This repo can be cloned as a submodule into any project in order to provide:

- helper functions
- grapher module
- metrics
- layers
- fid

These functions include things like:

- ones_like / zeros_like functions
- directory creation / csv appender
- expand / squeeze dims
- zero padding
- one_hot generataion
- normalization
- int_type / long_type / float_type for cuda vs. regular tensors

Include functions like:

- compute softmax / bce accuracies
- frechet distance calculations
- compute EWC
- compute FID

Include layers such as :

- dense with many sequential layers & bn
- conv stack
- dense / conv encoder + decoder stack
- bw2rgb module
- Identity, View layers, and more!

To compute FID use train_fid_model from fid.py. You can use a simple conv model or inceptionv3. The FID batch size can smaller than or equal to the model you train. For inceptionv3 you need small batch sizes unless you have a badass P100 or something.

After this you can use calculate_fid from metrics.

Currently there exists only a grapher, the visdom grapher. This helper utilizes a matplotlib style API for sending data to visdom.

from helpers.grapher import Grapher

grapher = Grapher("unique_experiment_uuid", "http://localhost", port=8097)
grapher.register_single(  # format is [xarr, yarr]
    {'loss': [[epoch], [loss.data[0]]]},
    plot_type='line'
)
grapher.show()