[FatimaFawaz] deepcognition.ai soda bottle competition (10h)

[shadiakiki1986]

Hi @FatimaFawaz

We already started building a model in deepcognition.ai competition (#2). The model is the same one that was demo'd for their "toilets" dataset. The most recent training achieved 14% accuracy on a validation dataset, which is not much better than chance (since the target is a classification with 8 options, random selection has a 12.5% accuracy). Please help us improve our model with the following steps:

• [x] warm up (~5h)
  • [x] play around with deepcognition.ai
  • [x] replicate the tutorial MNIST Handwritten Digit Classification (using DNN)
    • achieved 75% accuracy with 10 epochs
• [x] model version 1 (~5h)
  • [x] run the same model for MNIST above but on their "toilet dataset" (available under "public datasets")
    • 2018-07-02 achieved 50% validation accuracy with only 2 epochs so far due to memory / ETA problems
    • 2018-07-03 achieved 58% validation accuracy
    • will try to get better results
  • [x] run the same model on the soda bottle competition dataset
    • 2018-07-03 got 71% validation accuracy
  • [x] make a submission
    • got score = 25%
• [ ] ~model version 2: add a pre-trained layer such as VGG~ (0h)
  • this is to benefit from transfer learning (wikipedia)
  • example deepcognition.ai usage of VGG can be found here
  • Accuracy here should be better than random chance (12.5%) and better than version 0 of our model (14%)
  • More references:
    • machinelearningmastery.com
    • learnopencv.com: Transfer Learning using pre-trained models in Keras
    • learnopencv.com: Keras Tutorial : Fine-tuning using pre-trained models
  • validation accuracy = __
  • [ ] make a submission

[shadiakiki1986]

Hi @FatimaFawaz Can you please write here a summary of the steps that you've done on the deepcognition.ai soda bottle competition so far? Also, please export your models as python keras code and upload them into a folder on the ssoc18.teamshadi.net server so we can commit it to the repository

[FatimaFawaz]

So far:

• I've made the same model on the Soda Bottle Competition dataset
• Made a submission
• Added a pre-trained VGG layer
• Got a higher training accuracy but no validation accuracy (trained on 10 epochs)
• Did not submit it

[shadiakiki1986]

let's try to

• calculate a validation accuracy on your new model (with VGG)
• make a submission to compare it to the previous 25%

[FatimaFawaz]

I sent a message regarding this on Monday, achieved 32% accuracy with VGG.

[shadiakiki1986]

Yes I closed the issue because the competition itself is over too