def convert_randoms(paths, labels, resize_dims=None, warp_ok=False):
# ** Reads images from paths, randomly augments them and returns the images and # labels as a list. **
# paths - list of image file paths as strings
# labels - list or numpy array of labels as any data type
# resize_dims - integer tuple of dimensions to resize the images to
# warp_ok - optional boolean denoting if risizing should maintain aspect ratio
images = []
for i, path in enumerate(paths):
try:
if resize_dims and not warp_ok:
img = imanip.resize(path, maxsizes=resize_dims)
else:
img = mpimg.imread(path)
if resize_dims:
img = sci.imresize(img, resize_dims)
img = imanip.random_augment(img)
except OSError:
# Uses augmented version of next image in list
if i == 0:
if resize_dims and not warp_ok:
img = imanip.resize(paths[i + 1], maxsizes=resize_dims)
else:
img = mpimg.imread(paths[i + 1])
if resize_dims:
img = sci.imresize(img, resize_dims)
img = imanip.random_augment(img)
labels[i] = labels[i + 1]
# Uses most recent original image
elif i > 0:
img = imanip.random_augment(images[-1])
labels[i] = labels[i - 1]
images.append(img)
return images, labels
def convert_images(paths,
labels,
resize_dims=None,
warp_ok=False,
rpaths=[],
rlabels=[]):
# ** Reads in images from their paths, returns the images with their
# corresponding labels. **
# paths - list of file paths to the images as strings
# labels - a numpy array or list of the corresponding labels to the images
# resize_dims - tuple of integer output dimensions to resize image.
# (does not maintain aspect ratio)
# warp_ok - boolean to select if image resize should maintain the aspect ratio
# randomly_augment - optional boolean to add randomly rotated,
# translated, and scaled images to the output
# rpaths - list of image file paths as strings to be randomly augmented
# rlabels - list of corresponding labels to rpaths
if len(rpaths) > 0 and len(rlabels) > 0:
# pool = Pool(processes=1)
# result = pool.apply_async(convert_randoms, (rpaths,rlabels,resize_dims))
rand_imgs, rand_labels = convert_randoms(rpaths, rlabels, resize_dims)
images = []
for i, path in enumerate(paths):
try:
if resize_dims and not warp_ok:
img = imanip.resize(path, maxsizes=resize_dims)
else:
img = mpimg.imread(path)
if resize_dims:
img = sci.imresize(img, resize_dims)
except OSError:
# Uses augmented version of next image in list
if i == 0:
if resize_dims and not warp_ok:
img = imanip.resize(paths[i + 1], maxsizes=resize_dims)
else:
img = mpimg.imread(paths[i + 1])
if resize_dims:
img = sci.imresize(img, resize_dims)
img = imanip.random_augment(img)
labels[i] = labels[i + 1]
# Uses most recent original image
elif i > 0:
sub_index = -1
if randomly_augment:
sub_index = -2
img = imanip.random_augment(images[sub_index])
labels[i] = labels[i - sub_index]
images.append(img)
if len(rpaths) > 0 and len(rlabels) > 0:
# result.wait()
# rand_imgs, rand_labels = result.get()
images = images + rand_imgs
labels = np.concatenate([labels, rand_labels], axis=0)
return np.array(images, dtype=np.float32), labels
return np.array(images, dtype=np.float32), labels
## Resizing the image
new_img_shapes = [(299,299,3),(256,256,3)]
dest_folders = ['resized']
for new_img_shape,dest_folder in zip(new_img_shapes,dest_folders):
for i,path,label in zip(count(),image_paths,labels):
split_path = path.split('/')
new_path = 'size_'+str(new_img_shape[0])+'_'+split_path[-1]
new_path = '/'.join(['data/resized']+[str(label)]+[new_path])
add_flip = True
if label == 1:
add_flip = False
# Used to exclude corrupt data
try:
imanip.resize(path, maxsizes=new_img_shape,
save_path=new_path,
add_flip=add_flip)
except OSError:
print("Error at path " + path)
## Using resized image
resized_image_locations = ['data/resized/']
image_paths = []
labels = []
for i,root_path in enumerate(resized_image_locations):
new_paths, new_labels, n_classes = read_paths(root_path)
if len(new_paths) > 0:
image_paths += new_paths
labels += new_labels