Code from some models used in First Place's solution of Kaggle Data Science Challenge for EEF for ITA/Unifesp SJC students.
Kaggle Competition Link: https://www.kaggle.com/c/data-science-challenge-at-eef-2019/leaderboard
The code here is related to models I developed (before and after join "Los Bertones" team). Here you will encounter the following models:
The main repository of Los Bertones' team is: https://github.com/brunoklaus/EEF_2019
There is two main models: One is a stacking model using a diversity of 18 models. The other is a single network that uses all kinds of data to train a predictor.
- baseline.py -- Baseline with just tabular data final_model.py
- image_house_pred.py - Predictor using features extracted from images using NASNetMobile
- embedding_nlp_glove_300.py - GloVe Word Embeddings using description text data (length of 300)
- embedding_nlp_glove.py - Same as above, length = 500
- embeddings_nlp_fast_text.py - FastText Word Embeddings using description text data (length of 300)
- lstm_house_pred.py - Trainable embeddings, length of 500
- embeddings_source_300.py - Trainable embeddings, length of 300
- embeddings_source_attention_capsule.py - Trainable embeddings + Attention + Capsule Layer
- embeddings_source_attention_cnn.py - Trainable embeddings + Attention + CNN Layer
- tfidvectorizer_nlp.py - Using TF IDF features from description in a MLP network
- wang2vec_attention_capsule.py - Wang2Vec Word Embeddings + Attention + Capsule
- final_model.py -- Stacking model using 18 models previously trained
- final_model_merged.py -- Another model that shares embedding layer between all text data, uses image features and tabular data in a single network.
OBS: The other models with same nomenclature + title refers to the same network used by title text.
