https://github.com/dudeperf3ct/fellowship.ai-challenges
This repo contains implementations of the challenges from fellowship.ai. For more, visit here.
https://github.com/dudeperf3ct/fellowship.ai-challenges
few-shot-learning food-101 maml transfer-learning transfer-learning-nlp twitter-sentiment-analysis
Last synced: about 1 month ago
JSON representation
This repo contains implementations of the challenges from fellowship.ai. For more, visit here.
- Host: GitHub
- URL: https://github.com/dudeperf3ct/fellowship.ai-challenges
- Owner: dudeperf3ct
- Created: 2019-02-23T22:45:35.000Z (about 6 years ago)
- Default Branch: master
- Last Pushed: 2019-07-11T05:17:07.000Z (almost 6 years ago)
- Last Synced: 2025-03-28T15:21:28.919Z (about 2 months ago)
- Topics: few-shot-learning, food-101, maml, transfer-learning, transfer-learning-nlp, twitter-sentiment-analysis
- Language: Jupyter Notebook
- Homepage: https://fellowship.ai/challenge
- Size: 9.97 MB
- Stars: 6
- Watchers: 2
- Forks: 5
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
Awesome Lists containing this project
README
# Select Your Challenge Problem
## One-shot Learning
[Omniglot](https://github.com/brendenlake/omniglot), the “transpose” of MNIST, with 1623 character classes, each with 20 examples.
Use background set of [30 alphabets](https://github.com/brendenlake/omniglot/blob/master/python/images_background.zip) for training and evaluate on set of [20 alphabets](https://github.com/brendenlake/omniglot/blob/master/python/images_evaluation.zip). Refer to this [script](https://github.com/brendenlake/omniglot/blob/master/python/one-shot-classification/demo_classification.py) for sampling setup.
Report one-shot classification (20-way) results using a meta learning approach like [MAML](https://arxiv.org/pdf/1703.03400.pdf).
## Image Segmentation
Apply an automatic portrait segmentation model (aka image matting) to [celebrity face](http://mmlab.ie.cuhk.edu.hk/projects/CelebA.html) dataset.
## Food-101
[Food-101](https://www.vision.ee.ethz.ch/datasets_extra/food-101/) is a challenging vision problem, but everyone can relate to it. Recent SoTA is ~80% top-1, 90% top-5. These approaches rely on lots of TTA, large networks and even novel architectures.
Train a decent model >85% accuracy for top-1 for the test set, using a ResNet50 or smaller network with a reasonable set of augmentations.
## ULMFiT Sentiment
Apply a supervised or semi-supervised [ULMFiT](http://nlp.fast.ai/classification/2018/05/15/introducting-ulmfit.html) model to [Twitter US Airlines Sentiment](https://www.kaggle.com/crowdflower/twitter-airline-sentiment#Tweets.csv).