Skip to content

plot 2d T-SNE representations in space using bokeh

Notifications You must be signed in to change notification settings

bill-kalog/bokeh_plots

Repository files navigation

bokeh plot of a fully connected layer

toy example of plotting a fully connected layer as presented in the blog post Learning when to skim and when to read

Requirements

  • Python 3
  • sklearn
  • numpy
  • pandas
  • bokeh

instructions

the script innards.py expects a pandas dataframe similar to the one found in metrics.pkl. which could have been created from something like:

def save_info(x_dev, y_dev, y_net, prob_net, layer, path_):
    '''
    save test set info into a pandas dataframe and pickle it
    x_dev: sentences (list of strings)
    y_dev: sentiment label (list of ints) ex. 0 negative, 1 positive
    y_net: network output (list of ints) ex. 0 negative, 1 positive
    prob_net: networks probabilities/y
    layer: fully connected layer list of a vector list per sentence
    path_: path to save dataframe
    '''
    d = {'x_dev': x_dev, 'y_dev': y_dev, 'layer': layer,
         'y_net': y_net, 'prob_net': prob_net}
    df = pd.DataFrame(data=d)
    df.to_pickle(path_)

you can also check the jupyter notebook where a version of an LSTM fully connected layer from SST is plotted. So far the two plots of the paragraph Exploring the innards are only implemented

Usising the script innards_finegrained.py one can plot in space more than two classes. An example dataset exists at metrics_f.pkl and the program's output at plot_finegrained.html.

About

plot 2d T-SNE representations in space using bokeh

Topics

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published