Module `pysimt.layers.attention.hierarchical`

Expand source code

# -*- coding: utf-8 -*-
import torch
from torch import nn

from ...utils.nn import get_activation_fn


# Libovický, J., & Helcl, J. (2017). Attention Strategies for Multi-Source
# Sequence-to-Sequence Learning. In Proceedings of the 55th Annual Meeting of
# the Association for Computational Linguistics (Volume 2: Short Papers)
# (Vol. 2, pp. 196-202). [Code contributed by @jlibovicky]


class HierarchicalAttention(nn.Module):
    """Hierarchical attention over multiple modalities."""
    def __init__(self, ctx_dims, hid_dim, mid_dim, att_activ='tanh'):
        super().__init__()

        self.activ = get_activation_fn(att_activ)
        self.ctx_dims = ctx_dims
        self.hid_dim = hid_dim
        self.mid_dim = mid_dim

        self.ctx_projs = nn.ModuleList([
            nn.Linear(dim, mid_dim, bias=False) for dim in self.ctx_dims])
        self.dec_proj = nn.Linear(hid_dim, mid_dim, bias=True)
        self.mlp = nn.Linear(self.mid_dim, 1, bias=False)

    def forward(self, contexts, hid):
        dec_state_proj = self.dec_proj(hid)
        ctx_projected = torch.cat([
            p(ctx).unsqueeze(0) for p, ctx
            in zip(self.ctx_projs, contexts)], dim=0)
        energies = self.mlp(self.activ(dec_state_proj + ctx_projected))
        att_dist = nn.functional.softmax(energies, dim=0)

        ctxs_cat = torch.cat([c.unsqueeze(0) for c in contexts])
        joint_context = (att_dist * ctxs_cat).sum(0)

        return att_dist, joint_context

Classes

class HierarchicalAttention (ctx_dims, hid_dim, mid_dim, att_activ='tanh')

Hierarchical attention over multiple modalities.

Initializes internal Module state, shared by both nn.Module and ScriptModule.

Expand source code

class HierarchicalAttention(nn.Module):
    """Hierarchical attention over multiple modalities."""
    def __init__(self, ctx_dims, hid_dim, mid_dim, att_activ='tanh'):
        super().__init__()

        self.activ = get_activation_fn(att_activ)
        self.ctx_dims = ctx_dims
        self.hid_dim = hid_dim
        self.mid_dim = mid_dim

        self.ctx_projs = nn.ModuleList([
            nn.Linear(dim, mid_dim, bias=False) for dim in self.ctx_dims])
        self.dec_proj = nn.Linear(hid_dim, mid_dim, bias=True)
        self.mlp = nn.Linear(self.mid_dim, 1, bias=False)

    def forward(self, contexts, hid):
        dec_state_proj = self.dec_proj(hid)
        ctx_projected = torch.cat([
            p(ctx).unsqueeze(0) for p, ctx
            in zip(self.ctx_projs, contexts)], dim=0)
        energies = self.mlp(self.activ(dec_state_proj + ctx_projected))
        att_dist = nn.functional.softmax(energies, dim=0)

        ctxs_cat = torch.cat([c.unsqueeze(0) for c in contexts])
        joint_context = (att_dist * ctxs_cat).sum(0)

        return att_dist, joint_context

Ancestors

torch.nn.modules.module.Module

Class variables

var dump_patches : bool
var training : bool

Methods

def forward(self, contexts, hid) ‑> Callable[..., Any]

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the :class:Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

Expand source code

def forward(self, contexts, hid):
    dec_state_proj = self.dec_proj(hid)
    ctx_projected = torch.cat([
        p(ctx).unsqueeze(0) for p, ctx
        in zip(self.ctx_projs, contexts)], dim=0)
    energies = self.mlp(self.activ(dec_state_proj + ctx_projected))
    att_dist = nn.functional.softmax(energies, dim=0)

    ctxs_cat = torch.cat([c.unsqueeze(0) for c in contexts])
    joint_context = (att_dist * ctxs_cat).sum(0)

    return att_dist, joint_context