def __init__(self):
parser = argparse.ArgumentParser(
description='PyTorch hand pose Training')
args = parser.parse_args()
'''user setting'''
args.gpu_ids = '0'
args.use_net = ['hpe1_orig', 'hpe2'] #hpe1_orig, hig, hpe1, hpe2
forward_list = ['hpe1_orig', 'hpe2'] #hpe1_orig, hig_hpe1, hpe2
trained_modelFile_hpe1_orig = basedir + '..\\HPE\\output\\2020_2_26_11_29\\trainedModel_epoch99.pth.tar'
trained_modelFile_hpe2 = basedir + '..\\HPE2\\output\\2020_4_10_18_46\\trainedModel_epoch99.pth.tar'
'''common setting'''
out_list = {}
args.train_net = []
args.cgan = True
args.is_train = False
args.gan_mode = 'vanilla'
args.trainImageSize = 128
trainImageSize = 128
args.skeleton_pca_dim = 52
args.skeleton_orig_dim = 63
args.discriminator_reconstruction = False
args.test_batch = 1
hpe_numBlocks = 5
#device
#device = torch.device(torch.device("cuda:%s"%args.gpu_ids if torch.cuda.is_available() else "cpu")
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
args.device = device
cudnn.benchmark = True
#--HPE_orig net setting
if 'hpe1_orig' in args.use_net:
if not 'hpe1_orig' in locals():
hpe1_orig = dpnet(args.skeleton_pca_dim, hpe_numBlocks, device)
checkpoint = torch.load(
trained_modelFile_hpe1_orig,
map_location=lambda storage, loc: storage)
hpe1_orig.load_state_dict(checkpoint['state_dict'])
#--HPE2 net setting
if 'hpe2' in args.use_net:
if not 'hpe2' in locals():
hpe2 = dpnet(args.skeleton_pca_dim, hpe_numBlocks, device)
checkpoint = torch.load(
trained_modelFile_hpe2,
map_location=lambda storage, loc: storage)
hpe2.load_state_dict(checkpoint['state_dict_hpe2'])
with open(basedir + '..\\pca_52_by_957032_augmented_X2.pickle',
'rb') as f:
self.pca = pickle.load(f)
#--fusion net setting
#fusionnet=FusionNet(hpe1_orig,pix2pix,hpe1,hpe2,None,args)
fl = []
if 'hpe1_orig' in locals(): fl.append(hpe1_orig)
else: fl.append(None)
if 'pix2pix' in locals(): fl.append(pix2pix)
else: fl.append(None)
if 'hpe1' in locals(): fl.append(hpe1)
else: fl.append(None)
if 'hpe2' in locals(): fl.append(hpe2)
else: fl.append(None)
fusionnet = FusionNet(fl[0], fl[1], fl[2], fl[3], None, args)
self.fusionnet = fusionnet.to(device)
#--dataset of uvr (for utils)
camera_info = {}
camera_info['fx'] = 475.065948
camera_info['fy'] = 475.065857
camera_info['cx'] = 315.944855
camera_info['cy'] = 245.287079
camera_info['camerawidth'] = 640
camera_info['cameraheight'] = 480
camera_info['cube'] = np.asarray([250, 250, 250])
camera_info['trainImageSize'] = 128
self.trainImageSize = camera_info['trainImageSize']
self.cube = camera_info['cube']
camerawidth = camera_info['camerawidth']
cameraheight = camera_info['cameraheight']
jointnum = 21
d_maximum = 500
self.datasetloader_uvr = datasetloader_UVR('..', 0, camera_info, '..',
'..')
self.utils = self.datasetloader_uvr.utils
#--start
self.fusionnet.set_mode('eval')
self.ir_batch = np.zeros(
(1, 1, args.trainImageSize, args.trainImageSize), dtype=np.float32)
self.depth_batch = np.zeros(
(1, 1, args.trainImageSize, args.trainImageSize), dtype=np.float32)
camera_info['cameraheight']=240
elif dataset_version=='v2':
camera_info['fx']=475.065948
camera_info['fy']=475.065857
camera_info['cx']=315.944855
camera_info['cy']=245.287079
camera_info['camerawidth']=640
camera_info['cameraheight']=480
camera_info['cube']=np.asarray([250,250,250])
camera_info['trainImageSize']=128
trainImageSize=camera_info['trainImageSize']
cube=camera_info['cube']
utils=Utils(camera_info)
imgs_train=np.zeros((trainImageSize,2*trainImageSize),'float32')
datasetloader_uvr=datasetloader_UVR('..',0,camera_info,'..','..')
dataset_dict={}
dataset_dict['com']=[]
dataset_dict['window']=[]
#--start
progressbar=trange(frame_start,data_num//thread_num,leave=True)
for g in progressbar:
with Manager() as manager:
Mobj=manager.list([0 for x in range(thread_num)])
MpList=[]
depth_train_bag=[]
window_bag=[]
com_bag=[]
camera_info['fy'] = 475.065857
camera_info['cx'] = 315.944855
camera_info['cy'] = 245.287079
camera_info['camerawidth'] = 640
camera_info['cameraheight'] = 480
camera_info['cube'] = np.asarray([250, 250, 250])
camera_info['trainImageSize'] = 128
trainImageSize = camera_info['trainImageSize']
cube = camera_info['cube']
utils = Utils(camera_info)
imgs_train = np.zeros((trainImageSize, 2 * trainImageSize), 'float32')
#--dataset of uvr
datasetloader_uvr = datasetloader_UVR('..', data_num, camera_info, 'test',
'generator')
#--start
ground2d = {}
for frame in range(data_num):
depth = cv2.imread(load_filepath_img + 'depth%d.png' % frame, 2)
ir = cv2.imread(load_filepath_img + 'ir%d.png' % frame, 2)
depth_seg = depth.copy()
depth_seg[depth_seg > d_maximum] = 0
# preprocess (depth)
depth_train, depth_crop, com, window = datasetloader_uvr.preprocess_depth(
depth_seg, trainImageSize, cube)
# preprocess (ir)
trainImageSize = camera_info['trainImageSize']
cube = camera_info['cube']
#--number of datasets
traindataNum_uvr = 175000 #170000
traindataNum_blur_uvr = 0
traindataNum_vr20 = 0 #vr dataset (slow motion): 215
validateNum = 10000
validateNum_vr20 = 0 #vr test dataset: 2370
#--dataset of uvr
dataset_path_uvr = '/home/yong/ssd/dataset/preprocessed_HIG/'
datasetloader_uvr = {}
datasetloader_uvr['train'] = datasetloader_UVR(dataset_path_uvr,
traindataNum_uvr,
camera_info, 'train',
'generator')
datasetloader_uvr['train_blur'] = datasetloader_UVR(
dataset_path_uvr, traindataNum_blur_uvr, camera_info, 'train_blur',
'generator')
datasetloader_uvr['train'].set_datasetLoader(args.train_batch)
datasetloader_uvr['train_blur'].set_datasetLoader(args.train_batch)
#--dataset of VR20
camera_info_vr20 = {}
camera_info_vr20['fx'] = 475.065948 / 2
camera_info_vr20['fy'] = 475.065857 / 2
camera_info_vr20['cx'] = 315.944855 / 2
camera_info_vr20['cy'] = 245.287079 / 2
camera_info_vr20['camerawidth'] = int(640 / 2)
hpe.load_state_dict(checkpoint['state_dict_hpe1'])
hpe.eval()
img_hpe = np.zeros((1,1,trainImageSize,trainImageSize), dtype = np.float32)
with open(basedir+'..\\pca_52_by_957032_augmented_X2.pickle','rb') as f:
pca=pickle.load(f)
##--init realsense--##
if input_type=='sr300':
realsense=Realsense()
##--dataset--##
jointnum=21
datasetloader_uvr={}
datasetloader_uvr['train']=datasetloader_UVR(load_filepath_preprocess,data_num,camera_info,'train','generator')
datasetloader_uvr['train'].set_datasetLoader(args.test_batch)
datasetloader_uvr['train_blur']=datasetloader_UVR(load_filepath_preprocess,data_num,camera_info,'train_blur','generator')
datasetloader_uvr['train_blur'].set_datasetLoader(args.test_batch)
datasetloader_uvr['test']=datasetloader_UVR(load_filepath_preprocess,data_num,camera_info,'test','generator')
datasetloader_uvr['test'].set_datasetLoader(args.test_batch)
datasetloader_uvr['test_blur']=datasetloader_UVR(load_filepath_preprocess,data_num,camera_info,'test_blur','generator')
datasetloader_uvr['test_blur'].set_datasetLoader(args.test_batch)
utils=datasetloader_uvr[dataset_name].utils
d_maximum=500
if input_type=='dataset':
loadpath_dict="/home/yong/ssd/dataset/preprocessed_HIG/dataset_dict_"+dataset_name+".pickle"
with open(loadpath_dict,'rb') as f:
#--user setting
input_type = 'dataset_uvr' # 'sr300' 'dataset_uvr'
imgfile_path_orig = "/home/yong/ssd/dataset/depth_ir/" + dataset_name + "/"
#--common setting
ointNum = 21
fx = 475.065948
fy = 475.065857
cx = 315.944855
cy = 245.287079
cube = np.asarray([250, 250, 250])
utils = Utils(21, fx, fy, cx, cy, cube)
#--dataset of uvr
datasetloader_uvr = datasetloader_UVR(imgfile_path_orig, 0, 0)
#imgfile_path_orig="/home/yong/Downloads/saveData/uvrDataset - ablation/images/"
##--init realsense--##
if input_type == 'sr300':
pipeline, depth_scale = init_device()
#--start
try:
for frame in range(690, 1000):
#while(1):
#capture
if input_type == 'sr300':
depth_orig, ir_orig = read_frame_from_device(
pipeline, depth_scale)
else:
