# 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,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from typing import Sequence, Union
import torch.nn as nn
from monai.networks.blocks.patchembedding import PatchEmbeddingBlock
from monai.networks.blocks.transformerblock import TransformerBlock
from monai.networks.layers import Conv
from monai.utils import ensure_tuple_rep
__all__ = ["ViTAutoEnc"]
Vision Transformer (ViT), based on: "Dosovitskiy et al.,
An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale <https://arxiv.org/abs/2010.11929>"
Modified to also give same dimension outputs as the input size of the image
[docs] def __init__(
img_size: Union[Sequence[int], int],
patch_size: Union[Sequence[int], int],
out_channels: int = 1,
deconv_chns: int = 16,
hidden_size: int = 768,
mlp_dim: int = 3072,
num_layers: int = 12,
num_heads: int = 12,
pos_embed: str = "conv",
dropout_rate: float = 0.0,
spatial_dims: int = 3,
) -> None:
in_channels: dimension of input channels or the number of channels for input
img_size: dimension of input image.
patch_size: dimension of patch size.
hidden_size: dimension of hidden layer.
out_channels: number of output channels.
deconv_chns: number of channels for the deconvolution layers.
mlp_dim: dimension of feedforward layer.
num_layers: number of transformer blocks.
num_heads: number of attention heads.
pos_embed: position embedding layer type.
dropout_rate: faction of the input units to drop.
spatial_dims: number of spatial dimensions.
# for single channel input with image size of (96,96,96), conv position embedding and segmentation backbone
# It will provide an output of same size as that of the input
>>> net = ViTAutoEnc(in_channels=1, patch_size=(16,16,16), img_size=(96,96,96), pos_embed='conv')
# for 3-channel with image size of (128,128,128), output will be same size as of input
>>> net = ViTAutoEnc(in_channels=3, patch_size=(16,16,16), img_size=(128,128,128), pos_embed='conv')
self.patch_size = ensure_tuple_rep(patch_size, spatial_dims)
self.spatial_dims = spatial_dims
self.patch_embedding = PatchEmbeddingBlock(
self.blocks = nn.ModuleList(
[TransformerBlock(hidden_size, mlp_dim, num_heads, dropout_rate) for i in range(num_layers)]
self.norm = nn.LayerNorm(hidden_size)
new_patch_size =  * self.spatial_dims
conv_trans = Conv[Conv.CONVTRANS, self.spatial_dims]
# self.conv3d_transpose* is to be compatible with existing 3d model weights.
self.conv3d_transpose = conv_trans(hidden_size, deconv_chns, kernel_size=new_patch_size, stride=new_patch_size)
self.conv3d_transpose_1 = conv_trans(
in_channels=deconv_chns, out_channels=out_channels, kernel_size=new_patch_size, stride=new_patch_size
[docs] def forward(self, x):
x: input tensor must have isotropic spatial dimensions,
such as ``[batch_size, channels, sp_size, sp_size[, sp_size]]``.
spatial_size = x.shape[2:]
x = self.patch_embedding(x)
hidden_states_out = 
for blk in self.blocks:
x = blk(x)
x = self.norm(x)
x = x.transpose(1, 2)
d = [s // p for s, p in zip(spatial_size, self.patch_size)]
x = torch.reshape(x, [x.shape, x.shape, *d])
x = self.conv3d_transpose(x)
x = self.conv3d_transpose_1(x)
return x, hidden_states_out