Source code for torcheeg.models.transformer.simple_vit

from typing import Tuple

import numpy as np
import torch
from einops import rearrange
from einops.layers.torch import Rearrange
from torch import nn


def pair(t):
    return t if isinstance(t, tuple) else (t, t)


def get_emb(sin_inp):
    emb = torch.stack((sin_inp.sin(), sin_inp.cos()), dim=-1)
    return torch.flatten(emb, -2, -1)


class PositionEmbedding3D(nn.Module):
    def __init__(self, in_channels, temporature: float = 10000.0):
        super(PositionEmbedding3D, self).__init__()
        self.in_channels = in_channels
        self.temporature = temporature

        in_channels = int(np.ceil(in_channels / 6) * 2)
        if in_channels % 2:
            in_channels += 1
        self.in_channels = in_channels
        inv_freq = 1.0 / (temporature**(
            torch.arange(0, in_channels, 2).float() / in_channels))
        self.register_buffer("inv_freq", inv_freq)
        self.cached_penc = None

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        if len(x.shape) != 5:
            raise RuntimeError(
                "The input must be five-dimensional to perform thres-dimensional position embedding!"
            )

        if self.cached_penc is not None and self.cached_penc.shape == x.shape:
            return self.cached_penc

        self.cached_penc = None
        batch_size, a, b, c, orig_ch = x.shape
        pos_a = torch.arange(a, device=x.device).type(self.inv_freq.type())
        pos_b = torch.arange(b, device=x.device).type(self.inv_freq.type())
        pos_c = torch.arange(c, device=x.device).type(self.inv_freq.type())
        sin_inp_a = torch.einsum("i,j->ij", pos_a, self.inv_freq)
        sin_inp_b = torch.einsum("i,j->ij", pos_b, self.inv_freq)
        sin_inp_c = torch.einsum("i,j->ij", pos_c, self.inv_freq)
        emb_a = get_emb(sin_inp_a).unsqueeze(1).unsqueeze(1)
        emb_b = get_emb(sin_inp_b).unsqueeze(1)
        emb_c = get_emb(sin_inp_c)
        emb = torch.zeros((a, b, c, self.in_channels * 3),
                          device=x.device).type(x.type())
        emb[:, :, :, :self.in_channels] = emb_a
        emb[:, :, :, self.in_channels:2 * self.in_channels] = emb_b
        emb[:, :, :, 2 * self.in_channels:] = emb_c

        self.cached_penc = emb[None, :, :, :, :orig_ch].repeat(
            batch_size, 1, 1, 1, 1)
        return self.cached_penc


class FeedForward(nn.Module):
    def __init__(self, in_channels: int, hid_channels: int):
        super().__init__()
        self.net = nn.Sequential(
            nn.LayerNorm(in_channels),
            nn.Linear(in_channels, hid_channels),
            nn.GELU(),
            nn.Linear(hid_channels, in_channels),
        )

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        return self.net(x)


class Attention(nn.Module):
    def __init__(self,
                 hid_channels: int,
                 heads: int = 8,
                 head_channels: int = 64):
        super().__init__()
        inner_channels = head_channels * heads
        self.heads = heads
        self.scale = head_channels**-0.5
        self.norm = nn.LayerNorm(hid_channels)

        self.attend = nn.Softmax(dim=-1)

        self.to_qkv = nn.Linear(hid_channels, inner_channels * 3, bias=False)
        self.to_out = nn.Linear(inner_channels, hid_channels, bias=False)

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        x = self.norm(x)

        qkv = self.to_qkv(x).chunk(3, dim=-1)
        q, k, v = map(
            lambda t: rearrange(t, 'b n (h d) -> b h n d', h=self.heads), qkv)

        dots = torch.matmul(q, k.transpose(-1, -2)) * self.scale

        attn = self.attend(dots)

        out = torch.matmul(attn, v)
        out = rearrange(out, 'b h n d -> b n (h d)')
        return self.to_out(out)


class Transformer(nn.Module):
    def __init__(self, hid_channels: int, depth: int, heads: int,
                 head_channels: int, mlp_channels: int):
        super().__init__()
        self.layers = nn.ModuleList([])
        for _ in range(depth):
            self.layers.append(
                nn.ModuleList([
                    Attention(hid_channels,
                              heads=heads,
                              head_channels=head_channels),
                    FeedForward(hid_channels, mlp_channels)
                ]))

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        for attn, ff in self.layers:
            x = attn(x) + x
            x = ff(x) + x
        return x


class SimpleViT(nn.Module):
    r'''
    A Simple and Effective Vision Transformer (SimpleViT). The authors of Vision Transformer (ViT) present a few minor modifications and dramatically improve the performance of plain ViT models. For more details, please refer to the following information. 

    It is worth noting that this model is not designed for EEG analysis, but shows good performance and can serve as a good research start.

    - Paper: Beyer L, Zhai X, Kolesnikov A. Better plain ViT baselines for ImageNet-1k[J]. arXiv preprint arXiv:2205.01580, 2022.
    - URL: https://arxiv.org/abs/2205.01580
    - Related Project: https://github.com/lucidrains/vit-pytorch

    Below is a recommended suite for use in emotion recognition tasks:

    .. code-block:: python

        from torcheeg.datasets import DEAPDataset
        from torcheeg import transforms
        from torcheeg.models import SimpleViT
        from torch.utils.data import DataLoader

        dataset = DEAPDataset(io_path=f'./deap',
                    root_path='./data_preprocessed_python',
                    offline_transform=transforms.Compose([
                        transforms.MinMaxNormalize(axis=-1),
                        transforms.ToGrid(DEAP_CHANNEL_LOCATION_DICT)
                    ]),
                    online_transform=transforms.Compose([
                        transforms.ToTensor(),
                    ]),
                    label_transform=transforms.Compose([
                        transforms.Select('valence'),
                        transforms.Binary(5.0),
                    ]))

        model = SimpleViT(chunk_size=128,
                          grid_size=(9, 9),
                          t_patch_size=32,
                          num_classes=2)

        x, y = next(iter(DataLoader(dataset, batch_size=64)))
        model(x)

    It can also be used for the analysis of features such as DE, PSD, etc:

    .. code-block:: python

        from torcheeg.datasets import DEAPDataset
        from torcheeg import transforms
        from torcheeg.models import SimpleViT
        from torch.utils.data import DataLoader
        from torcheeg.datasets.constants import DEAP_CHANNEL_LOCATION_DICT


        dataset = DEAPDataset(root_path='./data_preprocessed_python',
                              offline_transform=transforms.Compose([
                                  transforms.BandDifferentialEntropy({
                                      "delta": [1, 4],
                                      "theta": [4, 8],
                                      "alpha": [8, 14],
                                      "beta": [14, 31],
                                      "gamma": [31, 49]
                                  }),
                                  transforms.ToGrid(DEAP_CHANNEL_LOCATION_DICT)
                              ]),
                              online_transform=transforms.Compose([
                                  transforms.ToTensor(),
                              ]),
                              label_transform=transforms.Compose([
                                  transforms.Select('valence'),
                                  transforms.Binary(5.0),
                              ]))

        model = SimpleViT(chunk_size=5,
                          grid_size=(9, 9),
                          t_patch_size=1,
                          num_classes=2)

        x, y = next(iter(DataLoader(dataset, batch_size=64)))
        model(x)

    Args:
        chunk_size (int): Number of data points included in each EEG chunk as training or test samples. (default: :obj:`128`)
        grid_size (tuple): Spatial dimensions of grid-like EEG representation. (default: :obj:`(9, 9)`)
        patch_size (tuple): The size (resolution) of each input patch. (default: :obj:`(3, 3)`)
        t_patch_size (int): The size of each input patch at the temporal (chunk size) dimension. (default: :obj:`32`)
        s_patch_size (tuple): The size (resolution) of each input patch at the spatial (grid size) dimension. (default: :obj:`(3, 3)`)
        hid_channels (int): The feature dimension of embeded patch. (default: :obj:`32`)
        depth (int): The number of attention layers for each transformer block. (default: :obj:`3`)
        heads (int): The number of attention heads for each attention layer. (default: :obj:`4`)
        head_channels (int): The dimension of each attention head for each attention layer. (default: :obj:`8`)
        mlp_channels (int): The number of hidden nodes in the fully connected layer of each transformer block. (default: :obj:`64`)
        num_classes (int): The number of classes to predict. (default: :obj:`2`)
    '''
    def __init__(self,
                 chunk_size: int = 128,
                 grid_size: Tuple[int, int] = (9, 9),
                 t_patch_size: int = 32,
                 s_patch_size: Tuple[int, int] = (3, 3),
                 hid_channels: int = 32,
                 depth: int = 3,
                 heads: int = 4,
                 head_channels: int = 8,
                 mlp_channels: int = 64,
                 num_classes: int = 2):
        super(SimpleViT, self).__init__()
        self.chunk_size = chunk_size
        self.grid_size = grid_size
        self.t_patch_size = t_patch_size
        self.s_patch_size = s_patch_size
        self.hid_channels = hid_channels
        self.depth = depth
        self.heads = heads
        self.head_channels = head_channels
        self.mlp_channels = mlp_channels
        self.num_classes = num_classes

        grid_height, grid_width = pair(grid_size)
        patch_height, patch_width = pair(s_patch_size)

        assert grid_height % patch_height == 0 and grid_width % patch_width == 0, f'EEG grid size {grid_size} must be divisible by the spatial patch size {s_patch_size}.'
        assert chunk_size % t_patch_size == 0, f'EEG chunk size {chunk_size} must be divisible by the temporal patch size {t_patch_size}.'

        patch_channels = t_patch_size * patch_height * patch_width

        self.to_patch_embedding = nn.Sequential(
            Rearrange('b (c p0) (h p1) (w p2) -> b c h w (p1 p2 p0)',
                      p0=t_patch_size,
                      p1=patch_height,
                      p2=patch_width),
            nn.Linear(patch_channels, hid_channels),
        )
        self.position_embedding = PositionEmbedding3D(hid_channels)

        self.transformer = Transformer(hid_channels, depth, heads,
                                       head_channels, mlp_channels)

        self.linear_head = nn.Sequential(nn.LayerNorm(hid_channels),
                                         nn.Linear(hid_channels, num_classes))

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        r'''
        Args:
            x (torch.Tensor): EEG signal representation, the ideal input shape is :obj:`[n, 128, 9, 9]`. Here, :obj:`n` corresponds to the batch size, :obj:`128` corresponds to :obj:`chunk_size`, and :obj:`(9, 9)` corresponds to :obj:`grid_size`.

        Returns:
            torch.Tensor[number of sample, number of classes]: the predicted probability that the samples belong to the classes.
        '''
        b, *_ = x.shape
        x = self.to_patch_embedding(x)
        pe = self.position_embedding(x)
        x = rearrange(x + pe, 'b ... d -> b (...) d')

        x = self.transformer(x)
        x = x.mean(dim=1)

        return self.linear_head(x)