willzhang6 / code-samples

Code samples from various machine learning and data science projects and competitions

Code samples

These are some code samples from various research projects, side projects and competitions.

Deep Learning

Variational Sparse Coding

Implementation of the "Variational Sparse Coding" paper as part of the ICLR 2019 Reproducibility Challenge. [code]

Neural Machine Translation

Sequence-to-sequence recurrent neural network (bidirectional LSTM) with Global Attention (Luong et al., 2015) and Beam Search implemented in PyTorch. ~41 BLEU in 110K-sentences English-Spanish corpus.

Pokédex: Transfer Learning for Image Classification

Pokémon image classification with transfer learning from ImageNet-pretrained MobileNet convolutional neural network (Howard et al., 2017). ~82% accuracy with 27 classes and 3.8K web-scraped images. Deployed in Flask with React JS for the SPA user interface. Presented at Infosoft 2017 and Hack Faire 2017. [slides] [demo]

Job2Vec: Job matching from word-embeddings

Information retrieval system between job descriptions and applicant profiles textual description matching based on Word2Vec and Doc2Vec (Le & Mikolov, 2014) semantic search and string matching algorithms for out-of-vocabulary misspelled words, constructed over an inverted index for efficient look-up. Presented at WAIMLAp 2017 and Hack Faire 2017. [poster]

Large-scale Text Classification

Commodity description classification using recurrent neural networks (bidirectional LSTM) implemented in PyTorch with FastText pretrained word embeddings (Joulin et al., 2016). ~92% top-5 accuracy with 3762 classes and 30.6M text descriptions.

Genomics time-series pattern-recognition with Convolutional Neural Networks

Convolutional neural networks architecture experimentation for genomic sequence pair binary classification with high imbalance (0.07% positive classes). ~78.5 F1-Score for ~200k pairs of sequences.

Fully-connected Autoencoder for MNIST

Fully-connected autoencoder for MNIST dataset with a bottleneck of size 20 implemented in PyTorch, based on DeepBayes 2018 practical assignment. 0.00069 L2 reconstruction loss + L1 regularization loss. t-SNE dimensionality reduction for bottleneck features visualization.

Data Science Competitions

Movistar Data Talent 2018

Main preprocessing and main cross validation loop with LightGBM

Data Science Game Finals 2018

LSTM conditioned on a structured embedding network implemented in PyTorch (refactored version)

Data Science Game Qualifiers 2018

LSTM conditioned on a structured embedding network implemented in PyTorch

Other competitions

• BBVA Challenge 2017
• DrivenData Competition
• Interbank Datathon
• Kaggle Bulldozers Competition
• Kaggle Homesite Insurance Competition
• Kaggle NYC Taxi Trip Duration Competition

Coursework

• CIFAR-10 with CNNs (Tensorflow)
• CelebA Image Generation with DCGANs (Tensorflow)
• Dogs vs Cats with CNNs + Transfer Learning (Keras)
• English-French Machine Translation with Seq2seq RNNs (Tensorflow)
• The Simpons TV Scripts Text Generation with LSTM RNNs (Tensorflow)
• LibriSpeech Speech Recognition with GRU RNNs and CNNs (Keras)

Miscellaneous

• Music automatic matching with audio descriptors
• Nearest Neighbor Collaborative Filtering
• Geospatial Visualization