def computeLinearSVMscore():
out_path = path_to_output_dir + '18_09_04'
strDims = 'dim100'
x_test, y_test = bastien_utils.load_data(PATH_TO_DATA,
train=False,
test=True,
subsetTest=False)
models = listdir(path_to_model_dir)
svm_score = np.zeros((8, 4))
svm_std = np.zeros((8, 4))
svm_best_params_C = np.zeros((8, 4))
objectives = [0.01, 0.05, 0.1, 0.2]
weights = [0.0001, 0.0005, 0.001, 0.005, 0.01, 0.05, 0.1, 0.5]
for m in models:
print('Model= ', m)
sp_o = m.split('_')[-2]
sp_w = m.split('_')[-3]
print('sp_w: ', sp_w, ' - sp_o: ', sp_o)
idx_w = weights.index(float(sp_w))
idx_o = objectives.index(float(sp_o))
print('idx_w: ', idx_w, ' - idx_o: ', idx_o, '\n')
shAE = AE_class.load(m)
res = shAE.best_linearSVM_classification_score(x_test,
y_test,
nb_values_C=30)
svm_score[idx_w, idx_o] = res[0]
svm_std[idx_w, idx_o] = res[1]
svm_best_params_C[idx_w, idx_o] = res[2]['C']
np.save(out_path + '_svm_score_' + strDims, svm_score)
np.save(out_path + '_svm_std_' + strDims, svm_std)
np.save(out_path + '_svm_best_param_C_' + strDims, svm_best_params_C)
def recomputeMaxApprox():
"""
Creating a super class for all auto-encoders would enable not to worry about the model being of a shallowAE or a AsymAE
Note that a shallowAE eventhough
"""
out_path = path_to_output_dir + '18_09_04'
strDims = 'dim100'
x_train, _, x_test, _ = bastien_utils.load_data(PATH_TO_DATA,
train=True,
test=True,
subsetTest=False)
models = listdir(path_to_model_dir)
max_approx_error_toOriginal_train = np.zeros((8, 4))
max_approx_error_toOriginal_test = np.zeros((8, 4))
max_approx_error_toRec_train = np.zeros((8, 4))
max_approx_error_toRec_test = np.zeros((8, 4))
objectives = [0.01, 0.05, 0.1, 0.2]
weights = [0.0001, 0.0005, 0.001, 0.005, 0.01, 0.05, 0.1, 0.5]
for m in models:
print('Model= ', m)
sp_o = m.split('_')[-2]
sp_w = m.split('_')[-3]
print('sp_w: ', sp_w, ' - sp_o: ', sp_o)
idx_w = weights.index(float(sp_w))
idx_o = objectives.index(float(sp_o))
print('idx_w: ', idx_w, ' - idx_o: ', idx_o, '\n')
shAE = AE_class.load(m)
max_approx_train = shAE.max_approximation_error(
x_train,
morphoMaths.dilatation,
original_images=x_train,
apply_to_bias=False,
SE_scale=1)
max_approx_error_toOriginal_train[idx_w, idx_o] = max_approx_train[0]
max_approx_error_toRec_train[idx_w, idx_o] = max_approx_train[1]
max_approx_test = shAE.max_approximation_error(x_test,
morphoMaths.dilatation,
original_images=x_test,
apply_to_bias=False,
SE_scale=1)
max_approx_error_toOriginal_test[idx_w, idx_o] = max_approx_test[0]
max_approx_error_toRec_test[idx_w, idx_o] = max_approx_test[1]
np.save(
out_path + '_training_max_approx_error_toOriginal_dilatation_' +
strDims, max_approx_error_toOriginal_train)
np.save(out_path + '_test_max_approx_error_toOriginal_dilation_' + strDims,
max_approx_error_toOriginal_test)
np.save(
out_path + '_training_max_approx_error_toRec_dilatation_' + strDims,
max_approx_error_toRec_train)
np.save(out_path + '_test_max_approx_error_toRec_dilation_' + strDims,
max_approx_error_toRec_test)
def test_Hoyer_Asym_AE(
AsymAE_class=Sparse_NonNeg_AsymAEinfoGAN_Hoyer_NonNegConstraint,
sparsity_weights=[1],
sparsity_objectives=[0.6],
latent_dimension=100,
nb_epochs=500,
nb_input_channels=1,
one_channel_output=True,
add_original_images=True,
AMD=False,
PADO=False,
AMD_step=1,
AMD_init_step=1,
svm=False,
path_to_dir="../Results/AsymAE_infoGAN/"):
original_images_train, _, original_images_test, y_test = bastien_utils.load_data(
PATH_TO_DATA, train=True, test=True, subsetTest=False)
if (nb_input_channels > 1):
if AMD:
if (add_original_images & (nb_input_channels > 2)):
x_train = morphoMaths.AMD_in_one_array(
original_images_train[:, :, :, 0],
levels=nb_input_channels - 2,
step=AMD_step,
init_step=AMD_init_step,
add_original_images=add_original_images)
x_test = morphoMaths.AMD_in_one_array(
original_images_test[:, :, :, 0],
levels=nb_input_channels - 2,
step=AMD_step,
init_step=AMD_init_step,
add_original_images=add_original_images)
path_to_dir = path_to_dir + "/SeveralChannels/WithAMD/"
else:
x_train = morphoMaths.AMD_in_one_array(
original_images_train[:, :, :, 0],
levels=nb_input_channels - 1,
step=AMD_step,
init_step=AMD_init_step,
add_original_images=add_original_images)
x_test = morphoMaths.AMD_in_one_array(
original_images_test[:, :, :, 0],
levels=nb_input_channels - 1,
step=AMD_step,
init_step=AMD_init_step,
add_original_images=add_original_images)
path_to_dir = path_to_dir + "/SeveralChannels/WithAMD_NoOriginals/"
x_train = bastien_utils.rescale_all_channels_between_0_and_1(
x_train)
x_test = bastien_utils.rescale_all_channels_between_0_and_1(x_test)
else:
if PADO:
if (add_original_images & (nb_input_channels > 2)):
x_train = morphoMaths.positive_decomposition_by_openings_by_rec(
original_images_train[:, :, :, 0],
levels=nb_input_channels - 2,
step=AMD_step,
init_step=AMD_init_step,
add_original_images=add_original_images)
x_test = morphoMaths.positive_decomposition_by_openings_by_rec(
original_images_test[:, :, :, 0],
levels=nb_input_channels - 2,
step=AMD_step,
init_step=AMD_init_step,
add_original_images=add_original_images)
path_to_dir = path_to_dir + "/SeveralChannels/WithPADO/"
else:
x_train = morphoMaths.positive_decomposition_by_openings_by_rec(
original_images_train[:, :, :, 0],
levels=nb_input_channels - 1,
step=AMD_step,
init_step=AMD_init_step,
add_original_images=add_original_images)
x_test = morphoMaths.positive_decomposition_by_openings_by_rec(
original_images_test[:, :, :, 0],
levels=nb_input_channels - 1,
step=AMD_step,
init_step=AMD_init_step,
add_original_images=add_original_images)
path_to_dir = path_to_dir + "/SeveralChannels/WithPADO_NoOriginals/"
x_train = bastien_utils.rescale_all_channels_between_0_and_1(
x_train)
x_test = bastien_utils.rescale_all_channels_between_0_and_1(
x_test)
else:
x_train = np.tile(x_train, (1, 1, 1, nb_input_channels))
x_test = np.tile(x_test, (1, 1, 1, nb_input_channels))
path_to_dir = path_to_dir + "/SeveralChannels/NoAMD/"
else:
x_train = original_images_train
x_test = original_images_test
if not one_channel_output:
original_images_train = x_train
original_images_test = x_test
d = datetime.date.today()
strDims = 'dim' + str(latent_dimension)
strDate = d.strftime("%y_%m_%d")
out_path = path_to_dir + "/Sparse_NonNeg/Hoyer_NonNegConstraint/TestOutputs/" + strDate
nb_sparsity_weights = len(sparsity_weights)
nb_sparsity_objectives = len(sparsity_objectives)
train_rec_errors = np.zeros((nb_sparsity_weights, nb_sparsity_objectives))
test_rec_errors = np.zeros((nb_sparsity_weights, nb_sparsity_objectives))
train_Hoyer_loss = np.zeros((nb_sparsity_weights, nb_sparsity_objectives))
test_Hoyer_loss = np.zeros((nb_sparsity_weights, nb_sparsity_objectives))
sparsity_train = np.zeros((nb_sparsity_weights, nb_sparsity_objectives))
sparsity_test = np.zeros((nb_sparsity_weights, nb_sparsity_objectives))
max_approx_error_toOriginal_train = np.zeros(
(nb_sparsity_weights, nb_sparsity_objectives))
max_approx_error_toOriginal_test = np.zeros(
(nb_sparsity_weights, nb_sparsity_objectives))
max_approx_error_toRec_train = np.zeros(
(nb_sparsity_weights, nb_sparsity_objectives))
max_approx_error_toRec_test = np.zeros(
(nb_sparsity_weights, nb_sparsity_objectives))
np.save(out_path + '_sparsityWeights_' + strDims, sparsity_weights)
np.save(out_path + '_sparsityObjectives_' + strDims, sparsity_objectives)
if svm:
SVM_classification_accuracy = np.zeros(
(nb_sparsity_weights, nb_sparsity_objectives))
for idx1, sp_w in enumerate(sparsity_weights):
for idx2, sp_o in enumerate(sparsity_objectives):
shAE = AsymAE_class(latent_dim=latent_dimension,
sparsity_weight=sp_w,
sparsity_objective=sp_o,
nb_input_channels=nb_input_channels,
one_channel_output=one_channel_output)
shAE.train(X_train=x_train,
X_train_expected_output=original_images_train,
nb_epochs=nb_epochs,
X_val=(x_test, original_images_test),
verbose=2)
shAE.save(path_to_model_directory=path_to_dir)
train_rec_errors[idx1, idx2] = shAE.reconstruction_error(
x_train, X_rec_th=original_images_train)
test_rec_errors[idx1, idx2] = shAE.reconstruction_error(
x_test, X_rec_th=original_images_test)
train_Hoyer_loss[idx1, idx2] = shAE.total_loss(
x_train,
X_rec_th=original_images_train) - train_rec_errors[idx1, idx2]
test_Hoyer_loss[idx1, idx2] = shAE.total_loss(
x_test, X_rec_th=original_images_test) - test_rec_errors[idx1,
idx2]
sparsity_train[idx1, idx2] = shAE.sparsity_measure(x_train)
sparsity_test[idx1, idx2] = shAE.sparsity_measure(x_test)
max_approx_train = shAE.max_approximation_error(
x_train,
morphoMaths.dilatation,
original_images=original_images_train,
apply_to_bias=False,
SE_scale=1)
max_approx_error_toOriginal_train[idx1, idx2] = max_approx_train[0]
max_approx_error_toRec_train[idx1, idx2] = max_approx_train[1]
max_approx_test = shAE.max_approximation_error(
x_test,
morphoMaths.dilatation,
original_images=original_images_test,
apply_to_bias=False,
SE_scale=1)
max_approx_error_toOriginal_test[idx1, idx2] = max_approx_test[0]
max_approx_error_toRec_test[idx1, idx2] = max_approx_test[1]
np.save(out_path + '_training_errors_' + strDims, train_rec_errors)
np.save(out_path + '_test_errors_' + strDims, test_rec_errors)
np.save(out_path + '_training_Hoyer_loss_' + strDims,
train_Hoyer_loss)
np.save(out_path + '_test_Hoyer_loss_' + strDims, test_Hoyer_loss)
np.save(out_path + '_training_sparsity_' + strDims, sparsity_train)
np.save(out_path + '_test_sparsity_' + strDims, sparsity_test)
np.save(
out_path +
'_training_max_approx_error_toOriginal_dilatation_' + strDims,
max_approx_error_toOriginal_train)
np.save(
out_path + '_test_max_approx_error_toOriginal_dilation_' +
strDims, max_approx_error_toOriginal_test)
np.save(
out_path + '_training_max_approx_error_toRec_dilatation_' +
strDims, max_approx_error_toRec_train)
np.save(
out_path + '_test_max_approx_error_toRec_dilation_' + strDims,
max_approx_error_toRec_test)
if svm:
SVM_classification_accuracy[
idx1, idx2] = shAE.best_SVM_classification_score(
x_test, y_test, nb_values_C=10, nb_values_gamma=10)[0]
np.save(out_path + '_svm_acc_' + strDims,
SVM_classification_accuracy)
if svm:
results = pd.DataFrame(
data={
'sparsity_weight':
np.repeat(sparsity_weights, len(sparsity_objectives)),
'sparsity_objective':
np.tile(sparsity_objectives, len(sparsity_weights)),
'training_error':
train_rec_errors.flatten(),
'test_error':
test_rec_errors.flatten(),
'training_Hoyer_loss':
train_Hoyer_loss.flatten(),
'test_Hoyer_loss':
test_Hoyer_loss.flatten(),
'training_sparsity':
sparsity_train.flatten(),
'test_sparsity':
sparsity_test.flatten(),
'training_max_approx_error_toOriginal_dilatation':
max_approx_error_toOriginal_train.flatten(),
'test_max_approx_error_toOriginal_dilatation':
max_approx_error_toOriginal_test.flatten(),
'training_max_approx_error_toRec_dilatation':
max_approx_error_toRec_train.flatten(),
'test_max_approx_error_toRec_dilatation':
max_approx_error_toRec_test.flatten(),
'SVM_classification_score':
SVM_classification_accuracy.flatten()
})
else:
results = pd.DataFrame(
data={
'sparsity_weight':
np.repeat(sparsity_weights, len(sparsity_objectives)),
'sparsity_objective':
np.tile(sparsity_objectives, len(sparsity_weights)),
'training_error':
train_rec_errors.flatten(),
'test_error':
test_rec_errors.flatten(),
'training_Hoyer_loss':
train_Hoyer_loss.flatten(),
'test_Hoyer_loss':
test_Hoyer_loss.flatten(),
'training_sparsity':
sparsity_train.flatten(),
'test_sparsity':
sparsity_test.flatten(),
'training_max_approx_error_toOriginal_dilatation':
max_approx_error_toOriginal_train.flatten(),
'test_max_approx_error_toOriginal_dilatation':
max_approx_error_toOriginal_test.flatten(),
'training_max_approx_error_toRec_dilatation':
max_approx_error_toRec_train.flatten(),
'test_max_approx_error_toRec_dilatation':
max_approx_error_toRec_test.flatten()
})
results.to_csv(out_path + 'results')
import numpy as np
import pandas as pd
from scipy import io
import metrics
import bastien_utils
FILE_PATH = "../"
x_test, y_test = bastien_utils.load_data(FILE_PATH,
train=False,
test=True,
subsetTest=False)
print('x_test shape:', x_test.shape)
data = io.loadmat('../Results/NMF/18_08_22_sparseNMF_dim100_spH0.6')
W = data['W']
H = data['H']
del data
print('W: ', W.shape)
print('H: ', H.shape)
atoms = W.transpose().reshape(100, 28, 28, 1)
H = H.transpose()
del W
svm_score, best_params = metrics.best_SVM_classification_score(
H, y_test, nb_values_C=10, nb_values_gamma=10)
SVM_best_C_parameter = best_params['C']
SVM_best_gamma_parameter = best_params['gamma']
results = pd.DataFrame(