Hey guys!
There’s a Kaggle competition that has started recently on landmarks detection. It is basically a large-scale image recognition challenge. In our Fastbook reading session too, we have looked at a few image recognition models. It’s a good opportunity to explore more image recognition models. Anyone participating in it ?
Starting this thread to start discussing ideas and share your approaches. It’s also my first time in a Kaggle competition. Experienced people can also share some tips for beginners like me to Kaggle on how to go about it. Looking forward to ideas from others on it.
Link to the competition - Google Landmark Recognition 2021 | Kaggle
3 Likes
@saiamrit It’s great to see that you’ll be participating. Just wanted to caution you against sharing the approaches here for a live competition, that’s against the terms of Kaggle comps since its private sharing. If you do chat about it here, please make sure you share it in the Kaggle forums as well to avoid getting into trouble 
Good Luck!
5 Likes
Also, just mentioning this in case you didn’t already know - it’s a pretty compute heavy competition 
3 Likes
Thanks @bhutanisanyam1 I was unaware of that. Thanks for letting me know. I will definitely take care of that.
1 Like
On thanks @amanarora , I also realise that sesing the size of the dataset but it seemed like a detection challenge which is a relatively simpler problem so went ahead with it. @bhutanisanyam1 Can you guys kindly suggest how to go about dealing such large datsets. which platform is preferred by you for such compute heavy competitions(google cloud/ AWS kind of servers or any other setup) ?
1 Like
Start with whatever you have and then work from there. Please remember setting up servers/services is a full time job and quickly becomes one, try to avoid that rabbit hole and focus on applying DL.
GCP is best for folks since it offers 300$ on signup IIRC but Kaggle kernels are also a great starting point
2 Likes
Thanks for your advise @bhutanisanyam1 . I am also planning to use kaggle kernels for the same as of now but would definitely look into setting up minimally on GCP and start experimenting. 
2 Likes
GCP also provides pre-configured VMs for deep learning (example: Deep Learning VM Images | Google Cloud) with PyTorch, CUDA, Jupyter lab, etc. pre-installed. I’ve had a great experience with this “single-click setup” so far.
The $300 signup credit would last for about a month with a T4 GPU (longer if you remember to shut down the VM when you’re not using it 
)
2 Likes
I’m sorry but this made me chuckle 
Looks like @bhutanisanyam1 having experience here 
1 Like