''' (Alessio Lugnan, end of July 2025) Here I try classification using only a single pixel at each time. In this version, each pixel defines a different dataset ''' import numpy as np # from matplotlib import pyplot as plt from sklearn.linear_model import LogisticRegression from sklearn.preprocessing import StandardScaler import mat73 import scipy.io as sio from skimage.measure import block_reduce # import time n_train=55000 # pix_cutoff = 80 # pixel index from which to start classification (main results were calculated using 80) pix_cutoff = 0 # pixel index from which to start classification (main results were calculated using 80) # ind_selected_pow = 5 n_ports = 6 n_classes = 10 freq_ind = np.arange(2, 12) # freq_ind = np.arange(5, 7) pow_ind = np.arange(1, 6) # pow_ind = np.arange(4, 5) n_freq = 11 n_pow = len(pow_ind) scaler = StandardScaler() # folder_path = '/media/alessio_lugnan/Elements/foradori_parsed/X4Y1_processed_110725/awg_mnist/' # AWGdata_path = '/media/alessio_lugnan/Elements/awg_mnist_010324.mat' # folder_path = '/media/alessio_lugnan/Elements/foradori_parsed/X4Y1_processed_100725/awg_mnist_double/' # AWGdata_path = '/media/alessio_lugnan/Elements/awg_mnist_double_140524.mat' # folder_path = '/media/alessio_lugnan/Elements/foradori_parsed/X4Y1_processed_100725/awg_mnist_box/' # AWGdata_path = '/media/alessio_lugnan/Elements/awg_mnist_box_100524.mat' # folder_path = '/media/alessio_lugnan/Elements/foradori_parsed/X4Y1_processed_100725/awg_mnist_double_box/' # AWGdata_path = '/media/alessio_lugnan/Elements/awg_mnist_double_box_140524.mat' folder_path = '/media/alessio_lugnan/Elements/foradori_parsed/X4Y1_processed_100725/awg_mnist_inverse/' AWGdata_path = '/media/alessio_lugnan/Elements/awg_mnist_inverse_280524.mat' # folder_path = '/media/alessio_lugnan/Elements/foradori_parsed/X4Y1_processed_100725/awg_mnist_double_inverse/' # AWGdata_path = '/media/alessio_lugnan/Elements/awg_mnist_double_inverse_280524.mat' # folder_path = '/media/alessio_lugnan/Elements/foradori_parsed/X4Y1_processed_110725/awg_fashion_mnist/' # AWGdata_path = '/media/alessio_lugnan/Elements/awg_fashion_mnist_040324.mat' # folder_path = '/media/alessio_lugnan/Elements/foradori_parsed/X4Y1_processed_110725/awg_fashion_mnist_double/' # AWGdata_path = '/media/alessio_lugnan/Elements/awg_fashion_mnist_double_280524.mat' # load and visualize original AWG pattern AWGdata = sio.loadmat(AWGdata_path) y = AWGdata['label'][0] # NB: here the 100 is a warm start index, and was already considered for the X array in the parsing code np.random.seed(seed=15) ind_start = np.arange(len(y)) ind_shuf = ind_start.copy() y = y[ind_shuf] i_port = 0 i_freq = 0 ind_port=1 ind_freq=3 # ind_pow=5 downsample_factor = 3 ind_selected_pow=5 def get_samplesets_1pix( n_train, folder_path, ind_shuf, y, n_ports, ind_selected_pow, downsample_factor=1, pix_cutoff=pix_cutoff ): '''Load and preprocess datasets from a given measurement. This version is for data parsed by Alessandro Foradori''' port_numbers = [1, 2, 3, 5, 7, 9] count = 0 for i_port in range(0,1): for i_freq in range(0,1): print(f'Load i_port {i_port} i_freq {i_freq}, for i_pow {ind_selected_pow-1}') filename = f'processed_iPort{port_numbers[i_port]}_iFreq{i_freq+1}_iPow{ind_selected_pow}_iSet0.npz' try: X = np.load(folder_path + filename)['features'] if downsample_factor != 1: X = block_reduce(X, block_size=(1,downsample_factor), func=np.mean, cval=np.mean(X)).astype('int16') X = X[ind_shuf] X = X[:,pix_cutoff:] n_pix = np.shape(X)[1] if count == 0: X_sp = X[:,:,None] else: X_sp = np.concatenate( [X_sp, X[:,:,None]], axis=2) count += 1 except: pass X_train = X_sp[:-10000, :].copy() X_test = X_sp[-10000:, :].copy() y_train = y[:-10000].copy() y_test = y[-10000:].copy() # split training and validation sets X_traintrain = X_train.copy()[:n_train] X_val = X_train.copy()[n_train:] y_traintrain = y_train.copy()[:n_train] y_val = y_train.copy()[n_train:] # return X_traintrain[:,:10], X_val[:,:10], X_test[:,:10], y_traintrain, y_val, y_test # line for debug return X_traintrain, X_val, X_test, y_traintrain, y_val, y_test, n_pix ############# check classification ################### # scores_sklearn = [ [] for i_pow in range(n_pow) ] scores_sklearn = [] for i_pow in range(4,5): X_traintrain, X_val, X_test, y_traintrain, y_val, y_test, n_pix = get_samplesets_1pix( n_train=n_train, folder_path=folder_path, ind_shuf=ind_shuf, y=y, n_ports=n_ports, ind_selected_pow=i_pow+1, downsample_factor=downsample_factor, pix_cutoff=pix_cutoff) # classifier = models.Sequential( # [ # layers.Dense(n_classes, activation='softmax'), # ] # ) # classifier.compile(optimizer='Adam', loss='sparse_categorical_crossentropy', metrics=['accuracy']) # scores = [] # t0 = time.time() # predictions_val = [] # predictions_test = [] # predictions_train = [] # for i_pix in range(n_pix): # epochs_info = classifier.fit(X_traintrain[:,i_pix,:], # y_traintrain, # batch_size=batch_size, # epochs=100, # callbacks=[early_stopping_callback], # validation_split=0.15, # verbose=0) # # classifier.summary() # _, score_train = classifier.evaluate(X_traintrain[:,i_pix,:], y_traintrain, verbose=0) # _, score_test = classifier.evaluate(X_test[:,i_pix,:], y_test, verbose=0) # _, score = classifier.evaluate(X_val[:,i_pix,:], y_val, verbose=0) # print(f'Scores for i_pix {i_pix}: val {score}, train {score_train}, test {score_test}') # scores.append( [score, score_train, score_test] ) # predictions_val.append( classifier.predict(X_val[:,i_pix,:], verbose=0 ) ) # predictions_test.append(classifier.predict(X_test[:, i_pix, :], verbose=0)) # predictions_train.append(classifier.predict(X_traintrain[:, i_pix, :], verbose=0)) # # scores = np.array(scores) # predictions_val = np.array(predictions_val) # predictions_test = np.array(predictions_test) # predictions_train = np.array(predictions_train) # t1 = time.time() # delta_time = t1-t0 # print(f'Best val accuracy is for pixel {np.argmax(scores[:,0])}, yelding {scores[np.argmax(scores[:,0])]}, over a time {delta_time}') ##### try with sklearn logistic reg classifier = LogisticRegression(penalty='l2', dual=False, tol=0.0001, C=1e6, fit_intercept=True, intercept_scaling=1, class_weight=None, random_state=None, solver='lbfgs', max_iter=10000, verbose=0, warm_start=False, n_jobs=None, l1_ratio=None) # t0 = time.time() predictions_skl_val = [] predictions_skl_test = [] predictions_skl_train = [] for i_pix in range(n_pix): epochs_info = classifier.fit(scaler.fit_transform(X_traintrain[:,i_pix,:]), y_traintrain) # classifier.summary() score_train = classifier.score(scaler.transform(X_traintrain[:,i_pix,:]), y_traintrain) score_test = classifier.score(scaler.transform(X_test[:,i_pix,:]), y_test) score = classifier.score(scaler.transform(X_val[:,i_pix,:]), y_val) print(f'Sklearn Scores for i_pix {i_pix}: val {score}, train {score_train}, test {score_test}') scores_sklearn.append( [score, score_train, score_test] ) # predictions_skl_val.append( classifier.predict_proba(scaler.transform(X_val[:,i_pix,:]) ) ) # predictions_skl_test.append(classifier.predict_proba(scaler.transform(X_test[:, i_pix, :]))) # predictions_skl_train.append(classifier.predict_proba(scaler.transform(X_traintrain[:, i_pix, :]))) scores_sklearn = np.array(scores_sklearn) # predictions_skl_val = np.array(predictions_skl_val) # predictions_skl_test = np.array(predictions_skl_test) # predictions_skl_train = np.array(predictions_skl_train) # t1 = time.time() # delta_time_sklearn = t1-t0 print(f'Best val accuracy for i_pow {i_pow} is for pixel {np.argmax(scores_sklearn[:,0])}, ' f'yelding {scores_sklearn[np.argmax(scores_sklearn[:,0])]}') scores_sklearn = np.array(scores_sklearn) # np.savez(f'Results_ML_singlePix2_ExpBASELINE_Mnist10ns_nTrain{n_train}_accu{scores_sklearn[np.argmax(scores_sklearn[:,2])][2]}', # scores_sklearn=scores_sklearn, # folder_path=folder_path, # n_train=n_train, pix_cutoff=pix_cutoff) np.savez(f'Results_ML_singlePix2_ExpBASELINE_NOpIXcUTOFF_Mnist10ns_nTrain{n_train}', scores_sklearn=scores_sklearn, folder_path=folder_path, n_train=n_train, pix_cutoff=pix_cutoff) # ind_best = np.unravel_index(np.argmax(scores_sklearn[:,:,2]), np.shape(scores_sklearn[:,:,2])) # score_tuple_best = scores_sklearn[ind_best] # print(f'Best score for i_pow {ind_best[0]}, pixel {ind_best[1]}: {score_tuple_best}') ############################ train classifier over 1st classifier's output ################################### # Xf_train = np.reshape(np.swapaxes(predictions_skl_train, 0, 1), # (len(predictions_skl_train[0]), len(predictions_skl_train)*len(predictions_skl_train[0,0]))) # Xf_val = np.reshape(np.swapaxes(predictions_skl_val, 0, 1), # (len(predictions_skl_val[0]), len(predictions_skl_val)*len(predictions_skl_val[0,0]))) # Xf_test = np.reshape(np.swapaxes(predictions_skl_test, 0, 1), # (len(predictions_skl_test[0]), len(predictions_skl_test)*len(predictions_skl_test[0,0]))) # # reg_strengths = np.logspace(-2,6,13) # scores_reg = [] # for i_reg in range(len(reg_strengths)): # classifier_f = LogisticRegression(penalty='l2', dual=False, tol=0.0001, C=1/reg_strengths[i_reg], fit_intercept=True, intercept_scaling=1, class_weight=None, random_state=None, # solver='lbfgs', max_iter=10000, verbose=0, warm_start=False, n_jobs=None, l1_ratio=None) # # t1_f = time.time() # epochs_info = classifier_f.fit(scaler.fit_transform(Xf_train), y_traintrain) # scoreF_train = classifier_f.score(scaler.transform(Xf_train), y_traintrain) # scoreF_test = classifier_f.score(scaler.transform(Xf_test), y_test) # scoreF = classifier_f.score(scaler.transform(Xf_val), y_val) # scores_reg.append(scoreF) # # t2_f = time.time() # print(f'Final reg. optimization with reg. strength {reg_strengths[i_reg]}: val {scoreF}, train {scoreF_train}, test {scoreF_test}')