def synthesise(request, model, image, feature):
if request.method == 'POST':
features_path = 'features/model_'+ str(model) + '_image_' + str(image) + '.dat'
clusters = read_clusters(features_path)
cluster = np.array(clusters[feature])
img_path = TestImage.objects.filter(id=image).first().image
test_model = TestModel.objects.filter(id=model).first()
architecture = str(test_model.architecture)
weights = str(test_model.weights)
labels = str(test_model.labels)
if architecture.split(".")[-1].lower() == "meta":
net = network.TensorFlowNet(architecture, './models/'+ str(test_model.user) +'_' + test_model.name + '/')
img = imread(img_path, mode='RGB')
img = imresize(img, (224, 224))
else:
net = network.CaffeNet(architecture, weights)
img, offset, resFac, newSize = utils.imgPreprocess(img_path=img_path)
net.set_new_size(newSize)
xmax, ymax, xmin, ymin = np.max(cluster[:,0]), np.max(cluster[:,1]), np.min(cluster[:,0]), np.min(cluster[:,1])
feature_img, _ = synthesis.synthesise_boundary(net, img, xmax, ymax, xmin, ymin)
mean = np.array([103.939, 116.779, 123.68])
feature_img[:,:,0] += mean[2]
feature_img[:,:,1] += mean[1]
feature_img[:,:,2] += mean[0]
num = FeatureImage.objects.filter(model__id=model,image__id=image,feature=feature).count()
# save synthesised image
feature_img = Image.fromarray(np.uint8(feature_img))
feature_path = 'synthesised_features/model_'+ str(model) + '_image_' + str(image) + '_' + str(feature) + '_' + str(num) + '.jpg'
feature_img.save(feature_path)
featureImage = FeatureImage(model = test_model, image=TestImage.objects.filter(id=image).first(), feature=feature, feature_image=feature_path)
featureImage.save()
return HttpResponse("{\"image\": " + feature_path +"}")
else:
return HttpResponse("{message: \"Invalid method.\"}", status=405)
def evaluate(request, model, image):
features_path = 'features/model_' + str(model) + '_image_' + str(
image) + '.dat'
body = json.loads(request.body.decode("utf-8"))
inactive_indices = body['inactiveFeatures']
clusters = read_clusters(features_path)
inactive_features = [clusters[i] for i in inactive_indices]
inactive_features = list(
map(
lambda cluster: list(
itertools.chain.from_iterable(
map(
lambda x:
[tuple(x + [0]),
tuple(x + [1]),
tuple(x + [2])], cluster))), inactive_features))
img_path = TestImage.objects.filter(id=image).first().image
test_model = TestModel.objects.filter(id=model).first()
architecture = str(test_model.architecture)
weights = str(test_model.weights)
labels = str(test_model.labels)
if architecture.split(".")[-1].lower() == "meta":
net = network.TensorFlowNet(
architecture,
'./models/' + str(test_model.user) + '_' + test_model.name + '/')
else:
net = network.CaffeNet(architecture, weights)
predictions, img = predictions_from_features(net, img_path,
inactive_features)
top_predictions = get_top_predictions(predictions, 5, labels)
# save modified image
img = Image.fromarray(np.uint8(img))
modification_path = 'features/model_' + str(model) + '_image_' + str(
image) + '_' + '_'.join(str(f) for f in inactive_indices) + '.jpg'
img.save(modification_path)
return HttpResponse("{\"predictions\":" + json.dumps(top_predictions) +
", \"image\": \"" + 'media/' + modification_path +
"\"}")
def analyse(request, model, image):
features_path = 'features/model_' + str(model) + '_image_' + str(
image) + '.dat'
clusters = read_clusters(features_path)
num_clusters = len(clusters)
img_path = TestImage.objects.filter(id=image).first().image
test_model = TestModel.objects.filter(id=model).first()
architecture = str(test_model.architecture)
weights = str(test_model.weights)
labels = str(test_model.labels)
if architecture.split(".")[-1].lower() == "meta":
net = network.TensorFlowNet(
architecture,
'./models/' + str(test_model.user) + '_' + test_model.name + '/')
else:
net = network.CaffeNet(architecture, weights)
predictions, _ = predictions_from_features(net, img_path, [])
basic_predictions = get_top_predictions(predictions, -1, labels)
predicted = basic_predictions[0]
# Find features for largest change
largest_change = 0
features = []
for i in range(1):
num = np.random.randint(low=0, high=2**num_clusters)
b = [num >> j & 1 for j in range(num.bit_length() - 1, -1, -1)]
b = [0] * (num_clusters - len(b)) + b
inactive_indices = list(itertools.compress(range(num_clusters), b))
inactive_features = [clusters[i] for i in inactive_indices]
inactive_features = list(
map(
lambda cluster: list(
itertools.chain.from_iterable(
map(
lambda x:
[tuple(x + [0]),
tuple(x + [1]),
tuple(x + [2])], cluster))), inactive_features))
predictions, img = predictions_from_features(net, img_path,
inactive_features)
change = predicted['value'] - predictions[predicted['index']]
if change > largest_change:
largest_change = change
features = inactive_indices
inactive_features = [clusters[i] for i in features]
inactive_features = list(
map(
lambda cluster: list(
itertools.chain.from_iterable(
map(
lambda x:
[tuple(x + [0]),
tuple(x + [1]),
tuple(x + [2])], cluster))), inactive_features))
predictions, img = predictions_from_features(net, img_path,
inactive_features)
img = Image.fromarray(np.uint8(img))
modification_path = 'features/model_' + str(model) + '_image_' + str(
image) + '_' + '_'.join(str(f) for f in features) + '.jpg'
img.save(modification_path)
top_predictions = get_top_predictions(predictions, 5, labels)
lc = {'features': features, 'predictions': top_predictions}
# Find most important feature
biggest_feature = 0
biggest_change = 0
for i in range(num_clusters):
cluster = clusters[i]
inactive_features = list(
itertools.chain.from_iterable(
map(lambda x: [tuple(x + [0]),
tuple(x + [1]),
tuple(x + [2])], cluster)))
predictions, _ = predictions_from_features(net, img_path,
[inactive_features])
change = predicted['value'] - predictions[predicted['index']]
if change > biggest_change:
biggest_change = change
biggest_feature = i
inactive_features = list(
itertools.chain.from_iterable(
map(lambda x: [tuple(x + [0]),
tuple(x + [1]),
tuple(x + [2])], clusters[biggest_feature])))
predictions, img = predictions_from_features(net, img_path,
[inactive_features])
img = Image.fromarray(np.uint8(img))
modification_path = 'features/model_' + str(model) + '_image_' + str(
image) + '_' + str(biggest_feature) + '.jpg'
img.save(modification_path)
top_predictions = get_top_predictions(predictions, 5, labels)
mi = {'feature': biggest_feature, 'predictions': top_predictions}
# Find minimal features required
selection = []
features_found = False
for num in range(num_clusters):
if num == 0:
inactive_features = list(itertools.chain.from_iterable(clusters))
inactive_features = list(
itertools.chain.from_iterable(
map(
lambda x:
[tuple(x + [0]),
tuple(x + [1]),
tuple(x + [2])], inactive_features)))
predictions, img = predictions_from_features(
net, img_path, [inactive_features])
top_predictions = get_top_predictions(predictions, 5, labels)
# No features are required
if top_predictions[0]['index'] == predicted['index']:
img = Image.fromarray(np.uint8(img))
modification_path = 'features/model_' + str(
model) + '_image_' + str(image) + '_' + '_'.join(
str(f) for f in range(num_clusters)) + '.jpg'
img.save(modification_path)
features_found = True
else:
combinations = itertools.combinations(range(num_clusters), num)
for selection in combinations:
inactive_indices = list(
set(range(num_clusters)) - set(selection))
inactive_features = [clusters[i] for i in inactive_indices]
inactive_features = list(
map(
lambda cluster: list(
itertools.chain.from_iterable(
map(
lambda x: [
tuple(x + [0]),
tuple(x + [1]),
tuple(x + [2])
], cluster))), inactive_features))
predictions, img = predictions_from_features(
net, img_path, inactive_features)
top_predictions = get_top_predictions(predictions, 5, labels)
if top_predictions[0]['index'] == predicted['index']:
img = Image.fromarray(np.uint8(img))
modification_path = 'features/model_' + str(
model) + '_image_' + str(image) + '_' + '_'.join(
str(f) for f in inactive_indices) + '.jpg'
img.save(modification_path)
features_found = True
break
if features_found:
break
mfRequired = {'features': selection, 'predictions': top_predictions}
# Find minimal to change
selection = []
features_found = False
for num in range(num_clusters):
if num == 0:
continue
combinations = itertools.combinations(range(num_clusters), num)
for selection in combinations:
inactive_indices = list(selection)
inactive_features = [clusters[i] for i in inactive_indices]
inactive_features = list(
map(
lambda cluster: list(
itertools.chain.from_iterable(
map(
lambda x: [
tuple(x + [0]),
tuple(x + [1]),
tuple(x + [2])
], cluster))), inactive_features))
predictions, img = predictions_from_features(
net, img_path, inactive_features)
top_predictions = get_top_predictions(predictions, 5, labels)
if top_predictions[0]['index'] != predicted['index']:
img = Image.fromarray(np.uint8(img))
modification_path = 'features/model_' + str(
model) + '_image_' + str(image) + '_' + '_'.join(
str(f) for f in inactive_indices) + '.jpg'
img.save(modification_path)
features_found = True
break
if features_found:
break
if not features_found:
selection = []
mfPerturbation = {'features': selection, 'predictions': top_predictions}
results = {
'originalClass': predicted['label'],
'lc': lc,
'mi': mi,
'mfRequired': mfRequired,
'mfPerturbation': mfPerturbation
}
return HttpResponse("{\"results\":" + json.dumps(results) + "}")
def index(request, model, image):
features_path = 'features/model_' + str(model) + '_image_' + str(
image) + '.dat'
feature_set = FeatureSet.objects.filter(model__id=model,
image__id=image).first()
unmodified_path = 'features/model_' + str(model) + '_image_' + str(
image) + '_.jpg'
if feature_set == None:
# carry out LRP and clustering and write to file
img_path = TestImage.objects.filter(id=image).first().image
test_model = TestModel.objects.filter(id=model).first()
architecture = str(test_model.architecture)
weights = str(test_model.weights)
labels = str(test_model.labels)
if architecture.split(".")[-1].lower() == "meta":
net = network.TensorFlowNet(
architecture, './models/' + str(test_model.user) + '_' +
test_model.name + '/')
img = imread(img_path, mode='RGB')
img = imresize(img, (224, 224))
im = Image.fromarray(np.uint8(img))
im.save(unmodified_path)
else:
net = network.CaffeNet(architecture, weights)
img = imread(img_path, mode='RGB')
img = 256 * utils.imageResize(img)
img = Image.fromarray(np.uint8(img))
img.save(unmodified_path)
img, offset, resFac, newSize = utils.imgPreprocess(
img_path=img_path)
net.set_new_size(newSize)
relevances = lrp.calculate_lrp_heatmap(net, img)
clusters = features.extract_features_from_relevances(relevances)
predictions, _ = predictions_from_features(net, img_path, [])
basic_predictions = get_top_predictions(predictions, -1, labels)
predicted = basic_predictions[0]
sorted_predictions = []
for i in range(len(clusters)):
cluster = list(
itertools.chain.from_iterable(
map(lambda x: [x + (0, ), x + (1, ), x + (2, )],
clusters[i])))
predictions, _ = predictions_from_features(net, img_path,
[cluster])
diff = predicted['value'] - predictions[predicted['index']]
sorted_predictions.append({'index': i, 'difference': diff})
sorted_predictions = sorted(sorted_predictions,
key=functools.cmp_to_key(compare))
sorted_predictions.reverse()
clusters = list(map(lambda x: clusters[x['index']],
sorted_predictions))
overlay_shape = (img.shape[0], img.shape[1], 4)
for i, cluster in enumerate(clusters):
overlay_img = np.zeros(overlay_shape)
for point in cluster:
overlay_img[point[0], point[1], 0] = 255
overlay_img[point[0], point[1], 3] = 128
overlay_img = Image.fromarray(np.uint8(overlay_img))
overlay_img.save('features/feature_model_' + str(model) +
'_image_' + str(image) + '_' + str(i) + '.png')
write_clusters(features_path, clusters)
feature_set = FeatureSet(
model=test_model,
image=TestImage.objects.filter(id=image).first(),
features=features_path)
feature_set.save()
# Get number of features and return features
with open(features_path) as f:
num_features = sum(1 for _ in f)
return HttpResponse(
"{\"features\":" + json.dumps(list(range(num_features))) +
", \"image\": \"" + 'media/' + unmodified_path +
"\",\"message\": \"features successfully retrieved.\"}")
from deep_introspection import network, utils
from scipy.misc import imread, imresize
net = network.CaffeNet(
'deep_introspection/test/VGG.prototxt',
'deep_introspection/test/VGG_ILSVRC_16_layers.caffemodel')
img, offset, resFac, newSize = utils.imgPreprocess(
img_path='deep_introspection/test/cat.jpg')
net.set_new_size(newSize)
tfNet = network.TensorFlowNet('deep_introspection/test/vgg16.meta',
'./deep_introspection/test/')
img1 = imread('deep_introspection/test/cat.jpg', mode='RGB')
img1 = imresize(img1, (224, 224))
def test_caffe_conv_weights_correct_shape():
weights = net.get_weights('conv1_1')
assert (weights.shape == (64, 3, 3, 3))
def test_caffe_fc_weights_correct_shape():
weights = net.get_weights('fc8')
assert (weights.shape == (1000, 4096))
def test_caffe_activations_correct_shape():
net.predict(img)
activations = net.get_activations('conv1_1')
assert (activations.shape == (64, 224, 224))