Telauges Alpha, finally gonna merge to Telauges
Switch branches/tags
Nothing to show
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Failed to load latest commit information.
exp_configs
scripts
telaugesa
.travis.yml
LICENSE
README.md
setup.py

README.md

Telauges Alpha

Currently, Max-pooling has some issues with cuDNN support, please use basic useage only

I finally decided to rewrite my coding style for neural networks. This code will be eventually merged to Telauges.

This code is heavily referred to three deep learning library

Requirements

General requirements can be achieved by Anaconda. More specifically, you need:

  • python 2.7
  • numpy
  • scipy
  • theano

Supported Features

  • Feedfward Layer

    • Identity Layer
    • Tanh Layer
    • Sigmoid Layer
    • ReLU Layer
    • Softmax Layer (classification layer)
    • Highway Identity Layer
    • Highway Tanh Layer
    • Highway Sigmoid Layer
    • Highway ReLU layer
  • ConvNet Layer

    • Identity Conv Layer
    • Tanh Conv Layer
    • Sigmoid Conv Layer
    • ReLU Conv Layer
    • Max-pooling Layer
    • Same size Max-pooling Layer
    • Flatten Layer
  • Recurrent Neural Network

    • Simple Recurrent Net [TODO]
    • Long-Short Term Memory [TODO]
    • Gated Recurrent Net [TODO]
  • Training Algorithm

    • Stochastic gradient descent (SGD) (Momentum and Nestrov Momentum)
    • Adagrad
    • Adadelta
    • RMSprop (Sugeest config: learning_rate=0.001, eps=1e-6, rho=0.9)
    • Adam
    • Dropout
  • Model

    • Feedforward Model
    • Auto-encoder Model
    • Convolutional Auto-encoder Model [TODO]
    • Mixed Layer Model [TODO]

Notes

NeuralTalk

I'm trying to reproduce the result of Deep Visual-Semantic Alignments for Generating Image Descriptions.

Here is some of my notes while playing with the dataset. There are descriptions of three datasets are available:

  • Flickr8k
  • Flickr30k
  • Microsoft COCO

The description file dataset.json is organized in this way (just a raw description). For each image

 -- image info
    |
    -- filename (string)
    -- imgid (int)
    -- sentences (dict)
       |
       -- tokens (dict, in order)
       -- raw (raw sentence)
       -- imgid (int, image that associates with the description)
       -- sentid (int, label of sentence)
       ...
    -- split (string, "train", "val", or "test")
    -- sentids (dict)

Contacts

Yuhuang Hu
No. 42, North, Flatland
Email: duguyue100@gmail.com