search

xslittlegrass / CarND-Vehicle-Detection

Vehicle detection using YOLO in Keras runs at 21FPS
389 stars 175 forks source link

weight- module #1

Closed Henrilin28 closed 7 years ago

Henrilin28 commented 7 years ago

Hi, I tried to reproduce you result but I found the yolo-tiny weight is not working.. Can you show me where you get the weights?

Best, Henry

xslittlegrass commented 7 years ago

Hi Henry,

The weight file can be downloaded from here https://drive.google.com/file/d/0B1tW_VtY7onibmdQWE1zVERxcjQ/view?usp=sharing

Let me know if you have problems.

Henrilin28 commented 7 years ago

HI @xslittlegrass,

It works but the same model have different result! I cant detect any car!

btw

add the following code can show the model layers

from IPython.display import SVG
from keras.utils.visualize_util import model_to_dot

SVG(model_to_dot(model).create(prog='dot', format='svg'))
xslittlegrass commented 7 years ago

Are you using the same test images in the test_images folder? Can you post your results?

Henrilin28 commented 7 years ago

Yes, I am using the same images you provided. 

Please find my result in attachment

--  Chuan-Heng (Henry) Lin Tel: +886 938-507-930  Email: imhenry@me.com ; Website: www.henrilin28.com

在 2017年2月24日 於 上午10:36:22, xslittlegrass (notifications@github.com) 寫下:

Are you using the same test images in the test_images folder? Can you post your results?

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

xslittlegrass commented 7 years ago

You attached the files in your email and I cannot see your them, could you attach them on this issue page on Github?

Henrilin28 commented 7 years ago

Here it is! 2017-02-24 10 37 20

xslittlegrass commented 7 years ago

It works fine on my system. I'm using tensorflow 0.12.0-rc0 on python 3.5.2. Could you update Keras to the newest version?

Henrilin28 commented 7 years ago

you use tensorflow for backend?

Henrilin28 commented 7 years ago

I use python 2.7!

xslittlegrass commented 7 years ago

Yes. Maybe the backend is the reason. The weights in these two backends are different, you need to convert them. There is a script in this wiki page on Keras that can help you to do that.

Henrilin28 commented 7 years ago

but you set the backend to "th", which means theano?

xslittlegrass commented 7 years ago

The backend and the image dimension ordering are two different settings. I set set_image_dim_ordering to "th". That's only for the image dimension ordering. For instance, whether the image is organized as [3,448,448] ("theano" ordering) or [448,448,3] ("tensorflow" ordering). But I'm still using tensorflow as the backend. See here. I did this way because the pretrained weight file in yolo is organized this way.