Can this model work when I don't give the bndbox?

Question

Can this model work when I don't give the bndbox?

zeroXscorpion7 opened this issue 4 years ago · comments

I am trying to use this model to Identify my pictures, if the bndbox is need?

Boshen Zhang · Answer 1 · Wed Jun 24 2020 16:52:17 GMT+0800 (China Standard Time)

Hi, @zeroXscorpion7 . You can try inference without bndbox, but, the performance can not be guaranteed, some skeleton flex is very likely to happen because of the mean/std shift.

zeroXscorpion7 · Answer 2 · Mon Jun 29 2020 14:02:01 GMT+0800 (China Standard Time)

I want to test the efficacy in real-time, how do I identify one picture at a time?

zeroXscorpion7 · Answer 3 · Mon Jun 29 2020 14:04:35 GMT+0800 (China Standard Time)

def main():

net = model.A2J_model(num_classes = keypointsNumber)
net.load_state_dict(torch.load(model_dir)) 
net = net.cuda()
net.eval()

post_precess = anchor.post_process(shape=[cropHeight//16,cropWidth//16],stride=16,P_h=None, P_w=None)

output = torch.FloatTensor()

data4DTemp = scio.loadmat(testingImageDir + str(1) + '.mat')['DepthNormal']       
depthTemp = data4DTemp[:,:,3]

img=depthTemp
heads = net(img)
pred_keypoints = post_precess(heads,voting=False)
output = torch.cat([output,pred_keypoints.data.cpu()], 0)
    
result = output.cpu().data.numpy()
assert np.shape(result)==np.shape(result), "source has different shape with target"
Test1_ = np.zeros(result.shape)
Test1_[:, 0] = result[:,1]
Test1_[:, 1] = result[:,0]
Test1_[:, 2] = result[:,2]
Test1 = Test1_
Test1[:,0] = Test1_[:,0]*(Bndbox[i,2]-Bndbox[i,0])/cropWidth + Bndbox[i,0]  # x
Test1[:,1] = Test1_[:,1]*(Bndbox[i,3]-Bndbox[i,1])/cropHeight + Bndbox[i,1]  # y
Test1[:,2] = Test1_[:,2]/depthFactor 
TestWorld = np.ones((len(Test1),keypointsNumber,3))    
TestWorld_tuple = pixel2world(Test1[:,0],Test1[:,1],Test1[:,2])
X=np.zeros((15),np.uint8)
Y=np.zeros((15),np.uint8)
for j in range(keypointsNumber):
    X[j],Y[j]=world2pixel(TestWorld[0,j,0],TestWorld[0,j,1],TestWorld[0,j,2])
IMGX=np.zeros((240,320,3),np.uint8)
cv2.line(IMGX,(X[0],Y[0]),(X[1],Y[1]),(0,0,255),2)
cv2.line(IMGX,(X[1],Y[1]),(X[2],Y[2]),(0,0,255),2)
cv2.line(IMGX,(X[1],Y[1]),(X[3],Y[3]),(0,0,255),2)
cv2.line(IMGX,(X[1],Y[1]),(X[8],Y[8]),(0,0,255),2)
cv2.line(IMGX,(X[2],Y[2]),(X[4],Y[4]),(0,0,255),2)
cv2.line(IMGX,(X[4],Y[4]),(X[6],Y[6]),(0,0,255),2)
cv2.line(IMGX,(X[3],Y[3]),(X[5],Y[5]),(0,0,255),2)
cv2.line(IMGX,(X[5],Y[5]),(X[7],Y[7]),(0,0,255),2)
cv2.line(IMGX,(X[8],Y[8]),(X[9],Y[9]),(0,0,255),2)
cv2.line(IMGX,(X[8],Y[8]),(X[10],Y[10]),(0,0,255),2)
cv2.line(IMGX,(X[9],Y[9]),(X[11],Y[11]),(0,0,255),2)
cv2.line(IMGX,(X[11],Y[11]),(X[13],Y[13]),(0,0,255),2)
cv2.line(IMGX,(X[10],Y[10]),(X[12],Y[12]),(0,0,255),2)
cv2.line(IMGX,(X[12],Y[12]),(X[14],Y[14]),(0,0,255),2)
for i in range(keypointsNumber):
    cv2.circle(IMGX,(X[i],Y[i]),4,(255,255,255),-1)
cv2.imshow('img',IMGX)
cv2.waitKey(0)
cv2.destroyAllWindows()

This is the code I edited, but it has some promble.
How do I edit it?

mstc-xqp · Answer 4 · Wed Sep 23 2020 20:52:55 GMT+0800 (China Standard Time)

def main():

net = model.A2J_model(num_classes = keypointsNumber)
net.load_state_dict(torch.load(model_dir)) 
net = net.cuda()
net.eval()

post_precess = anchor.post_process(shape=[cropHeight//16,cropWidth//16],stride=16,P_h=None, P_w=None)

output = torch.FloatTensor()

data4DTemp = scio.loadmat(testingImageDir + str(1) + '.mat')['DepthNormal']       
depthTemp = data4DTemp[:,:,3]

img=depthTemp
heads = net(img)
pred_keypoints = post_precess(heads,voting=False)
output = torch.cat([output,pred_keypoints.data.cpu()], 0)
    
result = output.cpu().data.numpy()
assert np.shape(result)==np.shape(result), "source has different shape with target"
Test1_ = np.zeros(result.shape)
Test1_[:, 0] = result[:,1]
Test1_[:, 1] = result[:,0]
Test1_[:, 2] = result[:,2]
Test1 = Test1_
Test1[:,0] = Test1_[:,0]*(Bndbox[i,2]-Bndbox[i,0])/cropWidth + Bndbox[i,0]  # x
Test1[:,1] = Test1_[:,1]*(Bndbox[i,3]-Bndbox[i,1])/cropHeight + Bndbox[i,1]  # y
Test1[:,2] = Test1_[:,2]/depthFactor 
TestWorld = np.ones((len(Test1),keypointsNumber,3))    
TestWorld_tuple = pixel2world(Test1[:,0],Test1[:,1],Test1[:,2])
X=np.zeros((15),np.uint8)
Y=np.zeros((15),np.uint8)
for j in range(keypointsNumber):
    X[j],Y[j]=world2pixel(TestWorld[0,j,0],TestWorld[0,j,1],TestWorld[0,j,2])
IMGX=np.zeros((240,320,3),np.uint8)
cv2.line(IMGX,(X[0],Y[0]),(X[1],Y[1]),(0,0,255),2)
cv2.line(IMGX,(X[1],Y[1]),(X[2],Y[2]),(0,0,255),2)
cv2.line(IMGX,(X[1],Y[1]),(X[3],Y[3]),(0,0,255),2)
cv2.line(IMGX,(X[1],Y[1]),(X[8],Y[8]),(0,0,255),2)
cv2.line(IMGX,(X[2],Y[2]),(X[4],Y[4]),(0,0,255),2)
cv2.line(IMGX,(X[4],Y[4]),(X[6],Y[6]),(0,0,255),2)
cv2.line(IMGX,(X[3],Y[3]),(X[5],Y[5]),(0,0,255),2)
cv2.line(IMGX,(X[5],Y[5]),(X[7],Y[7]),(0,0,255),2)
cv2.line(IMGX,(X[8],Y[8]),(X[9],Y[9]),(0,0,255),2)
cv2.line(IMGX,(X[8],Y[8]),(X[10],Y[10]),(0,0,255),2)
cv2.line(IMGX,(X[9],Y[9]),(X[11],Y[11]),(0,0,255),2)
cv2.line(IMGX,(X[11],Y[11]),(X[13],Y[13]),(0,0,255),2)
cv2.line(IMGX,(X[10],Y[10]),(X[12],Y[12]),(0,0,255),2)
cv2.line(IMGX,(X[12],Y[12]),(X[14],Y[14]),(0,0,255),2)
for i in range(keypointsNumber):
    cv2.circle(IMGX,(X[i],Y[i]),4,(255,255,255),-1)
cv2.imshow('img',IMGX)
cv2.waitKey(0)
cv2.destroyAllWindows()

This is the code I edited, but it has some promble.
How do I edit it?

do u solove it ?

zeroXscorpion7 · Answer 5 · Wed Sep 23 2020 21:12:36 GMT+0800 (China Standard Time)

I remove my_dataloader and enter depthTemp into dataPreprocess, then I use torch.from_numpy to make it to tensor

def dataPreprocess(img, depth_thres=0.4):

imageOutputs = np.ones((cropHeight, cropWidth, 1), dtype='float32') 

imCrop = img.copy()[:, :]
imgResize = cv2.resize(imCrop, (cropWidth, cropHeight), interpolation=cv2.INTER_NEAREST)
imgResize = np.asarray(imgResize,dtype = 'float32')  # H*W*C
imgResize = imgResize /5

imageOutputs[:,:,0] = imgResize
    
imageOutputs = np.asarray(imageOutputs)
imageNCHWOut = imageOutputs.transpose(2, 0, 1)  # [H, W, C] --->>>  [C, H, W]
imageNCHWOut = np.asarray(imageNCHWOut)

data = torch.from_numpy(imageNCHWOut)

return data

img=np.zeros((1,1,288,288),np.float32)
img[0,:,:,:]= dataPreprocess(depth_map, 0.4)
img=torch.from_numpy(img)
post_precess = anchor.post_process(shape=[cropHeight//16,cropWidth//16],stride=16,P_h=None, P_w=None)
output = torch.FloatTensor()

Like these.

mstc-xqp · Answer 6 · Wed Sep 23 2020 21:14:26 GMT+0800 (China Standard Time)

I remove my_dataloader and enter depthTemp into dataPreprocess, then I use torch.from_numpy to make it to tensor

def dataPreprocess(img, depth_thres=0.4):
imageOutputs = np.ones((cropHeight, cropWidth, 1), dtype='float32') 

imCrop = img.copy()[:, :]
imgResize = cv2.resize(imCrop, (cropWidth, cropHeight), interpolation=cv2.INTER_NEAREST)
imgResize = np.asarray(imgResize,dtype = 'float32')  # H*W*C
imgResize = imgResize /5

imageOutputs[:,:,0] = imgResize
    
imageOutputs = np.asarray(imageOutputs)
imageNCHWOut = imageOutputs.transpose(2, 0, 1)  # [H, W, C] --->>>  [C, H, W]
imageNCHWOut = np.asarray(imageNCHWOut)

data = torch.from_numpy(imageNCHWOut)

return data
img=np.zeros((1,1,288,288),np.float32)
img[0,:,:,:]= dataPreprocess(depth_map, 0.4)
img=torch.from_numpy(img)
post_precess = anchor.post_process(shape=[cropHeight//16,cropWidth//16],stride=16,P_h=None, P_w=None)
output = torch.FloatTensor()

Like these.

thank you !!
I am trying to use this model to Identify my pictures, if the bndbox is need?
i see u ask that .Did u try that? what is the performace

zeroXscorpion7 · Answer 7 · Wed Sep 23 2020 21:17:58 GMT+0800 (China Standard Time)

If you don't to use bndbox, you may train the data by yourself, or modify the train code.

mstc-xqp · Answer 8 · Wed Sep 23 2020 21:25:08 GMT+0800 (China Standard Time)

If you don't to use bndbox, you may train the data by yourself, or modify the train code.

I see. Can u share the train code? you train on ITOP or K2PHD?

I have seen some people use just use depth map to train a alphapose or openpose model .

zeroXscorpion7 · Answer 9 · Wed Sep 23 2020 21:27:58 GMT+0800 (China Standard Time)

import cv2
import torch
import torch.utils.data
import torch.optim.lr_scheduler as lr_scheduler
import numpy as np
import scipy.io as scio
import os
from PIL import Image
from torch.autograd import Variable
import model as model
import anchor as anchor
from tqdm import tqdm
import random_erasing
import logging
import time
import datetime
import random

os.environ["CUDA_VISIBLE_DEVICES"] = "0"

DataHyperParms

TrainImgFrames = 1000
TestImgFrames = 100
keypointsNumber = 15
cropWidth = 288
cropHeight = 288
batch_size = 64
learning_rate = 0.00035
Weight_Decay = 1e-4
nepoch = 35
RegLossFactor = 3
spatialFactor = 0.5
RandCropShift = 5
RandshiftDepth = 1
RandRotate = 180
RandScale = (1.0, 0.5)

randomseed = 12345
random.seed(randomseed)
np.random.seed(randomseed)
torch.manual_seed(randomseed)

save_dir = ''

try:
os.makedirs(save_dir)
except OSError:
pass

trainingImageDir = ''
testingImageDir = '' # mat images
keypointsfileTest = ''
keypointsfileTrain = ''
model_dir = ''
result_file = 'result_test.txt'

def pixel2world(x):
x[:, :, 0] = (x[:, :, 0] - 160.0) * x[:, :, 2] * 0.0035
x[:, :, 1] = (120.0 - x[:, :, 1]) * x[:, :, 2] * 0.0035
return x

def world2pixel(x):
x[:, :, 0] = 160.0 + x[:, :, 0] / (x[:, :, 2] * 0.0035)
x[:, :, 1] = 120.0 - x[:, :, 1] / (x[:, :, 2] * 0.0035)
return x

joint_id_to_name = {
0: 'Head',
1: 'Neck',
2: 'RShoulder',
3: 'LShoulder',
4: 'RElbow',
5: 'LElbow',
6: 'RHand',
7: 'LHand',
8: 'Torso',
9: 'RHip',
10: 'LHip',
11: 'RKnee',
12: 'LKnee',
13: 'RFoot',
14: 'LFoot',
}

loading GT keypoints and center points

keypointsWorldtest = scio.loadmat(keypointsfileTest)['keypoints3D'].astype(np.float32)
keypointsPixeltest = np.ones((len(keypointsWorldtest),15,2),dtype='float32')
keypointsPixeltest = world2pixel(keypointsWorldtest)

keypointsWorldtrain = scio.loadmat(keypointsfileTrain)['keypoints3D'].astype(np.float32)
keypointsPixeltrain = np.ones((len(keypointsWorldtrain),15,2),dtype='float32')
keypointsPixeltrain = world2pixel(keypointsWorldtrain)

def transform(img, label, matrix):
'''
img: [H, W] label, [N,2]
'''
img_out = cv2.warpAffine(img,matrix,(cropWidth,cropHeight))
label_out = np.ones((keypointsNumber, 3))
label_out[:,:2] = label[:,:2].copy()
label_out = np.matmul(matrix, label_out.transpose())
label_out = label_out.transpose()

return img_out, label_out

def dataPreprocess(index, img, keypointsUVD, depth_thres=0.4, augment=True):

imageOutputs = np.ones((cropHeight, cropWidth, 1), dtype='float32') 
labelOutputs = np.ones((keypointsNumber, 3), dtype = 'float32') 

if augment:
    RandomOffset_1 = np.random.randint(-1*RandCropShift,RandCropShift)
    RandomOffset_2 = np.random.randint(-1*RandCropShift,RandCropShift)
    RandomOffset_3 = np.random.randint(-1*RandCropShift,RandCropShift)
    RandomOffset_4 = np.random.randint(-1*RandCropShift,RandCropShift)
    RandomOffsetDepth = np.random.normal(0, RandshiftDepth, cropHeight*cropWidth).reshape(cropHeight,cropWidth) 
    RandomOffsetDepth[np.where(RandomOffsetDepth < RandshiftDepth)] = 0
    RandomRotate = np.random.randint(-1*RandRotate,RandRotate)
    RandomScale = np.random.rand()*RandScale[0]+RandScale[1]
    matrix = cv2.getRotationMatrix2D((cropWidth/2,cropHeight/2),RandomRotate,RandomScale)
else:
    RandomOffset_1, RandomOffset_2, RandomOffset_3, RandomOffset_4 = 0, 0, 0, 0
    RandomRotate = 0
    RandomScale = 1
    RandomOffsetDepth = 0
    matrix = cv2.getRotationMatrix2D((cropWidth/2,cropHeight/2),RandomRotate,RandomScale)

imCrop = img[:, :].copy()

imgResize = cv2.resize(imCrop, (cropWidth, cropHeight), interpolation=cv2.INTER_NEAREST)

imgResize = np.asarray(imgResize,dtype = 'float32')  # H*W*C
imgResize = imgResize / 5

## label
label_xy = np.ones((keypointsNumber, 2), dtype = 'float32') 
label_xy[:,0] = keypointsUVD[index,:,0].copy()*cropWidth/320 # x
label_xy[:,1] = keypointsUVD[index,:,1].copy()*cropHeight/240 # y

if augment:
    imgResize, label_xy = transform(imgResize, label_xy, matrix)  ## rotation, scale

imageOutputs[:,:,0] = imgResize

labelOutputs[:,1] = label_xy[:,0]
labelOutputs[:,0] = label_xy[:,1]
labelOutputs[:,2] = (keypointsUVD[index,:,2])*RandomScale   # Z  

imageOutputs = np.asarray(imageOutputs)
imageNCHWOut = imageOutputs.transpose(2, 0, 1)  # [H, W, C] --->>>  [C, H, W]
imageNCHWOut = np.asarray(imageNCHWOut)
labelOutputs = np.asarray(labelOutputs)

data, label = torch.from_numpy(imageNCHWOut), torch.from_numpy(labelOutputs)

return data, label

###################### Pytorch dataloader #################
class my_dataloader(torch.utils.data.Dataset):

def __init__(self, ImgDir, keypointsUVD, num, augment=True):

    self.ImgDir = ImgDir
    self.keypointsUVD = keypointsUVD
    self.num = num
    self.augment = augment
    self.randomErase = random_erasing.RandomErasing(probability = 0.5, sl = 0.02, sh = 0.4, r1 = 0.3, mean=[0])

def __getitem__(self, index):

    data4D = scio.loadmat(self.ImgDir + str(index+1) + '.mat')['DepthNormal']
    depth = data4D[:,:]

    data, label = dataPreprocess(index, depth, self.keypointsUVD, self.augment)

    if self.augment:
        data = self.randomErase(data)

    return data, label

def __len__(self):
    return self.num

train_image_datasets = my_dataloader(trainingImageDir, keypointsWorldtrain, TrainImgFrames, augment=True)
train_dataloaders = torch.utils.data.DataLoader(train_image_datasets, batch_size = batch_size,
shuffle = True, num_workers = 8)

test_image_datasets = my_dataloader(testingImageDir, keypointsWorldtest, TestImgFrames, augment=False)
test_dataloaders = torch.utils.data.DataLoader(test_image_datasets, batch_size = batch_size,
shuffle = False, num_workers = 8)

def train():

net = model.A2J_model(num_classes = keypointsNumber)
net = net.cuda()

post_precess = anchor.post_process(shape=[cropHeight//16,cropWidth//16],stride=16,P_h=None, P_w=None)
criterion = anchor.A2J_loss(shape=[cropHeight//16,cropWidth//16],thres = [16.0,32.0],stride=16,\
    spatialFactor=spatialFactor,img_shape=[cropHeight, cropWidth],P_h=None, P_w=None)
optimizer = torch.optim.Adam(net.parameters(), lr=learning_rate, weight_decay=Weight_Decay)
scheduler = lr_scheduler.StepLR(optimizer, step_size=10, gamma=0.2)

logging.basicConfig(format='%(asctime)s %(message)s', datefmt='%Y/%m/%d %H:%M:%S', \
                filename=os.path.join(save_dir, 'train.log'), level=logging.INFO)
logging.info('======================================================')

for epoch in range(nepoch):
    net = net.train()
    train_loss_add = 0.0
    Cls_loss_add = 0.0
    Reg_loss_add = 0.0
    timer = time.time()

    # Training loop
    for i, (img, label) in enumerate(train_dataloaders):

        torch.cuda.synchronize() 

        img, label = img.cuda(), label.cuda()
        heads  = net(img)
        #print(regression)     
        optimizer.zero_grad()  
        
        Cls_loss, Reg_loss = criterion(heads, label)

        loss = 1*Cls_loss + Reg_loss*RegLossFactor
        loss.backward()
        optimizer.step()

        torch.cuda.synchronize()
        
        train_loss_add = train_loss_add + (loss.item())*len(img)
        Cls_loss_add = Cls_loss_add + (Cls_loss.item())*len(img)
        Reg_loss_add = Reg_loss_add + (Reg_loss.item())*len(img)

        # printing loss info
        if i%10 == 0:
            print('epoch: ',epoch, ' step: ', i, 'Cls_loss ',Cls_loss.item(), 'Reg_loss ',Reg_loss.item(), ' total loss ',loss.item())

    scheduler.step(epoch)

    # time taken
    torch.cuda.synchronize()
    timer = time.time() - timer
    timer = timer / TrainImgFrames
    print('==> time to learn 1 sample = %f (ms)' %(timer*1000))

    train_loss_add = train_loss_add / TrainImgFrames
    Cls_loss_add = Cls_loss_add / TrainImgFrames
    Reg_loss_add = Reg_loss_add / TrainImgFrames
    print('mean train_loss_add of 1 sample: %f, #train_indexes = %d' %(train_loss_add, TrainImgFrames))
    print('mean Cls_loss_add of 1 sample: %f, #train_indexes = %d' %(Cls_loss_add, TrainImgFrames))
    print('mean Reg_loss_add of 1 sample: %f, #train_indexes = %d' %(Reg_loss_add, TrainImgFrames))

    Error_test = 0
    Error_train = 0
    Error_test_wrist = 0

    if (epoch % 1 == 0):  
        net = net.eval()
        output = torch.FloatTensor()
        outputTrain = torch.FloatTensor()

        for i, (img, label) in tqdm(enumerate(test_dataloaders)):
            with torch.no_grad():
                img, label = img.cuda(), label.cuda()       
                heads = net(img)  
                pred_keypoints = post_precess(heads, voting=False)
                output = torch.cat([output,pred_keypoints.data.cpu()], 0)

        result = output.cpu().data.numpy()
        Error_test = errorCompute(result,keypointsWorldtest,)
        print('epoch: ', epoch, 'Test error:', Error_test)
        saveNamePrefix = '%s/net_%d_wetD_' % (save_dir, epoch) + str(Weight_Decay) + '_depFact_' + str(spatialFactor) + '_RegFact_' + str(RegLossFactor) + '_rndShft_' + str(RandCropShift)
        torch.save(net.state_dict(), saveNamePrefix + '.pth')

    # log
    logging.info('Epoch#%d: total loss=%.4f, Cls_loss=%.4f, Reg_loss=%.4f, Err_test=%.4f, lr = %.6f'
    %(epoch, train_loss_add, Cls_loss_add, Reg_loss_add, Error_test, scheduler.get_lr()[0]))

def test():
net = model.A2J_model(num_classes = keypointsNumber)
net.load_state_dict(torch.load(model_dir))
net = net.cuda()
net.eval()

post_precess = anchor.post_process(shape=[cropHeight//16,cropWidth//16],stride=16,P_h=None, P_w=None)

output = torch.FloatTensor()
torch.cuda.synchronize() 
for i, (img, label) in tqdm(enumerate(test_dataloaders)):    
    with torch.no_grad():

        img, label = img.cuda(), label.cuda()    
        heads = net(img)  
        pred_keypoints = post_precess(heads,voting=False)
        output = torch.cat([output,pred_keypoints.data.cpu()], 0)
    
torch.cuda.synchronize()       

result = output.cpu().data.numpy()
writeTxt(result)
error = errorCompute(result, keypointsWorldtest)
print('Error:', error)

def errorCompute(source, target):
assert np.shape(source)==np.shape(target), "source has different shape with target"

Test1_ = source.copy()
target_ = target.copy()
Test1_[:, :, 0] = source[:,:,1]
Test1_[:, :, 1] = source[:,:,0]
Test1 = Test1_  # [x, y, z]

for i in range(len(Test1_)):

    Test1[i,:,0] = Test1_[i,:,0]*320/cropWidth  # x
    Test1[i,:,1] = Test1_[i,:,1]*240/cropHeight  # y
    Test1[i,:,2] = source[i,:,2]

labels = pixel2world(target_)
outputs = pixel2world(Test1.copy())

errors = np.sqrt(np.sum((labels - outputs) ** 2, axis=2))

return np.mean(errors)

def writeTxt(result):

resultUVD_ = result.copy()
resultUVD_[:, :, 0] = result[:,:,1]
resultUVD_[:, :, 1] = result[:,:,0]
resultUVD = resultUVD_  # [x, y, z]


for i in range(len(result)):

    resultUVD[i,:,0] = resultUVD_[i,:,0]*320/cropWidth  # x
    resultUVD[i,:,1] = resultUVD_[i,:,1]*240/cropHeight  # y
    resultUVD[i,:,2] = result[i,:,2]

resultReshape = resultUVD.reshape(len(result), -1)

with open(os.path.join(save_dir, result_file), 'w') as f:     
    for i in range(len(resultReshape)):
        for j in range(keypointsNumber*3):
            f.write(str(resultReshape[i, j])+' ')
        f.write('\n') 

f.close()

if name == 'main':
train()
test()

mstc-xqp · Answer 10 · Wed Sep 23 2020 21:32:40 GMT+0800 (China Standard Time)

import cv2
import torch
import torch.utils.data
import torch.optim.lr_scheduler as lr_scheduler
import numpy as np
import scipy.io as scio
import os
from PIL import Image
from torch.autograd import Variable
import model as model
import anchor as anchor
from tqdm import tqdm
import random_erasing
import logging
import time
import datetime
import random

os.environ["CUDA_VISIBLE_DEVICES"] = "0"

DataHyperParms

TrainImgFrames = 1000
TestImgFrames = 100
keypointsNumber = 15
cropWidth = 288
cropHeight = 288
batch_size = 64
learning_rate = 0.00035
Weight_Decay = 1e-4
nepoch = 35
RegLossFactor = 3
spatialFactor = 0.5
RandCropShift = 5
RandshiftDepth = 1
RandRotate = 180
RandScale = (1.0, 0.5)

randomseed = 12345
random.seed(randomseed)
np.random.seed(randomseed)
torch.manual_seed(randomseed)

save_dir = ''

try:
os.makedirs(save_dir)
except OSError:
pass

trainingImageDir = ''
testingImageDir = '' # mat images
keypointsfileTest = ''
keypointsfileTrain = ''
model_dir = ''
result_file = 'result_test.txt'

def pixel2world(x):
x[:, :, 0] = (x[:, :, 0] - 160.0) * x[:, :, 2] * 0.0035
x[:, :, 1] = (120.0 - x[:, :, 1]) * x[:, :, 2] * 0.0035
return x

def world2pixel(x):
x[:, :, 0] = 160.0 + x[:, :, 0] / (x[:, :, 2] * 0.0035)
x[:, :, 1] = 120.0 - x[:, :, 1] / (x[:, :, 2] * 0.0035)
return x

joint_id_to_name = {
0: 'Head',
1: 'Neck',
2: 'RShoulder',
3: 'LShoulder',
4: 'RElbow',
5: 'LElbow',
6: 'RHand',
7: 'LHand',
8: 'Torso',
9: 'RHip',
10: 'LHip',
11: 'RKnee',
12: 'LKnee',
13: 'RFoot',
14: 'LFoot',
}

loading GT keypoints and center points

keypointsWorldtest = scio.loadmat(keypointsfileTest)['keypoints3D'].astype(np.float32)
keypointsPixeltest = np.ones((len(keypointsWorldtest),15,2),dtype='float32')
keypointsPixeltest = world2pixel(keypointsWorldtest)

keypointsWorldtrain = scio.loadmat(keypointsfileTrain)['keypoints3D'].astype(np.float32)
keypointsPixeltrain = np.ones((len(keypointsWorldtrain),15,2),dtype='float32')
keypointsPixeltrain = world2pixel(keypointsWorldtrain)

def transform(img, label, matrix):
'''
img: [H, W] label, [N,2]
'''
img_out = cv2.warpAffine(img,matrix,(cropWidth,cropHeight))
label_out = np.ones((keypointsNumber, 3))
label_out[:,:2] = label[:,:2].copy()
label_out = np.matmul(matrix, label_out.transpose())
label_out = label_out.transpose()

return img_out, label_out

def dataPreprocess(index, img, keypointsUVD, depth_thres=0.4, augment=True):

imageOutputs = np.ones((cropHeight, cropWidth, 1), dtype='float32') 
labelOutputs = np.ones((keypointsNumber, 3), dtype = 'float32') 

if augment:
    RandomOffset_1 = np.random.randint(-1*RandCropShift,RandCropShift)
    RandomOffset_2 = np.random.randint(-1*RandCropShift,RandCropShift)
    RandomOffset_3 = np.random.randint(-1*RandCropShift,RandCropShift)
    RandomOffset_4 = np.random.randint(-1*RandCropShift,RandCropShift)
    RandomOffsetDepth = np.random.normal(0, RandshiftDepth, cropHeight*cropWidth).reshape(cropHeight,cropWidth) 
    RandomOffsetDepth[np.where(RandomOffsetDepth < RandshiftDepth)] = 0
    RandomRotate = np.random.randint(-1*RandRotate,RandRotate)
    RandomScale = np.random.rand()*RandScale[0]+RandScale[1]
    matrix = cv2.getRotationMatrix2D((cropWidth/2,cropHeight/2),RandomRotate,RandomScale)
else:
    RandomOffset_1, RandomOffset_2, RandomOffset_3, RandomOffset_4 = 0, 0, 0, 0
    RandomRotate = 0
    RandomScale = 1
    RandomOffsetDepth = 0
    matrix = cv2.getRotationMatrix2D((cropWidth/2,cropHeight/2),RandomRotate,RandomScale)

imCrop = img[:, :].copy()

imgResize = cv2.resize(imCrop, (cropWidth, cropHeight), interpolation=cv2.INTER_NEAREST)

imgResize = np.asarray(imgResize,dtype = 'float32')  # H*W*C
imgResize = imgResize / 5

## label
label_xy = np.ones((keypointsNumber, 2), dtype = 'float32') 
label_xy[:,0] = keypointsUVD[index,:,0].copy()*cropWidth/320 # x
label_xy[:,1] = keypointsUVD[index,:,1].copy()*cropHeight/240 # y

if augment:
    imgResize, label_xy = transform(imgResize, label_xy, matrix)  ## rotation, scale

imageOutputs[:,:,0] = imgResize

labelOutputs[:,1] = label_xy[:,0]
labelOutputs[:,0] = label_xy[:,1]
labelOutputs[:,2] = (keypointsUVD[index,:,2])*RandomScale   # Z  

imageOutputs = np.asarray(imageOutputs)
imageNCHWOut = imageOutputs.transpose(2, 0, 1)  # [H, W, C] --->>>  [C, H, W]
imageNCHWOut = np.asarray(imageNCHWOut)
labelOutputs = np.asarray(labelOutputs)

data, label = torch.from_numpy(imageNCHWOut), torch.from_numpy(labelOutputs)

return data, label

###################### Pytorch dataloader #################
class my_dataloader(torch.utils.data.Dataset):

def __init__(self, ImgDir, keypointsUVD, num, augment=True):

    self.ImgDir = ImgDir
    self.keypointsUVD = keypointsUVD
    self.num = num
    self.augment = augment
    self.randomErase = random_erasing.RandomErasing(probability = 0.5, sl = 0.02, sh = 0.4, r1 = 0.3, mean=[0])

def __getitem__(self, index):

    data4D = scio.loadmat(self.ImgDir + str(index+1) + '.mat')['DepthNormal']
    depth = data4D[:,:]

    data, label = dataPreprocess(index, depth, self.keypointsUVD, self.augment)

    if self.augment:
        data = self.randomErase(data)

    return data, label

def __len__(self):
    return self.num

train_image_datasets = my_dataloader(trainingImageDir, keypointsWorldtrain, TrainImgFrames, augment=True)
train_dataloaders = torch.utils.data.DataLoader(train_image_datasets, batch_size = batch_size,
shuffle = True, num_workers = 8)

test_image_datasets = my_dataloader(testingImageDir, keypointsWorldtest, TestImgFrames, augment=False)
test_dataloaders = torch.utils.data.DataLoader(test_image_datasets, batch_size = batch_size,
shuffle = False, num_workers = 8)

def train():

net = model.A2J_model(num_classes = keypointsNumber)
net = net.cuda()

post_precess = anchor.post_process(shape=[cropHeight//16,cropWidth//16],stride=16,P_h=None, P_w=None)
criterion = anchor.A2J_loss(shape=[cropHeight//16,cropWidth//16],thres = [16.0,32.0],stride=16,\
    spatialFactor=spatialFactor,img_shape=[cropHeight, cropWidth],P_h=None, P_w=None)
optimizer = torch.optim.Adam(net.parameters(), lr=learning_rate, weight_decay=Weight_Decay)
scheduler = lr_scheduler.StepLR(optimizer, step_size=10, gamma=0.2)

logging.basicConfig(format='%(asctime)s %(message)s', datefmt='%Y/%m/%d %H:%M:%S', \
                filename=os.path.join(save_dir, 'train.log'), level=logging.INFO)
logging.info('======================================================')

for epoch in range(nepoch):
    net = net.train()
    train_loss_add = 0.0
    Cls_loss_add = 0.0
    Reg_loss_add = 0.0
    timer = time.time()

    # Training loop
    for i, (img, label) in enumerate(train_dataloaders):

        torch.cuda.synchronize() 

        img, label = img.cuda(), label.cuda()
        heads  = net(img)
        #print(regression)     
        optimizer.zero_grad()  
        
        Cls_loss, Reg_loss = criterion(heads, label)

        loss = 1*Cls_loss + Reg_loss*RegLossFactor
        loss.backward()
        optimizer.step()

        torch.cuda.synchronize()
        
        train_loss_add = train_loss_add + (loss.item())*len(img)
        Cls_loss_add = Cls_loss_add + (Cls_loss.item())*len(img)
        Reg_loss_add = Reg_loss_add + (Reg_loss.item())*len(img)

        # printing loss info
        if i%10 == 0:
            print('epoch: ',epoch, ' step: ', i, 'Cls_loss ',Cls_loss.item(), 'Reg_loss ',Reg_loss.item(), ' total loss ',loss.item())

    scheduler.step(epoch)

    # time taken
    torch.cuda.synchronize()
    timer = time.time() - timer
    timer = timer / TrainImgFrames
    print('==> time to learn 1 sample = %f (ms)' %(timer*1000))

    train_loss_add = train_loss_add / TrainImgFrames
    Cls_loss_add = Cls_loss_add / TrainImgFrames
    Reg_loss_add = Reg_loss_add / TrainImgFrames
    print('mean train_loss_add of 1 sample: %f, #train_indexes = %d' %(train_loss_add, TrainImgFrames))
    print('mean Cls_loss_add of 1 sample: %f, #train_indexes = %d' %(Cls_loss_add, TrainImgFrames))
    print('mean Reg_loss_add of 1 sample: %f, #train_indexes = %d' %(Reg_loss_add, TrainImgFrames))

    Error_test = 0
    Error_train = 0
    Error_test_wrist = 0

    if (epoch % 1 == 0):  
        net = net.eval()
        output = torch.FloatTensor()
        outputTrain = torch.FloatTensor()

        for i, (img, label) in tqdm(enumerate(test_dataloaders)):
            with torch.no_grad():
                img, label = img.cuda(), label.cuda()       
                heads = net(img)  
                pred_keypoints = post_precess(heads, voting=False)
                output = torch.cat([output,pred_keypoints.data.cpu()], 0)

        result = output.cpu().data.numpy()
        Error_test = errorCompute(result,keypointsWorldtest,)
        print('epoch: ', epoch, 'Test error:', Error_test)
        saveNamePrefix = '%s/net_%d_wetD_' % (save_dir, epoch) + str(Weight_Decay) + '_depFact_' + str(spatialFactor) + '_RegFact_' + str(RegLossFactor) + '_rndShft_' + str(RandCropShift)
        torch.save(net.state_dict(), saveNamePrefix + '.pth')

    # log
    logging.info('Epoch#%d: total loss=%.4f, Cls_loss=%.4f, Reg_loss=%.4f, Err_test=%.4f, lr = %.6f'
    %(epoch, train_loss_add, Cls_loss_add, Reg_loss_add, Error_test, scheduler.get_lr()[0]))

def test():
net = model.A2J_model(num_classes = keypointsNumber)
net.load_state_dict(torch.load(model_dir))
net = net.cuda()
net.eval()

post_precess = anchor.post_process(shape=[cropHeight//16,cropWidth//16],stride=16,P_h=None, P_w=None)

output = torch.FloatTensor()
torch.cuda.synchronize() 
for i, (img, label) in tqdm(enumerate(test_dataloaders)):    
    with torch.no_grad():

        img, label = img.cuda(), label.cuda()    
        heads = net(img)  
        pred_keypoints = post_precess(heads,voting=False)
        output = torch.cat([output,pred_keypoints.data.cpu()], 0)
    
torch.cuda.synchronize()       

result = output.cpu().data.numpy()
writeTxt(result)
error = errorCompute(result, keypointsWorldtest)
print('Error:', error)

def errorCompute(source, target):
assert np.shape(source)==np.shape(target), "source has different shape with target"

Test1_ = source.copy()
target_ = target.copy()
Test1_[:, :, 0] = source[:,:,1]
Test1_[:, :, 1] = source[:,:,0]
Test1 = Test1_  # [x, y, z]

for i in range(len(Test1_)):

    Test1[i,:,0] = Test1_[i,:,0]*320/cropWidth  # x
    Test1[i,:,1] = Test1_[i,:,1]*240/cropHeight  # y
    Test1[i,:,2] = source[i,:,2]

labels = pixel2world(target_)
outputs = pixel2world(Test1.copy())

errors = np.sqrt(np.sum((labels - outputs) ** 2, axis=2))

return np.mean(errors)

def writeTxt(result):

resultUVD_ = result.copy()
resultUVD_[:, :, 0] = result[:,:,1]
resultUVD_[:, :, 1] = result[:,:,0]
resultUVD = resultUVD_  # [x, y, z]


for i in range(len(result)):

    resultUVD[i,:,0] = resultUVD_[i,:,0]*320/cropWidth  # x
    resultUVD[i,:,1] = resultUVD_[i,:,1]*240/cropHeight  # y
    resultUVD[i,:,2] = result[i,:,2]

resultReshape = resultUVD.reshape(len(result), -1)

with open(os.path.join(save_dir, result_file), 'w') as f:     
    for i in range(len(resultReshape)):
        for j in range(keypointsNumber*3):
            f.write(str(resultReshape[i, j])+' ')
        f.write('\n') 

f.close()

if name == 'main':
train()
test()
Thank you very much!! bro! mind i add your email or skype .my email is 15619217958@163.com

mstc-xqp · Answer 11 · Wed Sep 23 2020 21:43:14 GMT+0800 (China Standard Time)

If you don't to use bndbox, you may train the data by yourself, or modify the train code.

Hi bro.Can u share the itop model or k2phd model.
best regards!

Shreyas Narasimhiah Ramesh · Answer 12 · Sun Jun 12 2022 18:27:35 GMT+0800 (China Standard Time)

Hi @zeroXscorpion7 ,

Were you able to utilize this model to predict the Joints for a custom dataset?
I'm also trying to pass one depth frame along with the ITOP side dataset and change the mean value so that the input depth frame to the model matches with the ITOP_side dataset depth frame.
Unfortunately, the results are very bad.
Could you tell me if you were able to do something more on this?