def proofread():
# 开始测试。此句是防止改变权值。
net.eval()
rand = np.random.randint(0, len(image_list))
with open(cfg.OUTPUT.LOGGING, 'a') as f:
print('测试图片为:',
image_list[rand],
'\n对应Obj为：',
label_list[rand],
file=f)
test = SingleTest(img_path=image_path + image_list[rand])
image = test.output_data(img_path=image_path + image_list[rand])
image = image.expand(cfg.TEST.IMAGE_BATCH, cfg.TEST.IMAGE_CHANNEL,
cfg.TEST.IMAGE_LENGTH, cfg.TEST.IMAGE_HEIGHT)
image = image.to(device)
mkdir(cfg.TEST.SAVE_OBJ)
with torch.no_grad():
predict = net(image)[0]
vertics, faces = get_vertics(label_list[rand])
vertics = torch.tensor(
vertics.reshape(3, cfg.INPUT.VERTICS_NUM // 3).T.reshape(
cfg.INPUT.VERTICS_NUM))
predict = predict.cpu().numpy()
predict = torch.tensor(pca.inverse_transform(predict))
loss = criterion(vertics, predict)
with open(cfg.OUTPUT.LOGGING, 'a') as f:
print('Loss =', loss.item(), file=f)
# 输出为.obj文件
pre_vertics = []
for order in range(len(predict) // 3):
pre_vertics.append([
predict[3 * order], predict[3 * order + 1], predict[3 * order + 2]
])
writeObj(cfg.TEST.SAVE_OBJ + '/' + label_list[rand], pre_vertics, faces)
def proofread():
# 开始测试。此句是防止改变权值。
net.eval()
rand = np.random.randint(0, len(image_list))
with open(cfg.OUTPUT.LOGGING, 'a') as f:
print('测试图片为:', image_list[rand], '\n对应Obj为：', label_list[rand], file=f)
test = SingleTest(img_path=image_path + image_list[rand])
image = test.output_data(img_path=image_path + image_list[rand])
image = image.expand(cfg.TEST.IMAGE_BATCH, cfg.TEST.IMAGE_CHANNEL, cfg.TEST.IMAGE_LENGTH, cfg.TEST.IMAGE_HEIGHT)
image = image.to(device)
with torch.no_grad():
predict = net(image)[0]
vertics, faces = get_vertics(rand)
vertics = torch.Tensor(vertics).to(device)
loss = criterion(vertics, predict)
with open(cfg.OUTPUT.LOGGING, 'a') as f:
print('测试图片输出数据Loss =', loss.item(), file=f)
# 输出为.obj文件
mkdir(cfg.TEST.SAVE_OBJ)
predict = predict.cpu().numpy()
pre_vertics = []
for order in range(len(predict) // cfg.TEST.VERTICS_DIMENSION):
pre_vertics.append([predict[order], predict[order + len(predict) // cfg.TEST.VERTICS_DIMENSION],
predict[order + len(predict) // cfg.TEST.VERTICS_DIMENSION * 2]])
writeObj(cfg.TEST.SAVE_OBJ + '/' + label_list[rand], pre_vertics, faces)
def save(number):
mkdir(cfg.OUTPUT.PARAMETER)
torch.save(net.state_dict(), cfg.OUTPUT.PARAMETER + '/' + str(number + 1) + '_CNN.pkl',
_use_new_zipfile_serialization=False)
torch.save(optimizer.state_dict(), cfg.OUTPUT.PARAMETER + '/' + str(number + 1) + '_opt.pkl',
_use_new_zipfile_serialization=False)
print('\n网络参数已保存!\n')
def check_before_train(images, labels, faces, item):
restore(images).convert('L')
vertics = []
for i in range(len(labels) // 3):
vertics.append([
labels[i], labels[i + len(labels) // 3],
labels[i + 2 * len(labels) // 3]
])
mkdir(cfg.INPUT.CHECK)
writeObj(cfg.INPUT.CHECK + str(item) + '_check.obj', vertics, faces)
def whole_proofread():
net.eval()
for rand in range(len(image_list)):
print('测试图片为:', image_list[rand])
test = SingleTest(img_path=image_path + image_list[rand])
image = test.output_data(img_path=image_path + image_list[rand])
image = image.expand(cfg.TEST.IMAGE_BATCH, cfg.TEST.IMAGE_CHANNEL,
cfg.TEST.IMAGE_LENGTH, cfg.TEST.IMAGE_HEIGHT)
image = image.to(device)
mkdir(cfg.TEST.SAVE_OBJ)
with torch.no_grad():
predict = net(image)[0]
vertics, faces = get_vertics(0)
predict = predict.cpu().numpy()
predict = torch.tensor(pca.inverse_transform(predict))
# 输出为.obj文件
pre_vertics = []
for order in range(len(predict) // 3):
pre_vertics.append([
predict[3 * order], predict[3 * order + 1],
predict[3 * order + 2]
])
writeObj(cfg.TEST.SAVE_OBJ + '/' + str(rand) + '.obj', pre_vertics,
faces)
import sys
sys.path.insert(0, '../../')
import os
from ConvNet.config.defaults import get_cfg_defaults
from PIL import Image
from torchvision import transforms
from ConvNet.tools.preprocess_data import mkdir
if __name__ == '__main__':
print('The photo data is being resized...')
cfg = get_cfg_defaults()
cfg.merge_from_file(cfg.MODEL.CONFIG + os.listdir(cfg.MODEL.CONFIG)[-1])
transform = transforms.Compose([
transforms.Resize(cfg.DATASETS.TRANSFORM_RESIZE),
])
image_path = cfg.DATASETS.IMAGES_PATH
save_image_path = cfg.INPUT.SAVE_RESIZE_IMAGES
images = os.listdir(image_path)
mkdir(save_image_path)
for i in range(len(images)):
figure = Image.open(image_path + images[i]).convert('L')
new_figure = transform(figure)
new_figure.save(save_image_path + images[i])