详细记录solov2的ncnn实现和优化
详细记录solov2的ncnn实现和优化
允许在不修改内容的情况下转载本文
SOLOV2简介
solo大家都知道,核心思想是:将分割问题转化为位置分类问题,从而做到不需要anchor,不需要normalization,不需要bounding box detections的实例分割。
缘由
目前solo系列在github中没有转ncnn的项目
solov2的思想比较简单明了,但是转ncnn还是需要很多比较复杂操作,想在用ncnn实现更加了解solov2内部的算法流程,也熟悉ncnn的一些其他op操作
文字有很多干货: param手动修改、动态op、加速技巧 等等
1 、torch导出onnx
在导出onnx之前有几部操作:
- 修改split_feats : onnx不支持F.interpolate() 中 scale_factor < 1 ;因此对split_feats做了修改为nn.Maxpool 卷积,后面转ncnn手动修改;
def split_feats(self, feats):
return [
self.maxpool(feats[0]),
# F.interpolate(feats[0],scale_factor=0.5,mode='bilinear'),
feats[1],
feats[2],
feats[3],
F.interpolate(feats[4], size=feats[3].shape[-2:], mode='bilinear')]
- 添加输入参数:
添加onnx参数,金生coord改为手动输入
def forward(self, feats, eval=False,onnx=False):
if onnx:
feats,p3_input,p4_input,p5_input = feats
new_feats = self.split_feats(feats)
featmap_sizes = [featmap.size()[-2:] for featmap in new_feats]
if onnx:
new_feats[0] = [new_feats[0],p3_input]
new_feats[1] = [new_feats[1],p3_input]
new_feats[2] = [new_feats[2],p4_input]
new_feats[3] = [new_feats[3],p5_input]
new_feats[4] = [new_feats[4],p5_input]
- 实现转换代码
实现转换代码,并onnxsim优化,这里需要注意的是输入尽量小一点,因为solo的四个输出头的size比较小(40,36,24,16,12),输入较大时会使F.interpolate() 中 scale_factor < 1 ,input_size = 256时fpn最后一层上采样的scale_factor刚好可以等于2(256%64=0)
from torch import nn
import torch
from mmcv import Config
from mmdet.models.builder import build_backbone,build_head,build_neck
class SOLOV2(nn.Module):
def __init__(self,confg):
super(SOLOV2,self).__init__()
self.cfg = Config.fromfile(confg)
self.cfg = self.cfg["model"]
self.backbone = build_backbone(self.cfg["backbone"])
self.neck = build_neck(self.cfg["neck"])
self.bbox_head = build_head(self.cfg["bbox_head"])
def forward(self,x,p3_input,p4_input,p5_input):
x = self.backbone(x)
x = self.neck(x)
feature_pred,kernel_pred,cate_pred = self.bbox_head([x,p3_input,p4_input,p5_input],eval=True,onnx=True)
return feature_pred,kernel_pred,cate_pred
cfg = "./configs/solov2/solov2_r101_3x.py"
pretrained = "./weights/solov2.pth"
net = SOLOV2(cfg)
state_dict = torch.load(pretrained, map_location=lambda storage, loc: storage)["state_dict"]
res = net.load_state_dict(state_dict,strict=False)
net.eval()
input_size = 256
dummy_input1 = torch.randn(1, 3, input_size, input_size)
dummy_input2 = torch.randn(1, 2, input_size//8, input_size//8)
dummy_input3 = torch.randn(1, 2, input_size//16, input_size//16)
dummy_input4 = torch.randn(1, 2, input_size//32, input_size//32)
input_names = ["input","p3_input","p4_input","p5_input"]
output_names = ["feature_pred","kernel_pred1","kernel_pred2","kernel_pred3","kernel_pred4","kernel_pred5",
"cate_pred1","cate_pred2","cate_pred3","cate_pred4","cate_pred5" ]
save_name = "./onnx/solov2.onnx"
# torch.onnx.export(model, (dummy_input1, dummy_input2, dummy_input3), "C3AE.onnx", verbose=True, input_names=input_names, output_names=output_names)
torch.onnx.export(net, (dummy_input1,dummy_input2,dummy_input3,dummy_input4),save_name , verbose=False, input_names=input_names,
output_names=output_names)
import os
os.system("python -m onnxsim {0} {0}".format(save_name))
2、onnx转ncnn
- onnx2ncnn
./onnx2ncnn solov2.onnx solov2.param solov2.bin
- ncnn 参数手工优化
- 将之前转onnx时的maxpool改为上采样操作:
改为
2、将所有的kernel_pred和cate_pred之前的Interp参数改为对用的w和h:
改为
对应40尺寸的改好了,后面32、24、26、12也是类似的;
3、ncnn推理实现
- coord输入实现
对比torch半的coord实现:
x_range = torch.linspace(-1, 1, kernel_feat.shape[-1], device=kernel_feat.device)
y_range = torch.linspace(-1, 1, kernel_feat.shape[-2], device=kernel_feat.device)
y, x = torch.meshgrid(y_range, x_range)
y = y.expand([kernel_feat.shape[0], 1, -1, -1])
x = x.expand([kernel_feat.shape[0], 1, -1, -1])
coord_feat = torch.cat([x, y], 1)
因为对应stride只有8、16、32,因此ncnn需要生成3个对应尺寸的Mat:
size_t elemsize = sizeof(float);
ncnn::Mat x_p3;
ncnn::Mat x_p4;
ncnn::Mat x_p5;
// coord conv
int pw = int(target_w / 8);
int ph = int(target_h / 8);
x_p3.create(pw, ph, 2, elemsize);
float step_h = 2.f / (ph - 1);
float step_w = 2.f / (pw - 1);
for (int h = 0; h < ph; h++) {
for (int w = 0; w < pw; w++) {
x_p3.channel(0)[h * pw + w] = -1.f + step_w * (float) w;
x_p3.channel(1)[h * pw + w] = -1.f + step_h * (float) h;
pw = int(target_w / 16);
ph = int(target_h / 16);
x_p4.create(pw, ph, 2, elemsize);
step_h = 2.f / (ph - 1);
step_w = 2.f / (pw - 1);
for (int h = 0; h < ph; h++) {
for (int w = 0; w < pw; w++) {
x_p4.channel(0)[h * pw + w] = -1.f + step_w * (float) w;
x_p4.channel(1)[h * pw + w] = -1.f + step_h * (float) h;
pw = int(target_w / 32);
ph = int(target_h / 32);
x_p5.create(pw, ph, 2, elemsize);
step_h = 2.f / (ph - 1);
step_w = 2.f / (pw - 1);
for (int h = 0; h < ph; h++) {
for (int w = 0; w < pw; w++) {
x_p5.channel(0)[h * pw + w] = -1.f + step_w * (float) w;
x_p5.channel(1)[h * pw + w] = -1.f + step_h * (float) h;
将其作为输入:
ncnn::Extractor ex = solov2.create_extractor();
ex.input("input", in);
ex.input("p3_input", x_p3);
ex.input("p4_input", x_p4);
ex.input("p5_input", x_p5);
- 抽取feature_pred、kernerl_pred、cate_pred特征 :
ncnn::Mat feature_pred, cate_pred1, cate_pred2, cate_pred3, cate_pred4, cate_pred5, kernel_pred1, kernel_pred2, kernel_pred3, kernel_pred4, kernel_pred5;
ex.extract("cate_pred1", cate_pred1);
ex.extract("cate_pred2", cate_pred2);
ex.extract("cate_pred3", cate_pred3);
ex.extract("cate_pred4", cate_pred4);
ex.extract("cate_pred5", cate_pred5);
ex.extract("kernel_pred1", kernel_pred1);
ex.extract("kernel_pred2", kernel_pred2);
ex.extract("kernel_pred3", kernel_pred3);
ex.extract("kernel_pred4", kernel_pred4);
ex.extract("kernel_pred5", kernel_pred5);
ex.extract("feature_pred", feature_pred);
- 生成分割mask图
前面已经有了kernel_pred和feature_pred,接下来就是生成mask图,这就用到了low-level op api<这里说一句ncnn真灵活,谁用谁知道>,目的是实现torch中的动态conv2d,对应torch实现为:
ins_i = F.conv2d(feature_pred, kernel, groups=N).view(N, self.seg_num_grids[i] ** 2, h, w)
ncnn的low-level op api使用wiki:
创建ncnn::Option
ncnn::Option opt;
opt.num_threads = num_threads;
opt.use_fp16_storage = false;
opt.use_packing_layout = false;
实现分割的预测功能函数
static void ins_decode(const ncnn::Mat &kernel_pred, const ncnn::Mat &feature_pred, ncnn::Mat *ins_pred, int c_in, int c_out,
ncnn::Option &opt) {
ncnn::Layer *op = ncnn::create_layer("Convolution");
ncnn::ParamDict pd;
pd.set(0, c_out);
pd.set(1, 1);
pd.set(6, c_in * c_out);
op->load_param(pd);
ncnn::Mat weights[1];
weights[0].create(c_in * c_out);
float *kernel_pred_data = (float *) kernel_pred.data;
for (int i = 0; i < c_in * c_out; i++) {
weights[0][i] = kernel_pred_data[i];
op->load_model(ncnn::ModelBinFromMatArray(weights));
op->create_pipeline(opt);
ncnn::Mat temp_ins;
op->forward(feature_pred, temp_ins, opt);
*ins_pred = temp_ins;
op->destroy_pipeline(opt);
delete op;
将每个kernel_pre转换为分割图:
ncnn::Mat ins_pred1, ins_pred2, ins_pred3, ins_pred4, ins_pred5;
int c_in = feature_pred.c;
ins_decode(kernel_pred1, feature_pred, &ins_pred1, c_in, 40 * 40,opt);
ins_decode(kernel_pred2, feature_pred, &ins_pred2, c_in, 36 * 36,opt);
ins_decode(kernel_pred3, feature_pred, &ins_pred3, c_in, 24 * 24,opt);
ins_decode(kernel_pred4, feature_pred, &ins_pred4, c_in, 16 * 16,opt);
ins_decode(kernel_pred5, feature_pred, &ins_pred5, c_in, 12 * 12,opt);
- 后处理
有了这一切后就是后处理啦,贴代码:
功能函数
void generate_res(ncnn::Mat &cate_pred, ncnn::Mat &ins_pred, std::vector <std::vector<Object>> &objects, float cate_thresh,
float conf_thresh, int img_w, int img_h, int num_class, float stride) {
int w = cate_pred.w;
int h = cate_pred.h;
int w_ins = ins_pred.w;
int h_ins = ins_pred.h;
for (int q = 0; q < num_class; q++) {
const float *cate_ptr = cate_pred.channel(q);
for (int i = 0; i < h; i++) {
for (int j = 0; j < w; j++) {
int index = i * w + j;
float cate_socre = cate_ptr[index];
if (cate_socre < cate_thresh) {
continue;
const float *ins_ptr = ins_pred.channel(index);
cv::Mat mask(h_ins, w_ins, CV_32FC1);
float sum_mask = 0.f;
int count_mask = 0;
mask = cv::Scalar(0.f);
float *mp = (float *) mask.data;
for (int m = 0; m < w_ins * h_ins; m++) {
float mask_score = sigmoid(ins_ptr[m]);
if (mask_score > 0.5) {
mp[m] += mask_score;
sum_mask += mask_score;
count_mask++;
if (count_mask < stride) {
continue;
float mask_score = sum_mask / (float(count_mask) + 1e-6);
// float socre = mask_score * cate_socre;
float socre = mask_score * cate_socre;
if (socre < conf_thresh) {
continue;
cv::Mat mask2;
cv::resize(mask, mask2, cv::Size(img_w, img_h));
Object obj;
obj.mask = cv::Mat(img_h, img_w, CV_8UC1);
float sum_mask_y = 0.f;
float sum_mask_x = 0.f;
int area = 0;
obj.mask = cv::Scalar(0);
for (int y = 0; y < img_h; y++) {
const float *mp2 = mask2.ptr<const float>(y);
uchar *bmp = obj.mask.ptr<uchar>(y);
for (int x = 0; x < img_w; x++) {
if (mp2[x] > 0.5f) {
bmp[x] = 255;
sum_mask_y += (float) y;
sum_mask_x += (float) x;
area++;
} else bmp[x] = 0;
obj.cx = int(sum_mask_x / area);
obj.cy = int(sum_mask_y / area);
obj.label = q + 1;
obj.prob = socre;
objects[q].push_back(obj);
每个head的解码:
std::vector <std::vector<Object>> class_candidates;
class_candidates.resize(num_class);
generate_res(cate_pred1,ins_pred1,class_candidates,cate_thresh,confidence_thresh,img_w,img_h,num_class,8.f);
generate_res(cate_pred2,ins_pred2,class_candidates,cate_thresh,confidence_thresh,img_w,img_h,num_class,8.f);
generate_res(cate_pred3,ins_pred3,class_candidates,cate_thresh,confidence_thresh,img_w,img_h,num_class,16.f);
generate_res(cate_pred4,ins_pred4,class_candidates,cate_thresh,confidence_thresh,img_w,img_h,num_class,32.f);
generate_res(cate_pred5,ins_pred5,class_candidates,cate_thresh,confidence_thresh,img_w,img_h,num_class,32.f);
最后就是nms了,nms实现的比较低效:
objects.clear();
for (int i = 0; i < (int) class_candidates.size(); i++) {
std::vector <Object> &candidates = class_candidates[i];
qsort_descent_inplace(candidates);
std::vector<int> picked;
nms_sorted_segs(candidates, picked, nms_threshold,img_w,img_h);
for (int j = 0; j < (int) picked.size(); j++) {
int z = picked[j];
objects.push_back(candidates[z]);
qsort_descent_inplace(objects);
至此,已经基本完成了所有工作啦:来预测一张图看看:
可以看出来结果还是非常好的QAQ
速度优化
重中之重来啦,本人针对SOLO的特点做了一处优化,就是kernel_pred的部分,在kernel_pred1部分输出特征图为256x40x40,转换为cond2d的核就为1600x256,1600的通道在stride=4的特征图上耗时巨大,那怎么来优化呢。
SOLO预测的cate_pred1分支预测了40x40=1600个网格,但并不是所有网格都是需要的,只有大于预值得才需要,那么在kernel_pred1的输出部分仅仅只需要cate_pred1大于阈值对应的核就好,基于这个思想就能减少非常多的通道,进而减少计算量。
话不多说,具体这样实现:
- 首先获取cate_pred大于阈值的index:
功能函数为:
static void kernel_pick(const ncnn::Mat &cate_pred, std::vector<int> &picked, int num_class, float cate_thresh)
int w = cate_pred.w;
int h = cate_pred.h;
for (int q = 0; q < num_class; q++) {
const float *cate_ptr = cate_pred.channel(q);
for (int i = 0; i < h; i++) {
for (int j = 0; j < w; j++) {
int index = i * w + j;
float cate_score = cate_ptr[index];
if (cate_score < cate_thresh) {
continue;
else picked.push_back(index);
获取每个cate_pred分支的index结果:
std::vector<int> kernel_picked1, kernel_picked2, kernel_picked3, kernel_picked4, kernel_picked5;
kernel_pick(cate_pred1, kernel_picked1, num_class, cate_thresh);
kernel_pick(cate_pred2, kernel_picked2, num_class, cate_thresh);
kernel_pick(cate_pred3, kernel_picked3, num_class, cate_thresh);
kernel_pick(cate_pred4, kernel_picked4, num_class, cate_thresh);
kernel_pick(cate_pred5, kernel_picked5, num_class, cate_thresh);
- 在ins_decode功能函数中加入map,来保存index和ins_pred的通道索引的映射关系
static void ins_decode(const ncnn::Mat &kernel_pred, const ncnn::Mat &feature_pred, std::vector<int> &kernel_picked,
std::map<int, int> &kernel_map, ncnn::Mat *ins_pred, int c_in,
ncnn::Option &opt) {
std::set<int> kernel_pick_set;
kernel_pick_set.insert(kernel_picked.begin(), kernel_picked.end());
int c_out = kernel_pick_set.size();
if (c_out > 0) {
ncnn::Layer *op = ncnn::create_layer("Convolution");
ncnn::ParamDict pd;
pd.set(0, c_out);
pd.set(1, 1);
pd.set(6, c_in * c_out);
op->load_param(pd);
ncnn::Mat weights[1];
weights[0].create(c_in * c_out);
float *kernel_pred_data = (float *) kernel_pred.data;
std::set<int>::iterator pick_c;
int count_c = 0;
for (pick_c = kernel_pick_set.begin(); pick_c != kernel_pick_set.end(); pick_c++)
kernel_map[*pick_c] = count_c;
for (int j = 0; j < c_in; j++) {
weights[0][count_c * c_in + j] = kernel_pred_data[c_in * (*pick_c) + j];
count_c++;
op->load_model(ncnn::ModelBinFromMatArray(weights));
op->create_pipeline(opt);
ncnn::Mat temp_ins;
op->forward(feature_pred, temp_ins, opt);
*ins_pred = temp_ins;
op->destroy_pipeline(opt);
delete op;