#!/usr/bin/env python3
# -*- coding: utf-8 -*-
#
# MIT License
#
# Copyright (c) 2018 Kim Seonghyeon
#
# Permission is hereby granted, free of charge, to any person obtaining a copy
# of this software and associated documentation files (the "Software"), to deal
# in the Software without restriction, including without limitation the rights
# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
# copies of the Software, and to permit persons to whom the Software is
# furnished to do so, subject to the following conditions:
#
# The above copyright notice and this permission notice shall be included in all
# copies or substantial portions of the Software.
#
# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
# SOFTWARE.
#
# ------------------------------------------------------------------------------
#
# Copyright (C) IBM Corporation 2018
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""
s_control_unit.py:
- Implementation of the :py:class:`ControlUnit` for the ``S-MAC`` network (simplified MAC).
- Cf https://arxiv.org/abs/1803.03067 for the reference MAC paper (Hudson and Manning, ICLR 2018).
"""
__author__ = "Vincent Marois & T.S. Jayram"
import torch
from torch.nn import Module
from miprometheus.models.mac.utils_mac import linear
[docs]class ControlUnit(Module):
"""
Implementation of the :py:class:`ControlUnit` for the ``S-MAC`` model.
.. note::
This implementation is part of a simplified version of the MAC network, where modifications regarding \
the different units have been done to reduce the number of linear layers (and thus number of parameters).
This is part of a submission to the ViGIL workshop for NIPS 2018. Feel free to use this model and refer to it \
with the following BibTex:
::
@article{marois2018transfer,
title={On transfer learning using a MAC model variant},
author={Marois, Vincent and Jayram, TS and Albouy, Vincent and Kornuta, Tomasz and Bouhadjar, Younes and Ozcan, Ahmet S},
journal={arXiv preprint arXiv:1811.06529},
year={2018}
}
"""
[docs] def __init__(self, dim, max_step):
"""
Constructor for the :py:class:`ControlUnit`.
:param dim: global 'd' hidden dimension.
:type dim: int
:param max_step: maximum number of steps -> number of MAC cells in the network.
:type max_step: int
"""
# call base constructor
super(ControlUnit, self).__init__()
# define the linear layers (one per step) used to make the questions
# encoding
self.pos_aware_layers = torch.nn.ModuleList()
for _ in range(max_step):
self.pos_aware_layers.append(linear(2 * dim, dim, bias=True))
# define the linear layer used to create the cqi values
self.ctrl_question = linear(dim, dim, bias=True)
# define the linear layer used to create the attention weights. Should
# be one scalar weight per contextual word. No bias here.
self.attn = linear(dim, 1, bias=False)
self.step = 0
[docs] def forward(self, step, contextual_words, question_encoding, ctrl_state):
"""
Forward pass of the :py:class:`ControlUnit` for the ``S-MAC`` network.
:param step: index of the current MAC cell.
:type step: int
:param contextual_words: tensor of shape `[batch_size x maxQuestionLength x dim]` containing the words \
encodings ("representation of each word in the context of the question").
:type contextual_words: :py:class:`torch.Tensor`
:param question_encoding: question representation, of shape `[batch_size x 2*dim]`.
:type question_encoding: :py:class:`torch.Tensor`
:param ctrl_state: previous control state, of shape `[batch_size x dim]`
:type ctrl_state: :py:class:`torch.Tensor`
:return: new control state, `[batch_size x dim]` (:py:class:`torch.Tensor`)
"""
self.step = step
# select current 'position aware' linear layer & pass questions through
# it
pos_aware_question_encoding = self.pos_aware_layers[step](
question_encoding)
# create cqi values from projection of control state & question encoding (element-wise sum)
cqi = self.ctrl_question(ctrl_state) + pos_aware_question_encoding
# compute element-wise product between cqi & contextual words
# [batch_size x maxQuestionLength x dim]
context_ctrl = cqi.unsqueeze(1) * contextual_words
# compute attention weights
cai = self.attn(context_ctrl) # [batch_size x maxQuestionLength x 1]
# for the visualization
self.cvi = torch.nn.functional.softmax(cai, dim=1)
# compute next control state
# [batch_size x dim]
next_ctrl_state = (self.cvi * contextual_words).sum(1)
return next_ctrl_state