def __init__(self, X_train = None, Y_train = None , bins=256, show_train_acc=False, show_images=False, file = None):
if X_train != None :
print 'computing a priory log likelihoods...'
self.a_priory = np.sum(Y_train, axis = 0)/ Y_train.shape[0]
self.bins = bins
self.cat_num = Y_train.shape[1]
self.ccls_pos, self.ccls_neg = compute_ccls_per_category(self.bins, X_train, Y_train, show_images)
if show_train_acc :
print 'Train:'
prob_pos, prob_neg = compute_probabilities(self.bins, X_train, self.cat_num, self.ccls_pos, self.ccls_neg, self.a_priory)
train_output = prob_pos > prob_neg
cnn.print_error_rate_per_category(train_output, Y_train, thr = 0.5)
print
elif file != None :
with open(file, 'r') as f :
other = pickle.load(f)
self.a_priory = other.a_priory
self.bins = other.bins
self.cat_num = other.cat_num
self.ccls_pos = other.ccls_pos
self.ccls_neg = other.ccls_neg
else:
print "Tried to init an fih without train data or other fih or a pickle file specified"
def predict(self, X_test, show_accuracy = False, Y_test = None) :
output = np.zeros((X_test.shape[0], self.cat_num))
prob_pos, prob_neg = compute_probabilities(self.bins, X_test, self.cat_num, self.ccls_pos, self.ccls_neg, self.a_priory)
output = np.exp(prob_pos) / (np.exp(prob_pos) + np.exp(prob_neg))
output[np.isnan(output)] = 0 #This has been fixed by taking low probabilities instead of 0 probabilities, but this works better.
if show_accuracy :
print 'Test:'
if Y_test != None :
cnn.print_error_rate_per_category(output, Y_test, thr = 0.5)
else :
print 'Error: Could not print accuracy: Y_test not specified'
return output
def runNet_cat(imageType, saveDir):
name = saveDir + 'stressVSrelacs' + '_' + imageType + '.cnn'
temp = sys.stdout
with open(saveDir + 'stressVSrelacs' + '_' + imageType + '.txt', 'w') as sys.stdout:
X_train, Y_train_all, X_test, Y_test_all = cnn.load_normalized_data(imageType)
Y_train = cnn.get_categorical_data_stress(Y_train_all)
Y_test = cnn.get_categorical_data_stress(Y_test_all)
model = cnn.build_empty_softmax_model(X_train.shape, Y_train.shape)
model.fit(X_train, Y_train,
batch_size = 32, nb_epoch = 50,
validation_data= (X_test, Y_test))
cnn.save_weights(model, weights_filename = name)
output = model.predict(X_test)
cnn.print_error_rate_per_category(output, Y_test)
name = saveDir + 'categories' + '_' + imageType + '.cnn'
with open(saveDir + 'categories' + '_' + imageType + '.txt', 'w') as sys.stdout:
X_train, Y_train_all, X_test, Y_test_all = cnn.load_normalized_data(imageType)
cat_lists = [[6],[5],[11],[10],[4],[7],[9],[8]]
Y_train = cnn.get_categorical_data_cats(Y_train_all, cat_lists)
Y_test = cnn.get_categorical_data_cats(Y_test_all, cat_lists)
model = cnn.build_empty_softmax_model(X_train.shape, Y_train.shape)
model.fit(X_train, Y_train,
batch_size = 32, nb_epoch = 50,
validation_data= (X_test, Y_test))
cnn.save_weights(model, weights_filename = name)
output = model.predict(X_test)
cnn.print_error_rate_per_category(output, Y_test)
sys.stdout = temp
def predict(self, X_test, show_accuracy=False, Y_test=None):
output = np.zeros((X_test.shape[0], self.cat_num))
prob_pos, prob_neg = compute_probabilities(self.bins, X_test,
self.cat_num, self.ccls_pos,
self.ccls_neg,
self.a_priory)
output = np.exp(prob_pos) / (np.exp(prob_pos) + np.exp(prob_neg))
output[np.isnan(
output
)] = 0 #This has been fixed by taking low probabilities instead of 0 probabilities, but this works better.
if show_accuracy:
print 'Test:'
if Y_test != None:
cnn.print_error_rate_per_category(output, Y_test, thr=0.5)
else:
print 'Error: Could not print accuracy: Y_test not specified'
return output
def runNet(imageType, cats, saveDir):
# Make the filename
catName = str(cats)
name = saveDir + catName + '_' + imageType + '.cnn'
temp = sys.stdout
with open(saveDir + catName + '_' + imageType + '.txt', 'w') as sys.stdout:
X_train, Y_train_all, X_test, Y_test_all = cnn.load_normalized_data(imageType)
Y_train = Y_train_all[:, cats]
Y_test = Y_test_all[:, cats]
model = cnn.build(X_train, Y_train, X_test, Y_test, epochs = 50)
cnn.save_weights(model, weights_filename = name)
output = model.predict(X_test)
cnn.print_error_rate_per_category(output, Y_test)
sys.stdout = temp
def mlp(X_train, Y_train, X_test, Y_test) :
X_train = np.sum(X_train, axis = (1,3))
X_train = X_train/128
X_train -= 65
X_train /= 256
X_test = np.sum(X_test, axis = (1,3))
X_test = X_test/128
X_test -= 65
X_test /= 256
print 'Building model...'
model = build_empty_model(X_train.shape, Y_train.shape)
model.fit(X_train, Y_train, batch_size = 32, nb_epoch = 200, validation_data= (X_test, Y_test))
print 'predicting output'
output = model.predict(X_test)
cnn.print_error_rate_per_category(output, Y_test)
def bayes(X_train, Y_train, X_test, Y_test, pyramid_height = 1, max_bins = 256, show_train_acc = True) :
print 'computing a priory log likelihoods...'
a_priory = np.sum(Y_train, axis = 0)/ Y_train.shape[0]
output = np.zeros((pyramid_height , Y_test.shape[0], Y_test.shape[1]))
for n in range(pyramid_height ) :
bins = max_bins/ 2**n
print 'binsize: ' + str(bins)
ccls_pos, ccls_neg = compute_ccls_per_category(bins, X_train, Y_train)
print 'Test:'
prob_pos, prob_neg = compute_probabilities(bins, X_test, Y_test, ccls_pos, ccls_neg, a_priory)
output[n] = prob_pos > prob_neg
cnn.print_error_rate_per_category(output[n], Y_test, thr = 0.5)
if show_train_acc :
print 'Train:'
prob_pos, prob_neg = compute_probabilities(bins, X_train, Y_train, ccls_pos, ccls_neg, a_priory)
train_output = prob_pos > prob_neg
cnn.print_error_rate_per_category(train_output, Y_train, thr = 0.5)
print
if pyramid_height != 1 :
print 'Performance with all bin sizes together:'
compressed_output = np.sum(output, axis = 0)
compressed_output = compressed_output/pyramid_height
cnn.print_error_rate_per_category(compressed_output, Y_test, thr = 0.5)
def __init__(self,
X_train=None,
Y_train=None,
bins=256,
show_train_acc=False,
show_images=False,
file=None):
if X_train != None:
print 'computing a priory log likelihoods...'
self.a_priory = np.sum(Y_train, axis=0) / Y_train.shape[0]
self.bins = bins
self.cat_num = Y_train.shape[1]
self.ccls_pos, self.ccls_neg = compute_ccls_per_category(
self.bins, X_train, Y_train, show_images)
if show_train_acc:
print 'Train:'
prob_pos, prob_neg = compute_probabilities(
self.bins, X_train, self.cat_num, self.ccls_pos,
self.ccls_neg, self.a_priory)
train_output = prob_pos > prob_neg
cnn.print_error_rate_per_category(train_output,
Y_train,
thr=0.5)
print
elif file != None:
with open(file, 'r') as f:
other = pickle.load(f)
self.a_priory = other.a_priory
self.bins = other.bins
self.cat_num = other.cat_num
self.ccls_pos = other.ccls_pos
self.ccls_neg = other.ccls_neg
else:
print "Tried to init an fih without train data or other fih or a pickle file specified"
def mlp(X_train, Y_train, X_test, Y_test):
X_train = np.sum(X_train, axis=(1, 3))
X_train = X_train / 128
X_train -= 65
X_train /= 256
X_test = np.sum(X_test, axis=(1, 3))
X_test = X_test / 128
X_test -= 65
X_test /= 256
print 'Building model...'
model = build_empty_model(X_train.shape, Y_train.shape)
model.fit(X_train,
Y_train,
batch_size=32,
nb_epoch=200,
validation_data=(X_test, Y_test))
print 'predicting output'
output = model.predict(X_test)
cnn.print_error_rate_per_category(output, Y_test)
def bayes(X_train,
Y_train,
X_test,
Y_test,
pyramid_height=1,
max_bins=256,
show_train_acc=True,
show_images=False):
print 'computing a priory log likelihoods...'
a_priory = np.sum(Y_train, axis=0) / Y_train.shape[0]
output = np.zeros((pyramid_height, Y_test.shape[0], Y_test.shape[1]))
for n in range(pyramid_height):
bins = max_bins / 2**n
print 'binsize: ' + str(bins)
ccls_pos, ccls_neg = compute_ccls_per_category(bins, X_train, Y_train,
show_images)
print 'Test:'
prob_pos, prob_neg = compute_probabilities(bins, X_test,
Y_test.shape[1], ccls_pos,
ccls_neg, a_priory)
output[n] = prob_pos > prob_neg
cnn.print_error_rate_per_category(output[n], Y_test, thr=0.5)
if show_train_acc:
print 'Train:'
prob_pos, prob_neg = compute_probabilities(bins, X_train,
Y_train.shape[1],
ccls_pos, ccls_neg,
a_priory)
train_output = prob_pos > prob_neg
cnn.print_error_rate_per_category(train_output, Y_train, thr=0.5)
print
if pyramid_height != 1:
print 'Performance with all bin sizes together:'
compressed_output = np.sum(output, axis=0)
compressed_output = compressed_output / pyramid_height
cnn.print_error_rate_per_category(compressed_output, Y_test, thr=0.5)