search

dBeker / Faster-RCNN-TensorFlow-Python3

Tensorflow Faster R-CNN for Windows/Linux and Python 3 (3.5/3.6/3.7)
MIT License
609 stars 329 forks source link

there is a problem about convergence #58

Closed guokainew closed 5 years ago

guokainew commented 5 years ago

When i train my image use this Code,there is a problem about convergence: Training results shows that the 'total loss' is fluctuate around 2 ,and did not decrease anymore; Second, the 'loss_box' always decay to zero and it will rise a little only; Third, the training is always cut off, no reason; Fourth, the candidate boxes are generally small than the fact; Fiveth, when i run the demo there are undetectable object. So, what happen to my project, please reply to me as soon as possible,thank U!

morpheusthewhite commented 5 years ago
  1. I have enough information to help you, did you use this repository without changing anything?
  2. Same as 1
  3. What do you mean? Doesn't train start?
  4. That depends on the annotations you are using
  5. The demo detects only one object per image
guokainew commented 5 years ago

I am very pleasure to receive the reply from you, and i have solved this problem by some adjustment, thank yoo very much!

---Original--- From: "morpheusthewhite"notifications@github.com Date: Thu, May 23, 2019 17:32 PM To: "dBeker/Faster-RCNN-TensorFlow-Python3.5"Faster-RCNN-TensorFlow-Python3.5@noreply.github.com; Cc: "guokainew"2496296333@qq.com;"Author"author@noreply.github.com; Subject: Re: [dBeker/Faster-RCNN-TensorFlow-Python3.5] there is a problem about convergence (#58)

I have enough information to help you, did you use this repository without changing anything?

Same as 1

What do you mean? Doesn't train start?

That depends on the annotations you are using

The demo detects only one object per image

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or mute the thread.

morpheusthewhite commented 5 years ago

You're welcome

AICVHub commented 5 years ago

I am very pleasure to receive the reply from you, and i have solved this problem by some adjustment, thank yoo very much! … ---Original--- From: "morpheusthewhite"notifications@github.com Date: Thu, May 23, 2019 17:32 PM To: "dBeker/Faster-RCNN-TensorFlow-Python3.5"Faster-RCNN-TensorFlow-Python3.5@noreply.github.com; Cc: "guokainew"2496296333@qq.com;"Author"author@noreply.github.com; Subject: Re: [dBeker/Faster-RCNN-TensorFlow-Python3.5] there is a problem about convergence (#58) I have enough information to help you, did you use this repository without changing anything? Same as 1 What do you mean? Doesn't train start? That depends on the annotations you are using The demo detects only one object per image — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or mute the thread.

I meet the same problems! Glad to see you issue. How did you do to slove this problem? Could you teach me about it?