def main():
# div_fractions = [0.80, 0.0, 0.20] # Fractions to divide data into the train, valid, and test sets
train = gen_input.read_data_sets(DATA_PATH + "train_32x32.mat", [1, 0, 0], False, gen_input.reflect)
test = gen_input.read_data_sets(DATA_PATH + "test_32x32.mat", [0, 0, 1], False, gen_input.reflect)
#extra_train = load_extras(range(1,6))
Xtr = train.train.images.reshape(-1, IMAGE_DIMENSION, IMAGE_DIMENSION, 3)
Ytr = train.train.labels
#Xtr = np.concatenate((Xtr, extra_train.train.images.reshape(IMAGE_DIMENSION,IMAGE_DIMENSION,3)), axis=0)
#Ytr = np.concatenate((Ytr, extra_train.train.labels), axis=0)
Xte = test.test.images.reshape(-1, IMAGE_DIMENSION, IMAGE_DIMENSION, 3)
Yte = test.test.labels
print "Loaded data!"
# Convert to grayscale
Xtr = rgb_to_grayscale(Xtr).reshape(-1, IMAGE_DIMENSION * IMAGE_DIMENSION)
Xte = rgb_to_grayscale(Xte).reshape(-1, IMAGE_DIMENSION * IMAGE_DIMENSION)
# Xtr = normalize(Xtr)
# Xte = normalize(Xte)
# Y attribute stores data projects into PCA space using all eigen vectors.
# The eigen vectors are in decreasing order, so PCA().Y[:,0:x] returns the data projected to x-dimensional PCA space
pca_train = PCA(Xtr)
pca_test = PCA(Xte)
print "eigenvector top weights ", pca_train.fracs[0:20]
define_plot(x_title="K", y_title="Accuracy (%)",
title="Grayscale Test Accuracy vs K for Nearest Neighbors using PCA", label_prefix="K=")
print "1b"
plt.xlim((min(Ks) - 1, max(Ks) + 1))
min_accuracy = 1.0
max_accuracy = 0
all_accuracies = []
for reduced_dimension in REDUCED_DIMENSIONS:
print "starting with PCA dim ", reduced_dimension
Xtr = pca_train.Y[:, 0:reduced_dimension]
Xte = pca_test.Y[:, 0:reduced_dimension]
print "Done trimming to PCA dimension"
num_dimensions = Xtr.shape[1]
# Graph Input
pl_x_train = tf.placeholder("float", shape=[None, num_dimensions])
pl_x_test = tf.placeholder("float", shape=[num_dimensions])
# Nearest Neighbor calculation using L1 Norm Distance
# distance = tf.reduce_sum(tf.abs(tf.add(pl_x_train, tf.neg(pl_x_test))), reduction_indices=1)
# Nearest Neighbor calculation using L2 Norm Distance, sqrt necessary for Inverse Weight function
distance = tf.reduce_sum(tf.sqrt(tf.add(tf.mul(pl_x_train, pl_x_train), tf.mul(pl_x_test, pl_x_test))), reduction_indices=1)
neg_distance = tf.neg(distance) # MIN(distance) = MAX(neg_distance)
# Couldn't get this to work: Wanted to use top_k then use scipy's mode method in loop below
# largest_neg_distance, top_classes = tf.nn.top_k(neg_distances, K)
# Predict: Get index of the most frequent class (Nearest Neighbor)
# prediction = tf.argmin(distance, 0)
print "Init Session..."
# Get session ready
init = tf.initialize_all_variables()
session = tf.Session()
session.run(init)
print "Starting training/testing ", len(Xte), " examples"
accuracies = []
used_ks = []
for K in Ks:
print "Starting K = ", K
num_correct = 0
# loop over test data
for i in range(len(Xte)):
# Get nearest neighbor
# nn_index = session.run(prediction, feed_dict={pl_x_train: Xtr, pl_x_test: Xte[i, :]})
neg_distances = session.run(neg_distance, feed_dict={pl_x_train: Xtr, pl_x_test: Xte[i, :]})
top_classes_index = np.argpartition(neg_distances, -K)[-K:]
### For Inverse weighting
max_weight = 0
top_class = 0
inverse_weights = 1.0 / -1.0 * neg_distances[top_classes_index]
for cls in range(0,10):
ind,count = np.where(Ytr[top_classes_index] == [cls])
cls_total = np.sum(inverse_weights[ind])
if cls_total == max_weight:
top_class = cls
###
'''
### For weights = 1, (voting scheme)
top_class, count = mode(Ytr[top_classes_index])
top_class = top_class[0][0] # Unbox from matrix
###
'''
# Get nearest neighbor class label and compare it to its true label
if top_class == Yte[i][0]:
num_correct += 1
accuracy = float(num_correct) * 100.0 / len(Xte)
if accuracy > max_accuracy:
max_accuracy = accuracy
elif accuracy < min_accuracy:
min_accuracy = accuracy
print "Accuracy:", accuracy
accuracies.append(accuracy)
used_ks.append(K)
plot_accuracy(used_ks, accuracies)
all_accuracies.append(accuracies)
plot_accuracy(Ks, accuracies, label="PCA " + str(reduced_dimension))
print Ks
print all_accuracies
plt.ylim((min_accuracy - 0.1 * min_accuracy, max_accuracy + 0.1 * max_accuracy))
plt.legend(loc=0)
plt.ioff()
raw_input('Press Enter to exit.')
plt.savefig('./plots/knn/grayscale2.png', bbox_inches='tight')
def main():
trim = 4
train = gen_input.read_data_sets(DATA_PATH + "train_32x32.mat", [1, 0, 0], False, gen_input.reflect)
test = gen_input.read_data_sets(DATA_PATH + "test_32x32.mat", [0, 0, 1], False, gen_input.reflect)
#extra_train = load_extras(range(1,6))
Xtr = train.train.images.reshape(-1, IMAGE_DIMENSION, IMAGE_DIMENSION, 3)
Ytr = train.train.labels
#Xtr = np.concatenate((Xtr, extra_train.train.images), axis=0)
#Ytr = np.concatenate((Ytr, extra_train.train.labels), axis=0)
Xte = test.test.images.reshape(-1, IMAGE_DIMENSION, IMAGE_DIMENSION, 3)
Yte = test.test.labels
'''
mnist = input_data.read_data_sets("MNIST_data/", one_hot=False)
Xtr = mnist.train.images
Ytr = mnist.train.labels
Xte = mnist.test.images
Yte = mnist.test.labels
'''
# Ytr_2 = np.where(Ytr == 0)[0]
# Ytr_3 = np.where(Ytr == 7)[0]
# indices = np.concatenate((Ytr_2, Ytr_3))
# Ytr = Ytr[indices]
# Xtr = Xtr[indices,:]#,:,:]
# Yte_2 = np.where(Yte == 0)[0]
# Yte_3 = np.where(Yte == 7)[0]
# indices = np.concatenate((Yte_2, Yte_3))
# Yte = Yte[indices]
# Xte = Xte[indices,:]#,:,:]
print "Done loading data ", Xtr.shape[0], Xte.shape[0]
# Convert images to grayscale
t1 = feature_extraction.rgb_to_grayscale(Xtr)
t2 = feature_extraction.rgb_to_grayscale(Xte)
print "Done RGB"
for trim in [4,6,8,10,12,14]:
Xtr = t1[:,:,trim:-trim]
Xte = t2[:,:,trim:-trim]
'''
# Extract features
Xtr_feature, Xtr_feature_per_example, Xtr_duds = feature_extraction.all_extract_surf(Xtr)
Xte_feature, Xte_feature_per_example, Xte_duds = feature_extraction.all_extract_surf(Xte)
print "Done Feature Extraction ", Xtr_duds.shape, " ", Xte_duds.shape
'''
'''
Xtr_feature = np.delete(Xtr_feature, Xtr_duds, axis=0)
Xtr_feature_per_example = np.delete(Xtr_feature_per_example, Xtr_duds, axis=0)
Xte_feature = np.delete(Xte_feature, Xte_duds, axis=0)
Xte_feature_per_example = np.delete(Xte_feature_per_example, Xte_duds, axis=0)
Ytr = np.delete(Ytr, Xtr_duds, axis=0)
Yte = np.delete(Yte, Xte_duds, axis=0)
'''
# Extract HoG features
Xtr_feature, Xtr_feature_per_example = feature_extraction.all_extract_hog(Xtr)
Xte_feature, Xte_feature_per_example = feature_extraction.all_extract_hog(Xte)
print "Done Feature Extraction "
accuracies = {}
for CLUSTER_COUNT in NUM_CLUSTERS:
print "Starting config ", CLUSTER_COUNT
fb = FeatureBag(CLUSTER_COUNT)
#fb.set_data(Xtr_feature)
print "Begin Fitting"
fb.fit(Xtr_feature)
print "Done Fitting"
train_features = np.array(fb.predict_feature_vectors(Xtr_feature_per_example[0]))
for i in range(1, len(Xtr_feature_per_example)):
feature = Xtr_feature_per_example[i]
cluster_desc = fb.predict_feature_vectors(feature)
train_features = np.vstack((train_features, cluster_desc))
test_features = np.array(fb.predict_feature_vectors(Xte_feature_per_example[0]))
for i in range(1, len(Xte_feature_per_example)):
feature = Xte_feature_per_example[i]
cluster_desc = fb.predict_feature_vectors(feature)
test_features = np.vstack((test_features, cluster_desc))
print "Starting SVM"
svm = SVM()
svm.train(train_features, Ytr)
predictions = svm.predict(test_features).reshape(Yte.shape)
print "Done SVM predictions"
# Assume Yte is dense encoding, comparing elements not arrays for equality
accuracy = float(np.sum(predictions == Yte)) * 100.0 / Yte.shape[0]
print "trim ", trim
print "Accuracy = ", accuracy
accuracies[CLUSTER_COUNT] = accuracy
print accuracies
