https://github.com/hycis/transfer_learning

Mozi, Transfer Learning, Multi-Modal Learning, Theano
https://github.com/hycis/transfer_learning

Last synced: 4 days ago
JSON representation

Mozi, Transfer Learning, Multi-Modal Learning, Theano

• Host: GitHub
• URL: https://github.com/hycis/transfer_learning
• Owner: hycis
• Created: 2015-09-17T16:09:42.000Z (over 9 years ago)
• Default Branch: master
• Last Pushed: 2016-01-26T09:39:21.000Z (almost 9 years ago)
• Last Synced: 2024-12-27T14:39:27.778Z (20 days ago)
• Language: Python
• Homepage:
• Size: 998 KB
• Stars: 27
• Watchers: 7
• Forks: 11
• Open Issues: 1
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

README

        


# Transfer Learning with Theano

This package demonstrates how to build a transfer learning network effortlessly with [Theano](https://github.com/Theano/Theano.git) and [Mozi](https://github.com/hycis/Mozi).





Standard `transfer learning` also known as `multi-task learning` or `multi-modal learning` typically has multiple inputs and multiple outputs. And generally there are two types, `Type 1` (Fig 1) has a common shared representation layer where the objective is to map different modalities into a common representation space, while `Type 2` (Fig 2) tries to keep individual modality representation space separate, and finally concatenate them to pass through subsequent layers for fine-tuning.

The way to train `Type1` and `Type2` is also different. For `Type 1`, you first train `1 -> 3` (Fig 1a) then `2 -> 3` (Fig 1b), while for `Type 2`, you concatenate the outputs from `step 1` and `step 2` (Fig 2) before proceeding to `step 3` with the concatenated features, i.e  `1 + 2 -> 3`.

Checkout [Type 1 Model](type1_train.py) and [Type 2 Model](type2_train.py)