Source code for monai.networks.blocks.selfattention

# Copyright (c) MONAI Consortium
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# See the License for the specific language governing permissions and
# limitations under the License.

from __future__ import annotations

import torch
import torch.nn as nn

from monai.utils import optional_import

Rearrange, _ = optional_import("einops.layers.torch", name="Rearrange")

[docs]class SABlock(nn.Module): """ A self-attention block, based on: "Dosovitskiy et al., An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale <>" """
[docs] def __init__( self, hidden_size: int, num_heads: int, dropout_rate: float = 0.0, qkv_bias: bool = False, save_attn: bool = False, ) -> None: """ Args: hidden_size (int): dimension of hidden layer. num_heads (int): number of attention heads. dropout_rate (float, optional): faction of the input units to drop. Defaults to 0.0. qkv_bias (bool, optional): bias term for the qkv linear layer. Defaults to False. save_attn (bool, optional): to make accessible the attention matrix. Defaults to False. """ super().__init__() if not (0 <= dropout_rate <= 1): raise ValueError("dropout_rate should be between 0 and 1.") if hidden_size % num_heads != 0: raise ValueError("hidden size should be divisible by num_heads.") self.num_heads = num_heads self.out_proj = nn.Linear(hidden_size, hidden_size) self.qkv = nn.Linear(hidden_size, hidden_size * 3, bias=qkv_bias) self.input_rearrange = Rearrange("b h (qkv l d) -> qkv b l h d", qkv=3, l=num_heads) self.out_rearrange = Rearrange("b h l d -> b l (h d)") self.drop_output = nn.Dropout(dropout_rate) self.drop_weights = nn.Dropout(dropout_rate) self.head_dim = hidden_size // num_heads self.scale = self.head_dim**-0.5 self.save_attn = save_attn self.att_mat = torch.Tensor()
[docs] def forward(self, x): output = self.input_rearrange(self.qkv(x)) q, k, v = output[0], output[1], output[2] att_mat = (torch.einsum("blxd,blyd->blxy", q, k) * self.scale).softmax(dim=-1) if self.save_attn: # no gradients and new tensor; # self.att_mat = att_mat.detach() att_mat = self.drop_weights(att_mat) x = torch.einsum("bhxy,bhyd->bhxd", att_mat, v) x = self.out_rearrange(x) x = self.out_proj(x) x = self.drop_output(x) return x