def get_image_feature(file_path,dictionary,layer_num,K):
'''
Extracts the spatial pyramid matching feature.
[input]
* file_path: path of image file to read
* dictionary: numpy.ndarray of shape (K,3F)
* layer_num: number of spatial pyramid layers
* K: number of clusters for the word maps
[output]
* feature: numpy.ndarray of shape (K)
'''
# ----- TODO -----
#load image
image = skimage.io.imread(file_path)
image = image.astype('float')/255
# extract wordmap from image
wordmap = visual_words.get_visual_words(image,dictionary)
#compute SPM
feature = get_feature_from_wordmap_SPM(wordmap,layer_num,K)
#return computed feature
return feature
def predict_image(file_path, dictionary, layer_num, K, features, train_labels):
'''
Predicts the label using the trained system and extracted VGG-16 features
This is a function run by a subprocess.
[input]
* file_path: path of image file
* dictionary: numpy.ndarray of shape (K, 3F)
* layer_num: number of spatial pyramid layers
* K: number of clusters for the word maps
* features: trained features using SPM
* train_labels: trained set of labels
[output]
* predicted_label: int representing the predicted label
'''
global PROGRESS
with PROGRESS_LOCK:
PROGRESS += NPROC
print('Processing: %03d/160 | Image: %s' % (PROGRESS, file_path))
# Read image
image = imageio.imread('../data/' + file_path)
image = image.astype('float') / 255
# Create visual wordmap for the image
wordmap = visual_words.get_visual_words(image, dictionary)
# Compute features using SPM
word_hist = get_feature_from_wordmap_SPM(wordmap, layer_num, K)
# Find the predicted label using histogram intersection similarity
predicted_label = train_labels[np.argmax(
distance_to_set(word_hist, features))]
return predicted_label
def get_image_feature(file_path, dictionary, layer_num, K):
'''
Extracts the spatial pyramid matching feature.
[input]
* file_path: path of image file to read
* dictionary: numpy.ndarray of shape (K, 3F)
* layer_num: number of spatial pyramid layers
* K: number of clusters for the word maps
[output]
* feature: numpy.ndarray of shape (K*(4^layer_num-1)/3)
'''
image = skimage.io.imread(file_path)
image = image.astype('float') / 255
wordmap = visual_words.get_visual_words(image, dictionary)
histo = get_feature_from_wordmap_SPM(wordmap, layer_num, K)
return histo
def evaluate_recognition_system(opts, n_worker=8):
'''
Evaluates the recognition system for all test images and returns the confusion matrix.
[input]
* opts : options
* n_worker : number of workers to process in parallel
[output]
* conf: numpy.ndarray of shape (8,8)
* accuracy: accuracy of the evaluated system
'''
data_dir = opts.data_dir
out_dir = opts.out_dir
wrng = []
trained_system = np.load(join(out_dir, 'trained_system.npz'))
dictionary = trained_system['dictionary']
dict_size = len(dictionary)
# using the stored options in the trained system instead of opts.py
test_opts = copy(opts)
test_opts.K = dictionary.shape[0]
test_opts.L = trained_system['SPM_layer_num']
test_files = open(join(data_dir, 'test_files.txt')).read().splitlines()
test_labels = np.loadtxt(join(data_dir, 'test_labels.txt'), np.int32)
test_file_num = len(test_files)
# ----- TODO -----
conf_mat = np.zeros((8, 8), dtype=int)
test_pred_labl = np.zeros(test_file_num, dtype=int)
train_feats = trained_system['features']
train_labels = trained_system['labels']
T_img = len(test_files) #size of training data--> # of images, T
test_img_list = np.arange(T_img)
prcs = Pool(n_worker)
args = list(zip(test_img_list, test_files, test_labels))
feature_test = prcs.map(get_image_feature,
args) #create subprocesses to call ome_image
for ind_f in range(test_file_num):
img = Image.open("../data/" + (test_files[ind_f])) # Load test images
img = np.array(img).astype(
np.float32) / 255 # convert to 0-1 range values
wordmap_test = visual_words.get_visual_words(opts, img, dictionary)
# feature_test = get_feature_from_wordmap_SPM(opts,wordmap_test)
sim_dist = distance_to_set(feature_test[ind_f], train_feats)
labl_pred = train_labels[np.argmax(sim_dist)] #predicted label
test_pred_labl[ind_f] = labl_pred
labl_true = test_labels[ind_f]
conf_mat[labl_true, labl_pred] += 1
if labl_true != labl_pred:
wrng.append(test_files[ind_f])
per_acc = np.trace(conf_mat) / np.sum(conf_mat)
print('in evaluation')
print('L is', opts.L)
print('K is', opts.K)
print('alpha is', opts.alpha)
return conf_mat, per_acc
import skimage
if __name__ == '__main__':
num_cores = util.get_num_CPU()
path_img = "../data/kitchen/sun_aasmevtpkslccptd.jpg"
image = skimage.io.imread(path_img)
image = image.astype('float') / 255
filter_responses = visual_words.extract_filter_responses(image)
util.display_filter_responses(filter_responses)
visual_words.compute_dictionary(num_workers=num_cores)
dictionary = np.load('dictionary.npy')
img = visual_words.get_visual_words(image, dictionary)
#util.save_wordmap(wordmap, filename)
visual_recog.build_recognition_system(num_workers=num_cores)
conf, accuracy = visual_recog.evaluate_recognition_system(
num_workers=num_cores)
print(conf)
print(np.diag(conf).sum() / conf.sum())
vgg16 = torchvision.models.vgg16(pretrained=True).double()
vgg16.eval()
deep_recog.build_recognition_system(vgg16, num_workers=num_cores // 2)
conf = deep_recog.evaluate_recognition_system(vgg16,
num_workers=num_cores // 2)
print(conf)
print(np.diag(conf).sum() / conf.sum())
def main():
opts = get_opts()
print('L is', opts.L)
print('K is', opts.K)
print('alpha is', opts.alpha)
print()
# Q1.1
img_path = join(opts.data_dir, 'kitchen/sun_aasmevtpkslccptd.jpg')
img = Image.open(img_path)
img = np.array(img).astype(np.float32) / 255
filter_responses = visual_words.extract_filter_responses(opts, img)
# imageio.imsave('../results/filter_responses.jpg',filter_responses)
util.visualize_wordmap(img)
util.display_filter_responses(opts, filter_responses)
## # Q1.2
n_cpu = util.get_num_CPU()
visual_words.compute_dictionary(opts, n_worker=n_cpu)
dictionary = np.load(join(opts.out_dir, 'dictionary.npy'))
###
# ## Q1.3
img_path = join(opts.data_dir, 'kitchen/sun_aasmevtpkslccptd.jpg')
img = Image.open(img_path)
img = np.array(img).astype(np.float32) / 255
wordmap = visual_words.get_visual_words(opts, img, dictionary)
util.visualize_wordmap(img)
util.visualize_wordmap(wordmap)
#
img_path = join(opts.data_dir, 'waterfall/sun_bbeqjdnienanmmif.jpg')
img = Image.open(img_path)
img = np.array(img).astype(np.float32) / 255
wordmap = visual_words.get_visual_words(opts, img, dictionary)
util.visualize_wordmap(img)
util.visualize_wordmap(wordmap)
#
img_path = join(opts.data_dir, 'windmill/sun_bratfupeyvlazpba.jpg')
img = Image.open(img_path)
img = np.array(img).astype(np.float32) / 255
wordmap = visual_words.get_visual_words(opts, img, dictionary)
util.visualize_wordmap(img)
util.visualize_wordmap(wordmap)
img_path = join(opts.data_dir, 'desert/sun_adjlepvuitklskrz.jpg')
img = Image.open(img_path)
img = np.array(img).astype(np.float32) / 255
wordmap = visual_words.get_visual_words(opts, img, dictionary)
util.visualize_wordmap(img)
util.visualize_wordmap(wordmap)
#
# Q2.1-2.4
n_cpu = util.get_num_CPU()
visual_recog.build_recognition_system(opts, n_worker=n_cpu)
## Q2.5
n_cpu = util.get_num_CPU()
conf, accuracy = visual_recog.evaluate_recognition_system(opts,
n_worker=n_cpu)
print(conf)
print(accuracy)
np.savetxt(join(opts.out_dir, 'confmat.csv'),
conf,
fmt='%d',
delimiter=',')
np.savetxt(join(opts.out_dir, 'accuracy.txt'), [accuracy], fmt='%g')
# #Q1.1
filter_responses = visual_words.extract_filter_responses(image)
util.display_filter_responses(filter_responses)
# Q1.2
visual_words.compute_dictionary(num_workers=num_cores)
# Q1.3 visualize word map
dictionary = np.load('../code/dictionary.npy', allow_pickle=True)
train_data = np.load('../data/train_data.npz', allow_pickle=True)
train_name = train_data['image_names']
train_data = np.random.permutation(
train_name)[:3] # load 3 random images for visulize wordmap
for image in train_data:
image = imageio.imread('../data/' + image[0])
wordmap = visual_words.get_visual_words(image, dictionary)
f, axes = plt.subplots(1, 2)
f.set_size_inches(8, 8)
axes[0].imshow(image)
axes[0].set_title('original')
axes[1].imshow(wordmap)
axes[1].set_title('wordmap')
plt.show()
# Q2.1 plot feature of one word map
dictionary = np.load('../code/dictionary.npy', allow_pickle=True)
wordmap = visual_words.get_visual_words(image, dictionary)
dict_size = len(dictionary)
hist_all = visual_recog.get_feature_from_wordmap(wordmap, dict_size)
plt.hist(hist_all, dict_size)
plt.show()
def build_recognition_system_helper_func(img_path):
image = imageio.imread("../data/" + img_path)
wordmap = visual_words.get_visual_words(image, dictionary)
hist = get_feature_from_wordmap_SPM(wordmap, layers, 200)
return hist