„README.md“ ändern

README.md

@@ -1,20 +1,36 @@
-## Semesterproject of the lecture "Semesterproject Signal processing and Analysis of human brain potentials (eeg) WS 2020/21
+## Semesterproject of the lecture "Semesterproject Signal processing and Analysis of human brain potentials (eeg)" WS 2020/21
 
-This repository holds the code of the semesterproject as well as the report.
+This repository holds the code of the semesterproject as well as the report, created by Julius Voggesberger.
-The main files are 'preprocessing_and_cleaning.py', 'erp_analysis.py' and 'decoding_tf_analyis.py'.
+As the dataset for the project, the N170-dataset was chosen.
-The files hold:
+As the three subjects, to be manually pre-processed, the subjects 001, 003 and 014 were chosen.
-- preprocessing_and_cleaning.py : Holds the pre-processing pipeline of the project. By executing the file all subjects are pre-processed. Subjects 001, 003, 014 are pre-processed with manually selected pre-processing information, all other subjects are pre-processed with the given pre-processing information. Pre-processed cleaned data is saved in the BIDS file structure as 'sub-XXX_task-N170_cleaned.fif' where XXX is the subject number.
+The rest of the subjects were pre-processed with provided pre-processing information.
-Details can be found in the comments of the code.
-- erp_analysis.py : Holds the code for the erp-analysis. Computes the peak-differences and t-tests for several experimental contrasts. Details can be found in the comments of the code.
-- decoding_tf_analysis.py : Holds the code for the decoding and time-frequency analysis. Details can be found in the comments of the code.
 
-The folder 'utils' holds helper functions for some plots needed for the analysis and to load data, generate strings etc. and holds the code given in the lecture.
+### Structure
-The folder 'test' holds mostly unittests that test helper functions and one function which visually checks if N170 peaks are extracted correctly.
+```
+├── Dataset: The dataset of the project as well as the manually selected bad segments are stored here. 
+|   ├── n170: Store the dataset here. 
+|   └── preprocessed: Bad segments are stored here. 
+├── cached_data: Data that is generated in the analysis part is stored here. 
+|   ├── decoding_data: Results of the classifiers. 
+|   ├── erp_peaks: ERP peaks needed for the ERP analysis. 
+|   └── tf_data: Time-frequency data needed for the tf-analysis. 
+├── test: Contains unittests and one visual check. 
+├── utils: Contains helper methods 
+|   ├── ccs_eeg_semesterproject: Methods given in the lecture. 
+|   ├── ccs_eeg_utils_reduced: Method for reading in BIDS provided in the lecture. 
+|   ├── file_utils.py: Methods for reading in files and getting epochs. 
+|   └── plot_utils.py: Methods for manually created plots. 
+├── preprocessing_and_cleaning.py: The preprocessing pipeline. 
+├── erp_analysis.py: The ERP-Analysis and computation of ERP peaks.
+├── decoding_tf_analysis.py: Decoding and time-frequency analysis.
+└── semesterproject_report_voggesberger: The report of the project.
+```
 
-For the code to work properly, the N170 dataset needs to be provided.
+### Running the project
-When first running the analysis, it may take a while. After running it one time the data is cached, so that it can be reused if the analysis should be executed again. Be careful though, as a parameter has to be explicitly set in the code, so that the already computed data is used. This parameter is a boolean given to each analysis function which caches data.
+To run the project python 3.7 is required and anaconda recommended.\
-
+To ensure reproducability, randomstates were used for methods which are non-deterministic.
-This code was created using Python 3.7 and the following libraries:
+The randomstates used are either '123' or '1234'.\
+The following libraries are needed:
 - Matplotlib 3.3.3
 - MNE 0.22.0
 - MNE-Bids 0.6
@@ -22,3 +38,18 @@ This code was created using Python 3.7 and the following libraries:
 - Scikit-Learn 0.23.2
 - Pandas 1.2.0
 - Scipy 1.5.4
+
+For the code to work, the N170 dataset needs to be provided and put into the folder 'Dataset/n170/', so that the file structure 'Dataset/n170/sub-001', etc. exists.
+The pre-processed raw objects are saved in their respective subject folder, in 'Dataset/n170/'.
+When first running the analysis, it may take a while. 
+After running it one time the data is cached, so that it can be reused if the analysis should be executed again at a later time.
+For the cached data to be used, a boolean parameter has to be set in the respective analysis method.
+
+It may be necessary to set the parent directory 'semesterproject_lecture_eeg' as 'Sources Root' for the project, if pycharm is used as an IDE.
+
+### Parameters
+Parameters have to be changed manually in the code, if different settings want to be tried.
+
+### Visualisation
+The visualisation methods that were used to generate the visualisations in the report, are contained in the code, if they were created manually.
+If a visualisation method from mne was used to create the visualisation, it may exist in the code or not.

decoding_tf_analysis.py

@@ -183,7 +183,7 @@ def create_tfr(raw, condition, freqs, n_cycles, response='induced', baseline=Non
     return power
 
 
-def time_frequency(dataset, filename, compute_tfr=True):
+def time_frequency(dataset, filename, scaling='lin', compute_tfr=True):
     """
     Runs time frequency analysis
 
@@ -191,10 +191,13 @@ def time_frequency(dataset, filename, compute_tfr=True):
     :param filename: Filename of either the file from which the TFRs will be loaded
         or to which they will be saved
     :param compute_tfr: If True the TFRs will be created, else the TFRs will be loaded from a precomputed file
+    :param scaling: default 'lin' for linear scaling, else can be 'log' for logarithmic scaling
     """
     # Parameters
-    # freqs = np.linspace(0.1, 50, num=50) # Use this for linear space scaling
+    if scaling == 'lin':
-    freqs = np.logspace(*np.log10([0.1, 50]), num=50)
+        freqs = np.linspace(0.1, 50, num=50)  # Use this for linear space scaling
+    else:
+        freqs = np.logspace(*np.log10([0.1, 50]), num=50)
     n_cycles = freqs / 2
     cond1 = []
     cond2 = []
@@ -245,11 +248,12 @@ def time_frequency(dataset, filename, compute_tfr=True):
     F, clusters, cluster_p_values, h0 = mne.stats.permutation_cluster_test(
         [mne.grand_average(cond1).data, mne.grand_average(cond2).data], n_jobs=4, verbose='INFO',
         seed=123)
-    plot_tf_cluster(F, clusters, cluster_p_values, freqs, times)
+    plot_tf_cluster(F, clusters, cluster_p_values, freqs, times, scaling)
 
 
 if __name__ == '__main__':
     mne.set_log_level(verbose=VERBOSE_LEVEL)
     ds = 'N170'
     decoding(ds, 'faces_vs_cars', True)
-    time_frequency(ds, 'face_intact_vs_all_0.1_50hz_ncf2', True)
+    time_frequency(ds, 'face_intact_vs_all_0.1_50hz_ncf2', 'log', True)
+

utils/plot_utils.py

@@ -56,7 +56,7 @@ def plot_grand_average(dataset):
                                      linestyles=['solid', 'solid', 'dotted', 'dotted'])
 
 
-def plot_tf_cluster(F, clusters, cluster_p_values, freqs, times):
+def plot_tf_cluster(F, clusters, cluster_p_values, freqs, times, scaling='lin'):
     """
     Plot the F-Statistic values of permutation clusters with p-values <= 0.05 in color and > 0.05 in grey.
     Currently only works well for the linear scaling. For the logarithmic scaling a different x-axis has to be chosen
@@ -66,6 +66,7 @@ def plot_tf_cluster(F, clusters, cluster_p_values, freqs, times):
     :param cluster_p_values: p-values of the clusters
     :param freqs: frequency domain
     :param times: time domain
+    :param scaling: default 'lin' for linear scaling, else can be 'log' for logarithmic scaling
     """
     good_c = np.nan * np.ones_like(F)
     for clu, p_val in zip(clusters, cluster_p_values):
@@ -75,12 +76,19 @@ def plot_tf_cluster(F, clusters, cluster_p_values, freqs, times):
     bbox = [times[0], times[-1], freqs[0], freqs[-1]]
     plt.imshow(F, aspect='auto', origin='lower', cmap=cm.gray, extent=bbox, interpolation='None')
     a = plt.imshow(good_c, cmap=cm.RdBu_r, aspect='auto', origin='lower', extent=bbox, interpolation='None')
+
+    if scaling == 'log':
+        ticks = [1, 4, 8, 12, 14, 18, 22, 26, 30, 34, 38, 42, 46, 50]
+        labels = [round(freqs[i], 2) for i in range(len(freqs)) if i + 1 in ticks]
+        plt.yticks(ticks, labels)
+
     plt.colorbar(a)
     plt.xlabel('Time (s)')
     plt.ylabel('Frequency (Hz)')
     plt.show()
 
 
+
 def plot_oscillation_bands(condition):
     """
         Plot the oscillation bands for a given condition in the time from 130ms to 200ms