config_babi.m

% Copyright (c) 2015-present, Facebook, Inc.
% All rights reserved.
%
% This source code is licensed under the BSD-style license found in the
% LICENSE file in the root directory of this source tree. An additional grant 
% of patent rights can be found in the PATENTS file in the same directory.

batch_size = 32;
nhops = 3;
nepochs = 100;
lrate_decay_step = 25; % reduce learning rate by half every 25 epochs

% use 10% of training data for validation
train_range = 1:floor(0.9 * size(questions,2));
val_range = (floor(0.9 * size(questions,2))+1):size(questions,2);
enable_time = true; % add time embedings
use_bow = false; % use Bag-of-Words instead of Position-Encoding
share_type = 1; % 1: adjecent, 2: layer-wise weight tying 
randomize_time = 0.1; % amount of noise injected into time index
add_proj = false; % add linear layer between internal states
add_nonlin = false; % add non-linearity to internal states

config = {};
config.init_lrate = 0.01;
config.max_grad_norm = 40;
config.input_dim = 20;
config.out_dim = 20;

linear_start = true;
if linear_start
    ls_nepochs = 20;
    ls_lrate_decay_step = 21;
    ls_init_lrate = 0.01/2;
    config.init_lrate = 0.01/2;
end

config.sz = min(50, size(story,2));
config.voc_sz = length(dict);
config.bsz = batch_size;
config.max_words = size(story,1);
if enable_time 
   config.voc_sz = config.voc_sz + config.sz;
   config.max_words = config.max_words + 1; % +1 for time words
end