def _load_data_set(self):
self.images = []
self.kp2ds = []
self.boxs = []
clk = Clock()
print('start loading coco 2017 dataset.')
#anno_file_path = os.path.join(self.data_folder, 'annotations', 'person_keypoints_train2017.json')
anno_file_path = os.path.join(self.data_folder, 'annotations',
'person_keypoints_val2017.json')
with open(anno_file_path, 'r') as reader:
anno = json.load(reader)
def _hash_image_id_(image_id_to_info, coco_images_info):
for image_info in coco_images_info:
image_id = image_info['id']
image_name = image_info['file_name']
_anno = {}
#_anno['image_path'] = os.path.join(self.data_folder, 'images', 'train-valid2017', image_name)
#_anno['image_path'] = os.path.join(self.data_folder, 'images', 'train2017', image_name)
_anno['image_path'] = os.path.join(self.data_folder, 'images',
'val2017', image_name)
_anno['kps'] = []
_anno['box'] = []
assert not (image_id in image_id_to_info)
image_id_to_info[image_id] = _anno
images = anno['images']
image_id_to_info = {}
_hash_image_id_(image_id_to_info, images)
annos = anno['annotations']
for anno_info in annos:
self._handle_anno_info(anno_info, image_id_to_info)
for k, v in image_id_to_info.items():
self._handle_image_info_(v)
print('finished load coco 2017 dataset, total {} samples.'.format(
len(self.images)))
clk.stop()
from multiprocessing import Process, Array
from time import sleep
from timer import Clock
if __name__ == '__main__':
timer = Clock()
time = Array("i", 2)
clocker = Process(target=timer.increment, args=(time,))
clocker.start()
sleep(4)
print(str(time[0]).zfill(2)+":"+str(time[1]).zfill(2))
clocker.join()
tgt_dis_labels = torch.ones(config.batch_size, dtype=torch.long)
print('Initializing network ...')
if config.is_cuda:
net = Net(config).cuda()
else:
net = Net(config)
criterion_label = torch.nn.CrossEntropyLoss()
criterion_domain = torch.nn.CrossEntropyLoss()
# optimizer_F = torch.optim.SGD(net.feature_extractor.parameters(), lr=0.001, momentum=0.9)
# optimizer_D = torch.optim.SGD(net.adversarial_classifier.parameters(), lr=0.001, momentum=0.9)
optimizer_F = torch.optim.Adam(net.feature_extractor.parameters())
optimizer_D = torch.optim.Adam(net.adversarial_classifier.parameters())
print('Starting training ...')
clock_epoch = Clock(config.epochs)
for epoch in range(config.epochs):
print(' Epoch {}/{}'.format(epoch + 1, config.epochs))
step = 0
clock_batch = Clock(len(src_dataloader))
for src_data, tgt_data in zip(src_dataloader, cycle(tgt_dataloader)):
print(' Batch {}/{}'.format(step + 1, len(src_dataloader)))
src_inputs, src_labels = src_data
tgt_inputs, tgt_labels = tgt_data
# src_inputs = torch.autograd.Variable(src_inputs)
# tgt_inputs = torch.autograd.Variable(tgt_inputs)
if config.is_cuda:
src_inputs, src_labels = src_inputs.cuda(), src_labels.cuda()
tgt_inputs, tgt_labels = tgt_inputs.cuda(), tgt_labels.cuda()
else:
src_inputs, src_labels = src_inputs, src_labels
