diagram from https://arxiv.org/pdf/1705.02735.pdf#page=5

multimodal-diagram.tex

% Overview figure for
%   ``Combating human trafficking with multimodal deep models,''
% submitted to ACL 2017.
% Figure by William Chargin; send any comments to wchargin@gmail.com.
\documentclass[margin=2mm]{standalone}

\usepackage{amsmath}
\usepackage{amsfonts}

\usepackage{tikz}
\usetikzlibrary{positioning,decorations.pathreplacing}

\renewcommand{\Pr}{\textrm{P}}

\newcommand*{\vdotsalign}{\raisebox{1.25ex}{$\vdots$}}

\begin{document}

% Usage: \cube{origin}{width}{height}{extrudeX}{extrudeY}{label}{fill options}
% Construct a cube with front-top-left corner at #1, which should be a
% TikZ coordinate---e.g., (1cm, 2cm) or (my node). The label will be
% placed in the center of the cube. The fill options (e.g., fill=white)
% control the background color of the cube.
\def\cube#1#2#3#4#5#6#7{%
  \path[#7]
    #1 --
    ++(#4, #5) --
    ++(#2, 0) --
    ++(0, -#3) --
    ++(-#4, -#5) --
    ++(-#2, 0) --
    cycle
    ;
  \draw #1 -- ++(#2, 0) -- ++(0, -#3) -- ++(-#2, 0) -- cycle;
  \draw #1 -- ++(#4, #5) -- ++(#2, 0) -- ++(0, -#3) -- ++(-#4, -#5);
  \draw #1 ++(#2, 0) -- ++(#4, #5);
  \path #1 rectangle node[align=center, font=\small] {#6} ++(#2, -#3);
}
\def\fillcube#1#2#3#4#5#6{\cube{#1}{#2}{#3}{#4}{#5}{#6}{fill=white}}
\def\emptycube#1#2#3#4#5#6{\cube{#1}{#2}{#3}{#4}{#5}{#6}{opacity=0}}

% Usage: \boxedletter{x}. Emulate a fancy Unicode symbol, but with
% consistent height and mostly consistent spacing.
\def\boxedletter#1{\fbox{\textsc{\vphantom{m}\smash{#1}}}}

% Here are some styles for easy modification.
\tikzset{%
  neuron/.style={%
    circle,
    minimum width=2mm,
    inner sep=0,
    draw=black,
    fill=black!30!white,
  },
  convolution box/.style={%
    % This box is the ``conv'' prism inside the boxes.
    fill=black!15!white,
  },
  small label/.style={%
    font=\footnotesize,
    align=center,
  },
  sideways label/.style={%
    rotate=30,
    anchor=east,
    xshift=0.2cm,
    yshift=-0.1cm,
  },
}

\begin{tikzpicture}[>=stealth]
  % This scope block contains the upper-left portion of the figure.
  \begin{scope}
    \def\lstmSpacing{0.6cm}
    %
    % Change the words here to change the initial sentence.
    \foreach \word/\j in {%
      {d0ll@r}/0,%
      {to}/1,
      {\vdotsalign}/2,%
      {\boxedletter{g}\boxedletter{r}\boxedletter{8}}/3,%
      {skype}/4%
    }{%
      \node[left] at (0.5cm, -\lstmSpacing * \j) {\word};
    };
    %
    % usage: \makeLSTM{index}{LSTM label name}{neuron label name}
    \def\makeLSTM#1#2#3{%
      \def\j{#1}
      \def\lstmHeight{0.25cm}
      \def\lstmWidth{2cm}
      \path [fill=white, draw=black]
        (1cm, -\j * \lstmSpacing + \lstmHeight / 2) rectangle ++(\lstmWidth, -\lstmHeight);
      \draw [->] (0.5cm, -\j * \lstmSpacing) -- ++(0.5cm, 0);
      \foreach \x in {1.25cm,1.5cm,2.5cm,2.75cm}{%
        \draw (\x, -\j * \lstmSpacing + \lstmHeight / 2) -- ++(0, -\lstmHeight);
      }
      \node at (2cm, -\j * \lstmSpacing) {$\cdots\unkern$};
      \node [draw=black, fill=white, small label] (#2)
        at (4cm, -\j * \lstmSpacing) {LSTM};
      \draw [->] (1cm + \lstmWidth, -\j * \lstmSpacing) -- (#2);
      \node [neuron] (#3) at (5.5cm, -\j * \lstmSpacing) {};
    }
    \makeLSTM{0}{lstm 1}{lstm node 1}
    \makeLSTM{1}{lstm 2}{lstm node 2}
    \makeLSTM{3}{lstm 3}{lstm node 3}
    \makeLSTM{4}{lstm 4}{lstm node 4}
    \draw [decorate, decoration={brace, amplitude=4pt, raise=0.2cm}]
      (lstm node 1) -- (lstm node 4);
    \coordinate (lstm anchor) at (6cm, -2 * \lstmSpacing);
    \foreach \i in {1,2,3,4}{%
      \foreach \j in {1,2,3,4}{%
        \draw (lstm \i.east) -- (lstm node \j);
      }
    }
    \draw [->] (lstm 1) -- (lstm 2);
    \draw [->] (lstm 3) -- (lstm 4);
    \node [above=0.15cm of lstm node 1] {$h_l \in \mathbb{R}^{300}$};
    \node [align=center] at (2cm, 0.7cm) {Trafficking \\ embedding};
    \node at (2cm, -2 * \lstmSpacing) {\vdotsalign};
    \node at (4cm, -2 * \lstmSpacing) {\vdotsalign};
    %
    % Swap the comments on these next two lines to show a dotted arrow
    % between the LSTM gap, instead of just vdots.
    %\draw [dotted, ->] (lstm 2) -- (lstm 3);
    \node at (5.5cm, -2 * \lstmSpacing) {\vdotsalign};
    %
    % Swap the comments on these next two lines to change the node text.
    %\node [below=0 of lstm node 4, small label] {$300$ \\ sigmoid};
    \node [below=0 of lstm node 4, small label] {$300 \sigma$};
  \end{scope}

  % This scope block contains the lower-left portion of the figure.
  \begin{scope}[yshift=-4cm]
    %
    % Modify the following settings to change the appearance of the
    % images. In particular, you can make them darker by adjusting the
    % \imageIntensity.
    \def\imageWidth{0.4cm}
    \def\imageHeight{0.8cm}
    \def\imageColor{black}
    \def\imageIntensity{0.05}
    \foreach \i in {0,1,2,3,4}{%
      \pgfmathtruncatemacro{\iplusone}{\i + 1};
      \pgfmathtruncatemacro{\thisIntensity}{100*\imageIntensity*(\i+1)};
      \edef\thisImageColor{\imageColor!\thisIntensity!white},
    \node at (0, \i * -0.45cm + 0.25cm) {$\hat{\imath}_{\iplusone}$};
      \fill [
        color=\thisImageColor,
        draw=black,
        xshift=\i * 0.1cm + 0.2cm,
        yshift=\i * -0.3cm + 0.25cm, yslant=0.5,
      ]
        (0, 0) --
        ++(\imageWidth, 0) --
        ++(0, -\imageHeight) --
        ++(-\imageWidth, 0) --
        cycle
        ;
    }
    \foreach \i in {0,1,2,3,4}{%
      \pgfmathsetmacro{\x}{1.25 - 0.1*\i};
      \pgfmathsetmacro{\y}{-0.5 + 0.3*(\i-2)};
      \typeout{\y}
      \draw [->] (\x cm, \y cm) -- (1.75cm, \y cm);
    }
    \coordinate (TVGG origin) at (1.75cm, 0.5cm);
    \path (TVGG origin) ++ (2.25cm, 0.25cm) coordinate (TVGG top right);
    \path (TVGG origin) ++ (2.25cm, -1.75cm) coordinate (TVGG bottom right);
    \path (TVGG origin) ++ (1.75cm, 0) coordinate (TVGG front top right);
    \path (TVGG origin) ++ (1.75cm, -2cm) coordinate (TVGG front bottom right);
    \fillcube{(TVGG origin)}{1.75cm}{2cm}{0.5cm}{0.25cm}{Trafficking \\ VGG};
    \foreach \i in {1,2,3}{%
      \pgfmathsetmacro{\x}{4.5 + 0.5 * (\i - 1)};
      \node at (\x, -0.5cm) {\vdotsalign};
      \foreach \j in {1,2,4,5}{%
        \pgfmathsetmacro{\y}{-0.5 + 0.4 * (\j - 3)};
        \node [neuron] (node-\i-\j) at (\x, \y) {};
      }
    }
    \foreach \i/\direction in {1/below left,2/below,3/below right}{%
      % Here are two alternatives: one uses ``200 \sigma'', and the
      % other uses ``250 \\ sigmoid'' but requires that the labels be
      % positioned further from the nodes. Swap the comments to change
      % which one is in use.
      \node [below=0 of node-\i-1, small label, sideways label] {$200 \sigma$};
      %\node [\direction=0 of node-\i-1, small label] {$200$ \\ sigmoid};
    }
    \foreach \i in {1,2}{%
      \pgfmathtruncatemacro{\nexti}{\i + 1};
      \foreach \j in {1,2,4,5}{%
        \foreach \k in {1,2,4,5}{%
          \draw (node-\i-\j) -- (node-\nexti-\k);
        }
      }
    }
    \draw [dashed] (TVGG top right) -- (node-1-5);
    \draw [dashed] (TVGG bottom right) -- (node-1-1);
    \draw [dashed] (TVGG front top right) -- (node-1-5);
    \draw [dashed] (TVGG front bottom right) -- (node-1-1);
    %
    \coordinate (hV) at (5.75cm, -0.5cm);
    \node [below right] at (hV) {$h_v \in \mathbb{R}^{5 \times 200}$};
  \end{scope}

  % This scope block contains the right-hand portion of the figure.
  \begin{scope}[xshift=7cm]
    \begin{scope}[xshift=0.25cm,yshift=-1.75cm]
      \fillcube{(0, 0)}{1cm}{2cm}{0.5cm}{1cm}{};
      \cube{(0.425cm, 0.85cm)}{1cm}{0.3cm}{0.075cm}{0.15cm}{\footnotesize conv}{convolution box};
      \emptycube{(0, 0)}{1cm}{2cm}{0.5cm}{1cm}{};
      \node [above] at (1cm, 1cm) {$h_m \in \mathbb{R}^{5 \times 200 \times 300}$};
      %
      \coordinate (hM top right) at (3cm, 1cm);
      \draw [->] (lstm anchor) to[in=180, out=0] (-0.1cm, -0.9cm);
      \draw [->] (hV) to[in=180, out=0] (-0.1cm, -1.1cm);
      \node at (-0.5cm, -1cm) {$\otimes$};
      %
      \node [below, small label] at (0.5cm, -2cm) {$5 \times 200 \times 300$};
      %
      \draw [->] (1.5cm, -0.5cm) -- node[above, small label] {max \\ pooling} ++(1.2cm, 0);
      \fillcube{(2.7cm, 0cm)}{1cm}{1cm}{0.375cm}{0.75cm}{};
      \cube{(3.0cm, 0.6cm)}{1cm}{0.3cm}{0.075cm}{0.15cm}{\footnotesize conv}{convolution box};
      \emptycube{(2.7cm, 0cm)}{1cm}{1cm}{0.375cm}{0.75cm}{};
      \node [below, small label] at (3.2cm, -1cm) {$5 \times 100 \times 150$};
      %
      \draw [->] (4.075cm, 0) -- node[above, small label] {max \\ pooling} ++(1.2cm, 0);
      \draw [decorate, decoration={brace, amplitude=4pt, mirror, raise=-0.1cm}]
        (5.4cm, 1.5cm) -- ++(0, -3cm);
      \node [neuron] (output) at (6.2cm, 0) {};
      \foreach \i in {-4,-3,-2,2,3,4}{%
        \pgfmathsetmacro{\endY}{0.35*\i};
        \node [neuron] (dense-\i) at (5.65cm, \endY) {};
        \draw [->] (dense-\i) -- (output);
      }
      \node [below=0 of dense--4, small label] {$150$ linear};
      \node (dense dots) at (5.65cm, 0) {\vdotsalign};
      \draw [->, dotted] (dense dots) -- (output);
      \draw [->] (output) -- ++(1cm, 0);
      \node [below right=0 of output, small label] {$\Pr[\tau \mid h_m; W_d]$};
      \node [above right=0 of output] {$\sigma$};
    \end{scope}
  \end{scope}

\end{tikzpicture}



\end{document}