#!/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.
"""
write_unit.py: Implementation of the ``WriteUnit`` for the MAC network. Cf https://arxiv.org/abs/1803.03067 \
for the reference paper.
"""
__author__ = "Vincent Marois"
import torch
from torch.nn import Module
from miprometheus.models.mac.utils_mac import linear
[docs]class WriteUnit(Module):
"""
Implementation of the ``WriteUnit`` of the MAC network.
"""
[docs] def __init__(self, dim, self_attention=False, memory_gate=False):
"""
Constructor for the ``WriteUnit``.
:param dim: global 'd' hidden dimension
:type dim: int
:param self_attention: whether or not to use self-attention on the previous control states
:type self_attention: bool
:param memory_gate: whether or not to use memory gating.
:type memory_gate: bool
"""
# call base constructor
super(WriteUnit, self).__init__()
# linear layer for the concatenation of ri & mi-1
self.concat_layer = linear(2 * dim, dim, bias=True)
# self-attention & memory gating optional initializations
self.self_attention = self_attention
self.memory_gate = memory_gate
if self.self_attention:
self.attn = linear(dim, 1, bias=True)
self.mi_sa_proj = linear(dim, dim, bias=True)
self.mi_info_proj = linear(dim, dim, bias=True)
if self.memory_gate:
self.control = linear(dim, 1, bias=True)
[docs] def forward(self, memory_states, read_vector, ctrl_states):
"""
Forward pass of the ``WriteUnit``.
:param memory_states: All previous memory states, each of shape [batch_size x dim].
:type memory_states: list
:param read_vector: current read vector (output of the read unit), shape [batch_size x dim].
:type read_vector: torch.tensor
:param ctrl_states: All previous control states, each of shape [batch_size x dim].
:type ctrl_states: list
:return: current memory state, shape [batch_size x mem_dim]
"""
# retrieve the last memory state
memory_state = memory_states[-1]
# combine the new read vector with the prior memory state (w1)
mi_info = self.concat_layer(torch.cat([read_vector, memory_state], 1))
next_memory_state = mi_info # new memory state if no self-attention & memory-gating
if self.self_attention:
# compute attention weights from the relevance of each previous step to the current one (w2.1)
# [batch_size x dim x (i)], i: current step index (we count the initial control state c0)
controls_cat = torch.stack(ctrl_states[:-1], 2)
# [batch_size x dim x 1] * [batch_size x dim * (i)] -> [batch_size x dim * (i)]
attn = ctrl_states[-1].unsqueeze(2) * controls_cat
attn = self.attn(attn.permute(0, 2, 1)) # [batch_size x (i) x 1]
attn = torch.nn.functional.softmax(attn, dim=1).permute(
0, 2, 1) # [batch_size x 1 x (i)]
# compute weighted sum of the previous memory states (w2.2)
# [batch_size x dim x (i)], i: current step index (we count the initial memory state m0)
memories_cat = torch.stack(memory_states, dim=2)
mi_sa = (attn * memories_cat).sum(2) # [batch_size x dim]
# project both vector separately and element-wise sum (w2.3)
next_memory_state = self.mi_sa_proj(
mi_sa) + self.mi_info_proj(mi_info)
if self.memory_gate:
# project current control state (w3.1)
control = self.control(ctrl_states[-1])
# gating (w3.2)
gate = torch.nn.functional.sigmoid(control)
next_memory_state = gate * memory_state + \
(1 - gate) * next_memory_state
return next_memory_state