**完,只要讀取模型,然後進行**就ok了.
1.讀模型
# set up model
model = darknet(opt.model_config_path)
model.load_weights(opt.weight_path)
model.cuda()
model.eval() #
dataset = datasets(opt.valid)
dataloader = torch.utils.data.dataloader(
dataset, batch_size=opt.batch_size, shuffle=false)
for batch_i, (img_paths, input_imgs, targets) in enumerate(dataloader):
# configure input
input_imgs = input_imgs.type(tensor)
# get detections
with torch.no_grad():
detections = model(input_imgs)
detections = non_max_suppression(
detections, 80, 0.8, 0.4)
# log progress
# 繪製**和真值影象
draw_predict(input_imgs, detections, img_paths)
draw_predict(input_imgs, targets, img_paths, 'gt')
current_time = time.time()
inference_time = datetime.timedelta(seconds=current_time - prev_time)
prev_time = current_time
print ('\t+ batch %d, inference time: %s' % (batch_i, inference_time))
讀影象
for detection in detections:
if detection is none:
continue
img_path = img_paths[ii]
# 由於是torch.cuda()格式,需要轉化為陣列
image = imgs[ii].cpu().numpy() * 255
# opencv為gbr,所以要轉化通道 ,這裡有點bug,所以先存在讀取.
w, h, _ = image.shape
if types == 'pred':
# yolo**的為(x1,y1,x2,y2,c,score,label)
bboxs = detection[:, 0:4]
name = detection[:, -1].cpu().numpy()
score = detection[:, 5]
# print detection.shape
x1 = detection[:, 0]
y1 = detection[:, 1]
x2 = detection[:, 2]
y2 = detection[:, 3]
else:
# yolo的真值標籤為(label,x,y,w,h) 其中,x,y是bounding box矩陣中心
# 這裡求出x1,y1,x2,y2
target = np.zeros((50, 5))
target[:, 0] = detection[:, 0]
target[:, 1] = (detection[:, 1] - detection[:, 3] / 2.0) * w
target[:, 2] = (detection[:, 2] - detection[:, 4] / 2.0) * h
target[:, 3] = (detection[:, 1] + detection[:, 3] / 2.0) * w
target[:, 4] = (detection[:, 2] + detection[:, 4] / 2.0) * h
# 由於真值包含零填充,所以去除掉
detection = target[np.sum(target, axis=1) > 0]
if detection is none:
continue
bboxs = detection[:, 1:5]
name = detection[:, 0]
# score = 1.0
# print detection.shape
x1 = detection[:, 1]
y1 = detection[:, 2]
x2 = detection[:, 3]
y2 = detection[:, 4]
for i in range(len(name)):
xmin = int(round(float(x1[i])))
ymin = int(round(float(y1[i])))
xmax = int(round(float(x2[i])))
ymax = int(round(float(y2[i])))
# if score[i] <= 0.9:
# continue
if xmax < xmin or ymax < ymin:
continue
# 繪製bounding box
cv2.rectangle(image, (xmin, ymin),
(xmax, ymax), (0, 0, 255), 1)
# 繪製label
cv2.puttext(image, classname[int(name[i])], (xmin, ymin - 10),
cv2.font_hershey_******x, 1e-3 * image.shape[0], (0, 0, 255), 1)
# 繪製標籤
Deep Learning 從頭開始
deep learning已經火了好久，有些人已經在這裡面耕耘了好多年，而有些人才剛剛開始，比如本人。如何才能快速地進入這個領域在較短的時間內掌握deep learning最新的技術是值得思考的問題。就目前的情況看，通過網路上的課程及各種tutorials以及各種 來研究這個領域是最佳的途徑。經過一...
git從頭開始
當你本地修改了乙個檔案，而且該檔案被另乙個人修改，並push了，那麼 users terry workspace git練習 git practise git master git pull updating 67e4e18.cdbf666 error your local changes to t...
English 從頭開始
我們有好多事情都不能重新開始，比如我們的人生你沒有辦法把自己在塞回媽媽的肚子裡吧？比如我們的時間在此時此刻只有這乙個時間，全世界不會再有第二個。比如我們後悔的事情.雖然有很多我們無法改變的事情存在，但也有許多我們可以改變的事情。雖然我們沒有辦法從一歲開始重新開始，但我們可以掌握自己的人生，做自己想做...