def extract_features(layer, file_name):
global X
image = imread(file_name)
image.resize((image.shape[0], image.shape[1], 1))
# overfeat expects rgb, so replicate the grayscale values twice
image = np.repeat(image, 3, 2)
image = imresize(image, (231, 231)).astype(np.float32)
# numpy loads image with colors as last dimension, so transpose tensor
h = image.shape[0]
w = image.shape[1]
c = image.shape[2]
image = image.reshape(w * h, c)
image = image.transpose()
image = image.reshape(c, h, w)
b = overfeat.fprop(image)
b = b.flatten()
x = overfeat.get_output(layer)
# maxpooling
x = np.amax(np.amax(x, 1), 1)
x = x.reshape(x.shape[0], 1).T
if X is None:
X = x
else:
X = np.append(X, x, axis=0)
def get_overfeat_features(imgs, weights_path, typ, layer=None, cache=None):
"""Returns features at layer for given image(s) from OverFeat model.
Small (fast) network: 22 layers
Large (accurate) network: 25 layers
Args:
imgs: Iterable of images each of shape (h,w,c)
weights_path: Path to the OverFeat weights
typ: 0 for small, 1 for large version of OverFeat
layer: The layer to extract features from
cache: Dict containing descs/other cached values
"""
if cache is None:
cache = {}
if 'overfeat_descs' not in cache:
# Initialize network
print('Loading OverFeat ({}) model...'.format(typ))
overfeat.init(weights_path, typ)
# Determine feature layer if none specified
if layer is None:
if overfeat.get_n_layers() == 22: # small
layer = 19 # 16 also recommended
else: # large
# Layer used by Zhang et al.
layer = 22
# Determine resize dim
if typ == 0:
resize = 231 # small network
else:
resize = 221 # large network
# Allocate for feature descriptors
descs = []
# Run images through network
print('Running images through OverFeat, extracting features '
'at layer {}.'.format(layer))
for idx, img in enumerate(imgs):
if (idx + 1) % 100 == 0:
print('Processing image {}...'.format(idx + 1))
# Preprocess image
img = overfeat_preprocess(img, resize)
# Run through model
_ = overfeat.fprop(img)
# Retrieve feature output
desc = overfeat.get_output(layer)
descs.append(desc)
# Free network
overfeat.free()
# NumPy-ify
descs = np.asarray(descs)
cache.update(overfeat_descs=descs)
else:
descs = cache['overfeat_descs']
return descs, cache
def overfeatFeatures(image):
image = imresize(image, (231, 231)).astype(np.float32)
h = image.shape[0]
w = image.shape[1]
c = image.shape[2]
# numpy loads image with colors as last dimension, transpose tensor
image = image.reshape(w * h, c)
image = image.transpose()
image = image.reshape(c, h, w)
# run overfeat on the image
overfeat.fprop(image)
features = overfeat.get_output(19).copy()
return features
def _compute_features(self, images):
import overfeat # soft dep
features = []
for image in images:
image = self.prepare_image(image)
overfeat.fprop(image)
feat = overfeat.get_output(self.feature_layer)
if self.merge == 'maxmean':
feat = feat.max(2).mean(1)
elif self.merge == 'meanmax':
feat = feat.mean(2).max(1)
else:
feat = self.merge(feat)
features.append(feat)
return np.vstack(features)
def _compute_features(self, images):
import overfeat # soft dep
features = []
for image in images:
image = self.prepare_image(image)
overfeat.fprop(image)
feat = overfeat.get_output(self.feature_layer)
if self.merge == 'maxmean':
feat = feat.max(2).mean(1)
elif self.merge == 'meanmax':
feat = feat.mean(2).max(1)
else:
feat = self.merge(feat)
features.append(feat)
return np.vstack(features)
def extract_overfeat(image_path):
# if overfeat_initialized == None or not overfeat_initialized:
# overfeat_initialized = True
print "Overfeat: ", image_path
image = imread(image_path)
print "Image shape: ", image.shape
if len(image.shape) == 2 or image.shape[2] == 2:
image = skimage.color.gray2rgb(image)
elif image.shape[2] == 4:
image_rgb = numpy.zeros((image.shape[0], image.shape[1], 3),
numpy.uint8)
image_rgb[:, :, 0] = image[:, :, 0]
image_rgb[:, :, 1] = image[:, :, 1]
image_rgb[:, :, 2] = image[:, :, 2]
image = image_rgb
h0 = image.shape[0]
w0 = image.shape[1]
d0 = float(min(h0, w0))
image = image[int(round((h0 - d0) / 2.)):int(round((h0 - d0) / 2.) + d0),
int(round((w0 - d0) / 2.)):int(round((w0 - d0) / 2.) +
d0), :]
image = imresize(image, (231, 231)).astype(numpy.float32)
#image = cv2.resize(image, (231, 231)).astype(numpy.float32)
# numpy loads image with colors as last dimension, transpose tensor
h = image.shape[0]
w = image.shape[1]
c = image.shape[2]
image = image.reshape(w * h, c)
image = image.transpose()
image = image.reshape(c, h, w)
print "Image size :", image.shape
out_categories = overfeat.fprop(image)
#layer 21,22,23
layer_output = overfeat.get_output(20)
print "Layer size: ", layer_output.shape
layer_output = layer_output.flatten()
descriptors = []
descriptors.append(layer_output)
out_categories = out_categories.flatten()
top = [(out_categories[i], i) for i in xrange(len(out_categories))]
top.sort()
print "\nTop classes :"
for i in xrange(5):
print(overfeat.get_class_name(top[-(i + 1)][1]))
return descriptors
def extract_overfeat(image_path):
# if overfeat_initialized == None or not overfeat_initialized:
# overfeat_initialized = True
print "Overfeat: ", image_path
image = imread(image_path)
print "Image shape: ", image.shape
if len(image.shape) == 2 or image.shape[2] == 2:
image = skimage.color.gray2rgb(image)
elif image.shape[2] == 4:
image_rgb = numpy.zeros((image.shape[0],image.shape[1], 3), numpy.uint8)
image_rgb[:,:,0] = image[:,:,0]
image_rgb[:,:,1] = image[:,:,1]
image_rgb[:,:,2] = image[:,:,2]
image = image_rgb
h0 = image.shape[0]
w0 = image.shape[1]
d0 = float(min(h0, w0))
image = image[int(round((h0-d0)/2.)):int(round((h0-d0)/2.)+d0),
int(round((w0-d0)/2.)):int(round((w0-d0)/2.)+d0), :]
image = imresize(image, (231, 231)).astype(numpy.float32)
#image = cv2.resize(image, (231, 231)).astype(numpy.float32)
# numpy loads image with colors as last dimension, transpose tensor
h = image.shape[0]
w = image.shape[1]
c = image.shape[2]
image = image.reshape(w*h, c)
image = image.transpose()
image = image.reshape(c, h, w)
print "Image size :", image.shape
out_categories = overfeat.fprop(image)
#layer 21,22,23
layer_output = overfeat.get_output(20)
print "Layer size: ", layer_output.shape
layer_output = layer_output.flatten()
descriptors = []
descriptors.append(layer_output)
out_categories = out_categories.flatten()
top = [(out_categories[i], i) for i in xrange(len(out_categories))]
top.sort()
print "\nTop classes :"
for i in xrange(5):
print(overfeat.get_class_name(top[-(i+1)][1]))
return descriptors
def run_overfeat(image, layer=None):
'''
runs an image through overfeat. returns the 1,000-length likelihoods
vector and N-length layer in the net as copied and formatted numpy arrays.
layer
None: means return 4,096-length feature vector just prior to the output layer.
int: return that layer instead.
'''
if not layer:
# get layer just before output layer
feature_layer = overfeat.get_n_layers() - 2
overfeat_likelihoods = overfeat.fprop(image)
overfeat_features = overfeat.get_output(feature_layer)
# flatten and copy. NOTE: copy() is intentional, don't delete.
formatted_features = copy(overfeat_features.flatten())
formatted_likelihoods = copy(overfeat_likelihoods.flatten())
return formatted_likelihoods, formatted_features
def run_overfeat(image, layer=None):
'''
runs an image through overfeat. returns the 1,000-length likelihoods
vector and N-length layer in the net as copied and formatted numpy arrays.
layer
None: means return 4,096-length feature vector just prior to the output layer.
int: return that layer instead.
'''
if not layer:
# get layer just before output layer
feature_layer = overfeat.get_n_layers() - 2
overfeat_likelihoods = overfeat.fprop(image)
overfeat_features = overfeat.get_output(feature_layer)
# flatten and copy. NOTE: copy() is intentional, don't delete.
formatted_features = copy(overfeat_features.flatten())
formatted_likelihoods = copy(overfeat_likelihoods.flatten())
return formatted_likelihoods, formatted_features
def getDictOfFeatures(directory, dirname):
d = {}
for filename in os.listdir(directory):
if filename == ".directory":
continue
image = imread(os.path.join(directory, filename)).astype(numpy.float32)
# numpy loads image with colors as last dimension, transpose tensor
h = image.shape[0]
w = image.shape[1]
c = image.shape[2]
image = image.reshape(w * h, c)
image = image.transpose()
image = image.reshape(c, h, w)
# run overfeat on the image
overfeat.fprop(image)
features = overfeat.get_output(19)
d[filename] = features.copy()
pickle.dump(d, open(dirname + "_full.txt", "w"))
return d
def extract_feat(self, imgs, ftype):
'''
extract features
'''
if ftype is 'hog':
from skimage.feature import hog
L = np.sqrt(len(imgs[0])).astype(int)
X_im = [imresize(arr.reshape((L, L)), (64, 64)) for arr in imgs]
pool = Pool(processes=8)
X = pool.map(hog, X_im)
pool.close()
return np.asarray(X)
elif ftype is 'overfeat':
overfeat.init('OverFeat/data/default/net_weight_0', 1)
L = np.sqrt(len(imgs[0])).astype(int)
imgs_color = [imresize(arr.reshape((L, L)), (231, 231)) for arr in imgs]
if len(imgs_color[0].shape) != 3:
cmap = pl.get_cmap('jet')
imgs_color = [np.delete(cmap(im/255.), 3, 2) for im in imgs_color]
imgs_roll = [im.transpose((2, 0, 1)).astype(np.float32) for im in imgs_color]
feats = np.zeros((len(imgs_roll), 4096), dtype = float)
for i in range(len(imgs_roll)):
b = overfeat.fprop(imgs_roll[i])
f22 = overfeat.get_output(22)
f22 = np.asarray(f22).squeeze().astype(np.float)
feats[i, :] = f22
return feats
elif ftype is 'pix':
return imgs
else:
raise NameError('{0} is not implemented!'.format(ftype))
def getFeatures(self, layer_number): self.featuresAtLayer = overfeat.get_output(layer_number) self.allFeatures.append(self.featuresAtLayer) #pdb.set_trace() if self.debug == True: pdb.set_trace()
def process_through_net(photo, feature_layer=None):
if not feature_layer:
feature_layer = overfeat.get_n_layers() - 2
likelihoods = copy(overfeat.fprop(photo).flatten())
features = copy(overfeat.get_output(feature_layer).flatten())
return likelihoods, features
for i in range(len(training_images)):
print("Training Image %s of %s" % (i, len(training_images)))
if smush:
image = smush_overfeat_images(training_images[i])
else:
image = load_overfeat_image(training_images[i])
b = overfeat.fprop(image)
if layer == 'all':
# Calculate features for all layers at once
features = []
for n in range(num_layers):
features.append(deepcopy(overfeat.get_output(n)))
training_features.append(features)
else:
training_features.append(deepcopy(overfeat.get_output(layer)))
for i in range(len(testing_images)):
print("Testing Image %s of %s" % (i, len(testing_images)))
if smush:
image = smush_overfeat_images(testing_images[i])
else:
image = load_overfeat_image(testing_images[i])
b = overfeat.fprop(image)
def overfeatFeatures(image):
overfeat.fprop(image)
features = overfeat.get_output(19).copy()
return features
for i in range(len(training_images)):
print("Training Image %s of %s" % (i, len(training_images)))
if smush:
image = smush_overfeat_images(training_images[i])
else:
image = load_overfeat_image(training_images[i])
b = overfeat.fprop(image)
if layer == 'all':
# Calculate features for all layers at once
features = []
for n in range(num_layers):
features.append(deepcopy(overfeat.get_output(n)))
training_features.append(features)
else:
training_features.append(deepcopy(overfeat.get_output(layer)))
for i in range(len(testing_images)):
print("Testing Image %s of %s" % (i, len(testing_images)))
if smush:
image = smush_overfeat_images(testing_images[i])
else:
image = load_overfeat_image(testing_images[i])
b = overfeat.fprop(image)
# numpy loads image with colors as last dimension, transpose tensor h = image.shape[0] w = image.shape[1] c = image.shape[2] image = image.reshape(w*h, c) image = image.transpose() image = image.reshape(c, h, w) except: return None return image def print_features(image_path, features): line = image_path + '\t' + '\t'.join(map(str,features)) print(line) # initialize overfeat. Note that this takes time, so do it only once if possible overfeat.init(network_path, 0) for image_path in sys.stdin: image_path = image_path[:-1] # remove \n char image = read_image(image_path) # run overfeat on the image features = [] if not image is None: b = overfeat.fprop(image) features = overfeat.get_output(layer_index).flatten() print_features(image_path, features)
