Results from 18-hours of training a machine learning model | Airbnb Amenity Detection 7

I trained a machine learning model for 18-hours on ~35,000 images on a single GPU (P100) using transfer learning. These are the results.
Although we didn't reach Airbnb's requirement for mean average precision (metric for measuring object detection models), the project still goes on.
Next will be using the full model to build a small application around so others can use it and get it deployed.
Stay tuned for more.
All videos in this series:
Full playlist - dbourke.link/a...
Part 1 (overview) - • Using Machine Learning...
Part 2 (data exploration) - • Using Machine Learning...
Part 3 (data manipulation) - • Replicating Airbnb's M...
Part 4 (building a small dog model) - • I got Detectron2 worki...
Part 5 (experiment tracking) - • Every model I build ha...
Part 6 (modelling experimentation) - • 34,835 training images...
Part 7 (building a big dog model) - • Results from 18-hours ...
Part 8 (getting my modelling deployed) - • I got my machine learn...
Part 9 (project wrap up and retrospective) - • What could be improved...
Note: I am not affiliated with Airbnb nor are they involved with this series of videos in any way.
Links:
My machine learning course - dbourke.link/m...
All of my notes/ideas in Notion - dbourke.link/a...
Airbnb Amenity Detection article - / 05qperz3a4
Get email updates on my work - dbourke.link/n...
Support on Patreon - bit.ly/mrdbour...
Connect elsewhere:
Web - dbourke.link/web
Quora - dbourke.link/q...
Medium - dbourke.link/m...
Twitter - dbourke.link/t...
LinkedIn - dbourke.link/l...
#machinelearning #datascience #deeplearning

[@bhaveshwadhwani8841]

Nice work Buddy!!!

[@mrdbourke]

Thank you Bhavesh!

[@weeman297]

As a non-coder with slight coding concept knowledge, I wish I understood what you just said. I feel like you're making something great, but just don't know what. Liked the video though and good luck

[@mrdbourke]

Hey Chris, thank you for the kind words. The project is nothing over the top in greatness just a fun little side project to practice a few things/demo to my students how to work on their own projects.

[@mehdizobiri9267]

Good Job Daniel.

[@mrdbourke]

Thank you Mehdi!

[@puneetsingh5219]

15:14 Hand engineering hyper-params (and broke) gang rise

[@puneetsingh5219]

13:51 So here the whole architecture's weights are being updated every epoch. 13:57 This is fine-tuning, right?

[@rodrigocardoso7846]

Fine tuning is when you unfreeze the high level features from CNN

[@elakiouizouhaire9432]

Great work :-)

[@mrdbourke]

Thank you Zouhaire

[@elakiouizouhaire9432]

@@mrdbourke This kind of topics I like the most. For this purpose I made a youtube channel and created a blog website.

[@rahulpandit4565]

you can use YOLO-v3 for Better Detection

[@_ma7dev]

YOLOv3 is great for real-time object detection and you can find a lot of source code available online, but YOLOv3 doesn't give higher accuracy than other newer object detection models (e.g. Detectron, EfficientNet, etc.). In addition, YOLOv3 fails miserably to detect small objects or objects that have more complex features.

[@adeolaogunleye7965]

What about Faster RCNN NAS model from google object detection repo? I haven't tried Detectron before i git cloned the repo to check it out. Well good job.

[@mrdbourke]

Faster RCNN was one of the one’s I tried within Detectron2 but it didn’t perform as well as RetinaNet. And thank you.

[@adeolaogunleye7965]

@@mrdbourke How do you prevent people from upload a wrong image? I have experimented with Autoencoders as an anomaly detector. One should be able to train using all the bedroom images and MSE between predicted original images. Can one use Autoencoder to weed out the wrong images?

[@mj7146]

Hey Daniel doing ur Machine learning course, its great 😃. Whats ur opinion on the dataquest.io courses for data science/analysis/engineer?

[@mrdbourke]

Hey Mert, glad to hear you're enjoying it. Dataquest looks amazing, I've never personally used it but their articles are top knotch.

[@rodrigocardoso7846]

Only 18 hrs? Try use hyperopt to find the best hyperparameters

[@puneetsingh5219]

8:49 That's what stratified split is, right?

[@Vasko10100]

What are you using for deployment? And do you recommend AWS or Docker?

[@mrdbourke]

I’m using GCP’s App Engine + Docker, though AWS could be another option.

[@blacklabelmansociety]

Hey, Dan! Which software/app/website are you using for task management?

[@mrdbourke]

Pen and paper + Notion. Check my notes on the project here: dbourke.link/airbnb42days

[@yousefal-nazer3259]

How can I become machine learning Engineer and i can learn this data and solve problems then take project to work on it, what is the steps? And how many months it will take

[@mrdbourke]

One way is to check out my ML course: dbourke.link/mlcourse or check out my posts on machine learning at: www.mrdbourke.com There’s no way to measure how long it will take. It will take as long as you let it. The best way to get started is to start.

[@rahulpandit4565]

Bro which cuda version are you using because i have cuda 10.1 in colab and not work proper.it give me some time image-frame error please tell me .my project is steel defect detection and segmentation usig detectron-2 pytorch github.com/panditrahulsharma/severstal-steel-defect-detection-using-Detectron2-Pytorch

[@mrdbourke]

I used CUDA 10.0

[@OptimizeMyAirbnb]

This is fascinating!

[@gyimahfrancis60]

great

[@marcoasthetic06]

Did you use tensor flow or pytorch