MAHNOBDataset

class torcheeg.datasets.MAHNOBDataset(root_path: str = './Sessions', chunk_size: int = 128, sampling_rate: int = 128, overlap: int = 0, num_channel: int = 32, num_baseline: int = 30, baseline_chunk_size: int = 128, num_trial_sample: int = 30, online_transform: None | Callable = None, offline_transform: None | Callable = None, label_transform: None | Callable = None, before_trial: None | Callable = None, after_trial: None | Callable = None, after_session: None | Callable = None, after_subject: None | Callable = None, io_path: None | str = None, io_size: int = 1048576, io_mode: str = 'lmdb', num_worker: int = 0, verbose: bool = True)

MAHNOB-HCI is a multimodal database recorded in response to affective stimuli with the goal of emotion recognition and implicit tagging research. This class generates training samples and test samples according to the given parameters, and caches the generated results in a unified input and output format (IO). The relevant information of the dataset is as follows:

• Author: Soleymani et al.

• Year: 2011

• Download URL: https://mahnob-db.eu/hci-tagging/

• Reference: Soleymani M, Lichtenauer J, Pun T, et al. A multimodal database for affect recognition and implicit tagging[J]. IEEE transactions on affective computing, 2011, 3(1): 42-55.

• Stimulus: 20 videos from famous movies. Each video clip lasts 34-117 seconds (may not be an integer), in addition to 30 seconds before the beginning of the affective stimuli experience and another 30 seconds after the end.

• Signals: Electroencephalogram (32 channels at 512Hz), peripheral physiological signals (ECG, GSR, Temp, Resp at 256 Hz), and eye movement signals (at 60Hz) of 30-5=25 subjects (3 subjects with missing data records and 2 subjects with incomplete data records).

• Rating: Arousal, valence, control and predictability (all ona scale from 1 to 9).

In order to use this dataset, the download folder Sessions (Physiological files of emotion elicitation) is required, containing the following files:

• 1

  • Part_1_N_Trial1_emotion.bdf

  • session.xml

• …

• 3810

  • Part_30_S_Trial20_emotion.bdf

  • session.xml

An example dataset for CNN-based methods:

dataset = MAHNOBDataset(root_path='./Sessions',
                        offline_transform=transforms.Compose([
                            transforms.BandDifferentialEntropy(),
                            transforms.ToGrid(MAHNOB_CHANNEL_LOCATION_DICT)
                        ]),
                        online_transform=transforms.ToTensor(),
                        label_transform=transforms.Compose([
                            transforms.Select('feltVlnc'),
                            transforms.Binary(5.0),
                        ]))
print(dataset[0])
# EEG signal (torch.Tensor[4, 9, 9]),
# coresponding baseline signal (torch.Tensor[4, 9, 9]),
# label (int)

Another example dataset for CNN-based methods:

from torcheeg.datasets import MAHNOBDataset
from torcheeg import transforms

dataset = MAHNOBDataset(root_path='./Sessions',
                        online_transform=transforms.Compose([
                            transforms.To2d(),
                            transforms.ToTensor()
                        ]),
                        label_transform=transforms.Compose([
                            transforms.Select(['feltVlnc', 'feltArsl']),
                            transforms.Binary(5.0),
                            transforms.BinariesToCategory()
                        ]))
print(dataset[0])
# EEG signal (torch.Tensor[1, 32, 128]),
# coresponding baseline signal (torch.Tensor[1, 32, 128]),
# label (int)

An example dataset for GNN-based methods:

from torcheeg.datasets import MAHNOBDataset
from torcheeg import transforms
from torcheeg.datasets.constants.emotion_recognition.mahnob import MAHNOB_ADJACENCY_MATRIX
from torcheeg.transforms.pyg import ToG

dataset = MAHNOBDataset(root_path='./Sessions',
                        online_transform=transforms.Compose([
                            ToG(MAHNOB_ADJACENCY_MATRIX)
                        ]),
                        label_transform=transforms.Compose([
                            transforms.Select('feltArsl'),
                            transforms.Binary(5.0)
                        ]))
print(dataset[0])
# EEG signal (torch_geometric.data.Data),
# coresponding baseline signal (torch_geometric.data.Data),
# label (int)

Parameters:

• root_path (str) – Downloaded data files in bdf and xml (unzipped Sessions.zip) formats (default: './Sessions')

• chunk_size (int) – Number of data points included in each EEG chunk as training or test samples. If set to -1, the EEG signal of a trial is used as a sample of a chunk. (default: 128)

• sampling_rate (int) – The number of data points taken over a second. (default: 128)

• overlap (int) – The number of overlapping data points between different chunks when dividing EEG chunks. (default: 0)

• num_channel (int) – Number of channels used, of which the first 32 channels are EEG signals. (default: 32)

• num_baseline (int) – Number of baseline signal chunks used. (default: 30)

• baseline_chunk_size (int) – Number of data points included in each baseline signal chunk. The baseline signal in the MAHNOB dataset has a total of 512 (downsampled to 128) * 30 data points. (default: 128)

• num_trial_sample (int) – Number of samples picked from each trial. If set to -1, all samples in trials are used. (default: 30)

• online_transform (Callable, optional) – The transformation of the EEG signals and baseline EEG signals. The input is a np.ndarray, and the ouput is used as the first and second value of each element in the dataset. (default: None)

• offline_transform (Callable, optional) – The usage is the same as online_transform, but executed before generating IO intermediate results. (default: None)

• label_transform (Callable, optional) – The transformation of the label. The input is an information dictionary, and the ouput is used as the third value of each element in the dataset. (default: None)

• before_trial (Callable, optional) – The hook performed on the trial to which the sample belongs. It is performed before the offline transformation and thus typically used to implement context-dependent sample transformations, such as moving averages, etc. The input of this hook function is a 2D EEG signal with shape (number of electrodes, number of data points), whose ideal output shape is also (number of electrodes, number of data points).

• after_trial (Callable, optional) – The hook performed on the trial to which the sample belongs. It is performed after the offline transformation and thus typically used to implement context-dependent sample transformations, such as moving averages, etc. The input and output of this hook function should be a sequence of dictionaries representing a sequence of EEG samples. Each dictionary contains two key-value pairs, indexed by eeg (the EEG signal matrix) and key (the index in the database) respectively.

• io_path (str) – The path to generated unified data IO, cached as an intermediate result. If set to None, a random path will be generated. (default: None)

• io_size (int) – Maximum size database may grow to; used to size the memory mapping. If database grows larger than map_size, an exception will be raised and the user must close and reopen. (default: 1048576)

• io_mode (str) – Storage mode of EEG signal. When io_mode is set to lmdb, TorchEEG provides an efficient database (LMDB) for storing EEG signals. LMDB may not perform well on limited operating systems, where a file system based EEG signal storage is also provided. When io_mode is set to pickle, pickle-based persistence files are used. When io_mode is set to memory, memory are used. (default: lmdb)

• num_worker (int) – Number of subprocesses to use for data loading. 0 means that the data will be loaded in the main process. (default: 0)

• verbose (bool) – Whether to display logs during processing, such as progress bars, etc. (default: True)