ck090 / Google_Quick_Draw

A CNN implementation of Google's "Quick, Draw!"
MIT License
7 stars 3 forks source link

Memory Error #2

Open aqsasamreen opened 6 years ago

aqsasamreen commented 6 years ago

I am facing memory issue when train the system about 14 classes of images and i need to train the system for 345 classes of the image. Please suggest me something effective.

ck090 commented 6 years ago

Ohh thats a lot ..!! try reducing the batch size of the training step. But the accuracy will decrease. 345 classes of how many images in each class.

aqsasamreen commented 6 years ago

All classes of quick draw dataset.

On 2 April 2018 at 16:22, Chandra Kanth notifications@github.com wrote:

Ohh thats a lot ..!! try reducing the batch size of the training step. But the accuracy will decrease. 345 classes of how many images in each class.

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/ck090/Google_Quick_Draw/issues/2#issuecomment-377910313, or mute the thread https://github.com/notifications/unsubscribe-auth/AVyrSas395OhxvfF_NOCEh6ohGbpc73yks5tkgn8gaJpZM4TC5RQ .

ck090 commented 6 years ago

Well than it's hard to train that many using a normal CPU, you might have to use a GPU enabled system that is going to be capable of training on that many classes. Otherwise there is no other option but try something along the lines of Transfer learning!

aqsasamreen commented 6 years ago

But i need to train neural network on quickdraw dataset.

On 3 April 2018 at 10:52, Chandra Kanth notifications@github.com wrote:

Well than it's hard to train that many using a normal CPU, you might have to use a GPU enabled system that is going to be capable of training on that many classes. Otherwise there is no other option but try something along the lines of Transfer learning!

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/ck090/Google_Quick_Draw/issues/2#issuecomment-378136150, or mute the thread https://github.com/notifications/unsubscribe-auth/AVyrSfbVqWWpIanppvfX3OVsoauE5Wilks5tkw4ggaJpZM4TC5RQ .

ck090 commented 6 years ago

Sorry, did you try transfer learning!!

aqsasamreen commented 6 years ago

I have no idea about transfer learning.

On 3 April 2018 at 22:19, Chandra Kanth notifications@github.com wrote:

Sorry, did you try transfer learning!!

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/ck090/Google_Quick_Draw/issues/2#issuecomment-378328611, or mute the thread https://github.com/notifications/unsubscribe-auth/AVyrSWYvi7nRkcFZi5J4XOvL69sRyHVxks5tk68ugaJpZM4TC5RQ .

aqsasamreen commented 6 years ago

I split up the data and train more than one models.And now i want to load all models into single model for testing one new image but i am stuck and unable to solve the problem.Could you please suggest me something useful.

On 3 April 2018 at 22:21, Aqsa Samreen aqsa2014@namal.edu.pk wrote:

I have no idea about transfer learning.

On 3 April 2018 at 22:19, Chandra Kanth notifications@github.com wrote:

Sorry, did you try transfer learning!!

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/ck090/Google_Quick_Draw/issues/2#issuecomment-378328611, or mute the thread https://github.com/notifications/unsubscribe-auth/AVyrSWYvi7nRkcFZi5J4XOvL69sRyHVxks5tk68ugaJpZM4TC5RQ .

ck090 commented 6 years ago

What do you do or which college do you go to!