keke-nice
commented
6 months ago @keke-nice 您好，利用您提供的开源模型进行微调，测试的时候发现指标如下，请问是我测试的过程错了嘛。 注释： 1.大约有5个文件时间对不齐需要删除 2.有的检测不到lmk 请问这个会影响指标嘛，该如何清理出去呢。
import cv2 import os import numpy as np import shutil import pandas as pd import scipy.io as scio from scipy import interpolate import scipy.io as io
功能： 对每段视频计算平均HR 注释： 1.大约有5个文件时间对不齐需要删除 2.有的检测不到lmk def MyEval(HR_pr, HR_rel): HR_pr = np.array(HR_pr).reshape(-1) HR_rel = np.array(HR_rel).reshape(-1) temp = HR_pr - HR_rel me = np.mean(temp) std = np.std(temp) mae = np.sum(np.abs(temp))/len(temp) rmse = np.sqrt(np.sum(np.power(temp, 2))/len(temp)) mer = np.mean(np.abs(temp) / HR_rel) p = np.sum((HR_pr - np.mean(HR_pr))(HR_rel - np.mean(HR_rel))) / ( 0.01 + np.linalg.norm(HR_pr - np.mean(HR_pr), ord=2) np.linalg.norm(HR_rel - np.mean(HR_rel), ord=2)) print('| me: %.4f' % me, '| std: %.4f' % std, '| mae: %.4f' % mae, '| rmse: %.4f' % rmse, '| mer: %.4f' % mer, '| p: %.4f' % p ) return me, std, mae, rmse, mer, p
gt_name = 'Label/HR.mat' frames_num = 224 gt_av = [] pr_av = [] gt_ps = [] pr_ps = []
Idex_files = r'/data/dataset/STMap/PURE_Index/PURE_STMap50_Test' #change to your index file path pr_path = r'/data/ssl/rPPG-MAE/save_model/finetune_PURE/HR_pr.mat' rel_path = r'/data/ssl/rPPG-MAE/save_model/finetune_PURE/HR_rel.mat' pr = scio.loadmat(pr_path)['HR_pr'] pr = np.array(pr.astype('float32')).reshape(-1) rel = scio.loadmat(rel_path)['HR_rel'] rel = np.array(rel.astype('float32')).reshape(-1)
files_list = sorted(os.listdir(Idex_files)) temp = scio.loadmat(os.path.join(Idex_files, files_list[0])) lastPath = str(temp['Path'][0]) pr_temp = [] gt_temp = [] for HR_index in range(pr.size-1): temp = scio.loadmat(os.path.join(Idex_files, files_list[HR_index])) nowPath = str(temp['Path'][0]) Step_Index = int(temp['Step_Index']) a = pr[HR_index] b = rel[HR_index]
if lastPath != nowPath: print(')************') if pr_temp is None: print(nowPath) print(lastPath) pr_temp = [] gt_temp = [] else: # print('diff_gt', np.array(gt_temp[1:]) - np.array(gt_temp[:-1])) # print('diff', np.array(pr_temp[1:]) - np.array(pr_temp[:-1])) # print('aaa', np.array(pr_temp)-np.array(gt_temp)) # print('gt_temp', np.mean(np.array(pr_temp)-np.array(gt_temp))) pr_av.append(np.nanmean(pr_temp)) gt_av.append(np.nanmean(gt_temp)) print(len(gt_ps)) print(gt_temp) print(pr_temp) gt_ps.append(gt_temp) pr_ps.append(pr_temp) pr_temp = [] gt_temp = [] else: gt_path = os.path.join(nowPath, gt_name) gt = scio.loadmat(gt_path)['HR'] gt = np.array(gt.astype('float32')).reshape(-1) gt = np.nanmean(gt[Step_Index:Step_Index + frames_num]) gt = gt.astype('float32') pr_temp.append(pr[HR_index]) gt_temp.append(rel[HR_index]) lastPath = nowPathio.savemat('gt_ps.mat', {'HR': gt_ps}) io.savemat('pr_ps.mat', {'HR': pr_ps}) io.savemat('HR_rel.mat', {'HR': gt_av}) io.savemat('HR_pr.mat', {'HR': pr_av}) MyEval(gt_av, pr_av)
emm 理论上是不需要再处理的，直接用这个测就行
@keke-nice 您好，利用您提供的开源模型进行微调，测试的时候发现指标如下，请问是我测试的过程错了嘛。 注释： 1.大约有5个文件时间对不齐需要删除 2.有的检测不到lmk 请问这个会影响指标嘛，该如何清理出去呢。
import cv2
import os
import numpy as np
import shutil
import pandas as pd
import scipy.io as scio
from scipy import interpolate
import scipy.io as io
功能： 对每段视频计算平均HR 注释： 1.大约有5个文件时间对不齐需要删除 2.有的检测不到lmk def MyEval(HR_pr, HR_rel): HR_pr = np.array(HR_pr).reshape(-1) HR_rel = np.array(HR_rel).reshape(-1) temp = HR_pr - HR_rel me = np.mean(temp) std = np.std(temp) mae = np.sum(np.abs(temp))/len(temp) rmse = np.sqrt(np.sum(np.power(temp, 2))/len(temp)) mer = np.mean(np.abs(temp) / HR_rel) p = np.sum((HR_pr - np.mean(HR_pr))(HR_rel - np.mean(HR_rel))) / ( 0.01 + np.linalg.norm(HR_pr - np.mean(HR_pr), ord=2) np.linalg.norm(HR_rel - np.mean(HR_rel), ord=2)) print('| me: %.4f' % me, '| std: %.4f' % std, '| mae: %.4f' % mae, '| rmse: %.4f' % rmse, '| mer: %.4f' % mer, '| p: %.4f' % p ) return me, std, mae, rmse, mer, p
gt_name = 'Label/HR.mat' frames_num = 224 gt_av = [] pr_av = [] gt_ps = [] pr_ps = []
Idex_files = r'/data/dataset/STMap/PURE_Index/PURE_STMap50_Test' #change to your index file path pr_path = r'/data/ssl/rPPG-MAE/save_model/finetune_PURE/HR_pr.mat' rel_path = r'/data/ssl/rPPG-MAE/save_model/finetune_PURE/HR_rel.mat' pr = scio.loadmat(pr_path)['HR_pr'] pr = np.array(pr.astype('float32')).reshape(-1) rel = scio.loadmat(rel_path)['HR_rel'] rel = np.array(rel.astype('float32')).reshape(-1)
files_list = sorted(os.listdir(Idex_files)) temp = scio.loadmat(os.path.join(Idex_files, files_list[0])) lastPath = str(temp['Path'][0]) pr_temp = [] gt_temp = [] for HR_index in range(pr.size-1): temp = scio.loadmat(os.path.join(Idex_files, files_list[HR_index])) nowPath = str(temp['Path'][0]) Step_Index = int(temp['Step_Index']) a = pr[HR_index] b = rel[HR_index]
io.savemat('gt_ps.mat', {'HR': gt_ps}) io.savemat('pr_ps.mat', {'HR': pr_ps}) io.savemat('HR_rel.mat', {'HR': gt_av}) io.savemat('HR_pr.mat', {'HR': pr_av}) MyEval(gt_av, pr_av)