├── Baseline_Models.m
├── Concat_ANN.py
├── HDC_ANN.py
├── HDC_SNN.py
├── LICENSE
├── LSTM.py
├── README.md
├── config.py
├── create_figs_and_tables.m
├── create_train_test_split_MATLAB.py
├── data
    ├── README.md
    ├── motorway_dataset_window_64_proc_veh_DtA.pkl
    ├── secondary_dataset_window_64_proc_veh_DtA.pkl
    └── uah_dataset.mat
├── eval_baseline_models.m
├── getF1Score.m
├── main.py
├── model.py
├── requirements.txt
└── utils.py


/Baseline_Models.m:
--------------------------------------------------------------------------------
  1 | %% simple baseline predicition models
  2 | % scken, 2021
  3 | % Copyright (C) 2021 Chair of Automation Technology / TU Chemnitz
  4 | 
  5 | 
  6 | % parameter setup
  7 | dim = 576;
  8 | frac_scale = 6;
  9 | 
 10 | disp('----------------------------')
 11 | disp(['Baseline ' dataset])
 12 | disp(['Dim: ' num2str(dim)])
 13 | 
 14 | 
 15 | %% if dataset = full_crossval, create the training and test splits
 16 | 
 17 | if contains(dataset,'full_crossval')
 18 |     %%% HDC encoding with kNN classifier 
 19 |     
 20 |     % load the data with the python script 
 21 |     ret = system(['python3 create_train_test_split_MATLAB.py --dataset=' dataset ' --preproc=1 --input_dim=' num2str(dim) ' --scale=' num2str(frac_scale)]);
 22 |     if ret==0
 23 |         load('temp_data.mat')
 24 |     else
 25 |         disp('Data could not converted')
 26 |         return
 27 |     end
 28 |     delete('temp_data.mat')
 29 |     
 30 |     for i=1:size(X_train,2)
 31 |         %%%
 32 |         % HDC with k-NN
 33 | 
 34 |         % load data into item memory
 35 |         VSA = vsa_env('vsa','FHRR','dim',dim);
 36 |         VSA.add_vector('vec',X_train{i}','name',num2cell(num2str(Y_train{i})));
 37 | 
 38 |         % find k nearest neigbors 
 39 |         tic
 40 |         [~, l, s] = VSA.find_k_nearest(X_test{i}',3);
 41 |         pred = [];
 42 | 
 43 |         for c=1:size(l,2)
 44 |             temp = str2num(cell2mat(l(:,c)));
 45 |             pred(end+1) = mode(temp);
 46 |         end
 47 |         disp('Time for testing k-NN:')
 48 |         toc
 49 |         disp('Accuracy of HDC k-NN method: ')
 50 |         f1 = getF1Score(Y_test{i},pred);
 51 |         disp(f1)
 52 | 
 53 |     end
 54 | 
 55 |     %%%
 56 |     %  spectral features (FFT) with kNN
 57 | 
 58 |     ret = system(['python3 create_train_test_split_MATLAB.py --dataset=' dataset ' --preproc=0 --input_dim=' num2str(dim) ' --scale=' num2str(frac_scale)]);
 59 | 
 60 |     if ret==0
 61 |         load('temp_data.mat')
 62 |     else
 63 |         disp('Data could not converted')
 64 |         return
 65 |     end
 66 |     delete('temp_data.mat')
 67 | 
 68 |     for i=1:size(X_train,2)
 69 |         % fourier transformation
 70 |         X_train{i} = abs(fft(X_train{i},size(X_train{i},2),2));
 71 |         X_test{i} = abs(fft(X_test{i},size(X_test{i},2),2));
 72 | 
 73 |         % concat input
 74 |         X_train{i} = reshape(X_train{i},size(X_train{i},1),[]);
 75 |         X_test{i} = reshape(X_test{i},size(X_test{i},1),[]);
 76 | 
 77 |         Mdl = fitcknn(X_train{i},Y_train{i},'NumNeighbors',1,'Distance','Cityblock');
 78 | 
 79 |         % testing
 80 |         pred = predict(Mdl, X_test{i});
 81 | 
 82 |         disp('Accuracy of Spectral Features kNN method: ')
 83 |         f1 = getF1Score(Y_test{i},pred);
 84 |         disp(f1)
 85 | 
 86 |     end
 87 | end
 88 | 
 89 | %% HDC with SVM
 90 | 
 91 | ret = system(['python3 create_train_test_split_MATLAB.py --dataset=' dataset ' --preproc=1 --input_dim=' num2str(dim) ' --scale=' num2str(frac_scale)]);
 92 | 
 93 | if ret==0
 94 |     load('temp_data.mat')
 95 | else
 96 |     disp('Data could not converted')
 97 |     return
 98 | end
 99 | delete('temp_data.mat')
100 | 
101 | tic
102 | Mdl = fitcecoc(X_train,Y_train);
103 | disp('Time for training HDC-SVM:')
104 | toc
105 | 
106 | % testing
107 | tic
108 | pred = predict(Mdl, X_test);
109 | disp('Time for testing HDC-SVM:')
110 | toc
111 | f1 = getF1Score(Y_test,pred);
112 | disp('Accuracy of HDC SVM:')
113 | disp(f1)
114 | 
115 | % add result to table
116 | Result = table({'HDC-SVM'},f1,'VariableNames',{'Model','F1'});
117 | 
118 | %% HDC with k-NN
119 | 
120 | % load data into item memory
121 | VSA = vsa_env('vsa','FHRR','dim',dim);
122 | VSA.add_vector('vec',X_train','name',num2cell(num2str(Y_train)));
123 | 
124 | % find k nearest neigbors
125 | tic
126 | [~, l, s] = VSA.find_k_nearest(X_test',3);
127 | pred = [];
128 | 
129 | for c=1:size(l,2)
130 |     temp = str2num(cell2mat(l(:,c)));
131 |     pred(end+1) = mode(temp);
132 | end
133 | disp('Time for testing k-NN:')
134 | toc
135 | disp('Accuracy of HDC k-NN method: ')
136 | f1 = getF1Score(Y_test,pred);
137 | disp(f1)
138 | 
139 | % add to table
140 | % Result = table({'HDC-kNN'},acc,'VariableNames',{'Model','F1'});
141 | Result.Model{end+1} = 'HDC-kNN';
142 | Result.F1(end) = f1;
143 | 
144 | %% concat with SVM
145 | 
146 | ret = system(['python3 create_train_test_split_MATLAB.py --dataset=' dataset ' --preproc=0 --input_dim=' num2str(dim) ' --scale=' num2str(frac_scale)]);
147 | 
148 | if ret==0
149 |     load('temp_data.mat')
150 | else
151 |     disp('Data could not converted')
152 |     return
153 | end
154 | delete('temp_data.mat')
155 | 
156 | % concat input 
157 | X_train = reshape(X_train,size(X_train,1),[]);
158 | X_test = reshape(X_test,size(X_test,1),[]);
159 | 
160 | 
161 | Mdl = fitcecoc(X_train,Y_train);
162 | 
163 | % testing 
164 | pred = predict(Mdl, X_test);
165 | f1 = getF1Score(Y_test,pred);
166 | disp('Accuracy of Concat SVM method: ')
167 | disp(f1)
168 | 
169 | % add to table
170 | Result.Model{end+1} = 'Concat-SVM';
171 | Result.F1(end) = f1;
172 | 
173 | 
174 | %% concat with kNN 
175 | 
176 | % find optimal hyperparameter for concat model
177 | % rng(0)
178 | % Mdl_opt = fitcknn([X_train; X_test],[Y_train; Y_test],'OptimizeHyperparameters','auto',...
179 | %     'HyperparameterOptimizationOptions',...
180 | %     struct('AcquisitionFunctionName','expected-improvement-plus'))
181 | % 
182 | % Mdl = fitcknn(X_train,Y_train,'NumNeighbors',Mdl_opt.NumNeighbors,'Distance',Mdl_opt.Distance);
183 | 
184 | Mdl = fitcknn(X_train,Y_train,'NumNeighbors',3,'Distance','Cityblock');
185 | 
186 | % testing 
187 | pred = predict(Mdl, X_test);
188 | f1 = getF1Score(Y_test,pred);
189 | disp('Accuracy of Concat k-NN method: ')
190 | disp(f1)
191 | 
192 | % add to table
193 | Result.Model{end+1} = 'Concat-kNN';
194 | Result.F1(end) = f1;
195 | 
196 | %%  spectral features (FFT) with SVM 
197 | 
198 | ret = system(['python3 create_train_test_split_MATLAB.py --dataset=' dataset ' --preproc=0 --input_dim=' num2str(dim) ' --scale=' num2str(frac_scale)]);
199 | 
200 | if ret==0
201 |     load('temp_data.mat')
202 | else
203 |     disp('Data could not converted')
204 |     return
205 | end
206 | delete('temp_data.mat')
207 | 
208 | % fourier transformation
209 | X_train = abs(fft(X_train,size(X_train,2),2));
210 | X_test = abs(fft(X_test,size(X_test,2),2));
211 | 
212 | % concat input
213 | X_train = reshape(X_train,size(X_train,1),[]);
214 | X_test = reshape(X_test,size(X_test,1),[]);
215 | 
216 | tic
217 | % Mdl = fitcecoc(X_train,Y_train,'Learners',svm_template);
218 | Mdl = fitcecoc(X_train,Y_train);
219 | disp('Time for training SVM-Stat:')
220 | toc
221 | 
222 | % testing 
223 | tic
224 | pred = predict(Mdl, X_test);
225 | disp('Time for testing SVM-Stat:')
226 | toc 
227 | 
228 | disp('Accuracy of Spectral Features SVM method: ')
229 | f1 = getF1Score(Y_test,pred);
230 | disp(f1)
231 | 
232 | % add to table
233 | Result.Model{end+1} = 'Spect-SVM';
234 | Result.F1(end) = f1;
235 | 
236 | %%%
237 | %  spectral features (FFT) with kNN
238 | 
239 | ret = system(['python3 create_train_test_split_MATLAB.py --dataset=' dataset ' --preproc=0 --input_dim=' num2str(dim) ' --scale=' num2str(frac_scale)]);
240 | 
241 | if ret==0
242 |     load('temp_data.mat')
243 | else
244 |     disp('Data could not converted')
245 |     return
246 | end
247 | delete('temp_data.mat')
248 | 
249 | % fourier transformation
250 | X_train = abs(fft(X_train,size(X_train,2),2));
251 | X_test = abs(fft(X_test,size(X_test,2),2));
252 | 
253 | % concat input
254 | X_train = reshape(X_train,size(X_train,1),[]);
255 | X_test = reshape(X_test,size(X_test,1),[]);
256 | 
257 | 
258 | % find optimal hyperparameter for concat model
259 | % rng(0)
260 | % Mdl_opt = fitcknn([X_train; X_test],[Y_train; Y_test],'OptimizeHyperparameters','auto',...
261 | %     'HyperparameterOptimizationOptions',...
262 | %     struct('AcquisitionFunctionName','expected-improvement-plus'))
263 | % 
264 | % Mdl = fitcknn(X_train,Y_train,'NumNeighbors',Mdl_opt.NumNeighbors,'Distance',Mdl_opt.Distance);
265 | 
266 | Mdl = fitcknn(X_train,Y_train,'NumNeighbors',1,'Distance','Cityblock');
267 | 
268 | % testing 
269 | pred = predict(Mdl, X_test);
270 | 
271 | disp('Accuracy of Spectral Features kNN method: ')
272 | f1 = getF1Score(Y_test,pred);
273 | disp(f1)
274 | 
275 | % add to table
276 | Result.Model{end+1} = 'Spect-kNN';
277 | Result.F1(end) = f1;
278 | 
279 | %% print results
280 | 
281 | disp([dataset ' Dataset:'])
282 | disp(Result)
283 | 
284 | 


--------------------------------------------------------------------------------
/Concat_ANN.py:
--------------------------------------------------------------------------------
  1 | from __future__ import absolute_import
  2 | from __future__ import division
  3 | from __future__ import print_function
  4 | 
  5 | from config import Config
  6 | from utils import *
  7 | from model import HDC_ANN
  8 | from sklearn.metrics import classification_report, f1_score
  9 | from scipy.io import savemat, loadmat
 10 | from sklearn import metrics
 11 | from matplotlib import pyplot as plt
 12 | from tensorflow.keras.utils import to_categorical
 13 | from tensorflow.keras.callbacks import ModelCheckpoint
 14 | import logging
 15 | 
 16 | # config logger
 17 | logger = logging.getLogger('log')
 18 | 
 19 | def main_Concat_ANN(args):
 20 |     '''
 21 |     implementation of a network that uses the concatenate sequences of al variables
 22 |     - input size of the network is 64*9 (64 timesteps and 9 sensors) = 576
 23 |     - network is the same as in HDC approach
 24 |     '''
 25 |     config = Config()
 26 |     config.training_volume = args.training_volume
 27 |     config.input_dim = args.input_dim
 28 |     config.encoding_dim = args.encoding_dim
 29 |     if args.runtime_measurement:
 30 |         config.n_time_measures = 10
 31 |     else:
 32 |         config.n_time_measures = 1
 33 | 
 34 |     # load preprocessed data
 35 |     data = load_dataset(args.dataset,config)
 36 |     X_train = data[0]
 37 |     X_test = data[1]
 38 |     y_train = data[2]
 39 |     y_test = data[3]
 40 |     config = data[4]
 41 | 
 42 |     # if train test data not a list, create one
 43 |     if type(X_train)==list:
 44 |         print("given data is not a list")
 45 |         X_train_list = X_train
 46 |         X_test_list = X_test
 47 |         y_train_list = y_train
 48 |         y_test_list = y_test
 49 |     else:
 50 |         X_train_list =[X_train]
 51 |         X_test_list = [X_test]
 52 |         y_train_list = [y_train]
 53 |         y_test_list = [y_test]
 54 | 
 55 |     #######################################################################################
 56 |     # statistical iteration
 57 |     #######################################################################################
 58 |     acc_mean = []
 59 |     f1_mean = []
 60 | 
 61 |     for stat_it in range(args.stat_iterations):
 62 |         logger.info('Statistial iteration: ' + str(stat_it))
 63 | 
 64 |         # train for each element in list (that is why we need list form, even if it contains only one element)
 65 |         logger.info('Training data contains ' + str(len(X_train)) + ' training instances...')
 66 |         scores = []
 67 |         accs = []
 68 |         for it in range(len(X_train_list)):
 69 |             logger.info(('.......'))
 70 |             logger.info('instance ' + str(it) + ':')
 71 | 
 72 |             X_train = X_train_list[it]
 73 |             X_test = X_test_list[it]
 74 |             y_train = y_train_list[it]
 75 |             y_test = y_test_list[it]
 76 | 
 77 |             # use only fraction of training samples (if given)
 78 |             X_train = X_train[1:int(X_train.shape[0] * config.training_volume), :]
 79 |             y_train = y_train[1:int(y_train.shape[0] * config.training_volume), :]
 80 | 
 81 |             # concatenate the input data
 82 |             X_train = np.reshape(X_train, (X_train.shape[0], -1))
 83 |             X_test = np.reshape(X_test, (X_test.shape[0], -1))
 84 | 
 85 |             config.input_dim = X_train.shape[1]
 86 | 
 87 |             logger.info('Training dataset shape: ' + str(X_train.shape) + str(y_train.shape))
 88 |             logger.info('Test dataset shape: ' + str(X_test.shape) + str(y_test.shape))
 89 | 
 90 |             config.n_classes = len(np.unique(y_train))
 91 | 
 92 |             #######################################################################################
 93 |             # keras model training
 94 |             #######################################################################################
 95 | 
 96 |             model = HDC_ANN(config)
 97 |             model.summary()
 98 | 
 99 |             cb_time = TimingCallback()
100 |             weight_fn = "./weights/Concat_ANN/%s_weights.h5" % args.dataset
101 |             if not os.path.exists(weight_fn.rsplit('/', 1)[0]):
102 |                 os.makedirs(weight_fn.rsplit('/', 1)[0])
103 |             model_checkpoint = ModelCheckpoint(weight_fn, verbose=1, mode='auto',
104 |                                                monitor='loss', save_best_only=True, save_weights_only=True)
105 | 
106 |             # compile model
107 |             model.compile(optimizer='adam',
108 |                           loss='categorical_crossentropy',
109 |                           metrics=['accuracy'])
110 |             history = model.fit(X_train, to_categorical(y_train),
111 |                                 epochs=config.training_epochs,
112 |                                 batch_size=config.batch_size,
113 |                                 shuffle=True,
114 |                                 callbacks=[cb_time, model_checkpoint],
115 |                                 validation_data=(X_test, to_categorical(y_test)))
116 | 
117 |             # log training time
118 |             epoch_time = cb_time.logs
119 |             mean_epoch_time = np.mean(epoch_time)
120 |             overall_time = np.sum(epoch_time)
121 |             logger.info("Mean Epoch time: " + str(mean_epoch_time))
122 |             logger.info("overall training time: " + str(overall_time))
123 | 
124 |             # load the best model weights
125 |             model.load_weights(weight_fn)
126 | 
127 |             #############################################################################################
128 |             # evaluation of results
129 |             #############################################################################################
130 | 
131 |             # evaluate and print results
132 |             pred_test = model.predict(X_test)
133 |             pred_test_bool = np.argmax(pred_test, axis=1)
134 | 
135 |             logger.info('Accuracy on training data: ')
136 |             report = classification_report(y_test.astype(int), pred_test_bool, output_dict=True)
137 |             logger.info(classification_report(y_test.astype(int), pred_test_bool))
138 | 
139 |             accs.append((report['accuracy']))
140 | 
141 |             logger.info("Confusion matrix:")
142 |             confusion_matrix = metrics.confusion_matrix(y_test.astype(int), pred_test_bool)
143 |             logger.info(confusion_matrix)
144 | 
145 |             # f1 score
146 |             f1 = f1_score(y_test.astype(int), pred_test_bool, average='weighted')
147 |             scores.append(f1)
148 |             logger.info("F1 Score: " + str(f1))
149 | 
150 |             # save as mat files
151 |             save_dic = {"report": report, "confusion_matrix": confusion_matrix, "config": config, "pred": pred_test,
152 |                         "label": y_test, "f1": f1}
153 |             savemat("results/" + args.dataset + "/results_concatNet" + str(config.encoding_dim) + '_' +
154 |                     str(config.training_volume) + ".mat", save_dic)
155 | 
156 |             plt.plot(history.history['accuracy'])
157 |             plt.plot(history.history['val_accuracy'])
158 |             plt.title('model accuracy')
159 |             plt.ylabel('accuracy')
160 |             plt.xlabel('epoch')
161 |             plt.legend(['train', 'val'], loc='upper left')
162 |             plt.savefig("results/" + args.dataset + "/accuracy_history_concatNet" + str(config.encoding_dim) + '_' +
163 |                         str(config.training_volume) + "instance_" + str(it) + ".png")
164 |             plt.show()
165 | 
166 |             plt.plot(history.history['loss'])
167 |             plt.plot(history.history['val_loss'])
168 |             plt.title('model loss')
169 |             plt.ylabel('loss')
170 |             plt.xlabel('epoch')
171 |             plt.legend(['train', 'val'], loc='upper left')
172 |             plt.savefig("results/" + args.dataset + "/loss_history_concatNet" + str(config.encoding_dim) + '_' +
173 |                         str(config.training_volume) + "instance_" + str(it) + ".png")
174 |             plt.show()
175 | 
176 |         # add results to statistical result array
177 |         acc_mean.append(np.mean(accs))
178 |         f1_mean.append(np.mean(scores))
179 | 
180 |     logger.info('Accuracy results of statistical repetitions: ' + str(acc_mean))
181 |     logger.info('F1 scores of statistical repetitions: ' + str(f1_mean))
182 | 
183 |     # write all scores to extra file
184 |     logger.info('Mean Score: ' + str(np.mean(f1_mean)))
185 |     logger.info('Mean Accuracy: ' + str(np.mean(acc_mean)))
186 |     with open("results/results_" + args.dataset + "_Concat.txt", 'a') as file:
187 |         file.write(str(config.input_dim) + '\t'
188 |                    + str(config.encoding_dim) + '\t'
189 |                    + str(args.stat_iterations) + '\t'
190 |                    + str(round(np.mean(f1_mean), 3)) + '\t'
191 |                    + str(round(np.mean(acc_mean), 3)) + '\t'
192 |                    + str(round(np.std(f1_mean), 3)) + '\t'
193 |                    + str(round(np.std(acc_mean), 3)) + '\n'
194 |                    )


--------------------------------------------------------------------------------
/HDC_ANN.py:
--------------------------------------------------------------------------------
  1 | from __future__ import absolute_import
  2 | from __future__ import division
  3 | from __future__ import print_function
  4 | 
  5 | from config import Config
  6 | from utils import *
  7 | from model import HDC_ANN, HDC_ANN_tf
  8 | from sklearn.metrics import classification_report, f1_score
  9 | from sklearn.preprocessing import LabelEncoder
 10 | from scipy.io import savemat, loadmat
 11 | from sklearn import metrics
 12 | from matplotlib import pyplot as plt
 13 | from datetime import datetime
 14 | from time import time
 15 | from tensorflow.keras.utils import to_categorical
 16 | from tensorflow.keras.callbacks import ModelCheckpoint, ReduceLROnPlateau
 17 | import logging
 18 | 
 19 | # config logger
 20 | logger = logging.getLogger('log')
 21 | 
 22 | tf.compat.v1.disable_eager_execution()
 23 | 
 24 | def main_HDC(args):
 25 |     '''
 26 |     implementation of the HDC feed-forward network to predict the class of driving style
 27 |     - it uses preprocessed HDC encoding vectors
 28 |     '''
 29 |     # set config parameter
 30 |     config = Config()
 31 |     config.training_volume = args.training_volume
 32 |     config.input_dim = args.input_dim
 33 |     config.encoding_dim = args.encoding_dim
 34 |     config.scale = args.scale
 35 |     config.m = 0
 36 |     config.s = 1
 37 |     if args.runtime_measurement:
 38 |         config.n_time_measures = 10
 39 |     else:
 40 |         config.n_time_measures = 1
 41 |     # if dimension is smaller than 1000, set dropout to 0.5
 42 |     if args.input_dim<1000:
 43 |         config.dropout = 0.5
 44 | 
 45 |     # load data set
 46 |     data = load_dataset(args.dataset,config)
 47 |     X_train = data[0]
 48 |     X_test = data[1]
 49 |     y_train = data[2]
 50 |     y_test = data[3]
 51 |     config = data[4]
 52 | 
 53 |     # if train test data not a list, create one (full_crossval data set loading returns a list of splits - therefore we
 54 |     # are handling all training set as lists, even if they only contain one set)
 55 |     if type(X_train)==list:
 56 |         print("given data is not a list")
 57 |         X_train_list = X_train
 58 |         X_test_list = X_test
 59 |         y_train_list = y_train
 60 |         y_test_list = y_test
 61 |     else:
 62 |         X_train_list =[X_train]
 63 |         X_test_list = [X_test]
 64 |         y_train_list = [y_train]
 65 |         y_test_list = [y_test]
 66 | 
 67 |     #######################################################################################
 68 |     # statistical iteration
 69 |     #######################################################################################
 70 |     acc_mean = []
 71 |     f1_mean = []
 72 | 
 73 |     for stat_it in range(args.stat_iterations):
 74 |         logger.info('Statistial iteration: ' + str(stat_it))
 75 | 
 76 |         # train for each element in list (that is why we need list form, even if it contains only one element)
 77 |         logger.info('Training data contains ' + str(len(X_train_list)) + ' training instances...')
 78 |         scores = []
 79 |         accs = []
 80 |         for it in range(len(X_train_list)):
 81 |             logger.info(('.......'))
 82 |             logger.info('instance ' + str(it) + ':')
 83 | 
 84 |             X_train = X_train_list[it]
 85 |             X_test = X_test_list[it]
 86 |             y_train = y_train_list[it]
 87 |             y_test = y_test_list[it]
 88 | 
 89 |             # use only fraction of training samples (if given)
 90 |             X_train = X_train[1:int(X_train.shape[0] * config.training_volume), :]
 91 |             y_train = y_train[1:int(y_train.shape[0] * config.training_volume)]
 92 | 
 93 |             logger.info('Training dataset shape: ' + str(X_train.shape) + str(y_train.shape))
 94 |             logger.info('Test dataset shape: ' + str(X_test.shape) + str(y_test.shape))
 95 | 
 96 |             config.n_classes = len(np.unique(y_train))
 97 |             config.n_inputs = X_train.shape[2]
 98 |             config.n_steps = X_train.shape[1]
 99 |             config.train_count = len(X_train)
100 |             config.test_data_count = len(X_test)
101 | 
102 |             #######################################################################################
103 |             # create HDC vectors (encoding)
104 |             #######################################################################################
105 |             tf.compat.v1.disable_eager_execution()
106 |             # create HDC vectors
107 |             t_train, X_train_HDC, traces_train, init_vecs = create_HDC_vectors(config, X_train)
108 |             t_test, X_test_HDC, traces_test, init_vecs = create_HDC_vectors(config, X_test)
109 |             preprocessing_time = t_train+t_test
110 | 
111 |             # normalize HDC encodings
112 |             m = np.mean(X_train_HDC, axis=0)
113 |             s = np.std(X_train_HDC,axis=0)
114 |             config.m = m
115 |             config.s = s
116 |             X_train_HDC = np.divide(X_train_HDC - m,s)
117 |             X_test_HDC = np.divide(X_test_HDC - m,s)
118 | 
119 |             #######################################################################################
120 |             # keras model training
121 |             #######################################################################################
122 |             model = HDC_ANN(config)
123 |             model.summary()
124 | 
125 |             # Create a TensorBoard callback
126 |             logs = "logs/HDC_ts_" + datetime.now().strftime("%Y%m%d-%H%M%S")
127 |             logs_pre = "logs/HDC_ts_preproc_" + datetime.now().strftime("%Y%m%d-%H%M%S")
128 | 
129 |             cb_time = TimingCallback()
130 |             weight_fn = "./weights/HDC_ANN/%s_weights.h5" % args.dataset
131 |             if not os.path.exists(weight_fn.rsplit('/', 1)[0]):
132 |                 os.makedirs(weight_fn.rsplit('/', 1)[0])
133 |             model_checkpoint = ModelCheckpoint(weight_fn, verbose=1, mode='auto',
134 |                                                monitor='loss', save_best_only=True, save_weights_only=True)
135 |             # compile model
136 |             model.compile(optimizer='adam',
137 |                           loss='categorical_crossentropy',
138 |                           metrics=['accuracy'])
139 | 
140 |             if not args.test:
141 |                 # train model
142 |                 history = model.fit(X_train_HDC, to_categorical(y_train),
143 |                                     epochs=config.training_epochs,
144 |                                     batch_size=config.batch_size,
145 |                                     shuffle=True,
146 |                                     callbacks=[cb_time, model_checkpoint],
147 |                                     validation_data=(X_test_HDC, to_categorical(y_test)))
148 | 
149 |                 # log training time
150 |                 epoch_time = cb_time.logs
151 |                 mean_epoch_time = np.mean(epoch_time)
152 |                 overall_time = np.sum(epoch_time)
153 |                 logger.info("Mean Epoch time: " + str(mean_epoch_time))
154 |                 logger.info("overall training time: " + str(overall_time))
155 | 
156 |                 plt.plot(history.history['accuracy'])
157 |                 plt.plot(history.history['val_accuracy'])
158 |                 plt.title('model accuracy')
159 |                 plt.ylabel('accuracy')
160 |                 plt.xlabel('epoch')
161 |                 plt.legend(['train', 'val'], loc='upper left')
162 |                 plt.savefig("results/" + args.dataset + "/accuracy_history_HDC_" + str(config.input_dim) + "_" +
163 |                             str(config.scale) + '_' + str(config.encoding_dim) + '_' + str(config.training_volume) + "instance_" + str(it) + ".png")
164 |                 plt.show()
165 | 
166 |                 plt.plot(history.history['loss'])
167 |                 plt.plot(history.history['val_loss'])
168 |                 plt.title('model loss')
169 |                 plt.ylabel('loss')
170 |                 plt.xlabel('epoch')
171 |                 plt.legend(['train', 'val'], loc='upper left')
172 |                 plt.savefig("results/" + args.dataset + "/loss_history_HDC" + str(config.input_dim) + "_" +
173 |                             str(config.scale) + '_' + str(config.encoding_dim) + '_' + str(config.training_volume) + "instance_" + str(it) + ".png")
174 |                 plt.show()
175 | 
176 |             # load the best model weights
177 |             model.load_weights(weight_fn)
178 | 
179 |             #############################################################################################
180 |             # evaluation of results
181 |             #############################################################################################
182 |             # evaluate with tensorflow model (better comparability to LTSM TF model)
183 |             X = tf.compat.v1.placeholder(tf.float32, [None, config.n_steps, config.n_inputs], name="X")
184 |             # get weights of keras  model
185 |             weights = model.get_weights()
186 |             W = {'hidden': weights[0], 'output': weights[2]}
187 |             biases = {'hidden': weights[1], 'output': weights[3]}
188 |             # create TF model
189 |             tf_model = HDC_ANN_tf(X, config, init_vecs, W, biases)
190 |             t_i=[]
191 |             for i in range(config.n_time_measures):
192 |                 sess = tf.compat.v1.Session()
193 |                 t1 = time()
194 |                 pred_test = sess.run(tf_model, feed_dict={X: X_test})
195 |                 inference_time = time() - t1
196 |                 t_i.append(inference_time)
197 |             inference_time = np.median(t_i)
198 | 
199 |             logger.info("Preprocessing time for training: " + str(t_train))
200 |             logger.info("Inference time: " + str(inference_time))
201 |             logger.info("Inference time one sequence [ms]: " + str((inference_time*1000)/X_test.shape[0]))
202 | 
203 |             logger.info('Preprocessing time on dataset: ' + str(t_train) + ' + ' + str(t_test) + ' = ' + str(
204 |                 preprocessing_time))
205 |             pred_test_bool = np.argmax(pred_test, axis=1)
206 | 
207 |             logger.info('Accuracy on training data: ')
208 |             report = classification_report(y_test.astype(int), pred_test_bool, output_dict=True)
209 |             logger.info(classification_report(y_test.astype(int), pred_test_bool))
210 | 
211 |             accs.append((report['accuracy']))
212 | 
213 |             logger.info("Confusion matrix:")
214 |             confusion_matrix = metrics.confusion_matrix(y_test.astype(int), pred_test_bool)
215 |             logger.info(confusion_matrix)
216 | 
217 |             # f1 score
218 |             f1 = f1_score(y_test.astype(int), pred_test_bool, average='weighted')
219 |             scores.append(f1)
220 |             logger.info("F1 Score: " + str(f1))
221 | 
222 |         # add results to statistical result array
223 |         acc_mean.append(np.mean(accs))
224 |         f1_mean.append(np.mean(scores))
225 | 
226 |     # save as mat files
227 |     save_dic = {"report": report, "confusion_matrix": confusion_matrix, "config": config, "pred": pred_test,
228 |                 "label": y_test, "f1_mean": np.mean(f1_mean), "acc_mean": np.mean(acc_mean)}
229 |     savemat("results/" + args.dataset + "/results_HDC_" + str(config.input_dim) + "_" + str(config.scale) + "_" +
230 |             str(config.encoding_dim) + '_' + str(config.training_volume) + ".mat", save_dic)
231 | 
232 |     logger.info('Accuracy results of statistical repetitions: ' + str(acc_mean))
233 |     logger.info('F1 scores of statistical repetitions: ' + str(f1_mean))
234 | 
235 |     # write all scores to extra file
236 |     logger.info('Mean Score: ' + str(np.mean(f1_mean)))
237 |     logger.info('Mean Accuracy: ' + str(np.mean(acc_mean)))
238 |     with open("results/results_" + args.dataset + "_HDC.txt", 'a') as file:
239 |         file.write(str(config.input_dim) + '\t'
240 |                    + str(config.encoding_dim) + '\t'
241 |                    + str(config.scale) + '\t'
242 |                    + str(args.stat_iterations) + '\t'
243 |                    + str(round(np.mean(f1_mean),3)) + '\t'
244 |                    + str(round(np.mean(acc_mean),3)) + '\t'
245 |                    + str(round(np.std(f1_mean),3)) + '\t'
246 |                    + str(round(np.std(acc_mean),3)) + '\t'
247 |                    + str(args.training_volume) + '\n'
248 |                    )
249 | 
250 | 
251 | 
252 | 
253 | 
254 | 


--------------------------------------------------------------------------------
/HDC_SNN.py:
--------------------------------------------------------------------------------
  1 | from __future__ import absolute_import
  2 | from __future__ import division
  3 | from __future__ import print_function
  4 | 
  5 | from sklearn import metrics
  6 | from config import Config
  7 | from utils import *
  8 | from sklearn.metrics import classification_report
  9 | from scipy.io import savemat
 10 | from sklearn import metrics
 11 | import logging
 12 | from time import time
 13 | import sklearn
 14 | import nengo_dl
 15 | import nengo
 16 | 
 17 | physical_devices = tf.config.list_physical_devices('GPU')
 18 | try:
 19 |   tf.config.experimental.set_memory_growth(physical_devices[0], True)
 20 | except:
 21 |   # Invalid device or cannot modify virtual devices once initialized.
 22 |   pass
 23 | 
 24 | # config logger
 25 | logger = logging.getLogger('log')
 26 | 
 27 | tf.compat.v1.enable_eager_execution()
 28 | 
 29 | def main_SNN(args):
 30 |     config = Config()
 31 |     config.training_volume = args.training_volume
 32 |     config.input_dim = args.input_dim
 33 |     config.encoding_dim = args.encoding_dim
 34 |     config.scale = args.scale
 35 |     if args.runtime_measurement:
 36 |         config.n_time_measures = 10
 37 |     else:
 38 |         config.n_time_measures = 1
 39 | 
 40 |     # nego net params
 41 |     do_rate = 0.5
 42 |     num_epochs = 200
 43 |     enc_dim = 1000
 44 |     minibatch_size = 500
 45 |     seed = 0
 46 | 
 47 |     # load dataset
 48 |     data = load_dataset(args.dataset,config)
 49 |     X_train = data[0]
 50 |     X_test = data[1]
 51 |     y_train = data[2]
 52 |     y_test = data[3]
 53 |     config = data[4]
 54 | 
 55 |     # if train test data not a list, create one
 56 |     if type(X_train)==list:
 57 |         print("given data is not a list")
 58 |         X_train_list = X_train
 59 |         X_test_list = X_test
 60 |         y_train_list = y_train
 61 |         y_test_list = y_test
 62 |     else:
 63 |         X_train_list =[X_train]
 64 |         X_test_list = [X_test]
 65 |         y_train_list = [y_train]
 66 |         y_test_list = [y_test]
 67 | 
 68 |     #######################################################################################
 69 |     # statistical iteration
 70 |     #######################################################################################
 71 |     acc_mean = []
 72 |     f1_mean = []
 73 | 
 74 |     for stat_it in range(args.stat_iterations):
 75 |         logger.info('Statistial iteration: ' + str(stat_it))
 76 |         seed = stat_it
 77 | 
 78 |         # train for each element in list (that is why we need list form, even if it contains only one element)
 79 |         logger.info('Training data contains ' + str(len(X_train)) + ' training instances...')
 80 |         scores = []
 81 |         accs = []
 82 |         for it in range(len(X_train_list)):
 83 |             logger.info(('.......'))
 84 |             logger.info('instance ' + str(it) + ':')
 85 | 
 86 |             X_train = X_train_list[it]
 87 |             X_test = X_test_list[it]
 88 |             y_train = y_train_list[it]
 89 |             y_test = y_test_list[it]
 90 | 
 91 |             # use only fraction of training samples (if given)
 92 |             X_train = X_train[1:int(X_train.shape[0] * config.training_volume), :,:]
 93 |             y_train = y_train[1:int(y_train.shape[0] * config.training_volume)]
 94 | 
 95 |             y_train_oh = one_hot(y_train)
 96 |             y_test_oh = one_hot(y_test)
 97 | 
 98 |             # config.input_dim = X_train.shape[1]
 99 |             logger.info('Training dataset shape: ' + str(X_train.shape) + str(y_train.shape))
100 |             logger.info('Test dataset shape: ' + str(X_test.shape) + str(y_test.shape))
101 | 
102 |             config.n_classes = len(np.unique(y_train))
103 |             config.n_inputs = X_train.shape[2]
104 |             config.n_steps = X_train.shape[1]
105 | 
106 |             #######################################################################################
107 |             # create HDC vectors (encoding)
108 |             #######################################################################################
109 |             # create HDC vectors
110 |             t_train, X_train, traces_train, init_vecs = create_HDC_vectors(config, X_train)
111 |             t_test, X_test, traces_test, init_vecs = create_HDC_vectors(config, X_test)
112 |             preprocessing_time = t_train + t_test
113 | 
114 |             # normalize HDC encodings
115 |             m = np.mean(X_train, axis=0)
116 |             s = np.std(X_train,axis=0)
117 |             config.m = m
118 |             config.s = s
119 |             X_train = np.divide(X_train - m,s)
120 |             X_test = np.divide(X_test - m,s)
121 | 
122 |             #######################################################################################
123 |             # nengo model training
124 |             #######################################################################################
125 | 
126 |             net = nengo.Network(seed=seed + 1)
127 | 
128 |             with net:
129 |                 # set some default parameters for the neurons that will make
130 |                 # the training progress more smoothly
131 |                 net.config[nengo.Ensemble].max_rates = nengo.dists.Choice([100])
132 |                 net.config[nengo.Ensemble].intercepts = nengo.dists.Choice([0])
133 |                 net.config[nengo.Connection].synapse = None
134 |                 neuron_type = nengo.LIF(amplitude=0.01)
135 | 
136 |                 # this is an optimization to improve the training speed,
137 |                 # since we won't require stateful behaviour in this example
138 |                 nengo_dl.configure_settings(stateful=False)
139 | 
140 |                 # the input node that will be used to feed in input vectors
141 |                 inp = nengo.Node(np.zeros(config.input_dim))
142 | 
143 |                 x = nengo_dl.Layer(tf.keras.layers.Dropout(rate=do_rate))(inp)
144 |                 x = nengo_dl.Layer(neuron_type)(x)
145 | 
146 |                 x = nengo_dl.Layer(tf.keras.layers.Dense(units=enc_dim))(x)
147 |                 x = nengo_dl.Layer(neuron_type)(x)
148 | 
149 |                 out = nengo_dl.Layer(tf.keras.layers.Dense(units=len(y_train_oh[0])))(x)
150 | 
151 |                 # we'll create two different output probes, one with a filter
152 |                 # (for when we're simulating the network over time and
153 |                 # accumulating spikes), and one without (for when we're
154 |                 # training the network using a rate-based approximation)
155 |                 out_p = nengo.Probe(out, label="out_p")
156 |                 out_p_filt = nengo.Probe(out, synapse=0.1, label="out_p_filt")
157 | 
158 |             sim = nengo_dl.Simulator(net, minibatch_size=minibatch_size, device="/gpu:0")
159 | 
160 |             # run training
161 |             sim.compile(
162 |                 optimizer=tf.optimizers.RMSprop(0.001),
163 |                 loss={out_p: tf.losses.CategoricalCrossentropy(from_logits=True)},
164 |             )
165 | 
166 |             # add single timestep to training data
167 |             X_train = X_train[:, None, :]
168 |             y_train_oh = y_train_oh[:, None]
169 | 
170 |             # when testing our network with spiking neurons we will need to run it
171 |             # over time, so we repeat the input/target data for a number of
172 |             # timesteps.
173 |             n_steps = 30
174 |             X_test = np.tile(X_test[:, None, :], (1, n_steps, 1))
175 |             y_test_oh = np.tile(y_test_oh[:, None], (n_steps, 1))
176 | 
177 |             def classification_accuracy(y_true, y_pred):
178 |                 return tf.metrics.categorical_accuracy(y_true[:, -1], y_pred[:, -1])
179 | 
180 |             accuracy = sim.evaluate(X_test, {out_p_filt: y_test_oh}, verbose=0)["loss"],
181 |             print("Accuracy before training:", accuracy)
182 | 
183 |             cb_time = TimingCallback()
184 |             sim.fit(X_train,
185 |                     {out_p: y_train_oh},
186 |                     epochs=num_epochs,
187 |                     callbacks=[cb_time],
188 |                     )
189 |             # log training time
190 |             epoch_time = cb_time.logs
191 |             mean_epoch_time = np.mean(epoch_time)
192 |             training_time = np.sum(epoch_time)
193 | 
194 |             # save the parameters to file
195 |             # sim.save_params("./nengo_dl_params")
196 | 
197 |             #############################################################################################
198 |             # evaluation of results
199 |             #############################################################################################
200 | 
201 |             sim.compile(loss={out_p_filt: classification_accuracy})
202 | 
203 |             # runtime measurement
204 |             t=[]
205 |             for i in range(config.n_time_measures):
206 |                 t1 = time()
207 |                 accuracy = sim.evaluate(X_test, {out_p_filt: y_test_oh}, verbose=0)["loss"],
208 |                 inference_time = time() - t1
209 |                 t.append(inference_time)
210 |             inference_time = np.mean(t)
211 |             print("Accuracy after training:", accuracy)
212 |             accs.append(accuracy)
213 | 
214 |             sim2 = nengo_dl.Simulator(net, minibatch_size=1, device="/gpu:0")
215 | 
216 |             y_pred = sim2.predict(X_test)
217 |             y_pred_am = np.argmax(y_pred[out_p_filt][:,-1,:], axis=1)
218 |             y_pred_am.shape = (y_pred_am.shape[0],1)
219 |             f1_score = sklearn.metrics.f1_score(y_test,
220 |                                                 y_pred_am,
221 |                                                 average='micro')
222 |             f1_score_weighted = sklearn.metrics.f1_score(
223 |                 y_test,
224 |                 y_pred_am,
225 |                 average='weighted'
226 |             )
227 | 
228 |             logger.info("Training time: " + str(training_time))
229 |             logger.info("Mean epoch time: " + str(mean_epoch_time))
230 |             logger.info("Inference time: " + str(inference_time))
231 | 
232 |             logger.info("Preprocessing time for training: " + str(t_train))
233 |             logger.info("Preprocessing time for testing: " + str(t_test))
234 |             logger.info("Inference time one sequence [ms]: " + str((inference_time*1000+t_test*1000)/X_test.shape[0]))
235 | 
236 | 
237 |             print('Accuracy on training data: ')
238 |             report = classification_report(y_test, y_pred_am, output_dict=True)
239 |             logger.info(classification_report(y_test, y_pred_am))
240 | 
241 |             print("Confusion matrix:")
242 |             confusion_matrix = metrics.confusion_matrix(y_test, y_pred_am)
243 |             print(confusion_matrix)
244 | 
245 |             # f1 score
246 |             scores.append(f1_score_weighted)
247 |             logger.info("F1 Score: " + str(f1_score_weighted))
248 | 
249 |             # close simulator
250 |             sim.close()
251 |             sim2.close()
252 | 
253 | 
254 |         # add results to statistical result array
255 |         acc_mean.append(np.mean(accs))
256 |         f1_mean.append(np.mean(scores))
257 | 
258 |     # save as mat files
259 |     save_dic = {"report": report, "confusion_matrix": confusion_matrix, "config": config, "pred": y_pred,
260 |                 "label": y_test, "f1_mean": np.mean(f1_mean)}
261 |     savemat("results/" + args.dataset + "/results_Nengo_net_" + str(config.input_dim) + "_" + str(config.scale) + "_" +
262 |             str(config.encoding_dim) + '_' + str(config.training_volume) + ".mat", save_dic)
263 | 
264 |     logger.info('Accuracy results of statistical repetitions: ' + str(acc_mean))
265 |     logger.info('F1 scores of statistical repetitions: ' + str(f1_mean))
266 | 
267 |     # write all scores to extra file
268 |     logger.info('Mean Score: ' + str(np.mean(f1_mean)))
269 |     logger.info('Mean Accuracy: ' + str(np.mean(acc_mean)))
270 |     with open("results/results_" + args.dataset + "_SNN.txt", 'a') as file:
271 |         file.write(str(config.input_dim) + '\t'
272 |                    + str(config.encoding_dim) + '\t'
273 |                    + str(config.scale) + '\t'
274 |                    + str(args.stat_iterations) + '\t'
275 |                    + str(round(np.mean(f1_mean),3)) + '\t'
276 |                    + str(round(np.mean(acc_mean),3)) + '\t'
277 |                    + str(round(np.std(f1_mean),3)) + '\t'
278 |                    + str(round(np.std(acc_mean),3)) + '\n'
279 |                    )


--------------------------------------------------------------------------------
/LICENSE:
--------------------------------------------------------------------------------
  1 |                     GNU GENERAL PUBLIC LICENSE
  2 |                        Version 3, 29 June 2007
  3 | 
  4 |  Copyright (C) 2007 Free Software Foundation, Inc. <https://fsf.org/>
  5 |  Everyone is permitted to copy and distribute verbatim copies
  6 |  of this license document, but changing it is not allowed.
  7 | 
  8 |                             Preamble
  9 | 
 10 |   The GNU General Public License is a free, copyleft license for
 11 | software and other kinds of works.
 12 | 
 13 |   The licenses for most software and other practical works are designed
 14 | to take away your freedom to share and change the works.  By contrast,
 15 | the GNU General Public License is intended to guarantee your freedom to
 16 | share and change all versions of a program--to make sure it remains free
 17 | software for all its users.  We, the Free Software Foundation, use the
 18 | GNU General Public License for most of our software; it applies also to
 19 | any other work released this way by its authors.  You can apply it to
 20 | your programs, too.
 21 | 
 22 |   When we speak of free software, we are referring to freedom, not
 23 | price.  Our General Public Licenses are designed to make sure that you
 24 | have the freedom to distribute copies of free software (and charge for
 25 | them if you wish), that you receive source code or can get it if you
 26 | want it, that you can change the software or use pieces of it in new
 27 | free programs, and that you know you can do these things.
 28 | 
 29 |   To protect your rights, we need to prevent others from denying you
 30 | these rights or asking you to surrender the rights.  Therefore, you have
 31 | certain responsibilities if you distribute copies of the software, or if
 32 | you modify it: responsibilities to respect the freedom of others.
 33 | 
 34 |   For example, if you distribute copies of such a program, whether
 35 | gratis or for a fee, you must pass on to the recipients the same
 36 | freedoms that you received.  You must make sure that they, too, receive
 37 | or can get the source code.  And you must show them these terms so they
 38 | know their rights.
 39 | 
 40 |   Developers that use the GNU GPL protect your rights with two steps:
 41 | (1) assert copyright on the software, and (2) offer you this License
 42 | giving you legal permission to copy, distribute and/or modify it.
 43 | 
 44 |   For the developers' and authors' protection, the GPL clearly explains
 45 | that there is no warranty for this free software.  For both users' and
 46 | authors' sake, the GPL requires that modified versions be marked as
 47 | changed, so that their problems will not be attributed erroneously to
 48 | authors of previous versions.
 49 | 
 50 |   Some devices are designed to deny users access to install or run
 51 | modified versions of the software inside them, although the manufacturer
 52 | can do so.  This is fundamentally incompatible with the aim of
 53 | protecting users' freedom to change the software.  The systematic
 54 | pattern of such abuse occurs in the area of products for individuals to
 55 | use, which is precisely where it is most unacceptable.  Therefore, we
 56 | have designed this version of the GPL to prohibit the practice for those
 57 | products.  If such problems arise substantially in other domains, we
 58 | stand ready to extend this provision to those domains in future versions
 59 | of the GPL, as needed to protect the freedom of users.
 60 | 
 61 |   Finally, every program is threatened constantly by software patents.
 62 | States should not allow patents to restrict development and use of
 63 | software on general-purpose computers, but in those that do, we wish to
 64 | avoid the special danger that patents applied to a free program could
 65 | make it effectively proprietary.  To prevent this, the GPL assures that
 66 | patents cannot be used to render the program non-free.
 67 | 
 68 |   The precise terms and conditions for copying, distribution and
 69 | modification follow.
 70 | 
 71 |                        TERMS AND CONDITIONS
 72 | 
 73 |   0. Definitions.
 74 | 
 75 |   "This License" refers to version 3 of the GNU General Public License.
 76 | 
 77 |   "Copyright" also means copyright-like laws that apply to other kinds of
 78 | works, such as semiconductor masks.
 79 | 
 80 |   "The Program" refers to any copyrightable work licensed under this
 81 | License.  Each licensee is addressed as "you".  "Licensees" and
 82 | "recipients" may be individuals or organizations.
 83 | 
 84 |   To "modify" a work means to copy from or adapt all or part of the work
 85 | in a fashion requiring copyright permission, other than the making of an
 86 | exact copy.  The resulting work is called a "modified version" of the
 87 | earlier work or a work "based on" the earlier work.
 88 | 
 89 |   A "covered work" means either the unmodified Program or a work based
 90 | on the Program.
 91 | 
 92 |   To "propagate" a work means to do anything with it that, without
 93 | permission, would make you directly or secondarily liable for
 94 | infringement under applicable copyright law, except executing it on a
 95 | computer or modifying a private copy.  Propagation includes copying,
 96 | distribution (with or without modification), making available to the
 97 | public, and in some countries other activities as well.
 98 | 
 99 |   To "convey" a work means any kind of propagation that enables other
100 | parties to make or receive copies.  Mere interaction with a user through
101 | a computer network, with no transfer of a copy, is not conveying.
102 | 
103 |   An interactive user interface displays "Appropriate Legal Notices"
104 | to the extent that it includes a convenient and prominently visible
105 | feature that (1) displays an appropriate copyright notice, and (2)
106 | tells the user that there is no warranty for the work (except to the
107 | extent that warranties are provided), that licensees may convey the
108 | work under this License, and how to view a copy of this License.  If
109 | the interface presents a list of user commands or options, such as a
110 | menu, a prominent item in the list meets this criterion.
111 | 
112 |   1. Source Code.
113 | 
114 |   The "source code" for a work means the preferred form of the work
115 | for making modifications to it.  "Object code" means any non-source
116 | form of a work.
117 | 
118 |   A "Standard Interface" means an interface that either is an official
119 | standard defined by a recognized standards body, or, in the case of
120 | interfaces specified for a particular programming language, one that
121 | is widely used among developers working in that language.
122 | 
123 |   The "System Libraries" of an executable work include anything, other
124 | than the work as a whole, that (a) is included in the normal form of
125 | packaging a Major Component, but which is not part of that Major
126 | Component, and (b) serves only to enable use of the work with that
127 | Major Component, or to implement a Standard Interface for which an
128 | implementation is available to the public in source code form.  A
129 | "Major Component", in this context, means a major essential component
130 | (kernel, window system, and so on) of the specific operating system
131 | (if any) on which the executable work runs, or a compiler used to
132 | produce the work, or an object code interpreter used to run it.
133 | 
134 |   The "Corresponding Source" for a work in object code form means all
135 | the source code needed to generate, install, and (for an executable
136 | work) run the object code and to modify the work, including scripts to
137 | control those activities.  However, it does not include the work's
138 | System Libraries, or general-purpose tools or generally available free
139 | programs which are used unmodified in performing those activities but
140 | which are not part of the work.  For example, Corresponding Source
141 | includes interface definition files associated with source files for
142 | the work, and the source code for shared libraries and dynamically
143 | linked subprograms that the work is specifically designed to require,
144 | such as by intimate data communication or control flow between those
145 | subprograms and other parts of the work.
146 | 
147 |   The Corresponding Source need not include anything that users
148 | can regenerate automatically from other parts of the Corresponding
149 | Source.
150 | 
151 |   The Corresponding Source for a work in source code form is that
152 | same work.
153 | 
154 |   2. Basic Permissions.
155 | 
156 |   All rights granted under this License are granted for the term of
157 | copyright on the Program, and are irrevocable provided the stated
158 | conditions are met.  This License explicitly affirms your unlimited
159 | permission to run the unmodified Program.  The output from running a
160 | covered work is covered by this License only if the output, given its
161 | content, constitutes a covered work.  This License acknowledges your
162 | rights of fair use or other equivalent, as provided by copyright law.
163 | 
164 |   You may make, run and propagate covered works that you do not
165 | convey, without conditions so long as your license otherwise remains
166 | in force.  You may convey covered works to others for the sole purpose
167 | of having them make modifications exclusively for you, or provide you
168 | with facilities for running those works, provided that you comply with
169 | the terms of this License in conveying all material for which you do
170 | not control copyright.  Those thus making or running the covered works
171 | for you must do so exclusively on your behalf, under your direction
172 | and control, on terms that prohibit them from making any copies of
173 | your copyrighted material outside their relationship with you.
174 | 
175 |   Conveying under any other circumstances is permitted solely under
176 | the conditions stated below.  Sublicensing is not allowed; section 10
177 | makes it unnecessary.
178 | 
179 |   3. Protecting Users' Legal Rights From Anti-Circumvention Law.
180 | 
181 |   No covered work shall be deemed part of an effective technological
182 | measure under any applicable law fulfilling obligations under article
183 | 11 of the WIPO copyright treaty adopted on 20 December 1996, or
184 | similar laws prohibiting or restricting circumvention of such
185 | measures.
186 | 
187 |   When you convey a covered work, you waive any legal power to forbid
188 | circumvention of technological measures to the extent such circumvention
189 | is effected by exercising rights under this License with respect to
190 | the covered work, and you disclaim any intention to limit operation or
191 | modification of the work as a means of enforcing, against the work's
192 | users, your or third parties' legal rights to forbid circumvention of
193 | technological measures.
194 | 
195 |   4. Conveying Verbatim Copies.
196 | 
197 |   You may convey verbatim copies of the Program's source code as you
198 | receive it, in any medium, provided that you conspicuously and
199 | appropriately publish on each copy an appropriate copyright notice;
200 | keep intact all notices stating that this License and any
201 | non-permissive terms added in accord with section 7 apply to the code;
202 | keep intact all notices of the absence of any warranty; and give all
203 | recipients a copy of this License along with the Program.
204 | 
205 |   You may charge any price or no price for each copy that you convey,
206 | and you may offer support or warranty protection for a fee.
207 | 
208 |   5. Conveying Modified Source Versions.
209 | 
210 |   You may convey a work based on the Program, or the modifications to
211 | produce it from the Program, in the form of source code under the
212 | terms of section 4, provided that you also meet all of these conditions:
213 | 
214 |     a) The work must carry prominent notices stating that you modified
215 |     it, and giving a relevant date.
216 | 
217 |     b) The work must carry prominent notices stating that it is
218 |     released under this License and any conditions added under section
219 |     7.  This requirement modifies the requirement in section 4 to
220 |     "keep intact all notices".
221 | 
222 |     c) You must license the entire work, as a whole, under this
223 |     License to anyone who comes into possession of a copy.  This
224 |     License will therefore apply, along with any applicable section 7
225 |     additional terms, to the whole of the work, and all its parts,
226 |     regardless of how they are packaged.  This License gives no
227 |     permission to license the work in any other way, but it does not
228 |     invalidate such permission if you have separately received it.
229 | 
230 |     d) If the work has interactive user interfaces, each must display
231 |     Appropriate Legal Notices; however, if the Program has interactive
232 |     interfaces that do not display Appropriate Legal Notices, your
233 |     work need not make them do so.
234 | 
235 |   A compilation of a covered work with other separate and independent
236 | works, which are not by their nature extensions of the covered work,
237 | and which are not combined with it such as to form a larger program,
238 | in or on a volume of a storage or distribution medium, is called an
239 | "aggregate" if the compilation and its resulting copyright are not
240 | used to limit the access or legal rights of the compilation's users
241 | beyond what the individual works permit.  Inclusion of a covered work
242 | in an aggregate does not cause this License to apply to the other
243 | parts of the aggregate.
244 | 
245 |   6. Conveying Non-Source Forms.
246 | 
247 |   You may convey a covered work in object code form under the terms
248 | of sections 4 and 5, provided that you also convey the
249 | machine-readable Corresponding Source under the terms of this License,
250 | in one of these ways:
251 | 
252 |     a) Convey the object code in, or embodied in, a physical product
253 |     (including a physical distribution medium), accompanied by the
254 |     Corresponding Source fixed on a durable physical medium
255 |     customarily used for software interchange.
256 | 
257 |     b) Convey the object code in, or embodied in, a physical product
258 |     (including a physical distribution medium), accompanied by a
259 |     written offer, valid for at least three years and valid for as
260 |     long as you offer spare parts or customer support for that product
261 |     model, to give anyone who possesses the object code either (1) a
262 |     copy of the Corresponding Source for all the software in the
263 |     product that is covered by this License, on a durable physical
264 |     medium customarily used for software interchange, for a price no
265 |     more than your reasonable cost of physically performing this
266 |     conveying of source, or (2) access to copy the
267 |     Corresponding Source from a network server at no charge.
268 | 
269 |     c) Convey individual copies of the object code with a copy of the
270 |     written offer to provide the Corresponding Source.  This
271 |     alternative is allowed only occasionally and noncommercially, and
272 |     only if you received the object code with such an offer, in accord
273 |     with subsection 6b.
274 | 
275 |     d) Convey the object code by offering access from a designated
276 |     place (gratis or for a charge), and offer equivalent access to the
277 |     Corresponding Source in the same way through the same place at no
278 |     further charge.  You need not require recipients to copy the
279 |     Corresponding Source along with the object code.  If the place to
280 |     copy the object code is a network server, the Corresponding Source
281 |     may be on a different server (operated by you or a third party)
282 |     that supports equivalent copying facilities, provided you maintain
283 |     clear directions next to the object code saying where to find the
284 |     Corresponding Source.  Regardless of what server hosts the
285 |     Corresponding Source, you remain obligated to ensure that it is
286 |     available for as long as needed to satisfy these requirements.
287 | 
288 |     e) Convey the object code using peer-to-peer transmission, provided
289 |     you inform other peers where the object code and Corresponding
290 |     Source of the work are being offered to the general public at no
291 |     charge under subsection 6d.
292 | 
293 |   A separable portion of the object code, whose source code is excluded
294 | from the Corresponding Source as a System Library, need not be
295 | included in conveying the object code work.
296 | 
297 |   A "User Product" is either (1) a "consumer product", which means any
298 | tangible personal property which is normally used for personal, family,
299 | or household purposes, or (2) anything designed or sold for incorporation
300 | into a dwelling.  In determining whether a product is a consumer product,
301 | doubtful cases shall be resolved in favor of coverage.  For a particular
302 | product received by a particular user, "normally used" refers to a
303 | typical or common use of that class of product, regardless of the status
304 | of the particular user or of the way in which the particular user
305 | actually uses, or expects or is expected to use, the product.  A product
306 | is a consumer product regardless of whether the product has substantial
307 | commercial, industrial or non-consumer uses, unless such uses represent
308 | the only significant mode of use of the product.
309 | 
310 |   "Installation Information" for a User Product means any methods,
311 | procedures, authorization keys, or other information required to install
312 | and execute modified versions of a covered work in that User Product from
313 | a modified version of its Corresponding Source.  The information must
314 | suffice to ensure that the continued functioning of the modified object
315 | code is in no case prevented or interfered with solely because
316 | modification has been made.
317 | 
318 |   If you convey an object code work under this section in, or with, or
319 | specifically for use in, a User Product, and the conveying occurs as
320 | part of a transaction in which the right of possession and use of the
321 | User Product is transferred to the recipient in perpetuity or for a
322 | fixed term (regardless of how the transaction is characterized), the
323 | Corresponding Source conveyed under this section must be accompanied
324 | by the Installation Information.  But this requirement does not apply
325 | if neither you nor any third party retains the ability to install
326 | modified object code on the User Product (for example, the work has
327 | been installed in ROM).
328 | 
329 |   The requirement to provide Installation Information does not include a
330 | requirement to continue to provide support service, warranty, or updates
331 | for a work that has been modified or installed by the recipient, or for
332 | the User Product in which it has been modified or installed.  Access to a
333 | network may be denied when the modification itself materially and
334 | adversely affects the operation of the network or violates the rules and
335 | protocols for communication across the network.
336 | 
337 |   Corresponding Source conveyed, and Installation Information provided,
338 | in accord with this section must be in a format that is publicly
339 | documented (and with an implementation available to the public in
340 | source code form), and must require no special password or key for
341 | unpacking, reading or copying.
342 | 
343 |   7. Additional Terms.
344 | 
345 |   "Additional permissions" are terms that supplement the terms of this
346 | License by making exceptions from one or more of its conditions.
347 | Additional permissions that are applicable to the entire Program shall
348 | be treated as though they were included in this License, to the extent
349 | that they are valid under applicable law.  If additional permissions
350 | apply only to part of the Program, that part may be used separately
351 | under those permissions, but the entire Program remains governed by
352 | this License without regard to the additional permissions.
353 | 
354 |   When you convey a copy of a covered work, you may at your option
355 | remove any additional permissions from that copy, or from any part of
356 | it.  (Additional permissions may be written to require their own
357 | removal in certain cases when you modify the work.)  You may place
358 | additional permissions on material, added by you to a covered work,
359 | for which you have or can give appropriate copyright permission.
360 | 
361 |   Notwithstanding any other provision of this License, for material you
362 | add to a covered work, you may (if authorized by the copyright holders of
363 | that material) supplement the terms of this License with terms:
364 | 
365 |     a) Disclaiming warranty or limiting liability differently from the
366 |     terms of sections 15 and 16 of this License; or
367 | 
368 |     b) Requiring preservation of specified reasonable legal notices or
369 |     author attributions in that material or in the Appropriate Legal
370 |     Notices displayed by works containing it; or
371 | 
372 |     c) Prohibiting misrepresentation of the origin of that material, or
373 |     requiring that modified versions of such material be marked in
374 |     reasonable ways as different from the original version; or
375 | 
376 |     d) Limiting the use for publicity purposes of names of licensors or
377 |     authors of the material; or
378 | 
379 |     e) Declining to grant rights under trademark law for use of some
380 |     trade names, trademarks, or service marks; or
381 | 
382 |     f) Requiring indemnification of licensors and authors of that
383 |     material by anyone who conveys the material (or modified versions of
384 |     it) with contractual assumptions of liability to the recipient, for
385 |     any liability that these contractual assumptions directly impose on
386 |     those licensors and authors.
387 | 
388 |   All other non-permissive additional terms are considered "further
389 | restrictions" within the meaning of section 10.  If the Program as you
390 | received it, or any part of it, contains a notice stating that it is
391 | governed by this License along with a term that is a further
392 | restriction, you may remove that term.  If a license document contains
393 | a further restriction but permits relicensing or conveying under this
394 | License, you may add to a covered work material governed by the terms
395 | of that license document, provided that the further restriction does
396 | not survive such relicensing or conveying.
397 | 
398 |   If you add terms to a covered work in accord with this section, you
399 | must place, in the relevant source files, a statement of the
400 | additional terms that apply to those files, or a notice indicating
401 | where to find the applicable terms.
402 | 
403 |   Additional terms, permissive or non-permissive, may be stated in the
404 | form of a separately written license, or stated as exceptions;
405 | the above requirements apply either way.
406 | 
407 |   8. Termination.
408 | 
409 |   You may not propagate or modify a covered work except as expressly
410 | provided under this License.  Any attempt otherwise to propagate or
411 | modify it is void, and will automatically terminate your rights under
412 | this License (including any patent licenses granted under the third
413 | paragraph of section 11).
414 | 
415 |   However, if you cease all violation of this License, then your
416 | license from a particular copyright holder is reinstated (a)
417 | provisionally, unless and until the copyright holder explicitly and
418 | finally terminates your license, and (b) permanently, if the copyright
419 | holder fails to notify you of the violation by some reasonable means
420 | prior to 60 days after the cessation.
421 | 
422 |   Moreover, your license from a particular copyright holder is
423 | reinstated permanently if the copyright holder notifies you of the
424 | violation by some reasonable means, this is the first time you have
425 | received notice of violation of this License (for any work) from that
426 | copyright holder, and you cure the violation prior to 30 days after
427 | your receipt of the notice.
428 | 
429 |   Termination of your rights under this section does not terminate the
430 | licenses of parties who have received copies or rights from you under
431 | this License.  If your rights have been terminated and not permanently
432 | reinstated, you do not qualify to receive new licenses for the same
433 | material under section 10.
434 | 
435 |   9. Acceptance Not Required for Having Copies.
436 | 
437 |   You are not required to accept this License in order to receive or
438 | run a copy of the Program.  Ancillary propagation of a covered work
439 | occurring solely as a consequence of using peer-to-peer transmission
440 | to receive a copy likewise does not require acceptance.  However,
441 | nothing other than this License grants you permission to propagate or
442 | modify any covered work.  These actions infringe copyright if you do
443 | not accept this License.  Therefore, by modifying or propagating a
444 | covered work, you indicate your acceptance of this License to do so.
445 | 
446 |   10. Automatic Licensing of Downstream Recipients.
447 | 
448 |   Each time you convey a covered work, the recipient automatically
449 | receives a license from the original licensors, to run, modify and
450 | propagate that work, subject to this License.  You are not responsible
451 | for enforcing compliance by third parties with this License.
452 | 
453 |   An "entity transaction" is a transaction transferring control of an
454 | organization, or substantially all assets of one, or subdividing an
455 | organization, or merging organizations.  If propagation of a covered
456 | work results from an entity transaction, each party to that
457 | transaction who receives a copy of the work also receives whatever
458 | licenses to the work the party's predecessor in interest had or could
459 | give under the previous paragraph, plus a right to possession of the
460 | Corresponding Source of the work from the predecessor in interest, if
461 | the predecessor has it or can get it with reasonable efforts.
462 | 
463 |   You may not impose any further restrictions on the exercise of the
464 | rights granted or affirmed under this License.  For example, you may
465 | not impose a license fee, royalty, or other charge for exercise of
466 | rights granted under this License, and you may not initiate litigation
467 | (including a cross-claim or counterclaim in a lawsuit) alleging that
468 | any patent claim is infringed by making, using, selling, offering for
469 | sale, or importing the Program or any portion of it.
470 | 
471 |   11. Patents.
472 | 
473 |   A "contributor" is a copyright holder who authorizes use under this
474 | License of the Program or a work on which the Program is based.  The
475 | work thus licensed is called the contributor's "contributor version".
476 | 
477 |   A contributor's "essential patent claims" are all patent claims
478 | owned or controlled by the contributor, whether already acquired or
479 | hereafter acquired, that would be infringed by some manner, permitted
480 | by this License, of making, using, or selling its contributor version,
481 | but do not include claims that would be infringed only as a
482 | consequence of further modification of the contributor version.  For
483 | purposes of this definition, "control" includes the right to grant
484 | patent sublicenses in a manner consistent with the requirements of
485 | this License.
486 | 
487 |   Each contributor grants you a non-exclusive, worldwide, royalty-free
488 | patent license under the contributor's essential patent claims, to
489 | make, use, sell, offer for sale, import and otherwise run, modify and
490 | propagate the contents of its contributor version.
491 | 
492 |   In the following three paragraphs, a "patent license" is any express
493 | agreement or commitment, however denominated, not to enforce a patent
494 | (such as an express permission to practice a patent or covenant not to
495 | sue for patent infringement).  To "grant" such a patent license to a
496 | party means to make such an agreement or commitment not to enforce a
497 | patent against the party.
498 | 
499 |   If you convey a covered work, knowingly relying on a patent license,
500 | and the Corresponding Source of the work is not available for anyone
501 | to copy, free of charge and under the terms of this License, through a
502 | publicly available network server or other readily accessible means,
503 | then you must either (1) cause the Corresponding Source to be so
504 | available, or (2) arrange to deprive yourself of the benefit of the
505 | patent license for this particular work, or (3) arrange, in a manner
506 | consistent with the requirements of this License, to extend the patent
507 | license to downstream recipients.  "Knowingly relying" means you have
508 | actual knowledge that, but for the patent license, your conveying the
509 | covered work in a country, or your recipient's use of the covered work
510 | in a country, would infringe one or more identifiable patents in that
511 | country that you have reason to believe are valid.
512 | 
513 |   If, pursuant to or in connection with a single transaction or
514 | arrangement, you convey, or propagate by procuring conveyance of, a
515 | covered work, and grant a patent license to some of the parties
516 | receiving the covered work authorizing them to use, propagate, modify
517 | or convey a specific copy of the covered work, then the patent license
518 | you grant is automatically extended to all recipients of the covered
519 | work and works based on it.
520 | 
521 |   A patent license is "discriminatory" if it does not include within
522 | the scope of its coverage, prohibits the exercise of, or is
523 | conditioned on the non-exercise of one or more of the rights that are
524 | specifically granted under this License.  You may not convey a covered
525 | work if you are a party to an arrangement with a third party that is
526 | in the business of distributing software, under which you make payment
527 | to the third party based on the extent of your activity of conveying
528 | the work, and under which the third party grants, to any of the
529 | parties who would receive the covered work from you, a discriminatory
530 | patent license (a) in connection with copies of the covered work
531 | conveyed by you (or copies made from those copies), or (b) primarily
532 | for and in connection with specific products or compilations that
533 | contain the covered work, unless you entered into that arrangement,
534 | or that patent license was granted, prior to 28 March 2007.
535 | 
536 |   Nothing in this License shall be construed as excluding or limiting
537 | any implied license or other defenses to infringement that may
538 | otherwise be available to you under applicable patent law.
539 | 
540 |   12. No Surrender of Others' Freedom.
541 | 
542 |   If conditions are imposed on you (whether by court order, agreement or
543 | otherwise) that contradict the conditions of this License, they do not
544 | excuse you from the conditions of this License.  If you cannot convey a
545 | covered work so as to satisfy simultaneously your obligations under this
546 | License and any other pertinent obligations, then as a consequence you may
547 | not convey it at all.  For example, if you agree to terms that obligate you
548 | to collect a royalty for further conveying from those to whom you convey
549 | the Program, the only way you could satisfy both those terms and this
550 | License would be to refrain entirely from conveying the Program.
551 | 
552 |   13. Use with the GNU Affero General Public License.
553 | 
554 |   Notwithstanding any other provision of this License, you have
555 | permission to link or combine any covered work with a work licensed
556 | under version 3 of the GNU Affero General Public License into a single
557 | combined work, and to convey the resulting work.  The terms of this
558 | License will continue to apply to the part which is the covered work,
559 | but the special requirements of the GNU Affero General Public License,
560 | section 13, concerning interaction through a network will apply to the
561 | combination as such.
562 | 
563 |   14. Revised Versions of this License.
564 | 
565 |   The Free Software Foundation may publish revised and/or new versions of
566 | the GNU General Public License from time to time.  Such new versions will
567 | be similar in spirit to the present version, but may differ in detail to
568 | address new problems or concerns.
569 | 
570 |   Each version is given a distinguishing version number.  If the
571 | Program specifies that a certain numbered version of the GNU General
572 | Public License "or any later version" applies to it, you have the
573 | option of following the terms and conditions either of that numbered
574 | version or of any later version published by the Free Software
575 | Foundation.  If the Program does not specify a version number of the
576 | GNU General Public License, you may choose any version ever published
577 | by the Free Software Foundation.
578 | 
579 |   If the Program specifies that a proxy can decide which future
580 | versions of the GNU General Public License can be used, that proxy's
581 | public statement of acceptance of a version permanently authorizes you
582 | to choose that version for the Program.
583 | 
584 |   Later license versions may give you additional or different
585 | permissions.  However, no additional obligations are imposed on any
586 | author or copyright holder as a result of your choosing to follow a
587 | later version.
588 | 
589 |   15. Disclaimer of Warranty.
590 | 
591 |   THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY
592 | APPLICABLE LAW.  EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT
593 | HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY
594 | OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO,
595 | THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
596 | PURPOSE.  THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM
597 | IS WITH YOU.  SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF
598 | ALL NECESSARY SERVICING, REPAIR OR CORRECTION.
599 | 
600 |   16. Limitation of Liability.
601 | 
602 |   IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING
603 | WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS
604 | THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY
605 | GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE
606 | USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF
607 | DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD
608 | PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS),
609 | EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF
610 | SUCH DAMAGES.
611 | 
612 |   17. Interpretation of Sections 15 and 16.
613 | 
614 |   If the disclaimer of warranty and limitation of liability provided
615 | above cannot be given local legal effect according to their terms,
616 | reviewing courts shall apply local law that most closely approximates
617 | an absolute waiver of all civil liability in connection with the
618 | Program, unless a warranty or assumption of liability accompanies a
619 | copy of the Program in return for a fee.
620 | 
621 |                      END OF TERMS AND CONDITIONS
622 | 
623 |             How to Apply These Terms to Your New Programs
624 | 
625 |   If you develop a new program, and you want it to be of the greatest
626 | possible use to the public, the best way to achieve this is to make it
627 | free software which everyone can redistribute and change under these terms.
628 | 
629 |   To do so, attach the following notices to the program.  It is safest
630 | to attach them to the start of each source file to most effectively
631 | state the exclusion of warranty; and each file should have at least
632 | the "copyright" line and a pointer to where the full notice is found.
633 | 
634 |     <one line to give the program's name and a brief idea of what it does.>
635 |     Copyright (C) <year>  <name of author>
636 | 
637 |     This program is free software: you can redistribute it and/or modify
638 |     it under the terms of the GNU General Public License as published by
639 |     the Free Software Foundation, either version 3 of the License, or
640 |     (at your option) any later version.
641 | 
642 |     This program is distributed in the hope that it will be useful,
643 |     but WITHOUT ANY WARRANTY; without even the implied warranty of
644 |     MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
645 |     GNU General Public License for more details.
646 | 
647 |     You should have received a copy of the GNU General Public License
648 |     along with this program.  If not, see <https://www.gnu.org/licenses/>.
649 | 
650 | Also add information on how to contact you by electronic and paper mail.
651 | 
652 |   If the program does terminal interaction, make it output a short
653 | notice like this when it starts in an interactive mode:
654 | 
655 |     <program>  Copyright (C) <year>  <name of author>
656 |     This program comes with ABSOLUTELY NO WARRANTY; for details type `show w'.
657 |     This is free software, and you are welcome to redistribute it
658 |     under certain conditions; type `show c' for details.
659 | 
660 | The hypothetical commands `show w' and `show c' should show the appropriate
661 | parts of the General Public License.  Of course, your program's commands
662 | might be different; for a GUI interface, you would use an "about box".
663 | 
664 |   You should also get your employer (if you work as a programmer) or school,
665 | if any, to sign a "copyright disclaimer" for the program, if necessary.
666 | For more information on this, and how to apply and follow the GNU GPL, see
667 | <https://www.gnu.org/licenses/>.
668 | 
669 |   The GNU General Public License does not permit incorporating your program
670 | into proprietary programs.  If your program is a subroutine library, you
671 | may consider it more useful to permit linking proprietary applications with
672 | the library.  If this is what you want to do, use the GNU Lesser General
673 | Public License instead of this License.  But first, please read
674 | <https://www.gnu.org/licenses/why-not-lgpl.html>.
675 | 


--------------------------------------------------------------------------------
/LSTM.py:
--------------------------------------------------------------------------------
  1 | from __future__ import absolute_import
  2 | from __future__ import division
  3 | from __future__ import print_function
  4 | 
  5 | from config import Config
  6 | from utils import *
  7 | from model import HDC_ANN, LSTM_Network
  8 | from sklearn.metrics import classification_report, f1_score
  9 | from scipy.io import savemat, loadmat
 10 | from sklearn import metrics
 11 | import os
 12 | from time import time
 13 | from tensorflow.python.client import timeline
 14 | import logging
 15 | import matplotlib.pyplot as plt
 16 | 
 17 | # config logger
 18 | logger = logging.getLogger('log')
 19 | 
 20 | 
 21 | def main_LSTM(args):
 22 |     '''
 23 |     implementation of the original LSTM approach (https://github.com/KhaledSaleh/driving_behaviour_classification)
 24 |     '''
 25 |     # set config params specific for the original code
 26 |     config = Config()
 27 |     config.training_volume = args.training_volume
 28 |     config.input_dim = args.input_dim
 29 |     config.encoding_dim = args.encoding_dim
 30 |     config.scale = args.scale
 31 |     if args.runtime_measurement:
 32 |         config.n_time_measures = 10
 33 |     else:
 34 |         config.n_time_measures = 1
 35 | 
 36 |     # load preprocessed data
 37 |     data = load_dataset(args.dataset,config)
 38 |     X_train = data[0]
 39 |     X_test = data[1]
 40 |     y_train = data[2]
 41 |     y_test = data[3]
 42 |     config = data[4]
 43 | 
 44 |     logs = []
 45 |     # if train test data not a list, create one
 46 |     if type(X_train)==list:
 47 |         print("given data is not a list")
 48 |         X_train_list = X_train
 49 |         X_test_list = X_test
 50 |         y_train_list = y_train
 51 |         y_test_list = y_test
 52 |     else:
 53 |         X_train_list =[X_train]
 54 |         X_test_list = [X_test]
 55 |         y_train_list = [y_train]
 56 |         y_test_list = [y_test]
 57 | 
 58 |     #######################################################################################
 59 |     # statistical iteration
 60 |     #######################################################################################
 61 |     acc_mean = []
 62 |     f1_mean = []
 63 | 
 64 |     for stat_it in range(args.stat_iterations):
 65 |         logger.info('Statistial iteration: ' + str(stat_it))
 66 | 
 67 |         # train for each element in list (that is why we need list form, even if it contains only one element)
 68 |         logger.info('Training data contains ' + str(len(X_train_list)) + ' training instances...')
 69 |         scores = []
 70 |         accs = []
 71 |         for it in range(len(X_train_list)):
 72 |             logger.info(('.......'))
 73 |             logger.info('instance ' + str(it) + ':')
 74 | 
 75 |             X_train = X_train_list[it]
 76 |             X_test = X_test_list[it]
 77 |             y_train = y_train_list[it]
 78 |             y_test = y_test_list[it]
 79 | 
 80 |             # use only fraction of training samples (if given)
 81 |             X_train = X_train[1:int(X_train.shape[0] * config.training_volume), :]
 82 |             y_train = y_train[1:int(y_train.shape[0] * config.training_volume), :]
 83 | 
 84 |             config.n_inputs = X_train.shape[2]
 85 |             config.train_count = len(X_train)
 86 |             config.test_data_count = len(X_test)
 87 |             config.n_steps = len(X_train[0])
 88 |             config.n_classes = len(np.unique(y_train))
 89 | 
 90 |             logger.info('Training dataset shape: ' + str(X_train.shape) + str(y_train.shape))
 91 |             logger.info('Test dataset shape: ' + str(X_test.shape) + str(y_test.shape))
 92 |             graph = tf.Graph()
 93 |             with graph.as_default():
 94 | 
 95 |                 X = tf.compat.v1.placeholder(tf.float32, [None, config.n_steps, config.n_inputs], name="X")
 96 |                 Y = tf.compat.v1.placeholder(tf.float32, [None, config.n_classes], name="Y")
 97 | 
 98 |                 pred_Y = LSTM_Network(X, config)
 99 | 
100 |                 # Loss,optimizer,evaluation
101 |                 l2 = config.lambda_loss_amount * \
102 |                      sum(tf.nn.l2_loss(tf_var) for tf_var in tf.compat.v1.trainable_variables())
103 |                 # Softmax loss and L2
104 |                 cost = tf.reduce_mean(
105 |                     tf.nn.softmax_cross_entropy_with_logits(logits=pred_Y, labels=Y), name="cost") + l2
106 |                 optimizer = tf.compat.v1.train.AdamOptimizer(
107 |                     learning_rate=config.learning_rate).minimize(cost)
108 | 
109 |                 correct_pred = tf.equal(tf.argmax(pred_Y, 1), tf.argmax(Y, 1))
110 |                 accuracy = tf.reduce_mean(tf.cast(correct_pred, dtype=tf.float32))
111 | 
112 |                 saver = tf.compat.v1.train.Saver()
113 | 
114 |             with tf.compat.v1.Session(graph=graph, config=tf.compat.v1.ConfigProto(log_device_placement=False)) as sess:
115 |                 if not args.test:
116 |                     init_op = tf.compat.v1.global_variables_initializer()
117 |                     sess.run(init_op)
118 |                     best_accuracy = 0.0
119 |                     # Start training for each batch and loop epochs
120 |                     for i in range(config.training_epochs):
121 |                         starttime = time()
122 |                         for start, end in zip(range(0, config.train_count, config.batch_size),
123 |                                               range(config.batch_size, config.train_count + 1, config.batch_size)):
124 |                             sess.run(optimizer, feed_dict={X: X_train[start:end],
125 |                                                            Y: one_hot(y_train[start:end],config.n_classes)})
126 |                             saver.save(sess, os.path.join("./weights", 'LSTM_model'))
127 |                         # Test completely at every epoch: calculate accuracy
128 |                         pred_out, accuracy_out, loss_out = sess.run([pred_Y, accuracy, cost], feed_dict={
129 |                             X: X_test, Y: one_hot(y_test, config.n_classes)})
130 |                         logs.append(time() - starttime)
131 |                         print("Training iter: {},".format(i) + \
132 |                               " Test accuracy : {},".format(accuracy_out) + \
133 |                               " Loss : {}".format(loss_out))
134 |                         best_accuracy = max(best_accuracy, accuracy_out)
135 |                     print("")
136 |                     mean_epoch_time = np.mean(logs)
137 |                     overall_time = np.sum(logs)
138 |                     logger.info("Mean Epoch time: " + str(mean_epoch_time))
139 |                     logger.info("overall training time: " + str(overall_time))
140 |                     logger.info("Final test accuracy: {}".format(accuracy_out))
141 |                     logger.info("Best epoch's test accuracy: {}".format(best_accuracy))
142 | 
143 |                     print("")
144 |                 # start testing the trained model
145 |                 else:
146 |                     saver.restore(sess, os.path.join("./weights", 'LSTM_model'))
147 |                     t1 = time()
148 |                     pred_out, accuracy_out, loss_out = sess.run([pred_Y, accuracy, cost], feed_dict={
149 |                         X: X_test, Y: one_hot(y_test,config.n_classes)})
150 |                     inference_time = time() - t1
151 |                     print(" Test accuracy : {},".format(accuracy_out) + \
152 |                           " Loss : {}".format(loss_out))
153 | 
154 |             #############################################################################################
155 |             # evaluation of results
156 |             #############################################################################################
157 | 
158 |             pred_test_bool = pred_out.argmax(1)
159 | 
160 |             # runtime measurement
161 |             t=[]
162 |             traces = []
163 |             options = tf.compat.v1.RunOptions(trace_level=tf.compat.v1.RunOptions.FULL_TRACE)
164 |             run_metadata = tf.compat.v1.RunMetadata()
165 |             for i in range(config.n_time_measures):
166 |                 with tf.compat.v1.Session(graph=graph, config=tf.compat.v1.ConfigProto(log_device_placement=False)) as Sess:
167 |                     init_op = tf.compat.v1.global_variables_initializer()
168 |                     Sess.run(init_op)
169 |                     t1 = time()
170 |                     Sess.run([pred_Y, accuracy, cost], feed_dict={
171 |                         X: X_test, Y: one_hot(y_test, config.n_classes)}, options=options, run_metadata=run_metadata)
172 |                     inference_time = time() - t1
173 |                     fetched_timeline = timeline.Timeline(run_metadata.step_stats)
174 |                     chrome_trace = fetched_timeline.generate_chrome_trace_format()
175 |                     traces.append(chrome_trace)
176 |                     t.append(inference_time)
177 |             with open('./logs/LSTM_ts_preproc_timeline_test.json', 'w') as f:
178 |                 f.write(traces[-1])
179 |             inference_time = np.median(inference_time)
180 |             logger.info("Inference time: " + str(inference_time))
181 |             logger.info("Inference time of one sequence [ms]: " + str(inference_time*1000/X_test.shape[0]))
182 | 
183 |             logger.info('Accuracy on training data: ')
184 |             report = classification_report(y_test.astype(int), pred_test_bool, output_dict=True)
185 |             logger.info(classification_report(y_test.astype(int), pred_test_bool))
186 | 
187 |             accs.append((report['accuracy']))
188 | 
189 |             logger.info("Confusion matrix:")
190 |             confusion_matrix = metrics.confusion_matrix(y_test.astype(int), pred_test_bool)
191 |             logger.info(confusion_matrix)
192 | 
193 |             # f1 score
194 |             f1 = f1_score(y_test.astype(int), pred_test_bool, average='weighted')
195 |             scores.append(f1)
196 |             logger.info("F1 Score: " + str(f1))
197 | 
198 |         # add results to statistical result array
199 |         acc_mean.append(np.mean(accs))
200 |         f1_mean.append(np.mean(scores))
201 | 
202 |     # save as mat files
203 |     save_dic = {"report": report, "confusion_matrix": confusion_matrix, "config": config, "pred": pred_out,
204 |                 "label": y_test, "f1": np.mean(f1_mean), "acc_mean": np.mean(acc_mean)}
205 |     savemat("results/" + args.dataset + "/results_origNet_" + str(config.training_volume) + ".mat", save_dic)
206 | 
207 |     logger.info('Accuracy results of statistical repetitions: ' + str(acc_mean))
208 |     logger.info('F1 scores of statistical repetitions: ' + str(f1_mean))
209 | 
210 |     # write all scores to extra file
211 |     logger.info('Mean Score: ' + str(np.mean(f1_mean)))
212 |     logger.info('Mean Accuracy: ' + str(np.mean(acc_mean)))
213 |     with open("results/results_" + args.dataset + "_LSTM.txt", 'a') as file:
214 |         file.write(str(args.stat_iterations) + '\t'
215 |                    + str(round(np.mean(f1_mean), 3)) + '\t'
216 |                    + str(round(np.mean(acc_mean), 3)) + '\t'
217 |                    + str(round(np.std(f1_mean), 3)) + '\t'
218 |                    + str(round(np.std(acc_mean), 3)) + '\t'
219 |                    + str(args.training_volume) + '\n'
220 |                    )
221 | 
222 | 
223 | 
224 | 


--------------------------------------------------------------------------------
/README.md:
--------------------------------------------------------------------------------
 1 | # VSA for driving behaviour classification
 2 | 
 3 | This repository is mainly based on the code of https://github.com/KhaledSaleh/driving_behaviour_classification
 4 | 
 5 | It has 3 different models:
 6 | - a LSTM model (original model from [1])
 7 | - a feed-forward model (ANN) for HDC encodings
 8 | - a spiking neural model (SNN) for HDC encodings
 9 |  
10 | [1] K. Saleh, M. Hossny, and S. Nahavandi, “Driving behavior classification based on sensor data fusion using LSTM recurrent neural networks,” in International Conference on Intelligent Transportation Systems (ITSC), 2017.
11 | 
12 | Tested with the Python packages listed in requirements.txt.
13 | 
14 | ## Usage 
15 | * first, clone the Repo `git clone https://github.com/TUC-ProAut/HDC_driving_style_classification.git`
16 | 
17 | ### Train the networks (Python)
18 | 1. Run `python3 main.py --help` to check the available command line args.
19 | 2. Run ANN with HDC encodings:
20 |     * `python3 main.py --HDC_ANN True` (use --dataset argument to select between full, motorway, secondary or full_crossval) 
21 | 3. Run ANN with concatenated input sequences:
22 |     * `python3 main.py --Concat_ANN True`
23 | 4. Run the original LSTM model from https://github.com/KhaledSaleh/driving_behaviour_classification
24 |     * `python3 main.py --LSTM True`
25 | 5. Run SNN with HDC encodings:
26 |     * `python3 main.py --HDC_SNN True`
27 | 
28 | The results are written to the log file logs/main_log.log
29 | 
30 | ### Data efficiency experiment (Python)
31 | 1. Run `python3 main.py --data_efficiency True --HDC_ANN True` for the appropriate network as in section above 
32 | 
33 | ### (Optional) Hyper-parameter analysis for HDC encodings (Python)
34 | 1. Run `python3 main.py --hyperparams_experiment True --HDC_ANN True` 
35 | 
36 | ### Run Baseline models (MATLAB)
37 | 1. Run `eval_baseline_models.m`
38 | 
39 | 


--------------------------------------------------------------------------------
/config.py:
--------------------------------------------------------------------------------
 1 | import tensorflow as tf  
 2 | 
 3 | class Config(object):
 4 |     """
 5 |     define a class to store parameters,
 6 |     the input should be feature mat of training and testing
 7 |     """
 8 | 
 9 |     def __init__(self):
10 | 
11 |         # Trainging
12 |         self.learning_rate = 0.0025
13 |         self.lambda_loss_amount = 0.0015
14 |         self.training_epochs = 2000
15 |         self.batch_size = 1500
16 |         self.training_volume = 1
17 | 
18 |         # network structure
19 |         self.n_classes = 3
20 |         self.n_steps = 64
21 |         self.n_hidden = 100
22 |         self.dropout = 0.75
23 | 
24 | 
25 | 


--------------------------------------------------------------------------------
/create_figs_and_tables.m:
--------------------------------------------------------------------------------
  1 | %% create figures and tables 
  2 | % scken, 2021
  3 | % Copyright (C) 2021 Chair of Automation Technology / TU Chemnitz
  4 | 
  5 | clear all 
  6 | close all
  7 | 
  8 | n_dim = [512 1024 2048];
  9 | scale = [2 4 6 8];
 10 | encoding_dim = [20 40 60 80 100];
 11 | HDC_network = 'True';
 12 | dataset = 'full';
 13 | 
 14 | f1_tensor = zeros([3 3 3]);
 15 | 
 16 | %% iterate over all results of hyper-parameter analysis (table 3)
 17 | 
 18 | for d=1:numel(n_dim)
 19 |     for s=1:numel(scale)
 20 |         for e=1:numel(encoding_dim)
 21 |             % load file
 22 |             load(['results/' dataset '/results_HDC_' num2str(n_dim(d)) '_' num2str(scale(s)) '_' num2str(encoding_dim(e)) '_1.0.mat'])
 23 |             
 24 |             f1_tensor(d,s,e) = round(report.accuracy,2);
 25 |         end
 26 |     end
 27 | end
 28 | 
 29 | % create result table
 30 | subtable = {};
 31 | 
 32 | % create Rownames
 33 | rownames = {};
 34 | for s=1:numel(scale) 
 35 |     rownames{end +1} = ['scale = ' num2str(scale(s))];
 36 | end
 37 | 
 38 | % variable names 
 39 | varnames = {}; 
 40 | for e=1:numel(encoding_dim)
 41 |     varnames{end+1} = [num2str(encoding_dim(e))];
 42 | end
 43 | 
 44 | for d=1:numel(n_dim)
 45 |     subtable{d} = array2table(squeeze(f1_tensor(d,:,:)), ...
 46 |         'VariableNames',varnames);
 47 | end
 48 | 
 49 | tab = table(subtable{:},'VariableNames',{'\# Dimensions = 512', '\# Dimensions = 1024', '\# Dimensions = 2048'},'RowNames',rownames)
 50 | table2latex(tab,[3 2],0.5,'tables/results_vsa.tex')
 51 | 
 52 | %% evaluate model size (table 4)
 53 | 
 54 | param_array = zeros([numel(n_dim) numel(encoding_dim)]);
 55 | 
 56 | for d=1:numel(n_dim)
 57 |     for e=1:numel(encoding_dim)
 58 |         % load file
 59 |         load(['results/' dataset '/results_HDC_' num2str(n_dim(d)) '_8_' num2str(encoding_dim(e)) '_1.0.mat'])
 60 |         start_idx = findstr(model_summary,'Trainable params: ');
 61 |         start_idx = start_idx + 18;
 62 |         end_idx = findstr(model_summary,'Non-trainable');
 63 |         n_params = str2num(replace(model_summary(start_idx:end_idx-1),',',''));
 64 |         
 65 |         param_array(d,e) = n_params;
 66 |     end
 67 | end
 68 | 
 69 | 
 70 | %% plot different scales of fractional binding (fig. 2)
 71 | 
 72 | vsa = 'FHRR';
 73 | dim = 2048;
 74 | 
 75 | VSA = vsa_env('vsa',vsa,'dim',dim);
 76 | init_vector = VSA.add_vector();
 77 | values = -1.5:0.01:1.5;
 78 | line_type = {':';'-';'--';'-.'};
 79 | 
 80 | sim_values = zeros([numel(scale) numel(values)]);
 81 | 
 82 | leg = {};
 83 | 
 84 | for s=1:numel(scale)
 85 |     encoded_values = VSA.frac_binding(init_vector,values * scale(s));
 86 |     ref_value = VSA.frac_binding(init_vector,0);
 87 |     sim_values(s,:) = VSA.sim(ref_value, encoded_values);
 88 |     
 89 |     plot(values,sim_values(s,:),line_type{s})
 90 |     hold on 
 91 |     leg{end+1} = ['scaling = ' num2str(scale(s))];
 92 | end
 93 | 
 94 | grid on 
 95 | title('Similarity of encoded scalar value 0 to neighboring values')
 96 | xlabel('scalar value')
 97 | ylabel('similarity to encoded value 0')
 98 | legend(leg)
 99 | set(gcf,'color','w')
100 | saveas(gcf,'images/similarity_plot_frac_binding.png')
101 | export_fig('images/similarity_plot_frac_binding.pdf','-dpdf') 
102 | 
103 | 
104 | %% plot data efficiency (fig. 6)
105 | 
106 | figure()
107 | n_dim = 2048;
108 | scale = 10;
109 | encoding_dim = 40;
110 | dataset = 'full';
111 | training_volume = {'0.2'; '0.4'; '0.6'; '0.8'; '1.0'};
112 | f1_array_orig = [];
113 | f1_array_VSA = [];
114 | 
115 | % original network 
116 | for t=1:numel(training_volume)
117 |     % load file
118 |     load(['results/' dataset '/results_origNet_' training_volume{t} '.mat'])
119 | 
120 |     f1_array_orig(end+1) = report.accuracy;
121 | end
122 | 
123 | % VSA network 
124 | for t=1:numel(training_volume)
125 |     % load file
126 |     load(['results/' dataset '/results_HDC_' num2str(n_dim) '_' num2str(scale) '_' num2str(encoding_dim) '_' training_volume{t} '.mat'])
127 | 
128 |     f1_array_VSA(end+1) = report.accuracy;
129 | end
130 | 
131 | plot(str2num(cell2mat(training_volume)), f1_array_orig,'--','LineWidth',2)
132 | hold on 
133 | plot(str2num(cell2mat(training_volume)), f1_array_VSA,'-.','LineWidth',2)
134 | grid on 
135 | xlabel('training volume')
136 | ylabel('F_1 Score')
137 | title('Data efficiency')
138 | set(gcf,'color','w')
139 | set(gcf,'Position',[100 100 500 200])
140 | legend({'LSTM-ANN'; 'HDC-ANN'},'Location','SouthEast')
141 | saveas(gcf,'images/data_efficiency.png')
142 | export_fig('images/data_efficiency.pdf','-dpdf') 
143 | 
144 | 
145 | 
146 | 
147 | 


--------------------------------------------------------------------------------
/create_train_test_split_MATLAB.py:
--------------------------------------------------------------------------------
  1 | from __future__ import absolute_import
  2 | from __future__ import division
  3 | from __future__ import print_function
  4 | 
  5 | from argparse import ArgumentParser
  6 | from utils import *
  7 | from scipy.io import savemat, loadmat
  8 | from sklearn.model_selection import train_test_split
  9 | import numpy as np
 10 | from config import Config
 11 | 
 12 | 
 13 | if __name__ == '__main__':
 14 |     parser = ArgumentParser(description=__doc__)
 15 |     parser.add_argument('--dataset', '-d',
 16 |                         help='Which split of the dataset to train/test the model on?' \
 17 |                              '(i.e. full, motorway, secondary or full_crossval)',
 18 |                         default='full_crossval')
 19 |     parser.add_argument('--preproc',
 20 |                         default='1')
 21 |     parser.add_argument('--input_dim',
 22 |                         help='Defines the input dimension of the VSA model' \
 23 |                              '(possible values are 512, 1024 or 2048)',
 24 |                         default=2048)
 25 |     parser.add_argument('--scale',
 26 |                         help='scaling of the scalar encoding with fractional binding ' \
 27 |                              '(possible values are 2, 4, 6, 8 and 10)',
 28 |                         default=6)
 29 | 
 30 |     args = parser.parse_args()
 31 | 
 32 |     config = Config()
 33 |     config.input_dim = int(args.input_dim)
 34 |     config.scale = float(args.scale)
 35 | 
 36 |     if args.preproc=='1':
 37 |         data = load_dataset(args.dataset,config)
 38 |         if type(data[0]) == list:
 39 |             X_train = []
 40 |             X_test = []
 41 |             y_train = []
 42 |             y_test = []
 43 |             for i in range(len(data[0])):
 44 |                 config.n_time_measures = 1
 45 |                 config.n_inputs = data[0][i].shape[2]
 46 |                 config.n_steps = data[0][i].shape[1]
 47 |                 t_train, X_train_, traces_train, init_vecs = create_HDC_vectors(config, data[0][i])
 48 |                 t_test, X_test_, traces_test, init_vecs = create_HDC_vectors(config, data[1][i])
 49 | 
 50 |                 # normalize HDC encodings
 51 |                 m = np.mean(X_train_, axis=0)
 52 |                 s = np.std(X_train_, axis=0)
 53 |                 config.m = m
 54 |                 config.s = s
 55 |                 X_train_ = np.divide(X_train_ - m, s)
 56 |                 X_test_ = np.divide(X_test_ - m, s)
 57 | 
 58 |                 y_train_ = data[2][i]
 59 |                 y_test_ = data[3][i]
 60 | 
 61 |                 X_train.append(X_train_)
 62 |                 X_test.append(X_test_)
 63 |                 y_train.append(y_train_)
 64 |                 y_test.append(y_test_)
 65 | 
 66 |         else:
 67 |             config.n_time_measures = 1
 68 |             config.n_inputs = data[0].shape[2]
 69 |             config.n_steps = data[0].shape[1]
 70 |             t_train, X_train, traces_train, init_vecs = create_HDC_vectors(config, data[0])
 71 |             t_test, X_test, traces_test, init_vecs = create_HDC_vectors(config, data[1])
 72 | 
 73 |             # normalize HDC encodings
 74 |             m = np.mean(X_train, axis=0)
 75 |             s = np.std(X_train, axis=0)
 76 |             config.m = m
 77 |             config.s = s
 78 |             X_train = np.divide(X_train - m, s)
 79 |             X_test = np.divide(X_test - m, s)
 80 | 
 81 |             y_train = data[2]
 82 |             y_test = data[3]
 83 |     else:
 84 |         data = load_dataset(args.dataset, config)
 85 | 
 86 |         if type(data[0]) == list:
 87 |             X_train = []
 88 |             X_test = []
 89 |             y_train = []
 90 |             y_test = []
 91 |             for i in range(len(data[0])):
 92 |                 X_train_ = data[0][i]
 93 |                 X_test_ = data[1][i]
 94 |                 y_train_ = data[2][i]
 95 |                 y_test_ = data[3][i]
 96 | 
 97 |                 X_train.append(X_train_)
 98 |                 X_test.append(X_test_)
 99 |                 y_train.append(y_train_)
100 |                 y_test.append(y_test_)
101 |         else:
102 |             X_train = data[0]
103 |             X_test = data[1]
104 |             y_train = data[2]
105 |             y_test = data[3]
106 | 
107 |     savemat('temp_data.mat',{'X_train':X_train, 'X_test':X_test,'Y_train':y_train,'Y_test':y_test})
108 | 


--------------------------------------------------------------------------------
/data/README.md:
--------------------------------------------------------------------------------
 1 | # Data Description
 2 | The training/testing data has the shape (N x T x D) with labels of shape (N x C):
 3 | - N: The total number of the training/testing samples
 4 | - T: The times steps (64)
 5 | - D: The dimension of the data (9), which corresponds to [X_ACC, Y_ACC, Z_ACC, ROLL, PITCH, YAW, distance ahead, NoV, speed]
 6 | - C: The class of driving behaviour (1) [0-> normal, 1-> aggressive, 2-> drowsy]
 7 | 
 8 | ## Files 
 9 | 
10 | - .pkl files are the original preprocessed data sets of [1]
11 | - .mat files are the same data as in .pkl but in Matlab file format  
12 | 
13 | [1] K. Saleh, M. Hossny, and S. Nahavandi, “Driving behavior classification based on sensor data fusion using LSTM recurrent neural networks,” in International Conference on Intelligent Transportation Systems (ITSC), 2017.
14 | 


--------------------------------------------------------------------------------
/data/motorway_dataset_window_64_proc_veh_DtA.pkl:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/TUC-ProAut/HDC_driving_style_classification/21f957c484684641891a63c967fd99e5922c1e72/data/motorway_dataset_window_64_proc_veh_DtA.pkl


--------------------------------------------------------------------------------
/data/secondary_dataset_window_64_proc_veh_DtA.pkl:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/TUC-ProAut/HDC_driving_style_classification/21f957c484684641891a63c967fd99e5922c1e72/data/secondary_dataset_window_64_proc_veh_DtA.pkl


--------------------------------------------------------------------------------
/data/uah_dataset.mat:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/TUC-ProAut/HDC_driving_style_classification/21f957c484684641891a63c967fd99e5922c1e72/data/uah_dataset.mat


--------------------------------------------------------------------------------
/eval_baseline_models.m:
--------------------------------------------------------------------------------
 1 | %% classify all datasets with all baseline models 
 2 | % scken, 2021
 3 | % Copyright (C) 2021 Chair of Automation Technology / TU Chemnitz
 4 | 
 5 | clear all
 6 | close all
 7 | 
 8 | dim = 576;
 9 | frac_scale = 6;
10 | 
11 | 
12 | %% secondary 
13 | if 1
14 |     dataset = 'secondary';
15 |     Baseline_Models
16 | end
17 | 
18 | %% motorway
19 | if 1
20 |     dataset = 'motorway';
21 |     Baseline_Models
22 | end
23 | 
24 | %% full
25 | if 1
26 |     dataset = 'full';
27 |     Baseline_Models
28 | end
29 | 
30 | %% full dataset cross validation 
31 | if 1
32 |     dataset = 'full_crossval';
33 |     Baseline_Models
34 | end
35 | 
36 | 


--------------------------------------------------------------------------------
/getF1Score.m:
--------------------------------------------------------------------------------
 1 | function f1 = getF1Score(y_gt, y_pred)
 2 | % get the weighted F1 score
 3 | 
 4 | CM = confusionmat(double(y_gt),double(y_pred));
 5 | 
 6 | num_classes = size(CM,1);
 7 | [TP,TN,FP,FN,f1] = deal(zeros(num_classes,1));
 8 | for c = 1:num_classes
 9 |    TP(c) = CM(c,c);
10 |    tempMat = CM;
11 |    tempMat(:,c) = []; % remove column
12 |    tempMat(c,:) = []; % remove row
13 |    TN(c) = sum(sum(tempMat));
14 |    FP(c) = sum(CM(:,c))-TP(c);
15 |    FN(c) = sum(CM(c,:))-TP(c);
16 | end
17 | for c = 1:num_classes
18 |     f1(c) = 2*TP(c)/(2*TP(c) + FP(c) + FN(c));
19 | end
20 | 
21 | h=hist(double(y_gt),0:num_classes-1)';
22 | f1 = sum(f1.*(h/sum(h)));
23 | 
24 | end 


--------------------------------------------------------------------------------
/main.py:
--------------------------------------------------------------------------------
  1 | from __future__ import absolute_import
  2 | from __future__ import division
  3 | from __future__ import print_function
  4 | 
  5 | from argparse import ArgumentParser
  6 | from datetime import datetime
  7 | import logging
  8 | 
  9 | from utils import *
 10 | from HDC_ANN import main_HDC
 11 | from HDC_SNN import main_SNN
 12 | from LSTM import main_LSTM
 13 | from Concat_ANN import main_Concat_ANN
 14 | 
 15 | 
 16 | # config logger
 17 | logging.basicConfig(level=logging.INFO, format='%(message)s')
 18 | logger = logging.getLogger('log')
 19 | if not os.path.exists("./logs"):
 20 |     os.makedirs("./logs")
 21 | logger.addHandler(logging.FileHandler('./logs/main_log.log', 'a'))
 22 | stdout_handler = logging.StreamHandler()
 23 | logger.addHandler(stdout_handler)
 24 | 
 25 | 
 26 | if __name__ == '__main__':
 27 |     parser = ArgumentParser(description=__doc__)
 28 |     parser.add_argument('--dataset', '-d',
 29 |                         help='Which split of the dataset to train/test the model on?' \
 30 |                              '(i.e. full, motorway, secondary or full_crossval)',
 31 |                         default='full')
 32 |     parser.add_argument('--save_dir', '-s',
 33 |                         help='Directory of (to be)-saved model',
 34 |                         default='saves')
 35 |     parser.add_argument('--hyperparams_experiment', '-f',
 36 |                         help='Run complete experiments (hyper-parameter analysis of HDC approach) if set to true',
 37 |                         type=bool,
 38 |                         default=False)
 39 |     parser.add_argument('--data_efficiency',
 40 |                         help='Run data efficiency experiments if set to true',
 41 |                         type=bool,
 42 |                         default=False)
 43 |     parser.add_argument('--runtime_measurement',
 44 |                         help='Run inference time measurement with 10 repetitions.',
 45 |                         type=bool,
 46 |                         default=False)
 47 |     parser.add_argument('--test',
 48 |                         help='Start testing the saved model in $save_dir$ ' \
 49 |                              'othewrwise, it will start the training',
 50 |                         type=bool,
 51 |                         default=False)
 52 |     parser.add_argument('--input_dim',
 53 |                         help='Defines the input dimension of the HDC model' \
 54 |                              '(possible values are 512, 1024 or 2048)',
 55 |                         type=int,
 56 |                         default=2048)
 57 |     parser.add_argument('--scale',
 58 |                         help='scaling of the scalar encoding with fractional binding ' \
 59 |                              '(possible values are 2, 4, 6, 8 and 10)',
 60 |                         type=int,
 61 |                         default=6)
 62 |     parser.add_argument('--encoding_dim',
 63 |                         help='dimension of the first hidden layer (named encoding dimension)' \
 64 |                              ' possible values are 20, 40, 60, 80, 100)',
 65 |                         type=int,
 66 |                         default=40)
 67 |     parser.add_argument('--HDC_ANN',
 68 |                         help='Boolean value to train the HDC network (if true, the network ' \
 69 |                              'will be trained)',
 70 |                         type=bool,
 71 |                         default=False)
 72 |     parser.add_argument('--Concat_ANN',
 73 |                         help='Boolean value to train the concatenated network (if true, the network ' \
 74 |                              'will be trained)',
 75 |                         type=bool,
 76 |                         default=False)
 77 |     parser.add_argument('--LSTM',
 78 |                         help='Boolean value to train the LSTM network (if true, the network ' \
 79 |                              'will be trained)',
 80 |                         type=bool,
 81 |                         default=False)
 82 |     parser.add_argument('--HDC_SNN',
 83 |                         help='Boolean value to train the HDC network with SNN (if true, the network ' \
 84 |                              'will be trained)',
 85 |                         type=bool,
 86 |                         default=False)
 87 |     parser.add_argument('--training_volume',
 88 |                         help='To train with only a fraction of the training data. ' \
 89 |                              'Value is in range [0 1] (0 to 100 percentage).',
 90 |                         type=float,
 91 |                         default=1)
 92 |     parser.add_argument('--stat_iterations',
 93 |                         help="Number of repetitions",
 94 |                         default=1,
 95 |                         type=int)
 96 |     args = parser.parse_args()
 97 | 
 98 |     # check if result folder exists
 99 |     if not os.path.exists('results/full'):
100 |         os.makedirs('results/full')
101 |     if not os.path.exists('results/motorway'):
102 |         os.makedirs('results/motorway')
103 |     if not os.path.exists('results/secondary'):
104 |         os.makedirs('results/secondary')
105 |     if not os.path.exists('results/full_crossval'):
106 |         os.makedirs('results/full_crossval')
107 | 
108 |     # experiments to HDC classification
109 |     n_dim = [512, 1024, 2048]
110 |     scale = [2, 4, 6, 8, 10]
111 |     encoding_dim = [20, 40, 60, 80, 100]
112 | 
113 |     training_volume = 1.0
114 |     training_volume_range = [0.2, 0.4, 0.6, 0.8, 1.0]
115 | 
116 |     logger.info('_________________________' + str(datetime.now()))
117 | 
118 |     # HDC network
119 |     if args.HDC_ANN:
120 |         logger.info("---HDC Model---")
121 |         logger.info("- Dataset: " + args.dataset)
122 |         # multiple experiments based on the HDC approach
123 |         if args.hyperparams_experiment:
124 |             logger.info("##### Full experiment (hyper-parameter analysis)")
125 |             for d in range(len(n_dim)):
126 |                 for s in range(len(scale)):
127 |                     for e in range(len(encoding_dim)):
128 |                         args.input_dim = n_dim[d]
129 |                         args.scale = scale[s]
130 |                         args.encoding_dim = encoding_dim[e]
131 |                         args.training_volume = training_volume
132 | 
133 |                         logger.info("Training with " + str(n_dim[d]) + " " + str(scale[s]) + " " + str(
134 |                             encoding_dim[e]) + " training volume=" + str(training_volume))
135 |                         main_HDC(args)
136 |         elif args.data_efficiency:
137 |             logger.info("#### Training efficiency:")
138 |             logger.info("Config: input_dim = " + str(args.input_dim) + " scale = " + str(args.scale) +
139 |                         " encoding_dim = " + str(args.encoding_dim))
140 |             for t in range(len(training_volume_range)):
141 |                 args.training_volume = training_volume_range[t]
142 |                 logger.info("Training with training volume=" + str(training_volume_range[t]))
143 |                 main_HDC(args)
144 |         else:
145 |             logger.info("#### normal Training on " + args.dataset + ": ")
146 |             logger.info("Config: input_dim = " + str(args.input_dim) + " scale = " + str(args.scale) +
147 |                         " encoding_dim = " + str(args.encoding_dim) + " training_volume = " + str(args.training_volume))
148 |             main_HDC(args)
149 | 
150 |     # concatenate intput network
151 |     if args.Concat_ANN:
152 |         logger.info("---Concat Model---")
153 |         logger.info("- Dataset: " + args.dataset)
154 |         logger.info("#### normal Training on" + args.dataset + ": ")
155 |         main_Concat_ANN(args)
156 | 
157 |     # original LSTM network
158 |     if args.LSTM:
159 |         logger.info("---original LSTM Model---")
160 |         logger.info("- Dataset: " + args.dataset)
161 |         if args.data_efficiency:
162 |             logger.info("#### Training efficiency:")
163 |             for t in range(len(training_volume_range)):
164 |                 args.training_volume = training_volume_range[t]
165 |                 logger.info("Training with training volume=" + str(training_volume_range[t]))
166 |                 main_LSTM(args)
167 |         else:
168 |             logger.info("#### normal Training on " + args.dataset + ": ")
169 |             main_LSTM(args)
170 | 
171 |     # SNN network
172 |     if args.HDC_SNN:
173 |         logger.info("---SNN Model---")
174 |         logger.info("- Dataset: " + args.dataset)
175 |         # multiple experiments based on the HDC approach
176 |         if args.data_efficiency:
177 |             logger.info("#### Training efficiency:")
178 |             logger.info("Config: input_dim = " + str(args.input_dim) + " scale = " + str(args.scale) +
179 |                         " encoding_dim = " + str(args.encoding_dim) + " training_volume = " + str(args.training_volume))
180 |             for t in range(len(training_volume_range)):
181 |                 args.training_volume = training_volume_range[t]
182 |                 logger.info("Training with training volume=" + str(training_volume_range[t]))
183 |                 main_SNN(args)
184 |         else:
185 |             logger.info("#### normal Training on " + args.dataset + ": ")
186 |             logger.info("Config: input_dim = " + str(args.input_dim) + " scale = " + str(args.scale) +
187 |                         " encoding_dim = " + str(args.encoding_dim) + " training_volume = " + str(args.training_volume))
188 |             main_SNN(args)
189 | 


--------------------------------------------------------------------------------
/model.py:
--------------------------------------------------------------------------------
  1 | import numpy as np
  2 | import tensorflow as tf
  3 | import keras
  4 | from keras import layers
  5 | from keras.layers import *
  6 | import keras.backend as K
  7 | from utils import *
  8 | 
  9 | 
 10 | def HDC_ANN(config):
 11 |     """HDC feed-forward model
 12 |     """
 13 |     tf.config.optimizer.set_jit(True)
 14 |     encoding_dim = config.encoding_dim
 15 | 
 16 |     input = keras.Input(shape=(config.input_dim,))
 17 |     dropout = layers.Dropout(config.dropout)(input)
 18 |     encoded = layers.Dense(encoding_dim, activation='relu')(dropout)
 19 |     output = layers.Dense(config.n_classes, activation='softmax')(encoded)
 20 | 
 21 |     model = keras.Model(input, output)
 22 | 
 23 |     return model
 24 | 
 25 | def HDC_ANN_tf(input, config, init_vecs, W, biases):
 26 |     """ Tensorflow model of the HDC ANN """
 27 |     tf.config.optimizer.set_jit(True)
 28 | 
 29 |     # preprocessing
 30 |     preproc = HDC_tf_preproc(input, init_vecs)
 31 | 
 32 |     # normalize data
 33 |     norm_data = tf.divide(preproc - config.m,config.s)
 34 | 
 35 |     # neural network for classification
 36 |     #hidden = tf.nn.dropout(preproc, rate=config.dropout)
 37 |     hidden = tf.matmul(norm_data,W['hidden']) + biases['hidden']
 38 |     hidden = tf.nn.relu(hidden)
 39 | 
 40 |     output = tf.matmul(hidden,W['output']) + biases['output']
 41 |     output = tf.nn.softmax(output)
 42 | 
 43 |     print("total # of trainable parameters:" + str(
 44 |         np.sum([np.prod(v.get_shape().as_list()) for v in tf.compat.v1.trainable_variables()])))
 45 | 
 46 |     return output
 47 | 
 48 | def LSTM_Network(feature_mat, config):
 49 |     """model a LSTM Network, (borrowed from https://github.com/KhaledSaleh/driving_behaviour_classification)
 50 |       it stacks 2 LSTM layers, each layer has n_hidden=32 cells
 51 |        and 1 output layer, it is a full connet layer
 52 |       argument:
 53 |         feature_mat: ndarray fature matrix, shape=[batch_size,time_steps,n_inputs]
 54 |         config: class containing config of network
 55 |       return:
 56 |               : matrix  output shape [batch_size,n_classes]
 57 |     """
 58 |     W = {
 59 |         'hidden': tf.Variable(tf.random.normal([config.n_inputs, config.n_hidden]), name="W_hidden"),
 60 |         'output': tf.Variable(tf.random.normal([config.n_hidden, config.n_classes]), name="W_output")
 61 |     }
 62 |     biases = {
 63 |         'hidden': tf.Variable(tf.random.normal([config.n_hidden], mean=1.0), name="b_hidden"),
 64 |         'output': tf.Variable(tf.random.normal([config.n_classes]), name="b_output")
 65 |     }
 66 | 
 67 |     feature_mat = tf.transpose(feature_mat, [1, 0, 2])
 68 |     feature_mat = tf.reshape(feature_mat, [-1, config.n_inputs], name="features_reshape")
 69 |     hidden = tf.nn.relu(tf.matmul(
 70 |         feature_mat, W['hidden']
 71 |     ) + biases['hidden'])
 72 |     hidden = tf.split(hidden, config.n_steps, 0, name="input_hidden")
 73 |     lstm_cell = tf.compat.v1.nn.rnn_cell.BasicLSTMCell(config.n_hidden, forget_bias=1.0)
 74 |     lsmt_layers = tf.compat.v1.nn.rnn_cell.MultiRNNCell([lstm_cell] * 2)
 75 |     outputs, _ = tf.compat.v1.nn.static_rnn(lsmt_layers, hidden, dtype=tf.float32)
 76 |     lstm_last_output = outputs[-1]
 77 |     # Linear activation
 78 |     final_out = tf.add(tf.matmul(lstm_last_output, W['output']), biases['output'], name="logits")
 79 | 
 80 |     print("total # of trainable parameters:" + str(
 81 |         np.sum([np.prod(v.get_shape().as_list()) for v in tf.compat.v1.trainable_variables()])))
 82 | 
 83 |     return final_out
 84 | 
 85 | def HDC_tf_preproc(inputs, init_vecs):
 86 |     '''
 87 |     preprocessing function to create HDC vectors with tensorflow on GPU
 88 |     @param inputs: input tensor (#samples , #variables, #timesteps)
 89 |     @param init_vecs: initial hypervectors
 90 |     @return: context vectors
 91 |     '''
 92 |     init_vec = init_vecs['init_vec']
 93 |     sensor_ids = init_vecs['sensor_ids']
 94 |     timestamps = init_vecs['timestamps']
 95 |     scale = init_vecs['scale']
 96 | 
 97 |     tf.config.optimizer.set_jit(True)
 98 |     # fractional binding
 99 |     reshaped_input = tf.tile(tf.expand_dims(inputs, axis=3), [1,1,1,init_vec.shape[0]])
100 | 
101 |     expand_init_vec = tf.expand_dims(K.expand_dims(tf.transpose(init_vec * scale), axis=0), axis=0)
102 |     reshaped_init_vec = tf.tile(expand_init_vec, [1,reshaped_input.shape[1],1,1])
103 |     reshaped_init_vec = tf.tile(reshaped_init_vec, [1,1,reshaped_input.shape[2],1])
104 |     # fractional binding with scale
105 |     encoded_scalars = tf.math.multiply(reshaped_input, reshaped_init_vec)
106 | 
107 |     # bind to sensors
108 |     sensor_ids = tf.transpose(sensor_ids, (1, 0))
109 |     sensor_ids = tf.expand_dims(tf.expand_dims(sensor_ids, axis=0), axis=0)
110 |     sensor_ids = tf.tile(sensor_ids, [1,encoded_scalars.shape[1],1,1])
111 |     sensor_vals = tf.add(encoded_scalars, sensor_ids)
112 | 
113 |     # bundle all sensor vectors
114 |     vals_cos = tf.cos(sensor_vals)
115 |     vals_sin = tf.sin(sensor_vals)
116 |     sensor_bundle_cos = (tf.reduce_sum(vals_cos, axis=2))
117 |     sensor_bundle_sin = (tf.reduce_sum(vals_sin, axis=2))
118 |     sensor_bundle = tf.math.atan2(sensor_bundle_sin, sensor_bundle_cos)
119 | 
120 |     # encode temporal context
121 |     timestamps = tf.transpose(timestamps, (1, 0))
122 |     timestamps = tf.expand_dims(timestamps, axis=0)
123 |     context_vecs = tf.add(sensor_bundle, timestamps)
124 | 
125 |     # bundle temporal context
126 |     complex_context_cos = tf.cos(context_vecs)
127 |     complex_context_sin = tf.sin(context_vecs)
128 |     context_bundle_cos = (tf.reduce_sum(complex_context_cos, axis=1))
129 |     context_bundle_sin = (tf.reduce_sum(complex_context_sin, axis=1))
130 |     context_bundle = tf.math.atan2(context_bundle_sin, context_bundle_cos)
131 | 
132 |     return context_bundle


--------------------------------------------------------------------------------
/requirements.txt:
--------------------------------------------------------------------------------
1 | scipy==1.4.1
2 | nengo==3.1.0
3 | Keras==2.4.3
4 | tensorflow>=2.4.2
5 | nengo_dl==3.4.0
6 | matplotlib==3.1.2
7 | numpy==1.17.4
8 | scikit_learn==0.24.2
9 | 


--------------------------------------------------------------------------------
/utils.py:
--------------------------------------------------------------------------------
  1 | import numpy as np
  2 | import pickle
  3 | from sklearn.model_selection import train_test_split
  4 | from sklearn.model_selection import KFold
  5 | import os
  6 | import tensorflow as tf
  7 | from tensorflow.python.client import timeline
  8 | import io
  9 | from scipy.io import savemat, loadmat
 10 | from tensorflow.keras.callbacks import Callback
 11 | import time
 12 | from tensorflow.keras.layers import Input, Dense, LSTM, multiply, concatenate, Activation, Masking, Reshape
 13 | from tensorflow.keras.layers import Conv1D, BatchNormalization, GlobalAveragePooling1D, Permute, Dropout
 14 | from tensorflow.python.eager.context import context, EAGER_MODE, GRAPH_MODE, eager_mode, graph_mode
 15 | 
 16 | from model import HDC_tf_preproc
 17 | 
 18 | def extract_batch_size(_train, step, batch_size):
 19 |     # Function to fetch a "batch_size" amount of data from "(X|y)_train" data. 
 20 |     
 21 |     shape = list(_train.shape)
 22 |     shape[0] = batch_size
 23 |     batch_s = np.empty(shape)
 24 | 
 25 |     for i in range(batch_size):
 26 |         # Loop index
 27 |         index = ((step-1)*batch_size + i) % len(_train)
 28 |         batch_s[i] = _train[index] 
 29 | 
 30 |     return batch_s
 31 | 
 32 | def one_hot(y_,n_classes=-1):
 33 |     # Function to encode output labels from number indexes 
 34 |     # e.g.: [[5], [0], [3]] --> [[0, 0, 0, 0, 0, 1], [1, 0, 0, 0, 0, 0], [0, 0, 0, 1, 0, 0]]
 35 | 
 36 |     y_ = y_.reshape(len(y_))
 37 |     if n_classes==-1:
 38 |         n_values = np.max(y_) + 1
 39 |     else:
 40 |         n_values = n_classes
 41 |     return np.eye(n_values)[np.array(y_, dtype=np.int32)]  # Returns FLOATS
 42 | 
 43 | def load_motorway_dataset(data_path='data'):
 44 |     # Function to load the motorway dataset only 
 45 | 
 46 |     with open(os.path.join(data_path, 'motorway_dataset_window_64_proc_veh_DtA.pkl'), 'rb') as f:
 47 |         save = pickle.load(f, encoding='latin1')
 48 |         motorway_dataset = save['dataset']
 49 |         motorway_labels = save['labels']
 50 |         del save
 51 |         print('Motorway set', motorway_dataset.shape, motorway_labels.shape)
 52 | 
 53 |     X_train, X_test, y_train, y_test = train_test_split(motorway_dataset, motorway_labels, test_size=0.33, random_state=42)
 54 | 
 55 |     return X_train, X_test, y_train, y_test
 56 | 
 57 | def load_secondary_dataset(data_path='data'):
 58 |     # Function to load the secondary dataset only 
 59 | 
 60 |     with open(os.path.join(data_path,'secondary_dataset_window_64_proc_veh_DtA.pkl'), 'rb') as f:
 61 |         save = pickle.load(f, encoding='latin1')
 62 |         secondary_dataset = save['dataset']
 63 |         secondary_labels = save['labels']
 64 |         del save
 65 |         print('Secondary set', secondary_dataset.shape, secondary_labels.shape)
 66 | 
 67 |     X_train, X_test, y_train, y_test = train_test_split(secondary_dataset, secondary_labels, test_size=0.33, random_state=42)
 68 | 
 69 |     return X_train, X_test, y_train, y_test
 70 | 
 71 | def load_full_dataset(data_path='data'):
 72 |     # Function to load the full dataset (motorway+secondary roads)
 73 | 
 74 |     with open(os.path.join(data_path, 'motorway_dataset_window_64_proc_veh_DtA.pkl'), 'rb') as f:
 75 |         save = pickle.load(f, encoding='latin1')
 76 |         motorway_dataset = save['dataset']
 77 |         motorway_labels = save['labels']
 78 |         del save
 79 |         print('Motorway set', motorway_dataset.shape, motorway_labels.shape)
 80 | 
 81 |     with open(os.path.join(data_path,'secondary_dataset_window_64_proc_veh_DtA.pkl'), 'rb') as f:
 82 |         save = pickle.load(f, encoding='latin1')
 83 |         secondary_dataset = save['dataset']
 84 |         secondary_labels = save['labels']
 85 |         del save
 86 |         print('Secondary set', secondary_dataset.shape, secondary_labels.shape)
 87 | 
 88 |     dataset = np.concatenate((motorway_dataset,secondary_dataset), axis=0)
 89 |     labels = np.concatenate((motorway_labels,secondary_labels), axis=0)
 90 | 
 91 |     X_train, X_test, y_train, y_test = train_test_split(dataset, labels, test_size=0.33, random_state=42)
 92 | 
 93 |     return X_train, X_test, y_train, y_test
 94 | 
 95 | def get_model_summary(model):
 96 |     stream = io.StringIO()
 97 |     model.summary(print_fn=lambda x: stream.write(x + '\n'))
 98 |     summary_string = stream.getvalue()
 99 |     stream.close()
100 |     return summary_string
101 | 
102 | class TimingCallback(Callback):
103 |     def __init__(self):
104 |         self.logs=[]
105 |     def on_epoch_begin(self,epoch, logs={}):
106 |         self.starttime=time.time()
107 |     def on_epoch_end(self, epoch, logs={}):
108 |         self.logs.append(time.time()-self.starttime)
109 | 
110 | 
111 | def load_dataset(dataset,config):
112 |     """
113 |     load the specific data set (from the data/ folder)
114 |     @param dataset: specifies the data set [string]
115 |     @param config: configure struct with necessary parameters [struct]
116 |     @param hdc_encoded: if the data set to be loaded is to be HDC-coded [Bool]
117 |     @return: set of training and test data [list]
118 |     """
119 |     # load preprocessed data
120 |     if dataset == "full":
121 |         X_train, X_test, y_train, y_test = load_full_dataset()
122 |     elif dataset == "motorway":
123 |         X_train, X_test, y_train, y_test = load_motorway_dataset()
124 |     elif dataset == "secondary":
125 |         X_train, X_test, y_train, y_test = load_secondary_dataset()
126 |     elif dataset == "full_crossval":
127 |         data = loadmat('data/uah_dataset.mat')
128 |         mot_data = data['motorway_dataset']
129 |         sec_data = data['secondary_dataset']
130 |         mot_label = data['motorway_labels']
131 |         sec_label = data['secondary_labels']
132 | 
133 |         # creat cross-validation split (3-fold cross-validation) on the full dataset
134 |         k = 3
135 |         X_train = []
136 |         X_test = []
137 |         y_train = []
138 |         y_test = []
139 |         #info = data['data_info']
140 | 
141 |         kf = KFold(n_splits=k)
142 |         #motorway
143 |         for train_idx, test_idx in kf.split(mot_data):
144 |             X_train.append(mot_data[train_idx])
145 |             y_train.append(mot_label[train_idx])
146 |             X_test.append(mot_data[test_idx])
147 |             y_test.append(mot_label[test_idx])
148 |         k_idx = 0
149 |         # secondary
150 |         for train_idx, test_idx in kf.split(sec_data):
151 |             X_train[k_idx]=np.concatenate((X_train[k_idx],sec_data[train_idx]),0)
152 |             y_train[k_idx]=np.concatenate((y_train[k_idx],sec_label[train_idx]), 0)
153 |             X_test[k_idx]=np.concatenate((X_test[k_idx],sec_data[test_idx]),0)
154 |             y_test[k_idx]=np.concatenate((y_test[k_idx],sec_label[test_idx]), 0)
155 |             k_idx+=1
156 |     else:
157 |         print("No valid dataset argument was set!")
158 | 
159 |     return X_train, X_test, y_train, y_test, config
160 | 
161 | 
162 | def create_HDC_vectors(config, input):
163 |     """
164 |     create the HDC vectors from given input
165 |     @param config: config struct
166 |     @param input: inputs tensor with size m x t x v (m... number of samples, t... number of timesteps, v... number of variables)
167 |     """
168 |     with graph_mode():
169 |         tf.config.optimizer.set_jit(True)
170 |         # pre initialize vectors
171 |         init_vec = tf.random.uniform(shape=(config.input_dim, 1), minval=-np.pi, maxval=np.pi, seed=1,
172 |                                      dtype="float32")
173 |         sensor_ids = tf.random.uniform(shape=(config.input_dim, config.n_inputs), minval=-np.pi, maxval=np.pi,
174 |                                        seed=2,
175 |                                        dtype="float32")
176 |         timestamps = tf.random.uniform(shape=(config.input_dim, config.n_steps), minval=-np.pi, maxval=np.pi,
177 |                                        seed=3,
178 |                                        dtype="float32")
179 |         init_vecs = {'init_vec': init_vec, 'sensor_ids': sensor_ids, 'timestamps': timestamps, 'scale':config.scale}
180 | 
181 |         X = tf.compat.v1.placeholder(tf.float32, [None, config.n_steps, config.n_inputs], name="X")
182 |         preproc = HDC_tf_preproc(X, init_vecs)
183 |         t_proc = []
184 |         traces = []
185 | 
186 |         for i in range(config.n_time_measures):
187 |             sess = tf.compat.v1.Session()
188 |             options = tf.compat.v1.RunOptions(trace_level=tf.compat.v1.RunOptions.FULL_TRACE)
189 |             run_metadata = tf.compat.v1.RunMetadata()
190 |             t = time.perf_counter()
191 |             output = sess.run(preproc, feed_dict={X: input}, options=options, run_metadata=run_metadata)
192 |             t_proc.append((time.perf_counter() - t))
193 |             fetched_timeline = timeline.Timeline(run_metadata.step_stats)
194 |             chrome_trace = fetched_timeline.generate_chrome_trace_format()
195 |             traces.append(chrome_trace)
196 |         preprocessing_time = np.median(t_proc)
197 | 
198 |     return preprocessing_time, output, traces, init_vecs


--------------------------------------------------------------------------------