Kaggle IEEE Challenge: Need help (Practicing Lesson 2 v2)

Part 1 (2018)
1

Hi,
Yesterday, In an effort to master the concepts taught in lesson 2. I entered IEEE challenge on Kaggle. Competition is about identifying the camera from which photos were taken.
I have decided to use the pretrained resnet50 model from the fastai library.
I just followed the concepts/steps covered in lesson 2 (part1 v2). To my surprise, I was able to achieve 94% accuracy on a validation set. I, immediately, calculated the prediction on test data and submitted it to Kaggle, my accuracy on test data was only 21% :frowning:
My notebook is here.
Can anyone guide me, how can I get good accuracy on data set?

My guesses about why I am not getting good accuracy on test data set.

  • Different formates of train(JPEG) and test(tiff) data,
  • Not Suitable model: As resnet is used for object detection in image, it might not be good for detection of camera with which photos were taken

Can anyone guide me how can I move on with this challenge?

If any want to join my team please reply

2

The test set has some specific augmentations in it. It also has photos from different cameras. You would need to try to replicate that in your validation set.

2 Likes
3

@irshaduetian Can you share your code?..I am not able to see the code on git…I am also working on the same competition using the code that we learnt in lesson2…but I am finding it slightly difficult to work with the data…

1 Like
4

Hello, the link is present in the above post. But pasting here again https://github.com/irshadqemu/fast.ai_DeepLearning-1-MOOC/blob/master/lesson1-rxt50_iee.ipynb,
it is kind of rough but the code until the “Test Prediction” belong to the IEEE competition.
Code from In[77] to the end of the notebook don’t belong to this competition, need to remove that.

1 Like
5

Thanks…I could see the code…It has helped a lot in pre-processing the data for this competition…are you still working on this ?

6

No, I am not working on it anymore due to time constraints. If you could obtain some good results, do share them with me :slight_smile: Best of luck man.