def get_prediction(img, convolutional_model: BaselineModel,
tensorflow_session: tf.Session):
"""
Get the prediction for a given input image
:param convolutional_model: The convolutional neural network model
:param tensorflow_session: The tensorflow session
:param img: The image for which to generate the prediction
:return: The prediction
"""
data = numpy.asarray(img_crop(img, IMG_PATCH_SIZE, IMG_PATCH_SIZE))
data_indices = range(data.shape[0])
img_predictions = []
data_node = tf.placeholder(tf.float32,
shape=(None, EFFECTIVE_INPUT_SIZE,
EFFECTIVE_INPUT_SIZE, NUM_CHANNELS))
output = tf.nn.softmax(convolutional_model.model_func()(data_node))
for i in range(0, data.shape[0], BATCH_SIZE):
batch_data = data[data_indices[i:i + BATCH_SIZE]]
output_prediction = tensorflow_session.run(
output, feed_dict={data_node: batch_data})
img_predictions.append(output_prediction)
stacked_predictions = [
numpy.stack(batch_predictions_list)
for batch_predictions_list in img_predictions
]
stacked_batches = numpy.vstack(stacked_predictions)
return label_to_img(img.shape[0], img.shape[1], IMG_PATCH_SIZE,
IMG_PATCH_SIZE, stacked_batches)
def extract_all_labels(path, num_images=-1, convert_to_1hot=True):
"""
Extract from ground truth images the class labels and convert them
into a 1-hot matrix of the form [image index, label index
:param path: The common path for all images
:return: A tensor of 1-hot matrix representation of the class labels
"""
print("Extracting all labels from " + path)
image_count = num_images if num_images != -1 else len(
next(os.walk(path))[2])
gt_imgs = []
data = []
for filename in os.listdir(path):
print(filename)
gt_imgs.append(mpimg.imread(os.path.join(path, filename)))
data = []
if (len(gt_imgs) > 0):
IMG_WIDTH = gt_imgs[0].shape[0]
if IMG_WIDTH > EFFECTIVE_INPUT_SIZE:
# List formed by consecutive series of patches of each image (patches ordered in row order)
gt_patches = [
img_crop(i, IMG_PATCH_SIZE, IMG_PATCH_SIZE, is_2d=True)
for i in gt_imgs
]
# List of all the patches, ordered by image
data = numpy.asarray([
gt_patches[i][j] for i in range(len(gt_patches))
for j in range(len(gt_patches[i]))
])
else:
data = gt_imgs
# Compute the class label of each patch based on the mean value
if convert_to_1hot:
labels = numpy.asarray(
[value_to_class(numpy.mean(data[i])) for i in range(len(data))])
else:
labels = data
return labels.astype(numpy.float32)
def extract_labels(filename, permutations):
"""
Extract from ground truth images the class labels and convert them
into a 1-hot matrix of the form [image index, label index
:param filename: The common path for all images
:return: A tensor of 1-hot matrix representation of the class labels
"""
"""Extract the labels into a 1-hot matrix [image index, label index]."""
gt_imgs = []
for i in range(1, NUM_IMAGES + 1):
imageid = "satImage_%.3d" % i
image_filename = filename + imageid + ".png"
if os.path.isfile(image_filename):
print('Loading ' + image_filename)
img = mpimg.imread(image_filename)
gt_imgs.append(img)
else:
print('File ' + image_filename + ' does not exist')
gt_imgs = numpy.array(gt_imgs)
if SHUFFLE_DATA:
gt_imgs = gt_imgs[permutations]
# List formed by consecutive series of patches of each image (patches ordered in row order)
gt_patches = [
img_crop(gt_imgs[i], IMG_PATCH_SIZE, IMG_PATCH_SIZE, is_2d=True)
for i in range(NUM_IMAGES)
]
# List of all the patches, ordered by image
data = numpy.asarray([
gt_patches[i][j] for i in range(len(gt_patches))
for j in range(len(gt_patches[i]))
])
# Compute the class label of each patch based on the mean value
labels = numpy.asarray(
[value_to_class(numpy.mean(data[i])) for i in range(len(data))])
# Convert to dense 1-hot representation.
return labels.astype(numpy.float32)
def extract_data(filename, permutations):
"""
Extract the images into a 4D tensor [image index, y, x, channels]
Values are rescales from [0, 255] down to [-0.5, 0.5]
:param filename: The common path of all images
:return: The extracted 4D tensor
"""
imgs = []
for i in range(1, NUM_IMAGES + 1):
imageid = "satImage_%.3d" % i
image_filename = filename + imageid + ".png"
if os.path.isfile(image_filename):
print('Loading ' + image_filename)
img = mpimg.imread(image_filename)
imgs.append(img)
else:
print('File ' + image_filename + ' does not exist')
IMG_WIDTH = imgs[0].shape[0]
IMG_HEIGHT = imgs[0].shape[1]
N_PATCHES_PER_IMAGE = (IMG_WIDTH / IMG_PATCH_SIZE) * (IMG_HEIGHT /
IMG_PATCH_SIZE)
imgs = numpy.array(imgs)
if SHUFFLE_DATA:
imgs = imgs[permutations]
# List formed by consecutive series of patches of each image (patches ordered in row order)
img_patches = [
img_crop(imgs[i], IMG_PATCH_SIZE, IMG_PATCH_SIZE, is_2d=False)
for i in range(NUM_IMAGES)
]
# List of all the patches, ordered by image
data = [
img_patches[i][j] for i in range(len(img_patches))
for j in range(len(img_patches[i]))
]
return numpy.asarray(data)
def extract_all_data(path, num_images=-1):
"""
Extract the images into a 4D tensor [image index, y, x, channels]
Values are rescales from [0, 255] down to [-0.5, 0.5]
:param path: The common path of all images
:return: The extracted 4D tensor
"""
imgs = []
data = []
print("Extracting all patches from " + path)
for filename in os.listdir(path):
print(filename)
imgs.append(mpimg.imread(os.path.join(path, filename)))
data = []
if (len(imgs) > 0):
IMG_WIDTH = imgs[0].shape[0]
# Check whether the images are already cropped
if IMG_WIDTH > EFFECTIVE_INPUT_SIZE:
# List formed by consecutive series of patches of each image (patches ordered in row order)
img_patches = [
img_crop(i, IMG_PATCH_SIZE, IMG_PATCH_SIZE, is_2d=False)
for i in imgs
]
# List of all the patches, ordered by image
data = [
img_patches[i][j] for i in range(len(img_patches))
for j in range(len(img_patches[i]))
]
else:
data = imgs
print("Detected already cropped image")
return numpy.asarray(data)