chars74k on M7

General discussion about topics related to OpenMV.
ericchu923
Posts: 4
Joined: Mon May 06, 2019 2:09 am

chars74k on M7

Postby ericchu923 » Mon May 06, 2019 2:43 am

Hello,I use M7 board for Machine Learning.
I followed this direction to train smaller network:
https://github.com/kwagyeman/openmv/tre ... ml/cmsisnn
trained network size is 36KB.
when I train my network,accuracy is 0.874359:
Image
but on the M7,it seems not good on it:
Image
I only adjust conv1 and conv2 's num_output from 32 to 16.
Here's my all files for this network(on Google Drive):
https://drive.google.com/drive/folders/ ... sp=sharing
So my question is:how should I improve my network accuracy?
Any opinions are welcome,thank you.
User avatar
kwagyeman
Posts: 3207
Joined: Sun May 24, 2015 2:10 pm

Re: chars74k on M7

Postby kwagyeman » Mon May 06, 2019 11:53 am

Hi, I'm working on a tutorial showing how to run tensor flow models on the system. I hope to be done later this week.

As for the accuracy. Please understand the network needs the character by itself. You clearly have another character in the image. The network literally only learns what it sees in the training dataset. You cannot have an image with other noise there.
Nyamekye,
User avatar
iabdalkader
Posts: 918
Joined: Sun May 24, 2015 3:53 pm

Re: chars74k on M7

Postby iabdalkader » Mon May 06, 2019 2:33 pm

Also note that one image is not enough to test the accuracy, try a 100 or more.
ericchu923
Posts: 4
Joined: Mon May 06, 2019 2:09 am

Re: chars74k on M7

Postby ericchu923 » Tue May 07, 2019 2:34 am

kwagyeman wrote:
Mon May 06, 2019 11:53 am
Hi, I'm working on a tutorial showing how to run tensor flow models on the system. I hope to be done later this week.

As for the accuracy. Please understand the network needs the character by itself. You clearly have another character in the image. The network literally only learns what it sees in the training dataset. You cannot have an image with other noise there.
iabdalkader wrote:
Mon May 06, 2019 2:33 pm
Also note that one image is not enough to test the accuracy, try a 100 or more.
Thank you for your quick reply.
I'll try dataset's image for my network.
And I look forward to Mr.kwagyeman's TenserFlow tutorial.
If there are any problems, I will ask again.
Anyway,thanks for your help. :)

Return to “General Topics”

Who is online

Users browsing this forum: No registered users and 2 guests