ambitious-octopus / mi-eeg-1d-cnn Goto Github PK

A new approach based on a 10-layer one-dimensional convolution neural network (1D-CNN) to classify five brain states (four MI classes plus a 'baseline' class) using a data augmentation algorithm and a limited number of EEG channels. Paper: https://doi.org/10.1088/1741-2552/ac4430

License: GNU General Public License v3.0

Python 98.98% Dockerfile 1.02%

brain-computer-interface artificial-neural-networks neuroscience

mi-eeg-1d-cnn's People

Contributors

Stargazers

Watchers

Forkers

greeshma1987 yangqbo xjohnxjohn wartensie168 synizter hanjiaguan bugdaryan haoxiaodao ukaserge christosbouronikos igor-id jinbo4321 dawin2015 lkebir realzli yoorahi tuwenhao1

mi-eeg-1d-cnn's Issues

Test #2

Trained network, single subject, sliding windows [step 0.5 seconds]. Show Plot.

LSTM single-subject approach

Code

Generate data and understand how tf.data works
How to split train and test!
Clean the repo.

Theory:

Understand how the hell a LSTM works.

New strategy for real time classification

In order to achieve real-time classification, we are working on a new strategy:
Instead of creating a sliding window, the new strategy is to split the epoch into 2 (for now) and then train the net with a double number of examples. This implies no incoherent sample where there is more than one possible label.

Questions about the SavedModel

Hello, I am a college student. When I reproduce your code, I have not found a trained model. Where can I get the trained model?
Thank you very much

A problem with the path to manually download the dataset

Hello, I am a postgraduate student who just started studying BCI;
I tried running this code on my computer (Windows10, using anaconda to install python environment)；
I now have the Python environment installed, but I encountered the following error while performing this step (Generate the dataset):

FileNotFoundError: [WinError 3],The system could not find the specified path: './MI-EEG-1D-CNN-master/files/paper\FC1FC2'

I have downloaded and unpacked the dataset myself,I put this data set under the D:\CODE\MI-EEG-1D-CNN-master path.
I actually saw you talking about it in readme:Change the dataset path to the path of the dataset you downloaded, and you are ready!
I have tried several times and failed, so I come here for help. What should I do?
Sincerely hope to get your reply

2 second window, changelog

channgelog -> 08/02
learning_rate = 1e-4 -> 1e-3
Added callbacks, Early Stopping and Checkpoints
epochs = 150 -> 200

Prediction Time

Prediction time for the original HopefullNet with 4 second windows:
n = 10999

Mean -> 0.03883716804090982/s or 388.37168040909825/ms
STD -> 0.014576505798775771/s

Theoretical aspects, open discussion

First issue: Do we have a physical limit, in motor imagery?

Original HopefullNet Architecture

class HopefullNet(tf.keras.Model):
    """
    Original HopeFullNet
    """
    def __init__(self, inp_shape = (640,2)):
        super(HopefullNet, self).__init__()
        self.inp_shape = inp_shape

        self.kernel_size_0 = 20
        self.kernel_size_1 = 6
        self.drop_rate = 0.5

        self.conv1 = tf.keras.layers.Conv1D(filters=32,
                                            kernel_size=self.kernel_size_0,
                                            activation='relu',
                                            padding= "same",
                                            input_shape=self.inp_shape)
        self.batch_n_1 = tf.keras.layers.BatchNormalization()
        self.conv2 = tf.keras.layers.Conv1D(filters=32,
                                            kernel_size=self.kernel_size_0,
                                            activation='relu',
                                            padding= "valid")
        self.batch_n_2 = tf.keras.layers.BatchNormalization()
        self.spatial_drop_1 = tf.keras.layers.SpatialDropout1D(self.drop_rate)
        self.conv3 = tf.keras.layers.Conv1D(filters=32,
                                            kernel_size=self.kernel_size_1,
                                            activation='relu',
                                            padding= "valid")
        self.avg_pool1 = tf.keras.layers.AvgPool1D(pool_size=2)
        self.conv4 = tf.keras.layers.Conv1D(filters=32,
                                            kernel_size=self.kernel_size_1,
                                            activation='relu',
                                            padding= "valid")
        self.spatial_drop_2 = tf.keras.layers.SpatialDropout1D(self.drop_rate)
        self.flat = tf.keras.layers.Flatten()
        self.dense1 = tf.keras.layers.Dense(296, activation='relu')
        self.dropout1 = tf.keras.layers.Dropout(self.drop_rate)
        self.dense2 = tf.keras.layers.Dense(148, activation='relu')
        self.dropout2 = tf.keras.layers.Dropout(self.drop_rate)
        self.dense3 = tf.keras.layers.Dense(74, activation='relu')
        self.dropout3 = tf.keras.layers.Dropout(self.drop_rate)
        self.out = tf.keras.layers.Dense(5, activation='softmax')

To-Dos week April 26th

Generate for all subjects (for all runs) a database with window size 1/2 second and stride 1/2 and 1/4 second for all couples of specular channel in the corresponding hemisphere
Generate for all subjects (for all runs) a database with window size 2 second and stride 1/2 and 1/2 second for all couples of specular channel in the corresponding hemisphere

Question consultation

Dear authors:
Recently I have looking your this project and feel interested in it !could have a dataset of this project? Thanks a lot! I am looking forward to your reply Sincerely!

Best regards.

ICA on Imagery movement

ICA with template matching on runs [4,8,12] for all subjects.

ICA on Real Movement

ICA with template matching for runs [3,7,11]

Test

Test, Camillo suggests taking only the first 2 seconds of each stimulus.

Questions about the operation of script train_d.py and train_e.py

I'm sorry to bother you. I set up the required environment on a ubuntu18 computer a few days ago and also ran the tran_a,b,c scripts without any problems.

I use jupyter lab to view and run the code.train_a,b,c They all run smoothly as shown below:
They can train to complete.

When I run the scripts train_d and train_e, the script stops running when it prints out
" before oversampling=[ ] " (as if it had already finished).

Confusingly, there is no error reported either.

I am unable to solve this problem on my own for the time being and would appreciate your suggestions in your free time.

Additional note: I used pycharm to run these scripts in a windows 10 environment and they all worked fine.

Question about the train_test_spliter

Hi there, inspiring method and great paper
have been trying to apply your work on other dataset for couple of days, but cant achieve good results..
recheck the code, i think maybe it's a data-split problem.
for example, in this file, MI-EEG-1D-CNN/models/train_a.py, line 45

x is the loaded data, already shapes (events_num, 2, 640).
as we know, in one specific MI-task, different channel-couple in one ROI have similar behaviors,
in line 52, spliting reshape_x may split channel-couples in one task into train_set and test_set at same time, that maybe cause the acc rise not for the Model cause.

the data loading code of your work is a little bit hard for me to read, so i am trying to write my data loading function( humble one without base type event or SMOTE), which split data to train and test set first then reshaped it from (events_num, channels_num, 640) to (events_num, 2, 640) . then using HopefullNet to fit them, didn't end well.
i will paste my function below, after figure out how..

hope could get your respond, instruction about how to transfer HopefullNet to other dataset will be more than great.
best wishes

Installing the packages with the environment YML

Hi! I'm trying to replicate your results to also test & compare the model with some data I personally recorded.

I have a windows PC so I choose the second solution of cloning the repo and creating the environment through the YML file. But I run into some problems:

Firstly, most of the packages put into the YML file raise a PackageNotFound Error, this is strange bc for example one of those packages is matplotlib, this is the full Error:

ResolvePackageNotFound:                                                                                                   - sip==4.19.13=py38he6710b0_0                                                                                           - matplotlib==3.4.2=py38h06a4308_0                                                                                      - zstd==1.4.9=haebb681_0                                                                                                - libgfortran4==7.5.0=ha8ba4b0_17                                                                                       - lz4-c==1.9.3=h295c915_1                                                                                               - libxml2==2.9.12=h03d6c58_0                                                                                            - intel-openmp==2021.3.0=h06a4308_3350                                                                                  - libstdcxx-ng==9.3.0=hd4cf53a_17                                                                                       - mkl==2021.3.0=h06a4308_520                                                                                            - qt==5.9.7=h5867ecd_1                                                                                                  - gst-plugins-base==1.14.0=h8213a91_2                                                                                   - libpng==1.6.37=hbc83047_0                                                                                             - ca-certificates==2021.5.30=ha878542_0                                                                                 - pip==21.2.2=py38h06a4308_0                                                                                            - setuptools==58.0.4=py38h06a4308_0                                                                                     - fontconfig==2.13.1=h6c09931_0                                                                                         - ld_impl_linux-64==2.35.1=h7274673_9                                                                                   - libgomp==9.3.0=h5101ec6_17                                                                                            - xz==5.2.5=h7b6447c_0                                                                                                  - glib==2.69.1=h5202010_0                                                                                               - libgfortran-ng==7.5.0=ha8ba4b0_17                                                                                     - openssl==1.1.1l=h7f8727e_0                                                                                            - gstreamer==1.14.0=h28cd5cc_2                                                                                          - libuuid==1.0.3=h1bed415_2                                                                                             - libgcc-ng==9.3.0=h5101ec6_17                                                                                          - libtiff==4.2.0=h85742a9_0                                                                                             - readline==8.1=h27cfd23_0                                                                                              - brotli==1.0.9=he6710b0_2                                                                                              - mkl-service==2.4.0=py38h7f8727e_0                                                                                     - pyqt==5.9.2=py38h05f1152_4                                                                                            - dbus==1.13.18=hb2f20db_0                                                                                              - libwebp-base==1.2.0=h27cfd23_0                                                                                        - sqlite==3.36.0=hc218d9a_0                                                                                             - tornado==6.1=py38h27cfd23_0                                                                                           - pcre==8.45=h295c915_0                                                                                                 - libopenblas==0.3.13=h4367d64_0                                                                                        - matplotlib-base==3.4.2=py38hab158f2_0                                                                                 - freetype==2.10.4=h5ab3b9f_0                                                                                           - icu==58.2=he6710b0_3                                                                                                  - openjpeg==2.4.0=h3ad879b_0                                                                                            - zlib==1.2.11=h7b6447c_3                                                                                               - lcms2==2.12=h3be6417_0                                                                                                - certifi==2021.5.30=py38h578d9bd_0                                                                                     - libffi==3.3=he6710b0_2                                                                                                - ncurses==6.2=he6710b0_1                                                                                               - numpy-base==1.19.2=py38h75fe3a5_0                                                                                     - pillow==8.3.1=py38h2c7a002_0                                                                                          - kiwisolver==1.3.1=py38h2531618_0                                                                                      - _openmp_mutex==4.5=1_gnu                                                                                              - python==3.8.3=hcff3b4d_2                                                                                              - jpeg==9d=h7f8727e_0                                                                                                   - tk==8.6.11=h1ccaba5_0                                                                                                 - libxcb==1.14=h7b6447c_0                                                                                               - expat==2.4.1=h2531618_2                                                                                               - cudatoolkit==10.1.243=h6bb024c_0

(sorry, copying it from the command line must have breaking it into colums, but it was a print of each package under the next one)

Then I tried putting all packages under the pip line (I know this could raise some trouble in the future, but I was trying to al least make it run once, if that worked then I could have done it again but only putting under pip the ones Anaconda couldn't find -they are a lot though) but I got another Error:

Pip subprocess error:
ERROR: Invalid requirement: '_libgcc_mutex=0.1=main' (from line 1 of C:\Users\ottaaproject\Desktop\Lixi\MI GPU\MI-EEG-1D-CNN\docs\condaenv.amrlke0r.requirements.txt)
Hint: = is not a valid operator. Did you mean == ?
failed
CondaEnvException: Pip failed

On first glance I feel like I might be handling a different Anaconda version, I just updated to the latest release of anaconda but no changes happened. Hope you can help me solve this, I believe I could get a lot out of this comparison (and of course I can send you my results too!)

Thanks in advance. Greetings from Argentina!