# -------------------------------------------------------------------------------------------
# Copyright (c) Microsoft Corporation. All rights reserved.
# Licensed under the MIT License (MIT). See LICENSE in the repo root for license information.
# -------------------------------------------------------------------------------------------
from typing import Callable, Optional
import torch
import torch.nn as nn
[docs]class MLP(nn.Module):
"""
Fully connected layers to map between image embeddings and projection space where pairs of images are compared.
:param input_dim: Input embedding feature size
:param hidden_dim: Hidden layer size in MLP
:param output_dim: Output projection size
:param use_1x1_convs: Use 1x1 conv kernels instead of 2D linear transformations for speed and memory efficiency.
"""
def __init__(
self, input_dim: int, output_dim: int, hidden_dim: Optional[int] = None, use_1x1_convs: bool = False
) -> None:
super().__init__()
if use_1x1_convs:
linear_proj_1_args = {'in_channels': input_dim, 'out_channels': hidden_dim, 'kernel_size': 1, 'bias': False}
linear_proj_2_args = {'in_channels': hidden_dim, 'out_channels': output_dim, 'kernel_size': 1, 'bias': True}
normalisation_layer: Callable = nn.BatchNorm2d
projection_layer: Callable = nn.Conv2d
else:
linear_proj_1_args = {'in_features': input_dim, 'out_features': hidden_dim, 'bias': False}
linear_proj_2_args = {'in_features': hidden_dim, 'out_features': output_dim, 'bias': True}
normalisation_layer = nn.BatchNorm1d
projection_layer = nn.Linear
self.output_dim = output_dim
self.input_dim = input_dim
if hidden_dim is not None:
self.model = nn.Sequential(
projection_layer(**linear_proj_1_args),
normalisation_layer(hidden_dim),
nn.ReLU(inplace=True),
projection_layer(**linear_proj_2_args),
)
else:
self.model = nn.Linear(input_dim, output_dim) # type: ignore
[docs] def forward(self, x: torch.Tensor) -> torch.Tensor:
"""forward pass of the multi-layer perceptron"""
x = self.model(x)
return x
[docs]class MultiTaskModel(nn.Module):
"""Torch module for multi-task classification heads. We create a separate classification head
for each task and perform a forward pass on each head independently in forward(). Classification
heads are instances of `MLP`.
:param input_dim: Number of dimensions of the input feature map.
:param classifier_hidden_dim: Number of dimensions of hidden features in the MLP.
:param num_classes: Number of output classes per task.
:param num_tasks: Number of classification tasks or heads required.
"""
def __init__(self, input_dim: int, classifier_hidden_dim: Optional[int], num_classes: int, num_tasks: int):
super().__init__()
self.num_classes = num_classes
self.num_tasks = num_tasks
for task in range(num_tasks):
# TODO check if softmax not needed here.
setattr(self, "fc_" + str(task), MLP(input_dim, output_dim=num_classes, hidden_dim=classifier_hidden_dim))
[docs] def forward(self, x: torch.Tensor) -> torch.Tensor:
"""Returns [batch_size, num_tasks, num_classes] tensor of logits."""
batch_size = x.shape[0]
out = torch.zeros((batch_size, self.num_classes, self.num_tasks), dtype=x.dtype, device=x.device)
for task in range(self.num_tasks):
classifier = getattr(self, "fc_" + str(task))
out[:, :, task] = classifier(x)
return out