TextMLPEncoder

Description

The TextMLPEncoder class maps each vector or text representation to a fixed-width embedding with a multilayer perceptron. It is the text branch used by multimodal DynamicTarNetBase. Each hidden layer is followed by ReLU and optional dropout; the final projection to out_dim is linear.

Parameters

input_dim (int): width of the input representation.
hidden_dims (sequence of int, optional): hidden-layer widths. The default is (1024, 256).
out_dim (int, optional): output width. The default is 128.
dropout (float, optional): dropout probability after hidden layers. The default is 0.

Returns

forward(x) maps the tensor’s last dimension from input_dim to out_dim while preserving all leading dimensions. For example, both [B, input_dim] and [B, T, input_dim] inputs are accepted by the underlying linear layers. All input, hidden, and output widths must be positive; hidden_dims=() gives a single direct linear projection.

Example Usage

import torch
from gpi_pack.dyn_gpi import TextMLPEncoder

encoder = TextMLPEncoder(input_dim=4096, out_dim=128)
encoded = encoder(torch.randn(32, 4096))
print(encoded.shape)  # [32, 128]