Permalink
Switch branches/tags
Nothing to show
Commits on Jan 31, 2017
  1. Update README.md

    jakobgerstenlauer
    jakobgerstenlauer committed Jan 31, 2017
    remove some details
  2. Update

    jakobgerstenlauer
    jakobgerstenlauer committed Jan 31, 2017
    improve structure
  3. Update README.md

    jakobgerstenlauer
    jakobgerstenlauer committed Jan 31, 2017
    more complete background information
Commits on Jan 30, 2017
  1. print only values for the validation set

    jakobgerstenlauer
    jakobgerstenlauer committed Jan 30, 2017
    print only values of the validation set
  2. reverse previous commit

    jakobgerstenlauer
    jakobgerstenlauer committed Jan 30, 2017
  3. Include shuffle=True argument in model fitting

    jakobgerstenlauer
    jakobgerstenlauer committed Jan 30, 2017
    I again included the shuffle=True argument because else we can not see the accuracy for the validation set.
  4. Major changes of preprocessing and models

    jakobgerstenlauer
    jakobgerstenlauer committed Jan 30, 2017
    Additional output for -h flag.
    New comments throughout the code.
    Several changes in text preprocessing:
    1. Remove additional interpunctuation marks.
    2. Remove all tokens which are not true words but contain numbers using alpha().
    3. Increase the maximum sequence length to 500.
    4. Save the dictionary as a text file and as a binary file.
    5. Add more comments.
    
    Specification of the PLSTM and the LSTM model: 
    1. Reduce the number of neurons in the first hidden layer from 128 to 64.
    2. Create a separate dropout level and remove the dropout argument from all other layers.
    3. Remove the shuffle=TRUE argument because we want to use the same validation data 
    during all epochs and reduce the % of the data used for validation from 25% to 20%.
Commits on Jan 29, 2017
  1. removed print statements, added two new plots

    jakobgerstenlauer
    jakobgerstenlauer committed Jan 29, 2017
Commits on Jan 28, 2017
  1. Update plstm_validation.py

    jakobgerstenlauer
    jakobgerstenlauer committed Jan 28, 2017
    Added an additional flag t for dry run
    Now the script stores the results (accuracy and loss) of both models to a csv file
Commits on Jan 27, 2017
  1. Made python file a shell executable with two flags for epochs and dro…

    Jakob Gerstenlauer Jakob Gerstenlauer
    Jakob Gerstenlauer authored and Jakob Gerstenlauer committed Jan 27, 2017
    …p out.
    
    Modified the plots.
Commits on Jan 26, 2017
  1. Script runs 40 epochs and estimates loss and accuracy in the validati…

    Jakob Gerstenlauer Jakob Gerstenlauer
    Jakob Gerstenlauer authored and Jakob Gerstenlauer committed Jan 26, 2017
    …on data set, which is randomly chosen in each epoch and consists of 25% of the data.
    
    This script builds on Ahmads script plstm.py
  2. Work only with 1/3 of training data and use only 50% for training set.

    Jakob Gerstenlauer Jakob Gerstenlauer
    Jakob Gerstenlauer authored and Jakob Gerstenlauer committed Jan 26, 2017
    In order to speeep up model fitting and make it possible to decide how many epochs we need I included the following changes:
    1. Only work with the first  156060/3= 52020 lines of the training data set.
    2. In each epoch, the model will randomly split the data into a training and a validation set.
    Advantages: The model is much faster and we can check in the log statements when the
    accuracy in the validation set stabilizes. Then we can stop the simulation
    because the model overfits!
Commits on Jan 25, 2017
  1. Update train_model.py

    sabirdvd committed Jan 25, 2017
  2. Update README.md

    sabirdvd committed Jan 25, 2017
  3. Update

    sabirdvd committed Jan 25, 2017
  4. Update README.md

    sabirdvd committed Jan 25, 2017
  5. Update README.md

    sabirdvd committed Jan 25, 2017
  6. Update README.md

    sabirdvd committed Jan 25, 2017
  7. Update README.md

    sabirdvd committed Jan 25, 2017
  8. Add files via upload

    sabirdvd committed Jan 25, 2017
  9. Initial commit

    sabirdvd committed Jan 25, 2017