models/MLP.py

import torch
import torch.nn as nn


class MLP(nn.Module):
    def __init__(self, class_embeddings, input_size=256, hidden_sizes=[256, 256], output_size=1):
        super(MLP, self).__init__()

        self.class_embeddings = class_embeddings.clone().detach()
        #self.class_embeddings = nn.Parameter(class_embeddings)

        # Hidden layer
        layers = []
        in_size = input_size
        for hidden_size in hidden_sizes:
            layers.append(nn.Linear(in_size, hidden_size))
            layers.append(nn.ReLU())  # Apply ReLU activation function
            in_size = hidden_size

        # Output layer
        layers.append(nn.Linear(in_size, output_size))
        self.model = nn.Sequential(*layers)

    def forward(self, x):
        # Expand the dimensions of x to concatenate with class_embeddings; x has shape (batch_size, input_size)
        batch_size = x.size(0)
        # Here, concatenate each input x with all rows of class_embeddings
        x_expanded = x.unsqueeze(1).expand(batch_size, self.class_embeddings.size(0), -1)
        device = x_expanded.device
        embeddings_expanded = self.class_embeddings.unsqueeze(0).expand(batch_size, -1, -1).to(device)

        # Concatenate x and class_embeddings
        x_combined = torch.cat((x_expanded, embeddings_expanded), dim=-1)  # 沿着最后一个维度拼接

        # Flatten the concatenated tensor to pass it to subsequent layers
        x_combined = x_combined.view(batch_size * self.class_embeddings.size(0),
                                     -1)  # Flatten to (batch_size * n, input_size + h)
        # Forward pass through the network
        output = self.model(x_combined)

        # The returned output needs to be reshaped to (batch_size, n, output_size) to match the final output
        output = output.view(batch_size, self.class_embeddings.size(0))  # Reshape to (batch_size, n)
        return output