DI-engine/ding/torch_utils/network/merge.py

"""
This file provides an implementation of several different neural network modules that are used for merging and
transforming input data in various ways. The following components can be used when we are dealing with
data from multiple modes, or when we need to merge multiple intermediate embedded representations in
the forward process of a model.

The main classes defined in this code are:

    - BilinearGeneral: This class implements a bilinear transformation layer that applies a bilinear transformation to
        incoming data, as described in the "Multiplicative Interactions and Where to Find Them", published at ICLR 2020,
        https://openreview.net/forum?id=rylnK6VtDH. The transformation involves two input features and an output
        feature, and also includes an optional bias term.

    - TorchBilinearCustomized: This class implements a bilinear layer similar to the one provided by PyTorch
        (torch.nn.Bilinear), but with additional customizations. This class can be used as an alternative to the
        BilinearGeneral class.

    - TorchBilinear: This class is a simple wrapper around the PyTorch's built-in nn.Bilinear module. It provides the
        same functionality as PyTorch's nn.Bilinear but within the structure of the current module.

    - FiLM: This class implements a Feature-wise Linear Modulation (FiLM) layer. FiLM layers apply an affine
        transformation to the input data, conditioned on some additional context information.

    - GatingType: This is an enumeration class that defines different types of gating mechanisms that can be used in
        the modules.

    - SumMerge: This class provides a simple summing mechanism to merge input streams.

    - VectorMerge: This class implements a more complex merging mechanism for vector streams.
        The streams are first transformed using layer normalization, a ReLU activation, and a linear layer.
        Then they are merged either by simple summing or by using a gating mechanism.

The implementation of these classes involves PyTorch and Numpy libraries, and the classes use PyTorch's nn.Module as
the base class, making them compatible with PyTorch's neural network modules and functionalities.
These modules can be useful building blocks in more complex deep learning architectures.
"""

import enum
import math
from collections import OrderedDict
from typing import List, Dict, Tuple

import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch import Tensor


class BilinearGeneral(nn.Module):
    """
    Overview:
        Bilinear implementation as in: Multiplicative Interactions and Where to Find Them,
        ICLR 2020, https://openreview.net/forum?id=rylnK6VtDH.
    Interfaces:
        ``__init__``, ``forward``
    """

    def __init__(self, in1_features: int, in2_features: int, out_features: int):
        """
        Overview:
            Initialize the Bilinear layer.
        Arguments:
            - in1_features (:obj:`int`): The size of each first input sample.
            - in2_features (:obj:`int`): The size of each second input sample.
            - out_features (:obj:`int`): The size of each output sample.
        """

        super(BilinearGeneral, self).__init__()
        # Initialize the weight matrices W and U, and the bias vectors V and b
        self.W = nn.Parameter(torch.Tensor(out_features, in1_features, in2_features))
        self.U = nn.Parameter(torch.Tensor(out_features, in2_features))
        self.V = nn.Parameter(torch.Tensor(out_features, in1_features))
        self.b = nn.Parameter(torch.Tensor(out_features))
        self.in1_features = in1_features
        self.in2_features = in2_features
        self.out_features = out_features
        self.reset_parameters()

    def reset_parameters(self):
        """
        Overview:
            Initialize the parameters of the Bilinear layer.
        """

        stdv = 1. / np.sqrt(self.in1_features)
        self.W.data.uniform_(-stdv, stdv)
        self.U.data.uniform_(-stdv, stdv)
        self.V.data.uniform_(-stdv, stdv)
        self.b.data.uniform_(-stdv, stdv)

    def forward(self, x: torch.Tensor, z: torch.Tensor):
        """
        Overview:
            compute the bilinear function.
        Arguments:
            - x (:obj:`torch.Tensor`): The first input tensor.
            - z (:obj:`torch.Tensor`): The second input tensor.
        """

        # Compute the bilinear function
        # x^TWz
        out_W = torch.einsum('bi,kij,bj->bk', x, self.W, z)
        # x^TU
        out_U = z.matmul(self.U.t())
        # Vz
        out_V = x.matmul(self.V.t())
        # x^TWz + x^TU + Vz + b
        out = out_W + out_U + out_V + self.b
        return out


class TorchBilinearCustomized(nn.Module):
    """
    Overview:
        Customized Torch Bilinear implementation.
    Interfaces:
        ``__init__``, ``forward``
    """

    def __init__(self, in1_features: int, in2_features: int, out_features: int):
        """
        Overview:
            Initialize the Bilinear layer.
        Arguments:
            - in1_features (:obj:`int`): The size of each first input sample.
            - in2_features (:obj:`int`): The size of each second input sample.
            - out_features (:obj:`int`): The size of each output sample.
        """

        super(TorchBilinearCustomized, self).__init__()
        self.in1_features = in1_features
        self.in2_features = in2_features
        self.out_features = out_features
        self.weight = nn.Parameter(torch.Tensor(out_features, in1_features, in2_features))
        self.bias = nn.Parameter(torch.Tensor(out_features))
        self.reset_parameters()

    def reset_parameters(self):
        """
        Overview:
            Initialize the parameters of the Bilinear layer.
        """

        bound = 1 / math.sqrt(self.in1_features)
        nn.init.uniform_(self.weight, -bound, bound)
        nn.init.uniform_(self.bias, -bound, bound)

    def forward(self, x, z):
        """
        Overview:
            Compute the bilinear function.
        Arguments:
            - x (:obj:`torch.Tensor`): The first input tensor.
            - z (:obj:`torch.Tensor`): The second input tensor.
        """

        # Using torch.einsum for the bilinear operation
        out = torch.einsum('bi,oij,bj->bo', x, self.weight, z) + self.bias
        return out.squeeze(-1)


"""
Overview:
    Implementation of the Bilinear layer as in PyTorch:
    https://pytorch.org/docs/stable/generated/torch.nn.Bilinear.html#torch.nn.Bilinear
Arguments:
    - in1_features (:obj:`int`): The size of each first input sample.
    - in2_features (:obj:`int`): The size of each second input sample.
    - out_features (:obj:`int`): The size of each output sample.
    - bias (:obj:`bool`): If set to False, the layer will not learn an additive bias. Default: ``True``.
"""
TorchBilinear = nn.Bilinear


class FiLM(nn.Module):
    """
    Overview:
        Feature-wise Linear Modulation (FiLM) Layer.
        This layer applies feature-wise affine transformation based on context.
    Interfaces:
        ``__init__``, ``forward``
    """

    def __init__(self, feature_dim: int, context_dim: int):
        """
        Overview:
            Initialize the FiLM layer.
        Arguments:
            - feature_dim (:obj:`int`). The dimension of the input feature vector.
            - context_dim (:obj:`int`). The dimension of the input context vector.
        """

        super(FiLM, self).__init__()
        # Define the fully connected layer for context
        # The output dimension is twice the feature dimension for gamma and beta
        self.context_layer = nn.Linear(context_dim, 2 * feature_dim)

    def forward(self, feature: torch.Tensor, context: torch.Tensor):
        """
        Overview:
            Forward propagation.
        Arguments:
            - feature (:obj:`torch.Tensor`). The input feature, shape (batch_size, feature_dim).
            - context (:obj:`torch.Tensor`). The input context, shape (batch_size, context_dim).
        Returns:
            - conditioned_feature : torch.Tensor. The output feature after FiLM, shape (batch_size, feature_dim).
        """

        # Pass context through the fully connected layer
        out = self.context_layer(context)
        # Split the output into two parts: gamma and beta
        # The dimension for splitting is 1 (feature dimension)
        gamma, beta = torch.split(out, out.shape[1] // 2, dim=1)
        # Apply feature-wise affine transformation
        conditioned_feature = gamma * feature + beta
        return conditioned_feature


class GatingType(enum.Enum):
    """
    Overview:
        Enum class defining different types of tensor gating and aggregation in modules.
    """
    NONE = 'none'
    GLOBAL = 'global'
    POINTWISE = 'pointwise'


class SumMerge(nn.Module):
    """
    Overview:
        A PyTorch module that merges a list of tensors by computing their sum. All input tensors must have the same
        size. This module can work with any type of tensor (vector, units or visual).
    Interfaces:
        ``__init__``, ``forward``
    """

    def forward(self, tensors: List[Tensor]) -> Tensor:
        """
        Overview:
            Forward pass of the SumMerge module, which sums the input tensors.
        Arguments:
            - tensors (:obj:`List[Tensor]`): List of input tensors to be summed. All tensors must have the same size.
        Returns:
            - summed (:obj:`Tensor`): Tensor resulting from the sum of all input tensors.
        """
        # stack the tensors along the first dimension
        stacked = torch.stack(tensors, dim=0)

        # compute the sum along the first dimension
        summed = torch.sum(stacked, dim=0)
        # summed = sum(tensors)
        return summed


class VectorMerge(nn.Module):
    """
    Overview:
        Merges multiple vector streams. Streams are first transformed through layer normalization, relu, and linear
        layers, then summed. They don't need to have the same size. Gating can also be used before the sum.
    Interfaces:
        ``__init__``, ``encode``, ``_compute_gate``, ``forward``

    .. note::
        For more details about the gating types, please refer to the GatingType enum class.
    """

    def __init__(
        self,
        input_sizes: Dict[str, int],
        output_size: int,
        gating_type: GatingType = GatingType.NONE,
        use_layer_norm: bool = True,
    ):
        """
        Overview:
            Initialize the `VectorMerge` module.
        Arguments:
            - input_sizes (:obj:`Dict[str, int]`): A dictionary mapping input names to their sizes. \
                The size is a single integer for 1D inputs, or `None` for 0D inputs. \
                If an input size is `None`, we assume it's `()`.
            - output_size (:obj:`int`): The size of the output vector.
            - gating_type (:obj:`GatingType`): The type of gating mechanism to use. Default is `GatingType.NONE`.
            - use_layer_norm (:obj:`bool`): Whether to use layer normalization. Default is `True`.
        """
        super().__init__()
        self._input_sizes = OrderedDict(input_sizes)
        self._output_size = output_size
        self._gating_type = gating_type
        self._use_layer_norm = use_layer_norm

        if self._use_layer_norm:
            self._layer_norms = nn.ModuleDict()
        else:
            self._layer_norms = None

        self._linears = nn.ModuleDict()
        for name, size in self._input_sizes.items():
            linear_input_size = size if size > 0 else 1
            if self._use_layer_norm:
                self._layer_norms[name] = nn.LayerNorm(linear_input_size)
            self._linears[name] = nn.Linear(linear_input_size, self._output_size)

        self._gating_linears = nn.ModuleDict()
        if self._gating_type is GatingType.GLOBAL:
            self.gate_size = 1
        elif self._gating_type is GatingType.POINTWISE:
            self.gate_size = self._output_size
        elif self._gating_type is GatingType.NONE:
            self._gating_linears = None
        else:
            raise ValueError(f'Gating type {self._gating_type} is not supported')

        if self._gating_linears is not None:
            if len(self._input_sizes) == 2:
                # more efficient than the general version below
                for name, size in self._input_sizes.items():
                    gate_input_size = size if size > 0 else 1
                    gating_layer = nn.Linear(gate_input_size, self.gate_size)
                    torch.nn.init.normal_(gating_layer.weight, std=0.005)
                    torch.nn.init.constant_(gating_layer.bias, 0.0)
                    self._gating_linears[name] = gating_layer
            else:
                for name, size in self._input_sizes.items():
                    gate_input_size = size if size > 0 else 1
                    gating_layer = nn.Linear(gate_input_size, len(self._input_sizes) * self.gate_size)
                    torch.nn.init.normal_(gating_layer.weight, std=0.005)
                    torch.nn.init.constant_(gating_layer.bias, 0.0)
                    self._gating_linears[name] = gating_layer

    def encode(self, inputs: Dict[str, Tensor]) -> Tuple[List[Tensor], List[Tensor]]:
        """
        Overview:
            Encode the input tensors using layer normalization, relu, and linear transformations.
        Arguments:
            - inputs (:obj:`Dict[str, Tensor]`): The input tensors.
        Returns:
            - gates (:obj:`List[Tensor]`): The gate tensors after transformations.
            - outputs (:obj:`List[Tensor]`): The output tensors after transformations.
        """
        gates, outputs = [], []
        for name, size in self._input_sizes.items():
            feature = inputs[name]
            if size <= 0 and feature.dim() == 1:
                feature = feature.unsqueeze(-1)
            feature = feature.to(torch.float32)
            if self._use_layer_norm and name in self._layer_norms:
                feature = self._layer_norms[name](feature)
            feature = F.relu(feature)
            gates.append(feature)
            outputs.append(self._linears[name](feature))
        return gates, outputs

    def _compute_gate(
            self,
            init_gate: List[Tensor],
    ) -> List[Tensor]:
        """
        Overview:
            Compute the gate values based on the initial gate values.
        Arguments:
            - init_gate (:obj:`List[Tensor]`): The initial gate values.
        Returns:
            - gate (:obj:`List[Tensor]`): The computed gate values.
        """
        if len(self._input_sizes) == 2:
            gate = [self._gating_linears[name](y) for name, y in zip(self._input_sizes.keys(), init_gate)]
            gate = sum(gate)
            sigmoid = torch.sigmoid(gate)
            gate = [sigmoid, 1.0 - sigmoid]
        else:
            gate = [self._gating_linears[name](y) for name, y in zip(self._input_sizes.keys(), init_gate)]
            gate = sum(gate)
            gate = gate.reshape([-1, len(self._input_sizes), self.gate_size])
            gate = F.softmax(gate, dim=1)
            assert gate.shape[1] == len(self._input_sizes)
            gate = [gate[:, i] for i in range(len(self._input_sizes))]
        return gate

    def forward(self, inputs: Dict[str, Tensor]) -> Tensor:
        """
        Overview:
            Forward pass through the VectorMerge module.
        Arguments:
            - inputs (:obj:`Dict[str, Tensor]`): The input tensors.
        Returns:
            - output (:obj:`Tensor`): The output tensor after passing through the module.
        """
        gates, outputs = self.encode(inputs)
        if len(outputs) == 1:
            # Special case of 1-D inputs that do not need any gating.
            output = outputs[0]
        elif self._gating_type is GatingType.NONE:
            output = sum(outputs)
        else:
            gate = self._compute_gate(gates)
            data = [g * d for g, d in zip(gate, outputs)]
            output = sum(data)
        return output