This code provides an implementation of the research paper:

Learning the Latent "Look": Unsupervised Discovery of a Style-Coherent Embedding from Fashion Images
Wei-Lin Hsiao, Kristen Grauman
International Conference on Computer Vision (ICCV), 2017

See our project page for more detail.

This project is based on MALLET polylingual LDA implementation, with a few minor changes:
We smoothed the topic/word counts to topic/word probabilities.
For stability, after convergence, we run 100 more samplings to average the learnt topic/word probabilities.

To build a Mallet 2.0 development release, you must have the Apache ant build tool installed. From the command prompt, first change to the mallet directory, and then type ant

If ant finishes with "BUILD SUCCESSFUL", Mallet is now ready to use.

1. Prepare your documents. Every document is a single line in the file, every line is

docID language_name word_1 word_2 ...

The documents for each language will be in its own file. The Nth document in language 1 is assumed to have the same topic distribution (though a different vocabulary) as the Nth document in language 2, and so on. If a language lacks Nth document, just leave it blank as

docID language_name 

An example corpus can be downloaded here.

2. Import documents for each language. The token-regex is the appropriate one to use with our attribute vocabulary.

bin/mallet import-file --input <document_filename> --output <sequence_filename> --keep-sequence --token-regex '[\p{L}\p{N}_<>/-]+|[\p{P}]+'

3. Train a model.

bin/mallet run cc.mallet.topics.PolylingualTopicModel --language-inputs <language1_sequence> <language2_sequence> ... <languageN_sequence> --alpha <alpha_val> --beta <beta_val> --num-topics <number_of_topics> --output-average-doc-topics <theta_save_to_filename> --output-average-topic-keys <topic_top_words_save_to_filename> --output-average-key-probs <phi_save_to_filename>

For details about the commands please visit the API documentation and website at: http://mallet.cs.umass.edu/