def explain_image(self, labels, instance, column_name=None, num_features=100000,
num_samples=300, batch_size=200, hide_color=0):
"""Explain an image of a prediction.
It analyze the prediction by LIME, and returns a report of which words are most impactful
in contributing to certain labels.
Args:
labels: a list of labels to explain.
instance: the prediction instance. It needs to conform to model's input. Can be a csv
line string, or a dict.
column_name: which image column to explain. Can be None if there is only one image column
in the model input.
num_features: maximum number of areas (features) to analyze. Passed to
LIME LimeImageExplainer directly.
num_samples: size of the neighborhood to learn the linear model. Passed to
LIME LimeImageExplainer directly.
batch_size: size of batches passed to predict_fn. Passed to
LIME LimeImageExplainer directly.
hide_color: the color used to perturb images. Passed to
LIME LimeImageExplainer directly.
Returns:
A LIME's lime.explanation.Explanation.
Throws:
ValueError if the given image column is not found in model input or column_name is None
but there are multiple image columns in model input.
"""
from lime.lime_image import LimeImageExplainer
if len(self._image_columns) > 1 and not column_name:
raise ValueError('There are multiple image columns in the input of the model. ' +
'Please specify "column_name".')
elif column_name and column_name not in self._image_columns:
raise ValueError('Specified column_name "%s" not found in the model input.'
% column_name)
image_column_name = column_name if column_name else self._image_columns[0]
if isinstance(instance, six.string_types):
instance = next(csv.DictReader([instance], fieldnames=self._headers))
predict_fn = self._make_image_predict_fn(labels, instance, image_column_name)
explainer = LimeImageExplainer()
with file_io.FileIO(instance[image_column_name], 'rb') as fi:
im = Image.open(fi)
im.thumbnail((299, 299), Image.ANTIALIAS)
rgb_im = np.asarray(im.convert('RGB'))
exp = explainer.explain_instance(
rgb_im, predict_fn, labels=range(len(labels)), top_labels=None,
hide_color=hide_color, num_features=num_features,
num_samples=num_samples, batch_size=batch_size)
return exp
def lime(model, image):
explainer = LimeImageExplainer()
explanation = explainer.explain_instance(image,
model.predict,
hide_color=0,
top_labels=5,
num_samples=100)
temp, mask = explanation.get_image_and_mask(explanation.top_labels[0],
positive_only=True,
num_features=5,
hide_rest=True)
features = mark_boundaries(image, mask)
label = explanation.top_labels[0]
print(label)
result = CLASSES[label]
return label, temp, mask, features, result
def Lime(model, x, y, seed):
def Predict(Input):
return model.predict(
np.mean(Input.reshape((-1, 48, 48, 3)),
axis=3).reshape(-1, 48, 48, 1)).reshape((-1, 7))
def Segmentation(Input):
return slic(Input)
explainer = LimeImageExplainer()
explanation = explainer.explain_instance(image=x,
classifier_fn=Predict,
segmentation_fn=Segmentation,
random_seed=seed)
image, mask = explanation.get_image_and_mask(label=y,
positive_only=False,
hide_rest=False,
num_features=5,
min_weight=0.0)
return image
class LimeImageExplainer(ImageExplainer):
def __init__(self, model, label):
self.explainer = LimeImage()
self.model = model
self.label = label
def explain(self, instance, budget):
def classifier_fn(instances):
instances = np.moveaxis(instances, -1, 1)
instances = torch.tensor(instances).cuda().float()
with torch.no_grad():
output = self.model.predict_proba(instances)
return output.cpu().numpy()
instance = instance.double().detach().cpu().numpy()
instance = np.moveaxis(instance, 0, -1)
# Heuristic: each super pixel is about 5% of total pixels
# max ensures we return at least one superpixel
num_features = max(budget // 5, 1)
exp = self.explainer.explain_instance(instance, classifier_fn)
_, mask = exp.get_image_and_mask(
label=self.label,
positive_only=True,
# negative_only=False,
hide_rest=False,
num_features=num_features,
min_weight=0)
mask = mask.astype(np.float32)
# # only return above percentile
# top_percentile = np.percentile(mask, 100 - budget)
# mask[ mask < top_percentile ] = 0.0
# mask[ mask >= top_percentile ] = 1.0
return mask
class LimeExplainer:
def __init__(self, random_seed=42):
self.random_seed = random_seed
self.explainer = LimeImageExplainer(random_state=self.random_seed)
def explain(self,
image_array: np.array,
classifier_func: Callable,
top_labels: int = 2):
"""
Image (224x224) needs to be PyTorch Tensor.
:return:
"""
explanation = self.explainer.explain_instance(
image_array,
classifier_func,
top_labels=top_labels,
hide_color=0,
random_seed=self.random_seed,
num_samples=1000)
img, mask = explanation.get_image_and_mask(explanation.top_labels[0],
positive_only=False,
num_features=10,
hide_rest=False)
probabilities = classifier_func([image_array])
print(f"Top labels:")
for idx, (class_idx, class_probability) in enumerate(
zip(explanation.top_labels, probabilities), 1):
print(
f"\t{class_idx}) '{LABEL_MAPPING[class_idx]}: {class_probability.max()}'"
)
img_boundary = mark_boundaries(img / 255.0, mask)
return img_boundary
def explain_image(self,
labels,
instance,
column_name=None,
num_features=100000,
num_samples=300,
batch_size=200,
hide_color=0):
"""Explain an image of a prediction.
It analyze the prediction by LIME, and returns a report of which words are most impactful
in contributing to certain labels.
Args:
labels: a list of labels to explain.
instance: the prediction instance. It needs to conform to model's input. Can be a csv
line string, or a dict.
column_name: which image column to explain. Can be None if there is only one image column
in the model input.
num_features: maximum number of areas (features) to analyze. Passed to
LIME LimeImageExplainer directly.
num_samples: size of the neighborhood to learn the linear model. Passed to
LIME LimeImageExplainer directly.
batch_size: size of batches passed to predict_fn. Passed to
LIME LimeImageExplainer directly.
hide_color: the color used to perturb images. Passed to
LIME LimeImageExplainer directly.
Returns:
A LIME's lime.explanation.Explanation.
Throws:
ValueError if the given image column is not found in model input or column_name is None
but there are multiple image columns in model input.
"""
from lime.lime_image import LimeImageExplainer
if len(self._image_columns) > 1 and not column_name:
raise ValueError(
'There are multiple image columns in the input of the model. '
+ 'Please specify "column_name".')
elif column_name and column_name not in self._image_columns:
raise ValueError(
'Specified column_name "%s" not found in the model input.' %
column_name)
image_column_name = column_name if column_name else self._image_columns[
0]
if isinstance(instance, six.string_types):
instance = next(
csv.DictReader([instance], fieldnames=self._headers))
predict_fn = self._make_image_predict_fn(labels, instance,
image_column_name)
explainer = LimeImageExplainer()
with file_io.FileIO(instance[image_column_name], 'rb') as fi:
im = Image.open(fi)
im.thumbnail((299, 299), Image.ANTIALIAS)
rgb_im = np.asarray(im.convert('RGB'))
exp = explainer.explain_instance(rgb_im,
predict_fn,
labels=range(len(labels)),
top_labels=None,
hide_color=hide_color,
num_features=num_features,
num_samples=num_samples,
batch_size=batch_size)
return exp
x_pred = face_image(image, dog, 128, 128)
print(x_pred.shape)
# load_model
model = load_model('./project/project02/model_save/best_xception.hdf5')
try:
#--------------------------------------------------------------
from lime.lime_image import LimeImageExplainer
top_labels = 3
explainer = LimeImageExplainer()
explanation = explainer.explain_instance(x_pred[0],
model.predict,
hide_color=0,
top_labels=top_labels,
num_samples=1000)
from skimage.segmentation import mark_boundaries
temp, mask = explanation.get_image_and_mask(explanation.top_labels[0],
positive_only=True,
hide_rest=True)
plt.imshow(mark_boundaries(temp / 2 + 0.5, mask))
#--------------------------------------------------------------
# predict
prediction = model.predict(x_pred)
number = np.argmax(prediction, axis=1)
# 카테고리 불러오기
class NNModel(object):
"""docstring for NNModel"""
def __init__(self,
model,
logger,
datapath="/home/chris/data/ML/images/oxfordiiipets"):
super(NNModel, self).__init__()
self.datapath = datapath
self.model = model
self.logger = logger
def load_raw_data(self):
self.image_files = os.listdir(os.path.join(self.datapath, "images"))
# self.trimap_files = os.listdir(os.path.join(self.datapath,
# "annotations", "trimaps"))
# self.xml_files = os.listdir(os.path.join(self.datapath,
# "annotations", "xmls"))
def preprocess(self):
self.load_raw_data()
self.images, valid_files = transform_img_fn([
os.path.join(self.datapath, "images", i)
for i in self.image_files[:100]
])
self.valid = [self.image_files[i] for i in valid_files]
def create_model_explainer(self):
self.preds = self.model.predict(self.images)
self.decoded_preds = decode_predictions(self.preds, )
self.explainer = LimeImageExplainer()
def get_explanation(self, i):
title_text = self.valid[i].split("_")[:-1]
title_text = " ".join(title_text)
num_labels = 3
preds = self.decoded_preds[i]
explanation = self.explainer.explain_instance(self.images[i],
self.model.predict,
top_labels=num_labels,
hide_color=0,
num_samples=200)
fig, subplots = plt.subplots(1, num_labels)
fig.set_figheight(3)
fig.set_figwidth(9)
j = 0
for exp, ax in zip(explanation.local_exp.keys(), subplots):
temp, mask = explanation.get_image_and_mask(exp,
positive_only=False,
num_features=5,
hide_rest=False)
ax.imshow(mark_boundaries(temp / 2 + 0.5, mask))
ax.set_title(preds[j][1] + ", %.3f" % preds[j][2])
j += 1
plt.tight_layout()
return mpld3.fig_to_html(fig), title_text
def get_custom_explanation(self, filename):
img = transform_custom_image(os.path.join('tmp', filename))
num_labels = 3
preds = self.model.predict(img)
img = img.reshape(299, 299, 3)
decoded_preds = decode_predictions(preds, )
explanation = self.explainer.explain_instance(img,
self.model.predict,
top_labels=num_labels,
hide_color=0,
num_samples=200)
fig, subplots = plt.subplots(1, num_labels)
fig.set_figheight(3)
fig.set_figwidth(9)
j = 0
for exp, ax in zip(explanation.local_exp.keys(), subplots):
temp, mask = explanation.get_image_and_mask(exp,
positive_only=False,
num_features=5,
hide_rest=False)
ax.imshow(mark_boundaries(temp / 2 + 0.5, mask))
ax.set_title(decoded_preds[0][j][1] +
", %.3f" % decoded_preds[0][j][2])
j += 1
plt.tight_layout()
return mpld3.fig_to_html(fig), filename
class XDeepLimeImageExplainer(Explainer):
def __init__(self, predict_proba, class_names):
"""Init function.
# Arguments
predict_proba: Function. Classifier prediction probability function.
class_names: List. A list of class names, ordered according to whatever the classifier is using.
"""
Explainer.__init__(self, predict_proba, class_names)
# Initialize explainer
self.set_parameters()
def set_parameters(self, **kwargs):
"""Parameter setter for lime_text.
# Arguments
**kwargs: Parameters setter. For more detail, please check https://lime-ml.readthedocs.io/en/latest/index.html.
"""
self.explainer = LimeImageExplainer(**kwargs)
def explain(self, instance, top_labels=None, labels=(1, ), **kwargs):
"""Generate explanation for a prediction using certain method.
# Arguments
instance: Array. An image to be explained.
top_labels: Integer. Number of labels you care about.
labels: Tuple. Labels you care about, if top_labels is not none, it will be replaced by the predicted top labels.
**kwargs: Parameters setter. For more detail, please check https://lime-ml.readthedocs.io/en/latest/index.html.
"""
Explainer.explain(self, instance, top_labels=top_labels, labels=labels)
self.explanation = self.explainer.explain_instance(
instance,
self.predict_proba,
top_labels=top_labels,
labels=self.labels,
**kwargs)
def show_explanation(self,
deprocess=None,
positive_only=True,
num_features=5,
hide_rest=False):
"""Visualization of explanation of lime_image.
# Arguments
deprocess: Function. A function to deprocess the image.
positive_only: Boolean. Whether only show feature with positive weight.
num_features: Integer. Numbers of feature you care about.
hide_rest: Boolean. Whether to hide rest of the image.
"""
Explainer.show_explanation(self)
exp = self.explanation
labels = self.labels
print()
print("LIME Explanation")
print()
fig, axs = plt.subplots(nrows=1,
ncols=len(labels),
figsize=(len(labels) * 3, 3))
assert hasattr(labels, '__len__')
for idx in range(len(labels)):
# Inverse
label = labels[-idx - 1]
result = exp.intercept[label]
local_exp = exp.local_exp[label]
for item in local_exp:
result += item[1]
print("Explanation for label {}:".format(self.class_names[label]))
print("Local Prediction: {:.3f}".format(result))
print("Original Prediction: {:.3f}".format(
self.original_pred[label]))
print()
temp, mask = exp.get_image_and_mask(label,
positive_only=positive_only,
num_features=num_features,
hide_rest=hide_rest)
if deprocess is not None:
temp = deprocess(temp)
if len(labels) == 1:
axs.imshow(mark_boundaries(temp, mask), alpha=0.9)
axs.axis('off')
axs.set_title("{}".format(self.class_names[label]))
else:
axs[idx].imshow(mark_boundaries(temp, mask), alpha=0.9)
axs[idx].axis('off')
axs[idx].set_title("{}".format(self.class_names[label]))
plt.show()
def predict(request):
""" Predict - Show Image(with lime) and Probabilities """
import numpy as np
import matplotlib.pyplot as plt
from skimage.segmentation import mark_boundaries
from keras.preprocessing import image
from keras.models import load_model
from lime.lime_image import LimeImageExplainer
from lime.wrappers.scikit_image import SegmentationAlgorithm
if request.method == 'POST' and request.FILES['test']:
if not os.path.exists(os.path.join(STATIC_URL, 'img/test/')):
os.mkdir(os.path.join(STATIC_URL, 'img/test/'))
test = request.FILES['test']
with open(os.path.join(STATIC_URL, 'img/test/', 'test.jpg'),
'wb+') as destination:
for chunk in test.chunks():
destination.write(chunk)
img = image.load_img(os.path.join(STATIC_URL, 'img/test/', 'test.jpg'),
target_size=(128, 128))
img = image.img_to_array(img)
img = np.expand_dims(img, axis=0)
o_img = img / 255
t_img = o_img[0] #for lime (4D -> 3D)
t_img = t_img.astype('double')
model = cnn_model()
model.load_weights('./model/cnn_model.h5')
guess = np.argmax(model.predict(o_img), axis=-1)
out = 'dog' if guess == 1 else 'cat'
lime_explainer = LimeImageExplainer()
segmenter = SegmentationAlgorithm('slic',
n_segments=100,
compactness=1,
sigma=1)
explanation = lime_explainer.explain_instance(
t_img, model.predict, segmentation_fn=segmenter)
temp, mask = explanation.get_image_and_mask(
model.predict(o_img).argmax(axis=1)[0],
positive_only=True,
hide_rest=False)
fig = plt.figure()
plt.imshow(mark_boundaries(temp, mask))
plt.axis('off')
plt.savefig(os.path.join(STATIC_URL, 'img/test/', 'lime.jpg'), )
plt.close(fig)
context = {
'content': out,
'prob_cat': model.predict(o_img)[0][0],
'prob_dog': model.predict(o_img)[0][1],
}
return render(request, 'predict/predict.html', context)
return render(request, 'predict/predict.html', {'content': 'wrong access'})
class LimeCounterfactualImage(object):
def __init__(self):
"""
Initialise the LimeImageExplainer
Args:
None
"""
self.lime_explainer = LimeImageExplainer()
def explain_instance(self, image, predict_fn, step_size=1000):
"""
Return a grayscale image showing the counterfactual for a give input image
Args:
image (np.ndarray): Input image for which counterfactual is to be produced.
predict_fn (Callable): A function which gives the models predictions.
step_size (None or int): Number of random pixels to change everytime we
try to make a new counterfactual candidate.
"""
explanation = self.lime_explainer.explain_instance(
image.astype("uint8"),
predict_fn,
top_labels=2,
hide_color=0,
num_samples=10,
)
temp, mask = explanation.get_image_and_mask(
explanation.top_labels[0],
positive_only=False,
num_features=10,
hide_rest=False,
)
real_pred = explanation.top_labels[0]
new_img = image.copy()
all_positives_removed, out_idx, in_idx = set(), 0, 0
while real_pred == explanation.top_labels[0]:
positives = list()
for x, mrow in enumerate(mask):
for y, mcol in enumerate(mrow):
if mcol == 1:
positives.append((x, y))
new_mask = mask.copy()
new_img = new_img.copy()
to_change = random.sample(positives, min(step_size, len(positives)))
for p in to_change:
(
new_img[p[0]][p[1]][0],
new_img[p[0]][p[1]][1],
new_img[p[0]][p[1]][2],
) = (0, 0, 0)
all_positives_removed.add(p)
explanation = self.lime_explainer.explain_instance(
new_img.astype("uint8"),
predict_fn,
top_labels=2,
hide_color=0,
num_samples=10,
)
temp, mask = explanation.get_image_and_mask(
explanation.top_labels[0],
positive_only=False,
num_features=10,
hide_rest=False,
)
grayscale_img = image.copy().astype("uint8")
for out_idx, mrow in enumerate(grayscale_img):
for in_idx, mcol in enumerate(mrow):
if (out_idx, in_idx) not in all_positives_removed:
grayscale_img[out_idx][in_idx][0] = 0
grayscale_img[out_idx][in_idx][1] = 0
grayscale_img[out_idx][in_idx][2] = 0
return grayscale_img