Skip to content

Latest commit

 

History

History
135 lines (128 loc) · 4.41 KB

config.md

File metadata and controls

135 lines (128 loc) · 4.41 KB
Raw

Config

The config.ini files are parsed by the Python ConfigParser and so follow the .ini file format.

Model section

[model]
g_w = 8
k = 3
s = 2
hg_size = 128
hl_size = 128
g_size = 256
hidden_size = 256
glimpses = 6
num_classes = 10
loc_std = 0.1
  • g_w : int
    • size of glimpse window in pixels, default is 8 x 8
  • k : int
    • number of regions
  • s : int
    • scaling factor
  • hg_size : int
    • number of units in hidden layer that takes glimpse as input
  • hl_size : int
    • number of units in hidden layer that takes glimpse location as input
  • g_size : int
    • number of units in hidden layer that combines glimpse + location hidden layers
  • hidden_size : int
    • number of units in core network
  • glimpses : int
    • number of glimpses to take of each image
  • num_classes : int
    • number of classes in data
  • loc_std : float
    • standard deviation of distribution from which each glimpse location is drawn

Data section

[data]
root_results_dir = /home/you/Documents/data/RAM_output
data_dir = /home/you/Documents/data/mnist/raw
module = ram.dataset.mnist
train_size = 0.9
val_size = 0.1
  • root_results_dir : str
    • name of directory in which a new sub-directory will be created to hold results from run
  • data_dir : str
    • name of directory where data is saved
  • module : str
    • name of module that prepares and returns data sets, in absolute terms, i.e. package.module, package.subpackage.module, etc. Python will use importlib to import this module. This is a quick and dirty way to make it possible to use your own data with the library. The module must implement two functions, prep and get_split. The function should accept arguments and return variables like so:
    paths_dict = dataset_module.prep(download_dir=config.data.data_dir,
                                 train_size=config.data.train_size,
                                 val_size=config.data.val_size,
                                 output_dir=config.data.data_dir)
train_data, val_data = dataset_module.get_split(paths_dict, setname=['train', 'val'])
    For more detail and examples, see the ram.__main__ function and the ram.dataset.mnist module.
  • val_size : float
    • Amount of training data to use for a validation set
  • train_size : float
    • Amount of entire training data set to use as a subset for training

Train section

[train]
batch_size = 200
learning_rate = 1e-3
epochs = 200
optimizer = momentum
momentum = 0.9
checkpoint_prefix = ckpt
restore = False
save_log = Truesave_examples_every = 25
num_examples_to_save = 9
save_loss = True
save_train_inds = True
  • batch_size : int
    • number of samples in each batch fed to the network
  • learning_rate : float
    • learning rate used by optimizer during training
  • epochs : int
    • number of training epochs, i.e. number of times network sees entire training set
  • optimizer : str
    • one of {'adam', 'momentum', 'sgd'}; type of optimizer used
  • momentum: float
    • if optimizer is 'momentum', the value for 'momentum'
  • checkpoint_prefix : str
    • prefix for checkpoint files
  • restore : bool
    • if True, restore from previous checkpoint. Not used by command-line interface but useful if you want to train your own models with the Trainer class.
  • save_examples_every : int
    • if defined, save examples at an interval of this many epochs. Examples are glimpses seen by network, locations of gazes, fixations (locations converted from normalized to pixel co-ordinates), and images (from which glimpses were taken)
  • 'num_examples_to_save' : int
    • number of examples to save, if save_examples is True
  • 'save_loss' : bool if True, save records of loss from every epoch.
  • 'save_train_inds' : bool if True, save the indices of the training samples used for each epoch.

Test section

[test]
save_examples = True
num_examples_to_save = 15
  • save_examples : bool
    • if True, save examples. Examples are glimpses seen by network, locations of gazes, fixations (locations converted from normalized to pixel co-ordinates), and images (from which glimpses were taken)
  • 'num_examples_to_save' : int
    • number of examples to save, if save_examples is True

Misc section

[misc]
save_log = True
  • save_log : bool
    • if True, save logging to a text file.
  • random_seed : int
    • number used to seed random number generator, to make results reproducible across runs