def load_pytorch_model():
model = pose_estimation.PoseModel(num_point=19,
num_vector=19,
pretrained=True)
return model
def construct_model(args):
model = pose_estimation.PoseModel(num_point=19, num_vector=19, pretrained=True)
# state_dict = torch.load(args.pretrained)['state_dict']
# from collections import OrderedDict
# new_state_dict = OrderedDict()
# for k, v in state_dict.items():
# name = k[7:]
# new_state_dict[name] = v
# model.load_state_dict(new_state_dict)
# model.fc = nn.Linear(2048, 80)
model = torch.nn.DataParallel(model, device_ids=args.gpu).cuda()
return model
def construct_model(args):
if not args.snapshot:
model = pose_estimation.PoseModel(num_point=19, num_vector=19, pretrained=True)
else:
print('--------load model from {}----------------'.format(args.snapshot))
model = pose_estimation.PoseModel(num_point=19, num_vector=19, pretrained=True)
state_dict = torch.load(args.snapshot)['state_dict']
model.load_state_dict(state_dict)
# if not args.pretrained:
# model = pose_estimation.PoseModel(num_point=19, num_vector=19, pretrained=True)
# else:
# state_dict = torch.load(args.pretrained)['state_dict']
# from collections import OrderedDict
# new_state_dict = OrderedDict()
# for k, v in state_dict.items():
# name = k[7:]
# new_state_dict[name] = v
# model.load_state_dict(new_state_dict)
# os.environ['CUDA_VISIBLE_DEVICES'] = ','.join([str(gpu) for gpu in args.gpu])
# model = torch.nn.DataParallel(model, device_ids=range(len(args.gpu))).cuda()
model.cuda() # single gpu
return model
def construct_model(args):
model = pose_estimation.PoseModel(num_point=19, num_vector=19)
state_dict = torch.load(args.model)['state_dict']
from collections import OrderedDict
new_state_dict = OrderedDict()
for k, v in state_dict.items():
name = k[7:]
new_state_dict[name] = v
state_dict = model.state_dict()
state_dict.update(new_state_dict)
model.load_state_dict(state_dict)
model = model.cuda()
model.eval()
return model
def main():
import pose_estimation
#pytorch_model = '/home/xiangyu/data/pretrain/COCO/coco_pose_iter_440000.pth.tar'
pytorch_model = '/home/xiangyu/samsung_pose/experiments/baseline/60000.pth.tar'
model = pose_estimation.PoseModel(num_point=19, num_vector=19)
img_dir = '/home/xiangyu/data/coco/images/val2014/'
annFile = '/home/xiangyu/data/coco/annotations/person_keypoints_minival2014.json'
num_imgs = 50 #
orderCOCO = [
0, -1, 6, 8, 10, 5, 7, 9, 12, 14, 16, 11, 13, 15, 2, 1, 4, 3
] #[1, 0, 7, 9, 11, 6, 8, 10, 13, 15, 17, 12, 14, 16, 3, 2, 5, 4]
myjsonValidate = list(dict())
cocoGt = COCO(annFile)
img_names = cocoGt.imgs
# filter only person
cats = cocoGt.loadCats(cocoGt.getCatIds())
catIds = cocoGt.getCatIds(catNms=['person'])
imgIds = cocoGt.getImgIds(catIds=catIds)
#-------------------------- pytorch model------------------
state_dict = torch.load(pytorch_model)['state_dict']
model.load_state_dict(state_dict)
model = model.cuda()
model.eval()
#--------------------------------------------------------
#
for i in range(num_imgs):
print('{}/{}'.format(i, num_imgs))
img_info = cocoGt.loadImgs(imgIds[i])[0]
image_id = img_info['id']
oriImg = cv2.imread(os.path.join(img_dir, img_info['file_name']))
multiplier = [x * boxsize / oriImg.shape[0] for x in scale_search]
# apply model
candidate, subset, canvas = apply_model(oriImg, model, multiplier)
#cv2.imwrite(os.path.join('./result', img_info['file_name']), canvas)
for j in range(len(subset)):
category_id = 1
keypoints = np.zeros(51)
score = 0
for part in range(18):
if part == 1:
continue
index = int(subset[j][part])
if index > 0:
#realpart = orderCOCO[part] - 1
realpart = orderCOCO[part]
if realpart == -1:
continue
# if part == 0:
# keypoints[realpart * 3] = candidate[index][0] -0.5
# keypoints[realpart * 3 + 1] = candidate[index][1] -0.5
# keypoints[realpart * 3 + 2] = 1
# # score = score + candidate[index][2]
else:
keypoints[(realpart) * 3] = candidate[index][0]
keypoints[(realpart) * 3 + 1] = candidate[index][1]
keypoints[(realpart) * 3 + 2] = 1
# score = score + candidate[index][2]
keypoints_list = keypoints.tolist()
current_dict = {
'image_id': image_id,
'category_id': category_id,
'keypoints': keypoints_list,
'score': subset[j][-2]
}
myjsonValidate.append(current_dict)
#count = count + 1
import json
with open('evaluationResult.json', 'w') as outfile:
json.dump(myjsonValidate, outfile)
resJsonFile = 'evaluationResult.json'
cocoDt2 = cocoGt.loadRes(resJsonFile)
image_ids = []
for i in range(num_imgs):
img = cocoGt.loadImgs(imgIds[i])[0]
image_ids.append(img['id'])
# running evaluation
cocoEval = COCOeval(cocoGt, cocoDt2, 'keypoints')
cocoEval.params.imgIds = image_ids
cocoEval.evaluate()
cocoEval.accumulate()
k = cocoEval.summarize()
def main():
import pose_estimation
model = pose_estimation.PoseModel(num_point=19, num_vector=19)
img_dir = '/home/bst2017/workspace/data/coco/images/val2014/'
annFile = '/home/bst2017/workspace/data/coco/annotations/person_keypoints_minival2014.json'
num_imgs = 50 #50 # COCO 38%
orderCOCO = [
0, -1, 6, 8, 10, 5, 7, 9, 12, 14, 16, 11, 13, 15, 2, 1, 4, 3
] #[1, 0, 7, 9, 11, 6, 8, 10, 13, 15, 17, 12, 14, 16, 3, 2, 5, 4]
myjsonValidate = list(dict())
cocoGt = COCO(annFile)
img_names = cocoGt.imgs
# filter only person
cats = cocoGt.loadCats(cocoGt.getCatIds())
catIds = cocoGt.getCatIds(catNms=['person'])
imgIds = cocoGt.getImgIds(catIds=catIds)
#ids = list(cocoGt.imgs.keys())
#--------------------------------------------------------
#
for i in range(num_imgs):
print('{}/{}'.format(i, num_imgs))
img_info = cocoGt.loadImgs(imgIds[i])[0]
image_id = img_info['id']
oriImg = cv2.imread(os.path.join(img_dir, img_info['file_name']))
ann_ids = cocoGt.getAnnIds(imgIds=image_id)
img_anns = cocoGt.loadAnns(ann_ids)
candidate, subset, canvas = mechanism(oriImg, img_anns)
cv2.imwrite(os.path.join('./result', img_info['file_name']), canvas)
for j in range(len(subset)):
category_id = 1
keypoints = np.zeros(51)
score = 0
for part in range(18):
if part == 1:
continue
index = int(subset[j][part])
if index > 0:
#realpart = orderCOCO[part] - 1
realpart = orderCOCO[part]
if realpart == -1:
continue
# if part == 0:
# keypoints[realpart * 3] = candidate[index][0] -0.5
# keypoints[realpart * 3 + 1] = candidate[index][1] -0.5
# keypoints[realpart * 3 + 2] = 1
# # score = score + candidate[index][2]
else:
keypoints[(realpart) * 3] = candidate[index][0]
keypoints[(realpart) * 3 + 1] = candidate[index][1]
keypoints[(realpart) * 3 + 2] = 2
# score = score + candidate[index][2]
keypoints_list = keypoints.tolist()
current_dict = {
'image_id': image_id,
'category_id': category_id,
'keypoints': keypoints_list,
'score': subset[j][-2]
}
myjsonValidate.append(current_dict)
#count = count + 1
import json
with open('evaluationResult.json', 'w') as outfile:
json.dump(myjsonValidate, outfile)
resJsonFile = 'evaluationResult.json'
cocoDt2 = cocoGt.loadRes(resJsonFile)
image_ids = []
for i in range(num_imgs):
img = cocoGt.loadImgs(imgIds[i])[0]
image_ids.append(img['id'])
# running evaluation
cocoEval = COCOeval(cocoGt, cocoDt2, 'keypoints')
cocoEval.params.imgIds = image_ids
cocoEval.evaluate()
cocoEval.accumulate()
k = cocoEval.summarize()
