[i][1] - box[(i + 1) % pts_num][1]) *
(box[i][1] - box[(i + 1) % pts_num][1]));
}
area = fabs(float(area / 2.0));
distance = area * unclip_ratio / dist;
}
cv::RotatedRect Unclip(std::vector<std::vector<float>> box,
float unclip_ratio) {
float distance = 1.0;
GetContourArea(box, unclip_ratio, distance);
ClipperLib::ClipperOffset offset;
ClipperLib::Path p;
p << ClipperLib::IntPoint(static_cast<int>(box[0][0]),
static_cast<int>(box[0][1]))
<< ClipperLib::IntPoint(static_cast<int>(box[1][0]),
static_cast<int>(box[1][1]))
<< ClipperLib::IntPoint(static_cast<int>(box[2][0]),
static_cast<int>(box[2][1]))
<< ClipperLib::IntPoint(static_cast<int>(box[3][0]),
static_cast<int>(box[3][1]));
offset.AddPath(p, ClipperLib::jtRound, ClipperLib::etClosedPolygon);
ClipperLib::Paths soln;
offset.Execute(soln, distance);
std::vector<cv::Point2f> points;
for (int j = 0; j < soln.size(); j++) {
for (int i = 0; i < soln[soln.size() - 1].size(); i++) {
points.emplace_back(soln[j][i].X, soln[j][i].Y);
}
}
cv::RotatedRect res = cv::minAreaRect(points);
return res;
}
std::vector<std::vector<float>> Mat2Vector(cv::Mat mat) {
std::vector<std::vector<float>> img_vec;
std::vector<float> tmp;
for (int i = 0; i < mat.rows; ++i) {
tmp.clear();
for (int j = 0; j < mat.cols; ++j) {
tmp.push_back(mat.at<float>(i, j));
}
img_vec.push_back(tmp);
}
return img_vec;
}
bool XsortFp32(std::vector<float> a, std::vector<float> b) {
if (a[0] != b[0])
return a[0] < b[0];
return false;
}
bool XsortInt(std::vector<int> a, std::vector<int> b) {
if (a[0] != b[0])
return a[0] < b[0];
return false;
}
std::vector<std::vector<int>>
OrderPointsClockwise(std::vector<std::vector<int>> pts) {
std::vector<std::vector<int>> box = pts;
std::sort(box.begin(), box.end(), XsortInt);
std::vector<std::vector<int>> leftmost = {box[0], box[1]};
std::vector<std::vector<int>> rightmost = {box[2], box[3]};
if (leftmost[0][1] > leftmost[1][1])
std::swap(leftmost[0], leftmost[1]);
if (rightmost[0][1] > rightmost[1][1])
std::swap(rightmost[0], rightmost[1]);
std::vector<std::vector<int>> rect = {leftmost[0], rightmost[0], rightmost[1],
leftmost[1]};
return rect;
}
std::vector<std::vector<float>> GetMiniBoxes(cv::RotatedRect box, float &ssid) {
ssid = std::min(box.size.width, box.size.height);
cv::Mat points;
cv::boxPoints(box, points);
auto array = Mat2Vector(points);
std::sort(array.begin(), array.end(), XsortFp32);
std::vector<float> idx1 = array[0], idx2 = array[1], idx3 = array[2],
idx4 = array[3];
if (array[3][1] <= array[2][1]) {
idx2 = array[3];
idx3 = array[2];
} else {
idx2 = array[2];
idx3 = array[3];
}
if (array[1][1] <= array[0][1]) {
idx1 = array[1];
idx4 = array[0];
} else {
idx1 = array[0];
idx4 = array[1];
}
array[0] = idx1;
array[1] = idx2;
array[2] = idx3;
array[3] = idx4;
return array;
}
float BoxScoreFast(std::vector<std::vector<float>> box_array, cv::Mat pred) {
auto array = box_array;
int width = pred.cols;
int height = pred.rows;
float box_x[4] = {array[0][0], array[1][0], array[2][0], array[3][0]};
float box_y[4] = {array[0][1], array[1][1], array[2][1], array[3][1]};
int xmin = clamp(
static_cast<int>(std::floorf(*(std::min_element(box_x, box_x + 4)))), 0,
width - 1);
int xmax =
clamp(static_cast<int>(std::ceilf(*(std::max_element(box_x, box_x + 4)))),
0, width - 1);
int ymin = clamp(
static_cast<int>(std::floorf(*(std::min_element(box_y, box_y + 4)))), 0,
height - 1);
int ymax =
clamp(static_cast<int>(std::ceilf(*(std::max_element(box_y, box_y + 4)))),
0, height - 1);
cv::Mat mask;
mask = cv::Mat::zeros(ymax - ymin + 1, xmax - xmin + 1, CV_8UC1);
cv::Point root_point[4];
root_point[0] = cv::Point(static_cast<int>(array[0][0]) - xmin,
static_cast<int>(array[0][1]) - ymin);
root_point[1] = cv::Point(static_cast<int>(array[1][0]) - xmin,
static_cast<int>(array[1][1]) - ymin);
root_point[2] = cv::Point(static_cast<int>(array[2][0]) - xmin,
static_cast<int>(array[2][1]) - ymin);
root_point[3] = cv::Point(static_cast<int>(array[3][0]) - xmin,
static_cast<int>(array[3][1]) - ymin);
const cv::Point *ppt[1] = {root_point};
int npt[] = {4};
cv::fillPoly(mask, ppt, npt, 1, cv::Scalar(1));
cv::Mat croppedImg;
pred(cv::Rect(xmin, ymin, xmax - xmin + 1, ymax - ymin + 1))
.copyTo(croppedImg);
auto score = cv::mean(croppedImg, mask)[0];
return score;
}
float PolygonScoreAcc(std::vector<cv::Point> contour, cv::Mat pred) {
int width = pred.cols;
int height = pred.rows;
std::vector<float> box_x;
std::vector<float> box_y;
for (int i = 0; i < contour.size(); ++i) {
box_x.push_back(contour[i].x);
box_y.push_back(contour[i].y);
}
int xmin =
clamp(int(std::floor(*(std::min_element(box_x.begin(), box_x.end())))), 0,
width - 1);
int xmax =
clamp(int(std::ceil(*(std::max_element(box_x.begin(), box_x.end())))), 0,
width - 1);
int ymin =
clamp(int(std::floor(*(std::min_element(box_y.begin(), box_y.end())))), 0,
height - 1);
int ymax =
clamp(int(std::ceil(*(std::max_element(box_y.begin(), box_y.end())))), 0,
height - 1);
cv::Mat mask;
mask = cv::Mat::zeros(ymax - ymin + 1, xmax - xmin + 1, CV_8UC1);
cv::Point *rook_point = new cv::Point[contour.size()];
for (int i = 0; i < contour.size(); ++i) {
rook_point[i] = cv::Point(int(box_x[i]) - xmin, int(box_y[i]) - ymin);
}
const cv::Point *ppt[1] = {rook_point};
int npt[] = {int(contour.size())};
cv::fillPoly(mask, ppt, npt, 1, cv::Scalar(1));
cv::Mat croppedImg;
pred(cv::Rect(xmin, ymin, xmax - xmin + 1, ymax - ymin + 1))
.copyTo(croppedImg);
float score = cv::mean(croppedImg, mask)[0];
delete[] rook_point;
return score;
}
std::vector<std::vector<std::vector<int>>>
BoxesFromBitmap(const cv::Mat pred, const cv::Mat bitmap,
std::map<std::string, double> Config) {
const int min_size = 3;
const int max_candidates = 1000;
const float box_thresh = static_cast<float>(Config["det_db_box_thresh"]);
const float unclip_ratio = static_cast<float>(Config["det_db_unclip_ratio"]);
const int det_use_polygon_score = int(Config["det_use_polygon_score"]);
int width = bitmap.cols;
int height = bitmap.rows;
std::vector<std::vector<cv::Point>> contours;
std::vector<cv::Vec4i> hierarchy;
cv::findContours(bitmap, contours, hierarchy, cv::RETR_LIST,
cv::CHAIN_APPROX_SIMPLE);
int num_contours =
contours.size() >= max_candidates ? max_candidates : contours.size();
std::vector<std::vector<std::vector<int>>> boxes;
for (int i = 0; i < num_contours; i++) {
float ssid;
if (contours[i].size() <= 2)
continue;
cv::RotatedRect box = cv::minAreaRect(contours[i]);
auto array = GetMiniBoxes(box, ssid);
auto box_for_unclip = array;
// end get_mini_box
if (ssid < min_size) {
continue;
}
float score;
if (det_use_polygon_score) {
score = PolygonScoreAcc(contours[i], pred);
} else {
score = BoxScoreFast(array, pred);
}
// end box_score_fast
if (score < box_thresh)
continue;
// start for unclip
cv::RotatedRect points = Unclip(box_for_unclip, unclip_ratio);
if (points.size.height < 1.001 && points.size.width < 1.001)
continue;
// end for unclip
cv::RotatedRect clipbox = points;
auto cliparray = GetMiniBoxes(clipbox, ssid);
if (ssid < min_size + 2)
continue;
int dest_width = pred.cols;
int dest_height = pred.rows;
std::vector<std::vector<int>> intcliparray;
for (int num_pt = 0; num_pt < 4; num_pt++) {
std::vector<int> a{
static_cast<int>(clamp(
roundf(cliparray[num_pt][0] / float(width) * float(dest_width)),
float(0), float(dest_width))),
static_cast<int>(clamp(
roundf(cliparray[num_pt][1] / float(height) * float(dest_height)),
float(0), float(dest_height)))};
intcliparray.push_back(a);
}
boxes.push_back(intcliparray);
} // end for
return boxes;
}
std::vector<std::vector<std::vector<int>>>
FilterTagDetRes(std::vector<std::vector<std::vector<int>>> boxes, float ratio_h,
float ratio_w, cv::Mat srcimg) {
int oriimg_h = srcimg.rows;
int oriimg_w = srcimg.cols;
std::vector<std::vector<std::vector<int>>> root_points;
for (int n = 0; n < static_cast<int>(boxes.size()); n++) {
boxes[n] = OrderPointsClockwise(boxes[n]);
for (int m = 0; m < static_cast<int>(boxes[0].size()); m++) {
boxes[n][m][0] /= ratio_w;
boxes[n][m][1] /= ratio_h;
boxes[n][m][0] =
static_cast<int>(std::min(std::max(boxes[n][m][0], 0), oriimg_w - 1));
boxes[n][m][1] =
static_cast<int>(std::min(std::max(boxes[n][m][1], 0), oriimg_h - 1));
}
}
for (int n = 0; n < boxes.size(); n++) {
int rect_width, rect_height;
rect_width =
static_cast<int>(sqrt(pow(boxes[n][0][0] - boxes[n][1][0], 2) +
pow(boxes[n][0][1] - boxes[n][1][1], 2)));
rect_height =
static_cast<int>(sqrt(pow(boxes[n][0][0] - boxes[n][3][0], 2) +
pow(boxes[n][0][1] - boxes[n][3][1], 2)));
if (rect_width <= 4 || rect_height <= 4)
continue;
root_points.push_back(boxes[n]);
}
return root_points;
}
deploy/lite/db_post_process.h 0 → 100644
View file @ 19eb7eb8
// Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#pragma once
#include <math.h>
#include <iostream>
#include <map>
#include <vector>
#include "clipper.hpp"
#include "opencv2/core.hpp"
#include "opencv2/imgcodecs.hpp"
#include "opencv2/imgproc.hpp"
template <class T> T clamp(T x, T min, T max) {
if (x > max)
return max;
if (x < min)
return min;
return x;
}
std::vector<std::vector<float>> Mat2Vector(cv::Mat mat);
void GetContourArea(std::vector<std::vector<float>> box, float unclip_ratio,
float &distance);
cv::RotatedRect Unclip(std::vector<std::vector<float>> box, float unclip_ratio);
std::vector<std::vector<float>> Mat2Vector(cv::Mat mat);
bool XsortFp32(std::vector<float> a, std::vector<float> b);
bool XsortInt(std::vector<int> a, std::vector<int> b);
std::vector<std::vector<int>>
OrderPointsClockwise(std::vector<std::vector<int>> pts);
std::vector<std::vector<float>> GetMiniBoxes(cv::RotatedRect box, float &ssid);
float BoxScoreFast(std::vector<std::vector<float>> box_array, cv::Mat pred);
std::vector<std::vector<std::vector<int>>>
BoxesFromBitmap(const cv::Mat pred, const cv::Mat bitmap,
std::map<std::string, double> Config);
std::vector<std::vector<std::vector<int>>>
FilterTagDetRes(std::vector<std::vector<std::vector<int>>> boxes, float ratio_h,
float ratio_w, cv::Mat srcimg);
deploy/lite/ocr_db_crnn.cc 0 → 100644
View file @ 19eb7eb8
// Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#include "paddle_api.h" // NOLINT
#include <chrono>
#include "cls_process.h"
#include "crnn_process.h"
#include "db_post_process.h"
using namespace paddle::lite_api; // NOLINT
using namespace std;
// fill tensor with mean and scale and trans layout: nhwc -> nchw, neon speed up
void NeonMeanScale(const float *din, float *dout, int size,
const std::vector<float> mean,
const std::vector<float> scale) {
if (mean.size() != 3 || scale.size() != 3) {
std::cerr << "[ERROR] mean or scale size must equal to 3\n";
exit(1);
}
float32x4_t vmean0 = vdupq_n_f32(mean[0]);
float32x4_t vmean1 = vdupq_n_f32(mean[1]);
float32x4_t vmean2 = vdupq_n_f32(mean[2]);
float32x4_t vscale0 = vdupq_n_f32(scale[0]);
float32x4_t vscale1 = vdupq_n_f32(scale[1]);
float32x4_t vscale2 = vdupq_n_f32(scale[2]);
float *dout_c0 = dout;
float *dout_c1 = dout + size;
float *dout_c2 = dout + size * 2;
int i = 0;
for (; i < size - 3; i += 4) {
float32x4x3_t vin3 = vld3q_f32(din);
float32x4_t vsub0 = vsubq_f32(vin3.val[0], vmean0);
float32x4_t vsub1 = vsubq_f32(vin3.val[1], vmean1);
float32x4_t vsub2 = vsubq_f32(vin3.val[2], vmean2);
float32x4_t vs0 = vmulq_f32(vsub0, vscale0);
float32x4_t vs1 = vmulq_f32(vsub1, vscale1);
float32x4_t vs2 = vmulq_f32(vsub2, vscale2);
vst1q_f32(dout_c0, vs0);
vst1q_f32(dout_c1, vs1);
vst1q_f32(dout_c2, vs2);
din += 12;
dout_c0 += 4;
dout_c1 += 4;
dout_c2 += 4;
}
for (; i < size; i++) {
*(dout_c0++) = (*(din++) - mean[0]) * scale[0];
*(dout_c1++) = (*(din++) - mean[1]) * scale[1];
*(dout_c2++) = (*(din++) - mean[2]) * scale[2];
}
}
// resize image to a size multiple of 32 which is required by the network
cv::Mat DetResizeImg(const cv::Mat img, int max_size_len,
std::vector<float> &ratio_hw) {
int w = img.cols;
int h = img.rows;
float ratio = 1.f;
int max_wh = w >= h ? w : h;
if (max_wh > max_size_len) {
if (h > w) {
ratio = static_cast<float>(max_size_len) / static_cast<float>(h);
} else {
ratio = static_cast<float>(max_size_len) / static_cast<float>(w);
}
}
int resize_h = static_cast<int>(float(h) * ratio);
int resize_w = static_cast<int>(float(w) * ratio);
if (resize_h % 32 == 0)
resize_h = resize_h;
else if (resize_h / 32 < 1 + 1e-5)
resize_h = 32;
else
resize_h = (resize_h / 32 - 1) * 32;
if (resize_w % 32 == 0)
resize_w = resize_w;
else if (resize_w / 32 < 1 + 1e-5)
resize_w = 32;
else
resize_w = (resize_w / 32 - 1) * 32;
cv::Mat resize_img;
cv::resize(img, resize_img, cv::Size(resize_w, resize_h));
ratio_hw.push_back(static_cast<float>(resize_h) / static_cast<float>(h));
ratio_hw.push_back(static_cast<float>(resize_w) / static_cast<float>(w));
return resize_img;
}
cv::Mat RunClsModel(cv::Mat img, std::shared_ptr<PaddlePredictor> predictor_cls,
const float thresh = 0.9) {
std::vector<float> mean = {0.5f, 0.5f, 0.5f};
std::vector<float> scale = {1 / 0.5f, 1 / 0.5f, 1 / 0.5f};
cv::Mat srcimg;
img.copyTo(srcimg);
cv::Mat crop_img;
img.copyTo(crop_img);
cv::Mat resize_img;
int index = 0;
float wh_ratio =
static_cast<float>(crop_img.cols) / static_cast<float>(crop_img.rows);
resize_img = ClsResizeImg(crop_img);
resize_img.convertTo(resize_img, CV_32FC3, 1 / 255.f);
const float *dimg = reinterpret_cast<const float *>(resize_img.data);
std::unique_ptr<Tensor> input_tensor0(std::move(predictor_cls->GetInput(0)));
input_tensor0->Resize({1, 3, resize_img.rows, resize_img.cols});
auto *data0 = input_tensor0->mutable_data<float>();
NeonMeanScale(dimg, data0, resize_img.rows * resize_img.cols, mean, scale);
// Run CLS predictor
predictor_cls->Run();
// Get output and run postprocess
std::unique_ptr<const Tensor> softmax_out(
std::move(predictor_cls->GetOutput(0)));
auto *softmax_scores = softmax_out->mutable_data<float>();
auto softmax_out_shape = softmax_out->shape();
float score = 0;
int label = 0;
for (int i = 0; i < softmax_out_shape[1]; i++) {
if (softmax_scores[i] > score) {
score = softmax_scores[i];
label = i;
}
}
if (label % 2 == 1 && score > thresh) {
cv::rotate(srcimg, srcimg, 1);
}
return srcimg;
}
void RunRecModel(std::vector<std::vector<std::vector<int>>> boxes, cv::Mat img,
std::shared_ptr<PaddlePredictor> predictor_crnn,
std::vector<std::string> &rec_text,
std::vector<float> &rec_text_score,
std::vector<std::string> charactor_dict,
std::shared_ptr<PaddlePredictor> predictor_cls,
int use_direction_classify) {
std::vector<float> mean = {0.5f, 0.5f, 0.5f};
std::vector<float> scale = {1 / 0.5f, 1 / 0.5f, 1 / 0.5f};
cv::Mat srcimg;
img.copyTo(srcimg);
cv::Mat crop_img;
cv::Mat resize_img;
int index = 0;
for (int i = boxes.size() - 1; i >= 0; i--) {
crop_img = GetRotateCropImage(srcimg, boxes[i]);
if (use_direction_classify >= 1) {
crop_img = RunClsModel(crop_img, predictor_cls);
}
float wh_ratio =
static_cast<float>(crop_img.cols) / static_cast<float>(crop_img.rows);
resize_img = CrnnResizeImg(crop_img, wh_ratio);
resize_img.convertTo(resize_img, CV_32FC3, 1 / 255.f);
const float *dimg = reinterpret_cast<const float *>(resize_img.data);
std::unique_ptr<Tensor> input_tensor0(
std::move(predictor_crnn->GetInput(0)));
input_tensor0->Resize({1, 3, resize_img.rows, resize_img.cols});
auto *data0 = input_tensor0->mutable_data<float>();
NeonMeanScale(dimg, data0, resize_img.rows * resize_img.cols, mean, scale);
//// Run CRNN predictor
predictor_crnn->Run();
// Get output and run postprocess
std::unique_ptr<const Tensor> output_tensor0(
std::move(predictor_crnn->GetOutput(0)));
auto *predict_batch = output_tensor0->data<float>();
auto predict_shape = output_tensor0->shape();
// ctc decode
std::string str_res;
int argmax_idx;
int last_index = 0;
float score = 0.f;
int count = 0;
float max_value = 0.0f;
for (int n = 0; n < predict_shape[1]; n++) {
argmax_idx = int(Argmax(&predict_batch[n * predict_shape[2]],
&predict_batch[(n + 1) * predict_shape[2]]));
max_value =
float(*std::max_element(&predict_batch[n * predict_shape[2]],
&predict_batch[(n + 1) * predict_shape[2]]));
if (argmax_idx > 0 && (!(n > 0 && argmax_idx == last_index))) {
score += max_value;
count += 1;
str_res += charactor_dict[argmax_idx];
}
last_index = argmax_idx;
}
score /= count;
rec_text.push_back(str_res);
rec_text_score.push_back(score);
}
}
std::vector<std::vector<std::vector<int>>>
RunDetModel(std::shared_ptr<PaddlePredictor> predictor, cv::Mat img,
std::map<std::string, double> Config) {
// Read img
int max_side_len = int(Config["max_side_len"]);
int det_db_use_dilate = int(Config["det_db_use_dilate"]);
cv::Mat srcimg;
img.copyTo(srcimg);
std::vector<float> ratio_hw;
img = DetResizeImg(img, max_side_len, ratio_hw);
cv::Mat img_fp;
img.convertTo(img_fp, CV_32FC3, 1.0 / 255.f);
// Prepare input data from image
std::unique_ptr<Tensor> input_tensor0(std::move(predictor->GetInput(0)));
input_tensor0->Resize({1, 3, img_fp.rows, img_fp.cols});
auto *data0 = input_tensor0->mutable_data<float>();
std::vector<float> mean = {0.485f, 0.456f, 0.406f};
std::vector<float> scale = {1 / 0.229f, 1 / 0.224f, 1 / 0.225f};
const float *dimg = reinterpret_cast<const float *>(img_fp.data);
NeonMeanScale(dimg, data0, img_fp.rows * img_fp.cols, mean, scale);
// Run predictor
predictor->Run();
// Get output and post process
std::unique_ptr<const Tensor> output_tensor(
std::move(predictor->GetOutput(0)));
auto *outptr = output_tensor->data<float>();
auto shape_out = output_tensor->shape();
// Save output
float pred[shape_out[2] * shape_out[3]];
unsigned char cbuf[shape_out[2] * shape_out[3]];
for (int i = 0; i < int(shape_out[2] * shape_out[3]); i++) {
pred[i] = static_cast<float>(outptr[i]);
cbuf[i] = static_cast<unsigned char>((outptr[i]) * 255);
}
cv::Mat cbuf_map(shape_out[2], shape_out[3], CV_8UC1,
reinterpret_cast<unsigned char *>(cbuf));
cv::Mat pred_map(shape_out[2], shape_out[3], CV_32F,
reinterpret_cast<float *>(pred));
const double threshold = double(Config["det_db_thresh"]) * 255;
const double max_value = 255;
cv::Mat bit_map;
cv::threshold(cbuf_map, bit_map, threshold, max_value, cv::THRESH_BINARY);
if (det_db_use_dilate == 1) {
cv::Mat dilation_map;
cv::Mat dila_ele =
cv::getStructuringElement(cv::MORPH_RECT, cv::Size(2, 2));
cv::dilate(bit_map, dilation_map, dila_ele);
bit_map = dilation_map;
}
auto boxes = BoxesFromBitmap(pred_map, bit_map, Config);
std::vector<std::vector<std::vector<int>>> filter_boxes =
FilterTagDetRes(boxes, ratio_hw[0], ratio_hw[1], srcimg);
return filter_boxes;
}
std::shared_ptr<PaddlePredictor> loadModel(std::string model_file) {
MobileConfig config;
config.set_model_from_file(model_file);
std::shared_ptr<PaddlePredictor> predictor =
CreatePaddlePredictor<MobileConfig>(config);
return predictor;
}
cv::Mat Visualization(cv::Mat srcimg,
std::vector<std::vector<std::vector<int>>> boxes) {
cv::Point rook_points[boxes.size()][4];
for (int n = 0; n < boxes.size(); n++) {
for (int m = 0; m < boxes[0].size(); m++) {
rook_points[n][m] = cv::Point(static_cast<int>(boxes[n][m][0]),
static_cast<int>(boxes[n][m][1]));
}
}
cv::Mat img_vis;
srcimg.copyTo(img_vis);
for (int n = 0; n < boxes.size(); n++) {
const cv::Point *ppt[1] = {rook_points[n]};
int npt[] = {4};
cv::polylines(img_vis, ppt, npt, 1, 1, CV_RGB(0, 255, 0), 2, 8, 0);
}
cv::imwrite("./vis.jpg", img_vis);
std::cout << "The detection visualized image saved in ./vis.jpg" << std::endl;
return img_vis;
}
std::vector<std::string> split(const std::string &str,
const std::string &delim) {
std::vector<std::string> res;
if ("" == str)
return res;
char *strs = new char[str.length() + 1];
std::strcpy(strs, str.c_str());
char *d = new char[delim.length() + 1];
std::strcpy(d, delim.c_str());
char *p = std::strtok(strs, d);
while (p) {
string s = p;
res.push_back(s);
p = std::strtok(NULL, d);
}
return res;
}
std::map<std::string, double> LoadConfigTxt(std::string config_path) {
auto config = ReadDict(config_path);
std::map<std::string, double> dict;
for (int i = 0; i < config.size(); i++) {
std::vector<std::string> res = split(config[i], " ");
dict[res[0]] = stod(res[1]);
}
return dict;
}
int main(int argc, char **argv) {
if (argc < 5) {
std::cerr << "[ERROR] usage: " << argv[0]
<< " det_model_file cls_model_file rec_model_file image_path "
"charactor_dict\n";
exit(1);
}
std::string det_model_file = argv[1];
std::string rec_model_file = argv[2];
std::string cls_model_file = argv[3];
std::string img_path = argv[4];
std::string dict_path = argv[5];
//// load config from txt file
auto Config = LoadConfigTxt("./config.txt");
int use_direction_classify = int(Config["use_direction_classify"]);
auto start = std::chrono::system_clock::now();
auto det_predictor = loadModel(det_model_file);
auto rec_predictor = loadModel(rec_model_file);
auto cls_predictor = loadModel(cls_model_file);
auto charactor_dict = ReadDict(dict_path);
charactor_dict.insert(charactor_dict.begin(), "#"); // blank char for ctc
charactor_dict.push_back(" ");
cv::Mat srcimg = cv::imread(img_path, cv::IMREAD_COLOR);
auto boxes = RunDetModel(det_predictor, srcimg, Config);
std::vector<std::string> rec_text;
std::vector<float> rec_text_score;
RunRecModel(boxes, srcimg, rec_predictor, rec_text, rec_text_score,
charactor_dict, cls_predictor, use_direction_classify);
auto end = std::chrono::system_clock::now();
auto duration =
std::chrono::duration_cast<std::chrono::microseconds>(end - start);
//// visualization
auto img_vis = Visualization(srcimg, boxes);
//// print recognized text
for (int i = 0; i < rec_text.size(); i++) {
std::cout << i << "\t" << rec_text[i] << "\t" << rec_text_score[i]
<< std::endl;
}
std::cout << "花费了"
<< double(duration.count()) *
std::chrono::microseconds::period::num /
std::chrono::microseconds::period::den
<< "秒" << std::endl;
return 0;
}
\ No newline at end of file
deploy/lite/prepare.sh 0 → 100644
View file @ 19eb7eb8
#!/bin/bash
mkdir -p $1/demo/cxx/ocr/debug/
cp ../../ppocr/utils/ppocr_keys_v1.txt $1/demo/cxx/ocr/debug/
cp -r ./* $1/demo/cxx/ocr/
cp ./config.txt $1/demo/cxx/ocr/debug/
cp ../../doc/imgs/11.jpg $1/demo/cxx/ocr/debug/
echo "Prepare Done"
deploy/lite/readme.md 0 → 100644
View file @ 19eb7eb8
# 端侧部署
本教程将介绍基于[Paddle Lite](https://github.com/PaddlePaddle/Paddle-Lite) 在移动端部署PaddleOCR超轻量中文检测、识别模型的详细步骤。
Paddle Lite是飞桨轻量化推理引擎，为手机、IOT端提供高效推理能力，并广泛整合跨平台硬件，为端侧部署及应用落地问题提供轻量化的部署方案。
## 1. 准备环境
### 运行准备
- 电脑（编译Paddle Lite）
- 安卓手机（armv7或armv8）
### 1.1 准备交叉编译环境
交叉编译环境用于编译 Paddle Lite 和 PaddleOCR 的C++ demo。
支持多种开发环境，不同开发环境的编译流程请参考对应文档。
1. [Docker](https://paddle-lite.readthedocs.io/zh/latest/source_compile/compile_env.html#docker)
2. [Linux](https://paddle-lite.readthedocs.io/zh/latest/source_compile/compile_env.html#linux)
3. [MAC OS](https://paddle-lite.readthedocs.io/zh/latest/source_compile/compile_env.html#mac-os)
### 1.2 准备预测库
预测库有两种获取方式：
- 1. 直接下载，预测库下载链接如下：
| 平台 | 预测库下载链接 |
|---|---|
|Android|[arm7](https://github.com/PaddlePaddle/Paddle-Lite/releases/download/v2.9/inference_lite_lib.android.armv7.gcc.c++_shared.with_extra.with_cv.tar.gz) / [arm8](https://github.com/PaddlePaddle/Paddle-Lite/releases/download/v2.9/inference_lite_lib.android.armv8.gcc.c++_shared.with_extra.with_cv.tar.gz)|
|IOS|[arm7](https://github.com/PaddlePaddle/Paddle-Lite/releases/download/v2.9/inference_lite_lib.ios.armv7.with_cv.with_extra.with_log.tiny_publish.tar.gz) / [arm8](https://github.com/PaddlePaddle/Paddle-Lite/releases/download/v2.9/inference_lite_lib.ios.armv8.with_cv.with_extra.with_log.tiny_publish.tar.gz)|
注：1. 上述预测库为PaddleLite 2.9分支编译得到，有关PaddleLite 2.9 详细信息可参考 [链接](https://github.com/PaddlePaddle/Paddle-Lite/releases/tag/v2.9) 。
- 2. [推荐]编译Paddle-Lite得到预测库，Paddle-Lite的编译方式如下：
```
git clone https://github.com/PaddlePaddle/Paddle-Lite.git
cd Paddle-Lite
# 切换到Paddle-Lite release/v2.9 稳定分支
git checkout release/v2.9
./lite/tools/build_android.sh --arch=armv8 --with_cv=ON --with_extra=ON
```
注意：编译Paddle-Lite获得预测库时，需要打开`--with_cv=ON --with_extra=ON`两个选项，`--arch`表示`arm`版本，这里指定为armv8，
更多编译命令
介绍请参考 [链接](https://paddle-lite.readthedocs.io/zh/latest/source_compile/compile_andriod.html) 。
直接下载预测库并解压后，可以得到`inference_lite_lib.android.armv8/`文件夹，通过编译Paddle-Lite得到的预测库位于
`Paddle-Lite/build.lite.android.armv8.gcc/inference_lite_lib.android.armv8/`文件夹下。
预测库的文件目录如下：
```
inference_lite_lib.android.armv8/
|-- cxx C++ 预测库和头文件
| |-- include C++ 头文件
| | |-- paddle_api.h
| | |-- paddle_image_preprocess.h
| | |-- paddle_lite_factory_helper.h
| | |-- paddle_place.h
| | |-- paddle_use_kernels.h
| | |-- paddle_use_ops.h
| | `-- paddle_use_passes.h
| `-- lib C++预测库
| |-- libpaddle_api_light_bundled.a C++静态库
| `-- libpaddle_light_api_shared.so C++动态库
|-- java Java预测库
| |-- jar
| | `-- PaddlePredictor.jar
| |-- so
| | `-- libpaddle_lite_jni.so
| `-- src
|-- demo C++和Java示例代码
| |-- cxx C++ 预测库demo
| `-- java Java 预测库demo
```
## 2 开始运行
### 2.1 模型优化
Paddle-Lite 提供了多种策略来自动优化原始的模型，其中包括量化、子图融合、混合调度、Kernel优选等方法，使用Paddle-lite的opt工具可以自动
对inference模型进行优化，优化后的模型更轻量，模型运行速度更快。
如果已经准备好了 `.nb` 结尾的模型文件，可以跳过此步骤。
下述表格中也提供了一系列中文移动端模型：
|模型版本|模型简介|模型大小|检测模型|文本方向分类模型|识别模型|Paddle-Lite版本|
|---|---|---|---|---|---|---|
|V2.0|超轻量中文OCR 移动端模型|7.8M|[下载地址](https://paddleocr.bj.bcebos.com/dygraph_v2.0/lite/ch_ppocr_mobile_v2.0_det_opt.nb)|[下载地址](https://paddleocr.bj.bcebos.com/dygraph_v2.0/lite/ch_ppocr_mobile_v2.0_cls_opt.nb)|[下载地址](https://paddleocr.bj.bcebos.com/dygraph_v2.0/lite/ch_ppocr_mobile_v2.0_rec_opt.nb)|v2.9|
|V2.0(slim)|超轻量中文OCR 移动端模型|3.3M|[下载地址](https://paddleocr.bj.bcebos.com/dygraph_v2.0/lite/ch_ppocr_mobile_v2.0_det_slim_opt.nb)|[下载地址](https://paddleocr.bj.bcebos.com/dygraph_v2.0/lite/ch_ppocr_mobile_v2.0_cls_slim_opt.nb)|[下载地址](https://paddleocr.bj.bcebos.com/dygraph_v2.0/lite/ch_ppocr_mobile_v2.0_rec_slim_opt.nb)|v2.9|
如果直接使用上述表格中的模型进行部署，可略过下述步骤，直接阅读 [2.2节](#2.2与手机联调)。
如果要部署的模型不在上述表格中，则需要按照如下步骤获得优化后的模型。
模型优化需要Paddle-Lite的opt可执行文件，可以通过编译Paddle-Lite源码获得，编译步骤如下：
```
# 如果准备环境时已经clone了Paddle-Lite，则不用重新clone Paddle-Lite
git clone https://github.com/PaddlePaddle/Paddle-Lite.git
cd Paddle-Lite
git checkout release/v2.9
# 启动编译
./lite/tools/build.sh build_optimize_tool
```
编译完成后，opt文件位于`build.opt/lite/api/`下，可通过如下方式查看opt的运行选项和使用方式；
```
cd build.opt/lite/api/
./opt
```
|选项|说明|
|---|---|
|--model_dir|待优化的PaddlePaddle模型（非combined形式）的路径|
|--model_file|待优化的PaddlePaddle模型（combined形式）的网络结构文件路径|
|--param_file|待优化的PaddlePaddle模型（combined形式）的权重文件路径|
|--optimize_out_type|输出模型类型，目前支持两种类型：protobuf和naive_buffer，其中naive_buffer是一种更轻量级的序列化/反序列化实现。若您需要在mobile端执行模型预测，请将此选项设置为naive_buffer。默认为protobuf|
|--optimize_out|优化模型的输出路径|
|--valid_targets|指定模型可执行的backend，默认为arm。目前可支持x86、arm、opencl、npu、xpu，可以同时指定多个backend(以空格分隔)，Model Optimize Tool将会自动选择最佳方式。如果需要支持华为NPU（Kirin 810/990 Soc搭载的达芬奇架构NPU），应当设置为npu, arm|
|--record_tailoring_info|当使用 根据模型裁剪库文件 功能时，则设置该选项为true，以记录优化后模型含有的kernel和OP信息，默认为false|
`--model_dir`适用于待优化的模型是非combined方式，PaddleOCR的inference模型是combined方式，即模型结构和模型参数使用单独一个文件存储。
下面以PaddleOCR的超轻量中文模型为例，介绍使用编译好的opt文件完成inference模型到Paddle-Lite优化模型的转换。
```
# 【推荐】 下载PaddleOCR V2.0版本的中英文 inference模型
wget https://paddleocr.bj.bcebos.com/dygraph_v2.0/slim/ch_ppocr_mobile_v2.0_det_slim_infer.tar && tar xf ch_ppocr_mobile_v2.0_det_slim_infer.tar
wget https://paddleocr.bj.bcebos.com/dygraph_v2.0/slim/ch_ppocr_mobile_v2.0_rec_slim_infer.tar && tar xf ch_ppocr_mobile_v2.0_rec_slim_infer.tar
wget https://paddleocr.bj.bcebos.com/dygraph_v2.0/slim/ch_ppocr_mobile_v2.0_cls_slim_infer.tar && tar xf ch_ppocr_mobile_v2.0_cls_slim_infer.tar
# 转换V2.0检测模型
./opt --model_file=./ch_ppocr_mobile_v2.0_det_slim_infer/inference.pdmodel --param_file=./ch_ppocr_mobile_v2.0_det_slim_infer/inference.pdiparams --optimize_out=./ch_ppocr_mobile_v2.0_det_slim_opt --valid_targets=arm --optimize_out_type=naive_buffer
# 转换V2.0识别模型
./opt --model_file=./ch_ppocr_mobile_v2.0_rec_slim_infer/inference.pdmodel --param_file=./ch_ppocr_mobile_v2.0_rec_slim_infer/inference.pdiparams --optimize_out=./ch_ppocr_mobile_v2.0_rec_slim_opt --valid_targets=arm --optimize_out_type=naive_buffer
# 转换V2.0方向分类器模型
./opt --model_file=./ch_ppocr_mobile_v2.0_cls_slim_infer/inference.pdmodel --param_file=./ch_ppocr_mobile_v2.0_cls_slim_infer/inference.pdiparams --optimize_out=./ch_ppocr_mobile_v2.0_cls_slim_opt --valid_targets=arm --optimize_out_type=naive_buffer
```
转换成功后，inference模型目录下会多出`.nb`结尾的文件，即是转换成功的模型文件。
注意：使用paddle-lite部署时，需要使用opt工具优化后的模型。 opt 工具的输入模型是paddle保存的inference模型
<a name="2.2与手机联调"></a>
### 2.2 与手机联调
首先需要进行一些准备工作。
1. 准备一台arm8的安卓手机，如果编译的预测库和opt文件是armv7，则需要arm7的手机，并修改Makefile中`ARM_ABI = arm7`。
2. 打开手机的USB调试选项，选择文件传输模式，连接电脑。
3. 电脑上安装adb工具，用于调试。 adb安装方式如下：
3.1. MAC电脑安装ADB:
```
brew cask install android-platform-tools
```
3.2. Linux安装ADB
```
sudo apt update
sudo apt install -y wget adb
```
3.3. Window安装ADB
win上安装需要去谷歌的安卓平台下载adb软件包进行安装：[链接](https://developer.android.com/studio)
打开终端，手机连接电脑，在终端中输入
```
adb devices
```
如果有device输出，则表示安装成功。
```
List of devices attached
744be294 device
```
4. 准备优化后的模型、预测库文件、测试图像和使用的字典文件。
```
git clone https://github.com/PaddlePaddle/PaddleOCR.git
cd PaddleOCR/deploy/lite/
# 运行prepare.sh，准备预测库文件、测试图像和使用的字典文件，并放置在预测库中的demo/cxx/ocr文件夹下
sh prepare.sh /{lite prediction library path}/inference_lite_lib.android.armv8
# 进入OCR demo的工作目录
cd /{lite prediction library path}/inference_lite_lib.android.armv8/
cd demo/cxx/ocr/
# 将C++预测动态库so文件复制到debug文件夹中
cp ../../../cxx/lib/libpaddle_light_api_shared.so ./debug/
```
准备测试图像，以`PaddleOCR/doc/imgs/11.jpg`为例，将测试的图像复制到`demo/cxx/ocr/debug/`文件夹下。
准备lite opt工具优化后的模型文件，比如使用`ch_ppocr_mobile_v2.0_det_slim_opt.nb，ch_ppocr_mobile_v2.0_rec_slim_opt.nb, ch_ppocr_mobile_v2.0_cls_slim_opt.nb`，模型文件放置在`demo/cxx/ocr/debug/`文件夹下。
执行完成后，ocr文件夹下将有如下文件格式：
```
demo/cxx/ocr/
|-- debug/
| |--ch_ppocr_mobile_v2.0_det_slim_opt.nb 优化后的检测模型文件
| |--ch_ppocr_mobile_v2.0_rec_slim_opt.nb 优化后的识别模型文件
| |--ch_ppocr_mobile_v2.0_cls_slim_opt.nb 优化后的文字方向分类器模型文件
| |--11.jpg 待测试图像
| |--ppocr_keys_v1.txt 中文字典文件
| |--libpaddle_light_api_shared.so C++预测库文件
| |--config.txt 超参数配置
|-- config.txt 超参数配置
|-- cls_process.cc 方向分类器的预处理和后处理文件
|-- cls_process.h
|-- crnn_process.cc 识别模型CRNN的预处理和后处理文件
|-- crnn_process.h
|-- db_post_process.cc 检测模型DB的后处理文件
|-- db_post_process.h
|-- Makefile 编译文件
|-- ocr_db_crnn.cc C++预测源文件
```
#### 注意：
1. ppocr_keys_v1.txt是中文字典文件，如果使用的 nb 模型是英文数字或其他语言的模型，需要更换为对应语言的字典。
PaddleOCR 在ppocr/utils/下存放了多种字典，包括：
```
dict/french_dict.txt # 法语字典
dict/german_dict.txt # 德语字典
ic15_dict.txt # 英文字典
dict/japan_dict.txt # 日语字典
dict/korean_dict.txt # 韩语字典
ppocr_keys_v1.txt # 中文字典
...
```
2. `config.txt` 包含了检测器、分类器的超参数，如下：
```
max_side_len 960 # 输入图像长宽大于960时，等比例缩放图像，使得图像最长边为960
det_db_thresh 0.3 # 用于过滤DB预测的二值化图像，设置为0.-0.3对结果影响不明显
det_db_box_thresh 0.5 # DB后处理过滤box的阈值，如果检测存在漏框情况，可酌情减小
det_db_unclip_ratio 1.6 # 表示文本框的紧致程度，越小则文本框更靠近文本
use_direction_classify 0 # 是否使用方向分类器，0表示不使用，1表示使用
```
5. 启动调试
上述步骤完成后就可以使用adb将文件push到手机上运行，步骤如下：
```
# 执行编译，得到可执行文件ocr_db_crnn, 第一次执行此命令会下载opencv等依赖库，下载完成后，需要再执行一次
make -j
# 将编译的可执行文件移动到debug文件夹中
mv ocr_db_crnn ./debug/
# 将debug文件夹push到手机上
adb push debug /data/local/tmp/
adb shell
cd /data/local/tmp/debug
export LD_LIBRARY_PATH=${PWD}:$LD_LIBRARY_PATH
# 开始使用，ocr_db_crnn可执行文件的使用方式为:
# ./ocr_db_crnn 检测模型文件 方向分类器模型文件 识别模型文件 测试图像路径 字典文件路径
./ocr_db_crnn ch_ppocr_mobile_v2.0_det_slim_opt.nb ch_ppocr_mobile_v2.0_rec_slim_opt.nb ch_ppocr_mobile_v2.0_cls_slim_opt.nb ./11.jpg ppocr_keys_v1.txt
```
如果对代码做了修改，则需要重新编译并push到手机上。
运行效果如下：
<div align="center">
<img src="imgs/lite_demo.png" width="600">
</div>
## FAQ
Q1：如果想更换模型怎么办，需要重新按照流程走一遍吗？
A1：如果已经走通了上述步骤，更换模型只需要替换 .nb 模型文件即可，同时要注意更新字典
Q2：换一个图测试怎么做？
A2：替换debug下的.jpg测试图像为你想要测试的图像，adb push 到手机上即可
Q3：如何封装到手机APP中？
A3：此demo旨在提供能在手机上运行OCR的核心算法部分，PaddleOCR/deploy/android_demo是将这个demo封装到手机app的示例，供参考
deploy/lite/readme_en.md 0 → 100644
View file @ 19eb7eb8
# Tutorial of PaddleOCR Mobile deployment
This tutorial will introduce how to use [Paddle Lite](https://github.com/PaddlePaddle/Paddle-Lite) to deploy paddleOCR ultra-lightweight Chinese and English detection models on mobile phones.
paddle-lite is a lightweight inference engine for PaddlePaddle. It provides efficient inference capabilities for mobile phones and IoTs, and extensively integrates cross-platform hardware to provide lightweight deployment solutions for end-side deployment issues.
## 1. Preparation
### Preparation environment
- Computer (for Compiling Paddle Lite)
- Mobile phone (arm7 or arm8)
### 1.1 Prepare the cross-compilation environment
The cross-compilation environment is used to compile C++ demos of Paddle Lite and PaddleOCR.
Supports multiple development environments.
For the compilation process of different development environments, please refer to the corresponding documents.
1. [Docker](https://paddle-lite.readthedocs.io/zh/latest/source_compile/compile_env.html#docker)
2. [Linux](https://paddle-lite.readthedocs.io/zh/latest/source_compile/compile_env.html#linux)
3. [MAC OS](https://paddle-lite.readthedocs.io/zh/latest/source_compile/compile_env.html#mac-os)
### 1.2 Prepare Paddle-Lite library
There are two ways to obtain the Paddle-Lite library：
- 1. Download directly, the download link of the Paddle-Lite library is as follows：
| Platform | Paddle-Lite library download link |
|---|---|
|Android|[arm7](https://github.com/PaddlePaddle/Paddle-Lite/releases/download/v2.9/inference_lite_lib.android.armv7.gcc.c++_shared.with_extra.with_cv.tar.gz) / [arm8](https://github.com/PaddlePaddle/Paddle-Lite/releases/download/v2.9/inference_lite_lib.android.armv8.gcc.c++_shared.with_extra.with_cv.tar.gz)|
|IOS|[arm7](https://github.com/PaddlePaddle/Paddle-Lite/releases/download/v2.9/inference_lite_lib.ios.armv7.with_cv.with_extra.with_log.tiny_publish.tar.gz) / [arm8](https://github.com/PaddlePaddle/Paddle-Lite/releases/download/v2.9/inference_lite_lib.ios.armv8.with_cv.with_extra.with_log.tiny_publish.tar.gz)|
Note: 1. The above Paddle-Lite library is compiled from the Paddle-Lite 2.9 branch. For more information about Paddle-Lite 2.9, please refer to [link](https://github.com/PaddlePaddle/Paddle-Lite/releases/tag/v2.9).
- 2. [Recommended] Compile Paddle-Lite to get the prediction library. The compilation method of Paddle-Lite is as follows：
```
git clone https://github.com/PaddlePaddle/Paddle-Lite.git
cd Paddle-Lite
# Switch to Paddle-Lite release/v2.8 stable branch
git checkout release/v2.8
./lite/tools/build_android.sh --arch=armv8 --with_cv=ON --with_extra=ON
```
Note: When compiling Paddle-Lite to obtain the Paddle-Lite library, you need to turn on the two options `--with_cv=ON --with_extra=ON`, `--arch` means the `arm` version, here is designated as armv8,
More compilation commands refer to the introduction [link](https://paddle-lite.readthedocs.io/zh/latest/source_compile/compile_andriod.html) 。
After directly downloading the Paddle-Lite library and decompressing it, you can get the `inference_lite_lib.android.armv8/` folder, and the Paddle-Lite library obtained by compiling Paddle-Lite is located
`Paddle-Lite/build.lite.android.armv8.gcc/inference_lite_lib.android.armv8/` folder.
The structure of the prediction library is as follows:
```
inference_lite_lib.android.armv8/
|-- cxx C++ prebuild library
| |-- include C++
| | |-- paddle_api.h
| | |-- paddle_image_preprocess.h
| | |-- paddle_lite_factory_helper.h
| | |-- paddle_place.h
| | |-- paddle_use_kernels.h
| | |-- paddle_use_ops.h
| | `-- paddle_use_passes.h
| `-- lib C++ library
| |-- libpaddle_api_light_bundled.a C++ static library
| `-- libpaddle_light_api_shared.so C++ dynamic library
|-- java Java library
| |-- jar
| | `-- PaddlePredictor.jar
| |-- so
| | `-- libpaddle_lite_jni.so
| `-- src
|-- demo C++ and Java demo
| |-- cxx C++ demo
| `-- java Java demo
```
## 2 Run
### 2.1 Inference Model Optimization
Paddle Lite provides a variety of strategies to automatically optimize the original training model, including quantization, sub-graph fusion, hybrid scheduling, Kernel optimization and so on. In order to make the optimization process more convenient and easy to use, Paddle Lite provide opt tools to automatically complete the optimization steps and output a lightweight, optimal executable model.
If you have prepared the model file ending in .nb, you can skip this step.
The following table also provides a series of models that can be deployed on mobile phones to recognize Chinese. You can directly download the optimized model.
|Version|Introduction|Model size|Detection model|Text Direction model|Recognition model|Paddle-Lite branch|
|---|---|---|---|---|---|---|
|V2.0|extra-lightweight chinese OCR optimized model|7.8M|[download link](https://paddleocr.bj.bcebos.com/dygraph_v2.0/lite/ch_ppocr_mobile_v2.0_det_opt.nb)|[download link](https://paddleocr.bj.bcebos.com/dygraph_v2.0/lite/ch_ppocr_mobile_v2.0_cls_opt.nb)|[download link](https://paddleocr.bj.bcebos.com/dygraph_v2.0/lite/ch_ppocr_mobile_v2.0_rec_opt.nb)|v2.9|
|V2.0(slim)|extra-lightweight chinese OCR optimized model|3.3M|[download link](https://paddleocr.bj.bcebos.com/dygraph_v2.0/lite/ch_ppocr_mobile_v2.0_det_slim_opt.nb)|[download link](https://paddleocr.bj.bcebos.com/dygraph_v2.0/lite/ch_ppocr_mobile_v2.0_cls_slim_opt.nb)|[download link](https://paddleocr.bj.bcebos.com/dygraph_v2.0/lite/ch_ppocr_mobile_v2.0_rec_slim_opt.nb)|v2.9|
If you directly use the model in the above table for deployment, you can skip the following steps and directly read [Section 2.2](#2.2 Run optimized model on Phone).
If the model to be deployed is not in the above table, you need to follow the steps below to obtain the optimized model.
The `opt` tool can be obtained by compiling Paddle Lite.
```
git clone https://github.com/PaddlePaddle/Paddle-Lite.git
cd Paddle-Lite
git checkout release/v2.9
./lite/tools/build.sh build_optimize_tool
```
After the compilation is complete, the opt file is located under build.opt/lite/api/, You can view the operating options and usage of opt in the following ways:
```
cd build.opt/lite/api/
./opt
```
|Options|Description|
|---|---|
|--model_dir|The path of the PaddlePaddle model to be optimized (non-combined form)|
|--model_file|The network structure file path of the PaddlePaddle model (combined form) to be optimized|
|--param_file|The weight file path of the PaddlePaddle model (combined form) to be optimized|
|--optimize_out_type|Output model type, currently supports two types: protobuf and naive_buffer, among which naive_buffer is a more lightweight serialization/deserialization implementation. If you need to perform model prediction on the mobile side, please set this option to naive_buffer. The default is protobuf|
|--optimize_out|The output path of the optimized model|
|--valid_targets|The executable backend of the model, the default is arm. Currently it supports x86, arm, opencl, npu, xpu, multiple backends can be specified at the same time (separated by spaces), and Model Optimize Tool will automatically select the best method. If you need to support Huawei NPU (DaVinci architecture NPU equipped with Kirin 810/990 Soc), it should be set to npu, arm|
|--record_tailoring_info|When using the function of cutting library files according to the model, set this option to true to record the kernel and OP information contained in the optimized model. The default is false|
`--model_dir` is suitable for the non-combined mode of the model to be optimized, and the inference model of PaddleOCR is the combined mode, that is, the model structure and model parameters are stored in a single file.
The following takes the ultra-lightweight Chinese model of PaddleOCR as an example to introduce the use of the compiled opt file to complete the conversion of the inference model to the Paddle-Lite optimized model
```
# [Recommendation] Download the Chinese and English inference model of PaddleOCR V2.0
wget https://paddleocr.bj.bcebos.com/dygraph_v2.0/slim/ch_ppocr_mobile_v2.0_det_slim_infer.tar && tar xf ch_ppocr_mobile_v2.0_det_slim_infer.tar
wget https://paddleocr.bj.bcebos.com/dygraph_v2.0/slim/ch_ppocr_mobile_v2.0_rec_slim_infer.tar && tar xf ch_ppocr_mobile_v2.0_rec_slim_infer.tar
wget https://paddleocr.bj.bcebos.com/dygraph_v2.0/slim/ch_ppocr_mobile_v2.0_cls_slim_infer.tar && tar xf ch_ppocr_mobile_v2.0_cls_slim_infer.tar
# Convert V2.0 detection model
./opt --model_file=./ch_ppocr_mobile_v2.0_det_slim_infer/inference.pdmodel --param_file=./ch_ppocr_mobile_v2.0_det_slim_infer/inference.pdiparams --optimize_out=./ch_ppocr_mobile_v2.0_det_slim_opt --valid_targets=arm --optimize_out_type=naive_buffer
# Convert V2.0 recognition model
./opt --model_file=./ch_ppocr_mobile_v2.0_rec_slim_infer/inference.pdmodel --param_file=./ch_ppocr_mobile_v2.0_rec_slim_infer/inference.pdiparams --optimize_out=./ch_ppocr_mobile_v2.0_rec_slim_opt --valid_targets=arm --optimize_out_type=naive_buffer
# Convert V2.0 angle classifier model
./opt --model_file=./ch_ppocr_mobile_v2.0_cls_slim_infer/inference.pdmodel --param_file=./ch_ppocr_mobile_v2.0_cls_slim_infer/inference.pdiparams --optimize_out=./ch_ppocr_mobile_v2.0_cls_slim_opt --valid_targets=arm --optimize_out_type=naive_buffer
```
After the conversion is successful, there will be more files ending with `.nb` in the inference model directory, which is the successfully converted model file.
<a name="2.2 Run optimized model on Phone"></a>
### 2.2 Run optimized model on Phone
Some preparatory work is required first.
1. Prepare an Android phone with arm8. If the compiled prediction library and opt file are armv7, you need an arm7 phone and modify ARM_ABI = arm7 in the Makefile.
2. Make sure the phone is connected to the computer, open the USB debugging option of the phone, and select the file transfer mode.
3. Install the adb tool on the computer.
3.1. Install ADB for MAC:
```
brew cask install android-platform-tools
```
3.2. Install ADB for Linux
```
sudo apt update
sudo apt install -y wget adb
```
3.3. Install ADB for windows
To install on win, you need to go to Google's Android platform to download the adb package for installation：[link](https://developer.android.com/studio)
Verify whether adb is installed successfully
```
adb devices
```
If there is device output, it means the installation is successful。
```
List of devices attached
744be294 device
```
4. Prepare optimized models, prediction library files, test images and dictionary files used.
```
git clone https://github.com/PaddlePaddle/PaddleOCR.git
cd PaddleOCR/deploy/lite/
# run prepare.sh
sh prepare.sh /{lite prediction library path}/inference_lite_lib.android.armv8
#
cd /{lite prediction library path}/inference_lite_lib.android.armv8/
cd demo/cxx/ocr/
# copy paddle-lite C++ .so file to debug/ directory
cp ../../../cxx/lib/libpaddle_light_api_shared.so ./debug/
cd inference_lite_lib.android.armv8/demo/cxx/ocr/
cp ../../../cxx/lib/libpaddle_light_api_shared.so ./debug/
```
Prepare the test image, taking PaddleOCR/doc/imgs/11.jpg as an example, copy the image file to the demo/cxx/ocr/debug/ folder. Prepare the model files optimized by the lite opt tool, ch_det_mv3_db_opt.nb, ch_rec_mv3_crnn_opt.nb, and place them under the demo/cxx/ocr/debug/ folder.
The structure of the OCR demo is as follows after the above command is executed:
```
demo/cxx/ocr/
|-- debug/
| |--ch_ppocr_mobile_v2.0_det_slim_opt.nb Detection model
| |--ch_ppocr_mobile_v2.0_rec_slim_opt.nb Recognition model
| |--ch_ppocr_mobile_v2.0_cls_slim_opt.nb Text direction classification model
| |--11.jpg Image for OCR
| |--ppocr_keys_v1.txt Dictionary file
| |--libpaddle_light_api_shared.so C++ .so file
| |--config.txt Config file
|-- config.txt Config file
|-- cls_process.cc Pre-processing and post-processing files for the angle classifier
|-- cls_process.h
|-- crnn_process.cc Pre-processing and post-processing files for the CRNN model
|-- crnn_process.h
|-- db_post_process.cc Pre-processing and post-processing files for the DB model
|-- db_post_process.h
|-- Makefile
|-- ocr_db_crnn.cc C++ main code
```
#### 注意：
1. `ppocr_keys_v1.txt` is a Chinese dictionary file. If the nb model is used for English recognition or other language recognition, dictionary file should be replaced with a dictionary of the corresponding language. PaddleOCR provides a variety of dictionaries under ppocr/utils/, including:
```
dict/french_dict.txt # french
dict/german_dict.txt # german
ic15_dict.txt # english
dict/japan_dict.txt # japan
dict/korean_dict.txt # korean
ppocr_keys_v1.txt # chinese
```
2. `config.txt` of the detector and classifier, as shown below:
```
max_side_len 960 # Limit the maximum image height and width to 960
det_db_thresh 0.3 # Used to filter the binarized image of DB prediction, setting 0.-0.3 has no obvious effect on the result
det_db_box_thresh 0.5 # DDB post-processing filter box threshold, if there is a missing box detected, it can be reduced as appropriate
det_db_unclip_ratio 1.6 # Indicates the compactness of the text box, the smaller the value, the closer the text box to the text
use_direction_classify 0 # Whether to use the direction classifier, 0 means not to use, 1 means to use
```
5. Run Model on phone
After the above steps are completed, you can use adb to push the file to the phone to run, the steps are as follows:
```
# Execute the compilation and get the executable file ocr_db_crnn
# The first execution of this command will download dependent libraries such as opencv. After the download is complete, you need to execute it again
make -j
# Move the compiled executable file to the debug folder
mv ocr_db_crnn ./debug/
# Push the debug folder to the phone
adb push debug /data/local/tmp/
adb shell
cd /data/local/tmp/debug
export LD_LIBRARY_PATH=${PWD}:$LD_LIBRARY_PATH
# The use of ocr_db_crnn is:
# ./ocr_db_crnn Detection model file Orientation classifier model file Recognition model file Test image path Dictionary file path
./ocr_db_crnn ch_ppocr_mobile_v2.0_det_opt.nb ch_ppocr_mobile_v2.0_rec_opt.nb ch_ppocr_mobile_v2.0_cls_opt.nb ./11.jpg ppocr_keys_v1.txt
```
If you modify the code, you need to recompile and push to the phone.
The outputs are as follows:
<div align="center">
<img src="imgs/lite_demo.png" width="600">
</div>
## FAQ
Q1: What if I want to change the model, do I need to run it again according to the process?
A1: If you have performed the above steps, you only need to replace the .nb model file to complete the model replacement.
Q2: How to test with another picture?
A2: Replace the .jpg test image under ./debug with the image you want to test, and run adb push to push new image to the phone.
Q3: How to package it into the mobile APP?
A3: This demo aims to provide the core algorithm part that can run OCR on mobile phones. Further, PaddleOCR/deploy/android_demo is an example of encapsulating this demo into a mobile app for reference.
deploy/pdserving/README.md
View file @ 19eb7eb8
...@@ -30,38 +30,32 @@ The introduction and tutorial of Paddle Serving service deployment framework ref ...@@ -30,38 +30,32 @@ The introduction and tutorial of Paddle Serving service deployment framework ref
PaddleOCR operating environment and Paddle Serving operating environment are needed. PaddleOCR operating environment and Paddle Serving operating environment are needed.
1. Please prepare PaddleOCR operating environment reference [link](../../doc/doc_ch/installation.md). 1. Please prepare PaddleOCR operating environment reference [link](../../doc/doc_ch/installation.md).
Download the corresponding paddle whl package according to the environment, it is recommended to install version 2.0.1.
2. The steps of PaddleServing operating environment prepare are as follows: 2. The steps of PaddleServing operating environment prepare are as follows:
Install serving which used to start the service Install serving which used to start the service
``` ```
pip3 install paddle-serving-server==0.5.0 # for CPU pip3 install paddle-serving-server==0.6.1 # for CPU
pip3 install paddle-serving-server-gpu==0.5.0 # for GPU pip3 install paddle-serving-server-gpu==0.6.1 # for GPU
# Other GPU environments need to confirm the environment and then choose to execute the following commands # Other GPU environments need to confirm the environment and then choose to execute the following commands
pip3 install paddle-serving-server-gpu==0.5.0.post9 # GPU with CUDA9.0 pip3 install paddle-serving-server-gpu==0.6.1.post101 # GPU with CUDA10.1 + TensorRT6
pip3 install paddle-serving-server-gpu==0.5.0.post10 # GPU with CUDA10.0 pip3 install paddle-serving-server-gpu==0.6.1.post11 # GPU with CUDA11 + TensorRT7
pip3 install paddle-serving-server-gpu==0.5.0.post101 # GPU with CUDA10.1 + TensorRT6
pip3 install paddle-serving-server-gpu==0.5.0.post11 # GPU with CUDA10.1 + TensorRT7
``` ```
3. Install the client to send requests to the service 3. Install the client to send requests to the service
``` In [download link](https://github.com/PaddlePaddle/Serving/blob/develop/doc/LATEST_PACKAGES.md) find the client installation package corresponding to the python version.
pip3 install paddle-serving-client==0.5.0 # for CPU The python3.7 version is recommended here:
pip3 install paddle-serving-client-gpu==0.5.0 # for GPU ```
wget https://paddle-serving.bj.bcebos.com/test-dev/whl/paddle_serving_client-0.0.0-cp37-none-any.whl
pip3 install paddle_serving_client-0.0.0-cp37-none-any.whl
``` ```
4. Install serving-app 4. Install serving-app
``` ```
pip3 install paddle-serving-app==0.3.0 pip3 install paddle-serving-app==0.6.1
# fix local_predict to support load dynamic model
# find the install directoory of paddle_serving_app
vim /usr/local/lib/python3.7/site-packages/paddle_serving_app/local_predict.py
# replace line 85 of local_predict.py config = AnalysisConfig(model_path) with:
if os.path.exists(os.path.join(model_path, "__params__")):
config = AnalysisConfig(os.path.join(model_path, "__model__"), os.path.join(model_path, "__params__"))
else:
config = AnalysisConfig(model_path)
``` ```
**note:** If you want to install the latest version of PaddleServing, refer to [link](https://github.com/PaddlePaddle/Serving/blob/develop/doc/LATEST_PACKAGES.md). **note:** If you want to install the latest version of PaddleServing, refer to [link](https://github.com/PaddlePaddle/Serving/blob/develop/doc/LATEST_PACKAGES.md).
...@@ -74,38 +68,38 @@ When using PaddleServing for service deployment, you need to convert the saved i ...@@ -74,38 +68,38 @@ When using PaddleServing for service deployment, you need to convert the saved i
Firstly, download the [inference model](https://github.com/PaddlePaddle/PaddleOCR#pp-ocr-20-series-model-listupdate-on-dec-15) of PPOCR Firstly, download the [inference model](https://github.com/PaddlePaddle/PaddleOCR#pp-ocr-20-series-model-listupdate-on-dec-15) of PPOCR
``` ```
# Download and unzip the OCR text detection model # Download and unzip the OCR text detection model
wget https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_server_v2.0_det_infer.tar && tar xf ch_ppocr_server_v2.0_det_infer.tar wget https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_det_infer.tar && tar xf ch_ppocr_mobile_v2.0_det_infer.tar
# Download and unzip the OCR text recognition model # Download and unzip the OCR text recognition model
wget https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_server_v2.0_rec_infer.tar && tar xf ch_ppocr_server_v2.0_rec_infer.tar wget https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_rec_infer.tar && tar xf ch_ppocr_mobile_v2.0_rec_infer.tar
``` ```
Then, you can use installed paddle_serving_client tool to convert inference model to server model. Then, you can use installed paddle_serving_client tool to convert inference model to mobile model.
``` ```
# Detection model conversion # Detection model conversion
python3 -m paddle_serving_client.convert --dirname ./ch_ppocr_server_v2.0_det_infer/ \ python3 -m paddle_serving_client.convert --dirname ./ch_ppocr_mobile_v2.0_det_infer/ \
--model_filename inference.pdmodel \ --model_filename inference.pdmodel \
--params_filename inference.pdiparams \ --params_filename inference.pdiparams \
--serving_server ./ppocr_det_server_2.0_serving/ \ --serving_server ./ppocr_det_mobile_2.0_serving/ \
--serving_client ./ppocr_det_server_2.0_client/ --serving_client ./ppocr_det_mobile_2.0_client/
# Recognition model conversion # Recognition model conversion
python3 -m paddle_serving_client.convert --dirname ./ch_ppocr_server_v2.0_rec_infer/ \ python3 -m paddle_serving_client.convert --dirname ./ch_ppocr_mobile_v2.0_rec_infer/ \
--model_filename inference.pdmodel \ --model_filename inference.pdmodel \
--params_filename inference.pdiparams \ --params_filename inference.pdiparams \
--serving_server ./ppocr_rec_server_2.0_serving/ \ --serving_server ./ppocr_rec_mobile_2.0_serving/ \
--serving_client ./ppocr_rec_server_2.0_client/ --serving_client ./ppocr_rec_mobile_2.0_client/
``` ```
After the detection model is converted, there will be additional folders of `ppocr_det_server_2.0_serving` and `ppocr_det_server_2.0_client` in the current folder, with the following format: After the detection model is converted, there will be additional folders of `ppocr_det_mobile_2.0_serving` and `ppocr_det_mobile_2.0_client` in the current folder, with the following format:
``` ```
|- ppocr_det_server_2.0_serving/ |- ppocr_det_mobile_2.0_serving/
|- __model__ |- __model__
|- __params__ |- __params__
|- serving_server_conf.prototxt |- serving_server_conf.prototxt
|- serving_server_conf.stream.prototxt |- serving_server_conf.stream.prototxt
|- ppocr_det_server_2.0_client |- ppocr_det_mobile_2.0_client
|- serving_client_conf.prototxt |- serving_client_conf.prototxt
|- serving_client_conf.stream.prototxt |- serving_client_conf.stream.prototxt
...@@ -147,6 +141,89 @@ The recognition model is the same. ...@@ -147,6 +141,89 @@ The recognition model is the same.
After successfully running, the predicted result of the model will be printed in the cmd window. An example of the result is: After successfully running, the predicted result of the model will be printed in the cmd window. An example of the result is:
![](./imgs/results.png) ![](./imgs/results.png)
Adjust the number of concurrency in config.yml to get the largest QPS. Generally, the number of concurrent detection and recognition is 2:1
```
det:
concurrency: 8
...
rec:
concurrency: 4
...
```
Multiple service requests can be sent at the same time if necessary.
The predicted performance data will be automatically written into the `PipelineServingLogs/pipeline.tracer` file.
Tested on 200 real pictures, and limited the detection long side to 960. The average QPS on T4 GPU can reach around 23:
```
2021-05-13 03:42:36,895 ==================== TRACER ======================
2021-05-13 03:42:36,975 Op(rec):
2021-05-13 03:42:36,976 in[14.472382882882883 ms]
2021-05-13 03:42:36,976 prep[9.556855855855856 ms]
2021-05-13 03:42:36,976 midp[59.921905405405404 ms]
2021-05-13 03:42:36,976 postp[15.345945945945946 ms]
2021-05-13 03:42:36,976 out[1.9921216216216215 ms]
2021-05-13 03:42:36,976 idle[0.16254943864471572]
2021-05-13 03:42:36,976 Op(det):
2021-05-13 03:42:36,976 in[315.4468035714286 ms]
2021-05-13 03:42:36,976 prep[69.5980625 ms]
2021-05-13 03:42:36,976 midp[18.989535714285715 ms]
2021-05-13 03:42:36,976 postp[18.857803571428573 ms]
2021-05-13 03:42:36,977 out[3.1337544642857145 ms]
2021-05-13 03:42:36,977 idle[0.7477961159203756]
2021-05-13 03:42:36,977 DAGExecutor:
2021-05-13 03:42:36,977 Query count[224]
2021-05-13 03:42:36,977 QPS[22.4 q/s]
2021-05-13 03:42:36,977 Succ[0.9910714285714286]
2021-05-13 03:42:36,977 Error req[169, 170]
2021-05-13 03:42:36,977 Latency:
2021-05-13 03:42:36,977 ave[535.1678348214285 ms]
2021-05-13 03:42:36,977 .50[172.651 ms]
2021-05-13 03:42:36,977 .60[187.904 ms]
2021-05-13 03:42:36,977 .70[245.675 ms]
2021-05-13 03:42:36,977 .80[526.684 ms]
2021-05-13 03:42:36,977 .90[854.596 ms]
2021-05-13 03:42:36,977 .95[1722.728 ms]
2021-05-13 03:42:36,977 .99[3990.292 ms]
2021-05-13 03:42:36,978 Channel (server worker num[10]):
2021-05-13 03:42:36,978 chl0(In: ['@DAGExecutor'], Out: ['det']) size[0/0]
2021-05-13 03:42:36,979 chl1(In: ['det'], Out: ['rec']) size[6/0]
2021-05-13 03:42:36,979 chl2(In: ['rec'], Out: ['@DAGExecutor']) size[0/0]
```
## WINDOWS Users
Windows does not support Pipeline Serving, if we want to lauch paddle serving on Windows, we should use Web Service, for more infomation please refer to [Paddle Serving for Windows Users](https://github.com/PaddlePaddle/Serving/blob/develop/doc/WINDOWS_TUTORIAL.md)
**WINDOWS user can only use version 0.5.0 CPU Mode**
**Prepare Stage:**
```
pip3 install paddle-serving-server==0.5.0
pip3 install paddle-serving-app==0.3.1
```
1. Start Server
```
cd win
python3 ocr_web_server.py gpu(for gpu user)
or
python3 ocr_web_server.py cpu(for cpu user)
```
2. Client Send Requests
```
python3 ocr_web_client.py
```
<a name="faq"></a> <a name="faq"></a>
## FAQ ## FAQ
**Q1**: No result return after sending the request. **Q1**: No result return after sending the request.
......
deploy/pdserving/README_CN.md
View file @ 19eb7eb8
...@@ -29,41 +29,31 @@ PaddleOCR提供2种服务部署方式： ...@@ -29,41 +29,31 @@ PaddleOCR提供2种服务部署方式：
需要准备PaddleOCR的运行环境和Paddle Serving的运行环境。 需要准备PaddleOCR的运行环境和Paddle Serving的运行环境。
- 准备PaddleOCR的运行环境参考[链接](../../doc/doc_ch/installation.md) - 准备PaddleOCR的运行环境[链接](../../doc/doc_ch/installation.md)
根据环境下载对应的paddle whl包，推荐安装2.0.1版本
- 准备PaddleServing的运行环境，步骤如下 - 准备PaddleServing的运行环境，步骤如下
1. 安装serving，用于启动服务 1. 安装serving，用于启动服务
``` ```
pip3 install paddle-serving-server==0.5.0 # for CPU pip3 install paddle-serving-server==0.6.1 # for CPU
pip3 install paddle-serving-server-gpu==0.5.0 # for GPU pip3 install paddle-serving-server-gpu==0.6.1 # for GPU
# 其他GPU环境需要确认环境再选择执行如下命令 # 其他GPU环境需要确认环境再选择执行如下命令
pip3 install paddle-serving-server-gpu==0.5.0.post9 # GPU with CUDA9.0 pip3 install paddle-serving-server-gpu==0.6.1.post101 # GPU with CUDA10.1 + TensorRT6
pip3 install paddle-serving-server-gpu==0.5.0.post10 # GPU with CUDA10.0 pip3 install paddle-serving-server-gpu==0.6.1.post11 # GPU with CUDA11 + TensorRT7
pip3 install paddle-serving-server-gpu==0.5.0.post101 # GPU with CUDA10.1 + TensorRT6
pip3 install paddle-serving-server-gpu==0.5.0.post11 # GPU with CUDA10.1 + TensorRT7
``` ```
2. 安装client，用于向服务发送请求 2. 安装client，用于向服务发送请求
``` 在[下载链接](https://github.com/PaddlePaddle/Serving/blob/develop/doc/LATEST_PACKAGES.md)中找到对应python版本的client安装包，这里推荐python3.7版本：
pip3 install paddle-serving-client==0.5.0 # for CPU
pip3 install paddle-serving-client-gpu==0.5.0 # for GPU ```
wget https://paddle-serving.bj.bcebos.com/test-dev/whl/paddle_serving_client-0.0.0-cp37-none-any.whl
pip3 install paddle_serving_client-0.0.0-cp37-none-any.whl
``` ```
3. 安装serving-app 3. 安装serving-app
``` ```
pip3 install paddle-serving-app==0.3.0 pip3 install paddle-serving-app==0.6.1
```
**note:** 安装0.3.0版本的serving-app后，为了能加载动态图模型，需要修改serving_app的源码，具体为：
```
# 找到paddle_serving_app的安装目录，找到并编辑local_predict.py文件
vim /usr/local/lib/python3.7/site-packages/paddle_serving_app/local_predict.py
# 将local_predict.py 的第85行 config = AnalysisConfig(model_path) 替换为：
if os.path.exists(os.path.join(model_path, "__params__")):
config = AnalysisConfig(os.path.join(model_path, "__model__"), os.path.join(model_path, "__params__"))
else:
config = AnalysisConfig(model_path)
``` ```
**Note:** 如果要安装最新版本的PaddleServing参考[链接](https://github.com/PaddlePaddle/Serving/blob/develop/doc/LATEST_PACKAGES.md)。 **Note:** 如果要安装最新版本的PaddleServing参考[链接](https://github.com/PaddlePaddle/Serving/blob/develop/doc/LATEST_PACKAGES.md)。
...@@ -76,38 +66,38 @@ PaddleOCR提供2种服务部署方式： ...@@ -76,38 +66,38 @@ PaddleOCR提供2种服务部署方式：
首先，下载PPOCR的[inference模型](https://github.com/PaddlePaddle/PaddleOCR#pp-ocr-20-series-model-listupdate-on-dec-15) 首先，下载PPOCR的[inference模型](https://github.com/PaddlePaddle/PaddleOCR#pp-ocr-20-series-model-listupdate-on-dec-15)
``` ```
# 下载并解压 OCR 文本检测模型 # 下载并解压 OCR 文本检测模型
wget https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_server_v2.0_det_infer.tar && tar xf ch_ppocr_server_v2.0_det_infer.tar wget https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_det_infer.tar && tar xf ch_ppocr_mobile_v2.0_det_infer.tar
# 下载并解压 OCR 文本识别模型 # 下载并解压 OCR 文本识别模型
wget https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_server_v2.0_rec_infer.tar && tar xf ch_ppocr_server_v2.0_rec_infer.tar wget https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_rec_infer.tar && tar xf ch_ppocr_mobile_v2.0_rec_infer.tar
``` ```
接下来，用安装的paddle_serving_client把下载的inference模型转换成易于server部署的模型格式。 接下来，用安装的paddle_serving_client把下载的inference模型转换成易于server部署的模型格式。
``` ```
# 转换检测模型 # 转换检测模型
python3 -m paddle_serving_client.convert --dirname ./ch_ppocr_server_v2.0_det_infer/ \ python3 -m paddle_serving_client.convert --dirname ./ch_ppocr_mobile_v2.0_det_infer/ \
--model_filename inference.pdmodel \ --model_filename inference.pdmodel \
--params_filename inference.pdiparams \ --params_filename inference.pdiparams \
--serving_server ./ppocr_det_server_2.0_serving/ \ --serving_server ./ppocr_det_mobile_2.0_serving/ \
--serving_client ./ppocr_det_server_2.0_client/ --serving_client ./ppocr_det_mobile_2.0_client/
# 转换识别模型 # 转换识别模型
python3 -m paddle_serving_client.convert --dirname ./ch_ppocr_server_v2.0_rec_infer/ \ python3 -m paddle_serving_client.convert --dirname ./ch_ppocr_mobile_v2.0_rec_infer/ \
--model_filename inference.pdmodel \ --model_filename inference.pdmodel \
--params_filename inference.pdiparams \ --params_filename inference.pdiparams \
--serving_server ./ppocr_rec_server_2.0_serving/ \ --serving_server ./ppocr_rec_mobile_2.0_serving/ \
--serving_client ./ppocr_rec_server_2.0_client/ --serving_client ./ppocr_rec_mobile_2.0_client/
``` ```
检测模型转换完成后，会在当前文件夹多出`ppocr_det_server_2.0_serving` 和`ppocr_det_server_2.0_client`的文件夹，具备如下格式： 检测模型转换完成后，会在当前文件夹多出`ppocr_det_mobile_2.0_serving` 和`ppocr_det_mobile_2.0_client`的文件夹，具备如下格式：
``` ```
|- ppocr_det_server_2.0_serving/ |- ppocr_det_mobile_2.0_serving/
|- __model__ |- __model__
|- __params__ |- __params__
|- serving_server_conf.prototxt |- serving_server_conf.prototxt
|- serving_server_conf.stream.prototxt |- serving_server_conf.stream.prototxt
|- ppocr_det_server_2.0_client |- ppocr_det_mobile_2.0_client
|- serving_client_conf.prototxt |- serving_client_conf.prototxt
|- serving_client_conf.stream.prototxt |- serving_client_conf.stream.prototxt
...@@ -148,6 +138,86 @@ python3 -m paddle_serving_client.convert --dirname ./ch_ppocr_server_v2.0_rec_in ...@@ -148,6 +138,86 @@ python3 -m paddle_serving_client.convert --dirname ./ch_ppocr_server_v2.0_rec_in
成功运行后，模型预测的结果会打印在cmd窗口中，结果示例为： 成功运行后，模型预测的结果会打印在cmd窗口中，结果示例为：
![](./imgs/results.png) ![](./imgs/results.png)
调整 config.yml 中的并发个数获得最大的QPS, 一般检测和识别的并发数为2：1
```
det:
#并发数，is_thread_op=True时，为线程并发；否则为进程并发
concurrency: 8
...
rec:
#并发数，is_thread_op=True时，为线程并发；否则为进程并发
concurrency: 4
...
```
有需要的话可以同时发送多个服务请求
预测性能数据会被自动写入 `PipelineServingLogs/pipeline.tracer` 文件中。
在200张真实图片上测试，把检测长边限制为960。T4 GPU 上 QPS 均值可达到23左右：
```
2021-05-13 03:42:36,895 ==================== TRACER ======================
2021-05-13 03:42:36,975 Op(rec):
2021-05-13 03:42:36,976 in[14.472382882882883 ms]
2021-05-13 03:42:36,976 prep[9.556855855855856 ms]
2021-05-13 03:42:36,976 midp[59.921905405405404 ms]
2021-05-13 03:42:36,976 postp[15.345945945945946 ms]
2021-05-13 03:42:36,976 out[1.9921216216216215 ms]
2021-05-13 03:42:36,976 idle[0.16254943864471572]
2021-05-13 03:42:36,976 Op(det):
2021-05-13 03:42:36,976 in[315.4468035714286 ms]
2021-05-13 03:42:36,976 prep[69.5980625 ms]
2021-05-13 03:42:36,976 midp[18.989535714285715 ms]
2021-05-13 03:42:36,976 postp[18.857803571428573 ms]
2021-05-13 03:42:36,977 out[3.1337544642857145 ms]
2021-05-13 03:42:36,977 idle[0.7477961159203756]
2021-05-13 03:42:36,977 DAGExecutor:
2021-05-13 03:42:36,977 Query count[224]
2021-05-13 03:42:36,977 QPS[22.4 q/s]
2021-05-13 03:42:36,977 Succ[0.9910714285714286]
2021-05-13 03:42:36,977 Error req[169, 170]
2021-05-13 03:42:36,977 Latency:
2021-05-13 03:42:36,977 ave[535.1678348214285 ms]
2021-05-13 03:42:36,977 .50[172.651 ms]
2021-05-13 03:42:36,977 .60[187.904 ms]
2021-05-13 03:42:36,977 .70[245.675 ms]
2021-05-13 03:42:36,977 .80[526.684 ms]
2021-05-13 03:42:36,977 .90[854.596 ms]
2021-05-13 03:42:36,977 .95[1722.728 ms]
2021-05-13 03:42:36,977 .99[3990.292 ms]
2021-05-13 03:42:36,978 Channel (server worker num[10]):
2021-05-13 03:42:36,978 chl0(In: ['@DAGExecutor'], Out: ['det']) size[0/0]
2021-05-13 03:42:36,979 chl1(In: ['det'], Out: ['rec']) size[6/0]
2021-05-13 03:42:36,979 chl2(In: ['rec'], Out: ['@DAGExecutor']) size[0/0]
```
## WINDOWS用户
Windows用户不能使用上述的启动方式，需要使用Web Service，详情参见[Windows平台使用Paddle Serving指导](https://github.com/PaddlePaddle/Serving/blob/develop/doc/WINDOWS_TUTORIAL_CN.md)
**WINDOWS只能使用0.5.0版本的CPU模式**
准备阶段：
```
pip3 install paddle-serving-server==0.5.0
pip3 install paddle-serving-app==0.3.1
```
1. 启动服务端程序
```
cd win
python3 ocr_web_server.py gpu(使用gpu方式)
或者
python3 ocr_web_server.py cpu(使用cpu方式)
```
2. 发送服务请求
```
python3 ocr_web_client.py
```
<a name="FAQ"></a> <a name="FAQ"></a>
## FAQ ## FAQ
......
deploy/pdserving/config.yml
View file @ 19eb7eb8
#rpc端口, rpc_port和http_port不允许同时为空。当rpc_port为空且http_port不为空时，会自动将rpc_port设置为http_port+1 #rpc端口, rpc_port和http_port不允许同时为空。当rpc_port为空且http_port不为空时，会自动将rpc_port设置为http_port+1
rpc_port: 18090 rpc_port: 18091
#http端口, rpc_port和http_port不允许同时为空。当rpc_port可用且http_port为空时，不自动生成http_port #http端口, rpc_port和http_port不允许同时为空。当rpc_port可用且http_port为空时，不自动生成http_port
http_port: 9999 http_port: 9998
#worker_num, 最大并发数。当build_dag_each_worker=True时, 框架会创建worker_num个进程，每个进程内构建grpcSever和DAG #worker_num, 最大并发数。当build_dag_each_worker=True时, 框架会创建worker_num个进程，每个进程内构建grpcSever和DAG
##当build_dag_each_worker=False时，框架会设置主线程grpc线程池的max_workers=worker_num ##当build_dag_each_worker=False时，框架会设置主线程grpc线程池的max_workers=worker_num
worker_num: 20 worker_num: 10
#build_dag_each_worker, False，框架在进程内创建一条DAG；True，框架会每个进程内创建多个独立的DAG #build_dag_each_worker, False，框架在进程内创建一条DAG；True，框架会每个进程内创建多个独立的DAG
build_dag_each_worker: false build_dag_each_worker: False
dag: dag:
#op资源类型, True, 为线程模型；False，为进程模型 #op资源类型, True, 为线程模型；False，为进程模型
is_thread_op: False is_thread_op: False
#重试次数 #重试次数
retry: 1 retry: 10
#使用性能分析, True，生成Timeline性能数据，对性能有一定影响；False为不使用 #使用性能分析, True，生成Timeline性能数据，对性能有一定影响；False为不使用
use_profile: False use_profile: True
tracer: tracer:
interval_s: 10 interval_s: 10
op: op:
det: det:
#并发数，is_thread_op=True时，为线程并发；否则为进程并发 #并发数，is_thread_op=True时，为线程并发；否则为进程并发
concurrency: 4 concurrency: 8
#当op配置没有server_endpoints时，从local_service_conf读取本地服务配置 #当op配置没有server_endpoints时，从local_service_conf读取本地服务配置
local_service_conf: local_service_conf:
...@@ -34,18 +34,18 @@ op: ...@@ -34,18 +34,18 @@ op:
client_type: local_predictor client_type: local_predictor
#det模型路径 #det模型路径
model_config: /paddle/serving/models/det_serving_server/ #ocr_det_model model_config: ./ppocr_det_mobile_2.0_serving
#Fetch结果列表，以client_config中fetch_var的alias_name为准 #Fetch结果列表，以client_config中fetch_var的alias_name为准
fetch_list: ["save_infer_model/scale_0.tmp_1"] fetch_list: ["save_infer_model/scale_0.tmp_1"]
#计算硬件ID，当devices为""或不写时为CPU预测；当devices为"0", "0,1,2"时为GPU预测，表示使用的GPU卡 #计算硬件ID，当devices为""或不写时为CPU预测；当devices为"0", "0,1,2"时为GPU预测，表示使用的GPU卡
devices: "2" devices: "0"
ir_optim: True ir_optim: True
rec: rec:
#并发数，is_thread_op=True时，为线程并发；否则为进程并发 #并发数，is_thread_op=True时，为线程并发；否则为进程并发
concurrency: 1 concurrency: 4
#超时时间, 单位ms #超时时间, 单位ms
timeout: -1 timeout: -1
...@@ -60,12 +60,12 @@ op: ...@@ -60,12 +60,12 @@ op:
client_type: local_predictor client_type: local_predictor
#rec模型路径 #rec模型路径
model_config: /paddle/serving/models/rec_serving_server/ #ocr_rec_model model_config: ./ppocr_rec_mobile_2.0_serving
#Fetch结果列表，以client_config中fetch_var的alias_name为准 #Fetch结果列表，以client_config中fetch_var的alias_name为准
fetch_list: ["save_infer_model/scale_0.tmp_1"] #["ctc_greedy_decoder_0.tmp_0", "softmax_0.tmp_0"] fetch_list: ["save_infer_model/scale_0.tmp_1"]
#计算硬件ID，当devices为""或不写时为CPU预测；当devices为"0", "0,1,2"时为GPU预测，表示使用的GPU卡 #计算硬件ID，当devices为""或不写时为CPU预测；当devices为"0", "0,1,2"时为GPU预测，表示使用的GPU卡
devices: "2" devices: "0"
ir_optim: True ir_optim: True
deploy/pdserving/ocr_reader.py
View file @ 19eb7eb8
...@@ -21,7 +21,6 @@ import sys ...@@ -21,7 +21,6 @@ import sys
import argparse import argparse
import string import string
from copy import deepcopy from copy import deepcopy
import paddle
class DetResizeForTest(object): class DetResizeForTest(object):
...@@ -34,12 +33,12 @@ class DetResizeForTest(object): ...@@ -34,12 +33,12 @@ class DetResizeForTest(object):
elif 'limit_side_len' in kwargs: elif 'limit_side_len' in kwargs:
self.limit_side_len = kwargs['limit_side_len'] self.limit_side_len = kwargs['limit_side_len']
self.limit_type = kwargs.get('limit_type', 'min') self.limit_type = kwargs.get('limit_type', 'min')
elif 'resize_long' in kwargs: elif 'resize_short' in kwargs:
self.resize_type = 2
self.resize_long = kwargs.get('resize_long', 960)
else:
self.limit_side_len = 736 self.limit_side_len = 736
self.limit_type = 'min' self.limit_type = 'min'
else:
self.resize_type = 2
self.resize_long = kwargs.get('resize_long', 960)
def __call__(self, data): def __call__(self, data):
img = deepcopy(data) img = deepcopy(data)
...@@ -227,8 +226,6 @@ class CTCLabelDecode(BaseRecLabelDecode): ...@@ -227,8 +226,6 @@ class CTCLabelDecode(BaseRecLabelDecode):
super(CTCLabelDecode, self).__init__(config) super(CTCLabelDecode, self).__init__(config)
def __call__(self, preds, label=None, *args, **kwargs): def __call__(self, preds, label=None, *args, **kwargs):
if isinstance(preds, paddle.Tensor):
preds = preds.numpy()
preds_idx = preds.argmax(axis=2) preds_idx = preds.argmax(axis=2)
preds_prob = preds.max(axis=2) preds_prob = preds.max(axis=2)
text = self.decode(preds_idx, preds_prob, is_remove_duplicate=True) text = self.decode(preds_idx, preds_prob, is_remove_duplicate=True)
......
deploy/pdserving/pipeline_http_client.py
View file @ 19eb7eb8
...@@ -23,8 +23,8 @@ def cv2_to_base64(image): ...@@ -23,8 +23,8 @@ def cv2_to_base64(image):
return base64.b64encode(image).decode('utf8') return base64.b64encode(image).decode('utf8')
url = "http://127.0.0.1:9999/ocr/prediction" url = "http://127.0.0.1:9998/ocr/prediction"
test_img_dir = "../doc/imgs/" test_img_dir = "../../doc/imgs/"
for idx, img_file in enumerate(os.listdir(test_img_dir)): for idx, img_file in enumerate(os.listdir(test_img_dir)):
with open(os.path.join(test_img_dir, img_file), 'rb') as file: with open(os.path.join(test_img_dir, img_file), 'rb') as file:
image_data1 = file.read() image_data1 = file.read()
...@@ -36,5 +36,5 @@ for idx, img_file in enumerate(os.listdir(test_img_dir)): ...@@ -36,5 +36,5 @@ for idx, img_file in enumerate(os.listdir(test_img_dir)):
r = requests.post(url=url, data=json.dumps(data)) r = requests.post(url=url, data=json.dumps(data))
print(r.json()) print(r.json())
test_img_dir = "../doc/imgs/" test_img_dir = "../../doc/imgs/"
print("==> total number of test imgs: ", len(os.listdir(test_img_dir))) print("==> total number of test imgs: ", len(os.listdir(test_img_dir)))
deploy/pdserving/pipeline_rpc_client.py
View file @ 19eb7eb8
...@@ -23,7 +23,7 @@ import base64 ...@@ -23,7 +23,7 @@ import base64
import os import os
client = PipelineClient() client = PipelineClient()
client.connect(['127.0.0.1:18090']) client.connect(['127.0.0.1:18091'])
def cv2_to_base64(image): def cv2_to_base64(image):
...@@ -39,4 +39,3 @@ for img_file in os.listdir(test_img_dir): ...@@ -39,4 +39,3 @@ for img_file in os.listdir(test_img_dir):
for i in range(1): for i in range(1):
ret = client.predict(feed_dict={"image": image}, fetch=["res"]) ret = client.predict(feed_dict={"image": image}, fetch=["res"])
print(ret) print(ret)
#print(ret)
deploy/pdserving/web_service.py
View file @ 19eb7eb8
...@@ -11,10 +11,7 @@ ...@@ -11,10 +11,7 @@
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and # See the License for the specific language governing permissions and
# limitations under the License. # limitations under the License.
try: from paddle_serving_server.web_service import WebService, Op
from paddle_serving_server_gpu.web_service import WebService, Op
except ImportError:
from paddle_serving_server.web_service import WebService, Op
import logging import logging
import numpy as np import numpy as np
...@@ -48,28 +45,24 @@ class DetOp(Op): ...@@ -48,28 +45,24 @@ class DetOp(Op):
def preprocess(self, input_dicts, data_id, log_id): def preprocess(self, input_dicts, data_id, log_id):
(_, input_dict), = input_dicts.items() (_, input_dict), = input_dicts.items()
data = base64.b64decode(input_dict["image"].encode('utf8')) data = base64.b64decode(input_dict["image"].encode('utf8'))
self.raw_im = data
data = np.fromstring(data, np.uint8) data = np.fromstring(data, np.uint8)
# Note: class variables(self.var) can only be used in process op mode # Note: class variables(self.var) can only be used in process op mode
im = cv2.imdecode(data, cv2.IMREAD_COLOR) im = cv2.imdecode(data, cv2.IMREAD_COLOR)
self.im = im
self.ori_h, self.ori_w, _ = im.shape self.ori_h, self.ori_w, _ = im.shape
det_img = self.det_preprocess(im)
det_img = self.det_preprocess(self.im)
_, self.new_h, self.new_w = det_img.shape _, self.new_h, self.new_w = det_img.shape
print("det image shape", det_img.shape)
return {"x": det_img[np.newaxis, :].copy()}, False, None, "" return {"x": det_img[np.newaxis, :].copy()}, False, None, ""
def postprocess(self, input_dicts, fetch_dict, log_id): def postprocess(self, input_dicts, fetch_dict, log_id):
print("input_dicts: ", input_dicts)
det_out = fetch_dict["save_infer_model/scale_0.tmp_1"] det_out = fetch_dict["save_infer_model/scale_0.tmp_1"]
ratio_list = [ ratio_list = [
float(self.new_h) / self.ori_h, float(self.new_w) / self.ori_w float(self.new_h) / self.ori_h, float(self.new_w) / self.ori_w
] ]
dt_boxes_list = self.post_func(det_out, [ratio_list]) dt_boxes_list = self.post_func(det_out, [ratio_list])
dt_boxes = self.filter_func(dt_boxes_list[0], [self.ori_h, self.ori_w]) dt_boxes = self.filter_func(dt_boxes_list[0], [self.ori_h, self.ori_w])
out_dict = {"dt_boxes": dt_boxes, "image": self.im} out_dict = {"dt_boxes": dt_boxes, "image": self.raw_im}
print("out dict", out_dict["dt_boxes"])
return out_dict, None, "" return out_dict, None, ""
...@@ -83,35 +76,75 @@ class RecOp(Op): ...@@ -83,35 +76,75 @@ class RecOp(Op):
def preprocess(self, input_dicts, data_id, log_id): def preprocess(self, input_dicts, data_id, log_id):
(_, input_dict), = input_dicts.items() (_, input_dict), = input_dicts.items()
im = input_dict["image"] raw_im = input_dict["image"]
data = np.frombuffer(raw_im, np.uint8)
im = cv2.imdecode(data, cv2.IMREAD_COLOR)
dt_boxes = input_dict["dt_boxes"] dt_boxes = input_dict["dt_boxes"]
dt_boxes = self.sorted_boxes(dt_boxes) dt_boxes = self.sorted_boxes(dt_boxes)
feed_list = [] feed_list = []
img_list = [] img_list = []
max_wh_ratio = 0 max_wh_ratio = 0
for i, dtbox in enumerate(dt_boxes): ## Many mini-batchs, the type of feed_data is list.
boximg = self.get_rotate_crop_image(im, dt_boxes[i]) max_batch_size = 6 # len(dt_boxes)
img_list.append(boximg)
h, w = boximg.shape[0:2] # If max_batch_size is 0, skipping predict stage
wh_ratio = w * 1.0 / h if max_batch_size == 0:
max_wh_ratio = max(max_wh_ratio, wh_ratio) return {}, True, None, ""
_, w, h = self.ocr_reader.resize_norm_img(img_list[0], boxes_size = len(dt_boxes)
max_wh_ratio).shape batch_size = boxes_size // max_batch_size
rem = boxes_size % max_batch_size
imgs = np.zeros((len(img_list), 3, w, h)).astype('float32') for bt_idx in range(0, batch_size + 1):
for id, img in enumerate(img_list): imgs = None
norm_img = self.ocr_reader.resize_norm_img(img, max_wh_ratio) boxes_num_in_one_batch = 0
imgs[id] = norm_img if bt_idx == batch_size:
print("rec image shape", imgs.shape) if rem == 0:
feed = {"x": imgs.copy()} continue
return feed, False, None, "" else:
boxes_num_in_one_batch = rem
def postprocess(self, input_dicts, fetch_dict, log_id): elif bt_idx < batch_size:
rec_res = self.ocr_reader.postprocess(fetch_dict, with_score=True) boxes_num_in_one_batch = max_batch_size
res_lst = [] else:
for res in rec_res: _LOGGER.error("batch_size error, bt_idx={}, batch_size={}".
res_lst.append(res[0]) format(bt_idx, batch_size))
res = {"res": str(res_lst)} break
start = bt_idx * max_batch_size
end = start + boxes_num_in_one_batch
img_list = []
for box_idx in range(start, end):
boximg = self.get_rotate_crop_image(im, dt_boxes[box_idx])
img_list.append(boximg)
h, w = boximg.shape[0:2]
wh_ratio = w * 1.0 / h
max_wh_ratio = max(max_wh_ratio, wh_ratio)
_, w, h = self.ocr_reader.resize_norm_img(img_list[0],
max_wh_ratio).shape
imgs = np.zeros((boxes_num_in_one_batch, 3, w, h)).astype('float32')
for id, img in enumerate(img_list):
norm_img = self.ocr_reader.resize_norm_img(img, max_wh_ratio)
imgs[id] = norm_img
feed = {"x": imgs.copy()}
feed_list.append(feed)
return feed_list, False, None, ""
def postprocess(self, input_dicts, fetch_data, log_id):
res_list = []
if isinstance(fetch_data, dict):
if len(fetch_data) > 0:
rec_batch_res = self.ocr_reader.postprocess(
fetch_data, with_score=True)
for res in rec_batch_res:
res_list.append(res[0])
elif isinstance(fetch_data, list):
for one_batch in fetch_data:
one_batch_res = self.ocr_reader.postprocess(
one_batch, with_score=True)
for res in one_batch_res:
res_list.append(res[0])
res = {"res": str(res_list)}
return res, None, "" return res, None, ""
......
deploy/pdserving/win/ocr_reader.py 0 → 100644
View file @ 19eb7eb8
# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import cv2
import copy
import numpy as np
import math
import re
import sys
import argparse
import string
from copy import deepcopy
class DetResizeForTest(object):
def __init__(self, **kwargs):
super(DetResizeForTest, self).__init__()
self.resize_type = 0
if 'image_shape' in kwargs:
self.image_shape = kwargs['image_shape']
self.resize_type = 1
elif 'limit_side_len' in kwargs:
self.limit_side_len = kwargs['limit_side_len']
self.limit_type = kwargs.get('limit_type', 'min')
elif 'resize_short' in kwargs:
self.limit_side_len = 736
self.limit_type = 'min'
else:
self.resize_type = 2
self.resize_long = kwargs.get('resize_long', 960)
def __call__(self, data):
img = deepcopy(data)
src_h, src_w, _ = img.shape
if self.resize_type == 0:
img, [ratio_h, ratio_w] = self.resize_image_type0(img)
elif self.resize_type == 2:
img, [ratio_h, ratio_w] = self.resize_image_type2(img)
else:
img, [ratio_h, ratio_w] = self.resize_image_type1(img)
return img
def resize_image_type1(self, img):
resize_h, resize_w = self.image_shape
ori_h, ori_w = img.shape[:2] # (h, w, c)
ratio_h = float(resize_h) / ori_h
ratio_w = float(resize_w) / ori_w
img = cv2.resize(img, (int(resize_w), int(resize_h)))
return img, [ratio_h, ratio_w]
def resize_image_type0(self, img):
"""
resize image to a size multiple of 32 which is required by the network
args:
img(array): array with shape [h, w, c]
return(tuple):
img, (ratio_h, ratio_w)
"""
limit_side_len = self.limit_side_len
h, w, _ = img.shape
# limit the max side
if self.limit_type == 'max':
if max(h, w) > limit_side_len:
if h > w:
ratio = float(limit_side_len) / h
else:
ratio = float(limit_side_len) / w
else:
ratio = 1.
else:
if min(h, w) < limit_side_len:
if h < w:
ratio = float(limit_side_len) / h
else:
ratio = float(limit_side_len) / w
else:
ratio = 1.
resize_h = int(h * ratio)
resize_w = int(w * ratio)
resize_h = int(round(resize_h / 32) * 32)
resize_w = int(round(resize_w / 32) * 32)
try:
if int(resize_w) <= 0 or int(resize_h) <= 0:
return None, (None, None)
img = cv2.resize(img, (int(resize_w), int(resize_h)))
except:
print(img.shape, resize_w, resize_h)
sys.exit(0)
ratio_h = resize_h / float(h)
ratio_w = resize_w / float(w)
# return img, np.array([h, w])
return img, [ratio_h, ratio_w]
def resize_image_type2(self, img):
h, w, _ = img.shape
resize_w = w
resize_h = h
# Fix the longer side
if resize_h > resize_w:
ratio = float(self.resize_long) / resize_h
else:
ratio = float(self.resize_long) / resize_w
resize_h = int(resize_h * ratio)
resize_w = int(resize_w * ratio)
max_stride = 128
resize_h = (resize_h + max_stride - 1) // max_stride * max_stride
resize_w = (resize_w + max_stride - 1) // max_stride * max_stride
img = cv2.resize(img, (int(resize_w), int(resize_h)))
ratio_h = resize_h / float(h)
ratio_w = resize_w / float(w)
return img, [ratio_h, ratio_w]
class BaseRecLabelDecode(object):
""" Convert between text-label and text-index """
def __init__(self, config):
support_character_type = [
'ch', 'en', 'EN_symbol', 'french', 'german', 'japan', 'korean',
'it', 'xi', 'pu', 'ru', 'ar', 'ta', 'ug', 'fa', 'ur', 'rs', 'oc',
'rsc', 'bg', 'uk', 'be', 'te', 'ka', 'chinese_cht', 'hi', 'mr',
'ne', 'EN'
]
character_type = config['character_type']
character_dict_path = config['character_dict_path']
use_space_char = True
assert character_type in support_character_type, "Only {} are supported now but get {}".format(
support_character_type, character_type)
self.beg_str = "sos"
self.end_str = "eos"
if character_type == "en":
self.character_str = "0123456789abcdefghijklmnopqrstuvwxyz"
dict_character = list(self.character_str)
elif character_type == "EN_symbol":
# same with ASTER setting (use 94 char).
self.character_str = string.printable[:-6]
dict_character = list(self.character_str)
elif character_type in support_character_type:
self.character_str = ""
assert character_dict_path is not None, "character_dict_path should not be None when character_type is {}".format(
character_type)
with open(character_dict_path, "rb") as fin:
lines = fin.readlines()
for line in lines:
line = line.decode('utf-8').strip("\n").strip("\r\n")
self.character_str += line
if use_space_char:
self.character_str += " "
dict_character = list(self.character_str)
else:
raise NotImplementedError
self.character_type = character_type
dict_character = self.add_special_char(dict_character)
self.dict = {}
for i, char in enumerate(dict_character):
self.dict[char] = i
self.character = dict_character
def add_special_char(self, dict_character):
return dict_character
def decode(self, text_index, text_prob=None, is_remove_duplicate=False):
""" convert text-index into text-label. """
result_list = []
ignored_tokens = self.get_ignored_tokens()
batch_size = len(text_index)
for batch_idx in range(batch_size):
char_list = []
conf_list = []
for idx in range(len(text_index[batch_idx])):
if text_index[batch_idx][idx] in ignored_tokens:
continue
if is_remove_duplicate:
# only for predict
if idx > 0 and text_index[batch_idx][idx - 1] == text_index[
batch_idx][idx]:
continue
char_list.append(self.character[int(text_index[batch_idx][
idx])])
if text_prob is not None:
conf_list.append(text_prob[batch_idx][idx])
else:
conf_list.append(1)
text = ''.join(char_list)
result_list.append((text, np.mean(conf_list)))
return result_list
def get_ignored_tokens(self):
return [0] # for ctc blank
class CTCLabelDecode(BaseRecLabelDecode):
""" Convert between text-label and text-index """
def __init__(
self,
config,
#character_dict_path=None,
#character_type='ch',
#use_space_char=False,
**kwargs):
super(CTCLabelDecode, self).__init__(config)
def __call__(self, preds, label=None, *args, **kwargs):
preds_idx = preds.argmax(axis=2)
preds_prob = preds.max(axis=2)
text = self.decode(preds_idx, preds_prob, is_remove_duplicate=True)
if label is None:
return text
label = self.decode(label)
return text, label
def add_special_char(self, dict_character):
dict_character = ['blank'] + dict_character
return dict_character
class CharacterOps(object):
""" Convert between text-label and text-index """
def __init__(self, config):
self.character_type = config['character_type']
self.loss_type = config['loss_type']
if self.character_type == "en":
self.character_str = "0123456789abcdefghijklmnopqrstuvwxyz"
dict_character = list(self.character_str)
elif self.character_type == "ch":
character_dict_path = config['character_dict_path']
self.character_str = ""
with open(character_dict_path, "rb") as fin:
lines = fin.readlines()
for line in lines:
line = line.decode('utf-8').strip("\n").strip("\r\n")
self.character_str += line
dict_character = list(self.character_str)
elif self.character_type == "en_sensitive":
# same with ASTER setting (use 94 char).
self.character_str = string.printable[:-6]
dict_character = list(self.character_str)
else:
self.character_str = None
assert self.character_str is not None, \
"Nonsupport type of the character: {}".format(self.character_str)
self.beg_str = "sos"
self.end_str = "eos"
if self.loss_type == "attention":
dict_character = [self.beg_str, self.end_str] + dict_character
self.dict = {}
for i, char in enumerate(dict_character):
self.dict[char] = i
self.character = dict_character
def encode(self, text):
"""convert text-label into text-index.
input:
text: text labels of each image. [batch_size]
output:
text: concatenated text index for CTCLoss.
[sum(text_lengths)] = [text_index_0 + text_index_1 + ... + text_index_(n - 1)]
length: length of each text. [batch_size]
"""
if self.character_type == "en":
text = text.lower()
text_list = []
for char in text:
if char not in self.dict:
continue
text_list.append(self.dict[char])
text = np.array(text_list)
return text
def decode(self, text_index, is_remove_duplicate=False):
""" convert text-index into text-label. """
char_list = []
char_num = self.get_char_num()
if self.loss_type == "attention":
beg_idx = self.get_beg_end_flag_idx("beg")
end_idx = self.get_beg_end_flag_idx("end")
ignored_tokens = [beg_idx, end_idx]
else:
ignored_tokens = [char_num]
for idx in range(len(text_index)):
if text_index[idx] in ignored_tokens:
continue
if is_remove_duplicate:
if idx > 0 and text_index[idx - 1] == text_index[idx]:
continue
char_list.append(self.character[text_index[idx]])
text = ''.join(char_list)
return text
def get_char_num(self):
return len(self.character)
def get_beg_end_flag_idx(self, beg_or_end):
if self.loss_type == "attention":
if beg_or_end == "beg":
idx = np.array(self.dict[self.beg_str])
elif beg_or_end == "end":
idx = np.array(self.dict[self.end_str])
else:
assert False, "Unsupport type %s in get_beg_end_flag_idx"\
% beg_or_end
return idx
else:
err = "error in get_beg_end_flag_idx when using the loss %s"\
% (self.loss_type)
assert False, err
class OCRReader(object):
def __init__(self,
algorithm="CRNN",
image_shape=[3, 32, 320],
char_type="ch",
batch_num=1,
char_dict_path="./ppocr_keys_v1.txt"):
self.rec_image_shape = image_shape
self.character_type = char_type
self.rec_batch_num = batch_num
char_ops_params = {}
char_ops_params["character_type"] = char_type
char_ops_params["character_dict_path"] = char_dict_path
char_ops_params['loss_type'] = 'ctc'
self.char_ops = CharacterOps(char_ops_params)
self.label_ops = CTCLabelDecode(char_ops_params)
def resize_norm_img(self, img, max_wh_ratio):
imgC, imgH, imgW = self.rec_image_shape
if self.character_type == "ch":
imgW = int(32 * max_wh_ratio)
h = img.shape[0]
w = img.shape[1]
ratio = w / float(h)
if math.ceil(imgH * ratio) > imgW:
resized_w = imgW
else:
resized_w = int(math.ceil(imgH * ratio))
resized_image = cv2.resize(img, (resized_w, imgH))
resized_image = resized_image.astype('float32')
resized_image = resized_image.transpose((2, 0, 1)) / 255
resized_image -= 0.5
resized_image /= 0.5
padding_im = np.zeros((imgC, imgH, imgW), dtype=np.float32)
padding_im[:, :, 0:resized_w] = resized_image
return padding_im
def preprocess(self, img_list):
img_num = len(img_list)
norm_img_batch = []
max_wh_ratio = 0
for ino in range(img_num):
h, w = img_list[ino].shape[0:2]
wh_ratio = w * 1.0 / h
max_wh_ratio = max(max_wh_ratio, wh_ratio)
for ino in range(img_num):
norm_img = self.resize_norm_img(img_list[ino], max_wh_ratio)
norm_img = norm_img[np.newaxis, :]
norm_img_batch.append(norm_img)
norm_img_batch = np.concatenate(norm_img_batch)
norm_img_batch = norm_img_batch.copy()
return norm_img_batch[0]
def postprocess_old(self, outputs, with_score=False):
rec_res = []
rec_idx_lod = outputs["ctc_greedy_decoder_0.tmp_0.lod"]
rec_idx_batch = outputs["ctc_greedy_decoder_0.tmp_0"]
if with_score:
predict_lod = outputs["softmax_0.tmp_0.lod"]
for rno in range(len(rec_idx_lod) - 1):
beg = rec_idx_lod[rno]
end = rec_idx_lod[rno + 1]
if isinstance(rec_idx_batch, list):
rec_idx_tmp = [x[0] for x in rec_idx_batch[beg:end]]
else: #nd array
rec_idx_tmp = rec_idx_batch[beg:end, 0]
preds_text = self.char_ops.decode(rec_idx_tmp)
if with_score:
beg = predict_lod[rno]
end = predict_lod[rno + 1]
if isinstance(outputs["softmax_0.tmp_0"], list):
outputs["softmax_0.tmp_0"] = np.array(outputs[
"softmax_0.tmp_0"]).astype(np.float32)
probs = outputs["softmax_0.tmp_0"][beg:end, :]
ind = np.argmax(probs, axis=1)
blank = probs.shape[1]
valid_ind = np.where(ind != (blank - 1))[0]
score = np.mean(probs[valid_ind, ind[valid_ind]])
rec_res.append([preds_text, score])
else:
rec_res.append([preds_text])
return rec_res
def postprocess(self, outputs, with_score=False):
preds = outputs["save_infer_model/scale_0.tmp_1"]
try:
preds = preds.numpy()
except:
pass
preds_idx = preds.argmax(axis=2)
preds_prob = preds.max(axis=2)
text = self.label_ops.decode(
preds_idx, preds_prob, is_remove_duplicate=True)
return text
deploy/pdserving/win/ocr_web_client.py 0 → 100644
View file @ 19eb7eb8
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# -*- coding: utf-8 -*-
import requests
import json
import cv2
import base64
import os, sys
import time
def cv2_to_base64(image):
#data = cv2.imencode('.jpg', image)[1]
return base64.b64encode(image).decode(
'utf8') #data.tostring()).decode('utf8')
headers = {"Content-type": "application/json"}
url = "http://127.0.0.1:9292/ocr/prediction"
test_img_dir = "../../../doc/imgs/"
for idx, img_file in enumerate(os.listdir(test_img_dir)):
with open(os.path.join(test_img_dir, img_file), 'rb') as file:
image_data1 = file.read()
image = cv2_to_base64(image_data1)
for i in range(1):
data = {"feed": [{"image": image}], "fetch": ["save_infer_model/scale_0.tmp_1"]}
r = requests.post(url=url, headers=headers, data=json.dumps(data))
print(r.json())
test_img_dir = "../../../doc/imgs/"
print("==> total number of test imgs: ", len(os.listdir(test_img_dir)))
deploy/pdserving/win/ocr_web_server.py 0 → 100644
View file @ 19eb7eb8
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from paddle_serving_client import Client
import cv2
import sys
import numpy as np
import os
from paddle_serving_client import Client
from paddle_serving_app.reader import Sequential, URL2Image, ResizeByFactor
from paddle_serving_app.reader import Div, Normalize, Transpose
from paddle_serving_app.reader import DBPostProcess, FilterBoxes, GetRotateCropImage, SortedBoxes
from ocr_reader import OCRReader
try:
from paddle_serving_server_gpu.web_service import WebService
except ImportError:
from paddle_serving_server.web_service import WebService
from paddle_serving_app.local_predict import LocalPredictor
import time
import re
import base64
class OCRService(WebService):
def init_det_debugger(self, det_model_config):
self.det_preprocess = Sequential([
ResizeByFactor(32, 960), Div(255),
Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]), Transpose(
(2, 0, 1))
])
self.det_client = LocalPredictor()
if sys.argv[1] == 'gpu':
self.det_client.load_model_config(
det_model_config, use_gpu=True, gpu_id=0)
elif sys.argv[1] == 'cpu':
self.det_client.load_model_config(det_model_config)
self.ocr_reader = OCRReader(
char_dict_path="../../../ppocr/utils/ppocr_keys_v1.txt")
def preprocess(self, feed=[], fetch=[]):
data = base64.b64decode(feed[0]["image"].encode('utf8'))
data = np.fromstring(data, np.uint8)
im = cv2.imdecode(data, cv2.IMREAD_COLOR)
ori_h, ori_w, _ = im.shape
det_img = self.det_preprocess(im)
_, new_h, new_w = det_img.shape
det_img = det_img[np.newaxis, :]
det_img = det_img.copy()
det_out = self.det_client.predict(
feed={"x": det_img}, fetch=["save_infer_model/scale_0.tmp_1"], batch=True)
filter_func = FilterBoxes(10, 10)
post_func = DBPostProcess({
"thresh": 0.3,
"box_thresh": 0.5,
"max_candidates": 1000,
"unclip_ratio": 1.5,
"min_size": 3
})
sorted_boxes = SortedBoxes()
ratio_list = [float(new_h) / ori_h, float(new_w) / ori_w]
dt_boxes_list = post_func(det_out["save_infer_model/scale_0.tmp_1"], [ratio_list])
dt_boxes = filter_func(dt_boxes_list[0], [ori_h, ori_w])
dt_boxes = sorted_boxes(dt_boxes)
get_rotate_crop_image = GetRotateCropImage()
img_list = []
max_wh_ratio = 0
for i, dtbox in enumerate(dt_boxes):
boximg = get_rotate_crop_image(im, dt_boxes[i])
img_list.append(boximg)
h, w = boximg.shape[0:2]
wh_ratio = w * 1.0 / h
max_wh_ratio = max(max_wh_ratio, wh_ratio)
if len(img_list) == 0:
return [], []
_, w, h = self.ocr_reader.resize_norm_img(img_list[0],
max_wh_ratio).shape
imgs = np.zeros((len(img_list), 3, w, h)).astype('float32')
for id, img in enumerate(img_list):
norm_img = self.ocr_reader.resize_norm_img(img, max_wh_ratio)
imgs[id] = norm_img
feed = {"x": imgs.copy()}
fetch = ["save_infer_model/scale_0.tmp_1"]
return feed, fetch, True
def postprocess(self, feed={}, fetch=[], fetch_map=None):
rec_res = self.ocr_reader.postprocess(fetch_map, with_score=True)
res_lst = []
for res in rec_res:
res_lst.append(res[0])
res = {"res": res_lst}
return res
ocr_service = OCRService(name="ocr")
ocr_service.load_model_config("../ppocr_rec_mobile_2.0_serving")
ocr_service.prepare_server(workdir="workdir", port=9292)
ocr_service.init_det_debugger(det_model_config="../ppocr_det_mobile_2.0_serving")
if sys.argv[1] == 'gpu':
ocr_service.set_gpus("0")
ocr_service.run_debugger_service(gpu=True)
elif sys.argv[1] == 'cpu':
ocr_service.run_debugger_service()
ocr_service.run_web_service()
deploy/slim/prune/README.md
View file @ 19eb7eb8
...@@ -23,13 +23,13 @@ ...@@ -23,13 +23,13 @@
```bash ```bash
git clone https://github.com/PaddlePaddle/PaddleSlim.git git clone https://github.com/PaddlePaddle/PaddleSlim.git
cd PaddleSlim
git checkout develop git checkout develop
cd Paddleslim
python3 setup.py install python3 setup.py install
``` ```
### 2. 获取预训练模型 ### 2. 获取预训练模型
模型裁剪需要加载事先训练好的模型，PaddleOCR也提供了一系列(模型)[../../../doc/doc_ch/models_list.md]，开发者可根据需要自行选择模型或使用自己的模型。 模型裁剪需要加载事先训练好的模型，PaddleOCR也提供了一系列[模型](../../../doc/doc_ch/models_list.md)，开发者可根据需要自行选择模型或使用自己的模型。
### 3. 敏感度分析训练 ### 3. 敏感度分析训练
...@@ -49,14 +49,14 @@ python3 setup.py install ...@@ -49,14 +49,14 @@ python3 setup.py install
进入PaddleOCR根目录，通过以下命令对模型进行敏感度分析训练： 进入PaddleOCR根目录，通过以下命令对模型进行敏感度分析训练：
```bash ```bash
python3.7 deploy/slim/prune/sensitivity_anal.py -c configs/det/ch_ppocr_v2.0/ch_det_mv3_db_v2.0.yml -o Global.pretrain_weights="your trained model" python3.7 deploy/slim/prune/sensitivity_anal.py -c configs/det/ch_ppocr_v2.0/ch_det_mv3_db_v2.0.yml -o Global.pretrained_model="your trained model" Global.save_model_dir=./output/prune_model/
``` ```
### 4. 导出模型、预测部署 ### 4. 导出模型、预测部署
在得到裁剪训练保存的模型后，我们可以将其导出为inference_model： 在得到裁剪训练保存的模型后，我们可以将其导出为inference_model：
```bash ```bash
pytho3.7 deploy/slim/prune/export_prune_model.py -c configs/det/ch_ppocr_v2.0/ch_det_mv3_db_v2.0.yml -o Global.pretrain_weights=./output/det_db/best_accuracy Global.save_inference_dir=inference_model pytho3.7 deploy/slim/prune/export_prune_model.py -c configs/det/ch_ppocr_v2.0/ch_det_mv3_db_v2.0.yml -o Global.pretrained_model=./output/det_db/best_accuracy Global.save_inference_dir=./prune/prune_inference_model
``` ```
inference model的预测和部署参考： inference model的预测和部署参考：
......
deploy/slim/prune/README_en.md
View file @ 19eb7eb8
...@@ -22,15 +22,15 @@ Five steps for OCR model prune: ...@@ -22,15 +22,15 @@ Five steps for OCR model prune:
```bash ```bash
git clone https://github.com/PaddlePaddle/PaddleSlim.git git clone https://github.com/PaddlePaddle/PaddleSlim.git
cd PaddleSlim
git checkout develop git checkout develop
cd Paddleslim
python3 setup.py install python3 setup.py install
``` ```
### 2. Download Pretrain Model ### 2. Download Pretrain Model
Model prune needs to load pre-trained models. Model prune needs to load pre-trained models.
PaddleOCR also provides a series of (models)[../../../doc/doc_en/models_list_en.md]. Developers can choose their own models or use their own models according to their needs. PaddleOCR also provides a series of [models](../../../doc/doc_en/models_list_en.md). Developers can choose their own models or use their own models according to their needs.
### 3. Pruning sensitivity analysis ### 3. Pruning sensitivity analysis
...@@ -54,7 +54,7 @@ Enter the PaddleOCR root directory，perform sensitivity analysis on the model w ...@@ -54,7 +54,7 @@ Enter the PaddleOCR root directory，perform sensitivity analysis on the model w
```bash ```bash
python3.7 deploy/slim/prune/sensitivity_anal.py -c configs/det/ch_ppocr_v2.0/ch_det_mv3_db_v2.0.yml -o Global.pretrain_weights="your trained model" python3.7 deploy/slim/prune/sensitivity_anal.py -c configs/det/ch_ppocr_v2.0/ch_det_mv3_db_v2.0.yml -o Global.pretrained_model="your trained model" Global.save_model_dir=./output/prune_model/
``` ```
...@@ -63,7 +63,7 @@ python3.7 deploy/slim/prune/sensitivity_anal.py -c configs/det/ch_ppocr_v2.0/ch_ ...@@ -63,7 +63,7 @@ python3.7 deploy/slim/prune/sensitivity_anal.py -c configs/det/ch_ppocr_v2.0/ch_
We can export the pruned model as inference_model for deployment: We can export the pruned model as inference_model for deployment:
```bash ```bash
python deploy/slim/prune/export_prune_model.py -c configs/det/ch_ppocr_v2.0/ch_det_mv3_db_v2.0.yml -o Global.pretrain_weights=./output/det_db/best_accuracy Global.test_batch_size_per_card=1 Global.save_inference_dir=inference_model python deploy/slim/prune/export_prune_model.py -c configs/det/ch_ppocr_v2.0/ch_det_mv3_db_v2.0.yml -o Global.pretrained_model=./output/det_db/best_accuracy Global.save_inference_dir=./prune/prune_inference_model
``` ```
Reference for prediction and deployment of inference model: Reference for prediction and deployment of inference model:
......
deploy/slim/prune/sensitivity_anal.py
View file @ 19eb7eb8
...@@ -75,7 +75,7 @@ def main(config, device, logger, vdl_writer): ...@@ -75,7 +75,7 @@ def main(config, device, logger, vdl_writer):
model = build_model(config['Architecture']) model = build_model(config['Architecture'])
flops = paddle.flops(model, [1, 3, 640, 640]) flops = paddle.flops(model, [1, 3, 640, 640])
logger.info(f"FLOPs before pruning: {flops}") logger.info("FLOPs before pruning: {}".format(flops))
from paddleslim.dygraph import FPGMFilterPruner from paddleslim.dygraph import FPGMFilterPruner
model.train() model.train()
...@@ -106,8 +106,8 @@ def main(config, device, logger, vdl_writer): ...@@ -106,8 +106,8 @@ def main(config, device, logger, vdl_writer):
def eval_fn(): def eval_fn():
metric = program.eval(model, valid_dataloader, post_process_class, metric = program.eval(model, valid_dataloader, post_process_class,
eval_class) eval_class, False)
logger.info(f"metric['hmean']: {metric['hmean']}") logger.info("metric['hmean']: {}".format(metric['hmean']))
return metric['hmean'] return metric['hmean']
params_sensitive = pruner.sensitive( params_sensitive = pruner.sensitive(
...@@ -123,16 +123,17 @@ def main(config, device, logger, vdl_writer): ...@@ -123,16 +123,17 @@ def main(config, device, logger, vdl_writer):
# calculate pruned params's ratio # calculate pruned params's ratio
params_sensitive = pruner._get_ratios_by_loss(params_sensitive, loss=0.02) params_sensitive = pruner._get_ratios_by_loss(params_sensitive, loss=0.02)
for key in params_sensitive.keys(): for key in params_sensitive.keys():
logger.info(f"{key}, {params_sensitive[key]}") logger.info("{}, {}".format(key, params_sensitive[key]))
#params_sensitive = {}
#for param in model.parameters():
# if 'transpose' not in param.name and 'linear' not in param.name:
# params_sensitive[param.name] = 0.1
plan = pruner.prune_vars(params_sensitive, [0]) plan = pruner.prune_vars(params_sensitive, [0])
for param in model.parameters():
if ("weights" in param.name and "conv" in param.name) or (
"w_0" in param.name and "conv2d" in param.name):
logger.info(f"{param.name}: {param.shape}")
flops = paddle.flops(model, [1, 3, 640, 640]) flops = paddle.flops(model, [1, 3, 640, 640])
logger.info(f"FLOPs after pruning: {flops}") logger.info("FLOPs after pruning: {}".format(flops))
# start train # start train
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment