[mickmek] deepcognition.ai soda bottle competition (18h)

[shadiakiki1986]

@MickMek here's your issue to coordinate about deepcognition.ai model

• [x] warm up (3h)
  • try out deepcognition.ai
• [x] first model version including pre-trained VGG (5h)
  • achieved 18% accuracy on validation set
• [x] model version 2 (5h)
  • move cropping earlier
  • reduce the image sizes to have faster computation
  • [x] submission score: 34%
  • epochs = 10, batch size = 32, 1 hour of training time, train/validation/test: 80%/10%/10%
  • [x] export keras code from deepcognition and upload on server
• [x] model version 3 (5h)
  • thoughts to improve the model:
    • simplify the model
    • train on full dataset for the submission
    • either fix or remove cropping (current setup just crops boundary 1 pixel)
    • circular convolution matrix
  • [x] submission score: 65% (rank: 36 as of 2018-07-09)

[MickMek]

Update: Completed model2 with around 30% testing accuracy (cropping moved and some modifications to the CNN topology). Submitted the second version of the model to the deepcognition site.

[MickMek]

Update: Model version 3 achieved 99% training accuracy and 82% validation accuracy (training for 10 epochs). Next step is to reduce overfitting with batch_norms and dropouts, before testing "deeper" networks. Submitted the third version of the model to the deep cognition site (still no results for the 3 submitted models)

[shadiakiki1986]

I think that since we did not receive any submission result by email, send them a message via their contact button asking them why you're not getting any results by email.

On Wed, Jul 4, 2018, 12:49 MickMek notifications@github.com wrote:

Update: Model version 3 achieved 99% training accuracy and 82% validation accuracy (training for 10 epochs). Next step is to reduce overfitting with batch_norms and dropouts, before testing "deeper" networks. Submitted the third version of the model to the deep cognition site (still no results for the 3 submitted models)

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/teamshadi/ssoc18/issues/6#issuecomment-402427428, or mute the thread https://github.com/notifications/unsubscribe-auth/AIAOhN2iW6qQ9U_mj5QBMTtdz38rwxDvks5uDI-YgaJpZM4U_W8S .

[shadiakiki1986]

@MickMek updated issue as discussed over the phone

[MickMek]

Update: Improved model (optimized parameters and topology) to attain 97% validation accuracy, and scores 62.5% on deep cognition. The code is now on the server

[shadiakiki1986]

Wow! I see you're number 36 on the leader board. Well done

On Mon, Jul 9, 2018, 12:41 MickMek notifications@github.com wrote:

Update: Improved model (optimized parameters and topology) to attain 97% validation accuracy, and scores 62.5% on deep cognition. The code is now on the server

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/teamshadi/ssoc18/issues/6#issuecomment-403421674, or mute the thread https://github.com/notifications/unsubscribe-auth/AIAOhEEVT5ERb6eXY5yBsiiFIC0cbW-jks5uEyVkgaJpZM4U_W8S .