def get_datagenerators_folders(image_path_224,
image_path_224_crop,
image_path_331,
image_path_331_crop,
tta=False,
batch_size=100):
"""
This function produces a list of datagenerators from paths to all 4 preprocessed
versions of images
Parameters:
image_path_* (string): path to corresponding image data set
tta (boolean): switch to turn on test time augmentation, default is off
batch_size (int): mini-batch size for making predictions
Returns:
datagen_list (list): a list of test data generators.
"""
img_util_224 = imgUtils(224)
test_datagen_224_tta, test_datagen_224_notta = img_util_224.dataGen(
rotation=15, h_shift=0.05, w_shift=0.05)
img_util_331 = imgUtils(331)
test_datagen_331_tta, test_datagen_331_notta = img_util_331.dataGen(
rotation=15, h_shift=0.05, w_shift=0.05)
if tta:
test_datagen_224 = test_datagen_224_tta
test_datagen_331 = test_datagen_331_tta
else:
test_datagen_224 = test_datagen_224_notta
test_datagen_331 = test_datagen_331_notta
test_generator_224 = img_util_224.testgenerator_from_folder(
batch_size=batch_size, test=test_datagen_224, data_dir=image_path_224)
test_generator_224_crop = img_util_224.testgenerator_from_folder(
batch_size=batch_size,
test=test_datagen_224,
data_dir=image_path_224_crop)
test_generator_331 = img_util_331.testgenerator_from_folder(
batch_size=batch_size, test=test_datagen_331, data_dir=image_path_331)
test_generator_331_crop = img_util_331.testgenerator_from_folder(
batch_size=batch_size,
test=test_datagen_331,
data_dir=image_path_331_crop)
datagenlist = [
test_generator_224, test_generator_224_crop, test_generator_331,
test_generator_331_crop
]
datagenlist = datagenlist * 6
return datagenlist
def get_datagenerators_file(image_path_uncrop, image_path_crop, tta=False):
"""
This function creates a list of image data generators corresponds to each model
used to ensemble with different types of augmentation (cropped or uncopped).
Parameters:
image_path_crop/uncrop (string): the path to images with different
augmentations.
tta (boolean): a flag to check if test-time augmentation should be applied
or not.
Returns:
datagenlist (list): a list of image data generators corresponds to each model
used to ensemble.
"""
img_util_224 = imgUtils(224)
img_util_331 = imgUtils(331)
img_224 = img_util_224.proc_img(image_path_uncrop)
img_331 = img_util_331.proc_img(image_path_uncrop)
img_224_crop = img_util_224.proc_img(image_path_crop)
img_331_crop = img_util_331.proc_img(image_path_crop)
test_datagen_tta, test_datagen_no_tta = img_util_224.dataGen(rotation=15,
h_shift=0.05,
w_shift=0.05)
if tta:
test_datagen = test_datagen_tta
else:
test_datagen = test_datagen_no_tta
test_generator_224 = test_datagen.flow(img_224,
batch_size=1,
shuffle=False)
test_generator_224_crop = test_datagen.flow(img_224_crop,
batch_size=1,
shuffle=False)
test_generator_331 = test_datagen.flow(img_331,
batch_size=1,
shuffle=False)
test_generator_331_crop = test_datagen.flow(img_331_crop,
batch_size=1,
shuffle=False)
datagenlist = [
test_generator_224, test_generator_224_crop, test_generator_331,
test_generator_331_crop
]
datagenlist = datagenlist * 6
return datagenlist
model_name = args.model_name
img_size = args.img_size
weights = os.path.normpath(args.weight_path)
hyperparameters = args.hyperparameters
exp_name = model_name + '_' + str(img_size)
output_path = args.output
if output_path is not None:
output_path = os.path.normpath(output_path)
else:
output_path = os.getcwd()
train_dir, valid_dir, weight_dir1, weight_dir2, img_dir1, img_dir2 = make_path(
data_path, output_path, exp_name)
img_proc = imgUtils(img_size)
train_idg, val_idg = img_proc.dataGen(rotation_range, height_shift,
width_shift)
train_gen, val_gen = img_proc.generator(batch_size, train_idg, val_idg,
train_dir, valid_dir)
if hyperparameters is not None:
hyperparameters = pickle.load(open(hyperparameters, "rb"))
lr = hyperparameters['learning_rate']
momentum = hyperparameters['momentum']
dropout_rate = hyperparameters['dropout_rate']
else:
lr = 0.001
momentum = 0.9
dropout_rate = 0.3
def get_datagenerators_folder(image_path_224,
image_path_224_crop,
image_path_331,
image_path_331_crop,
tta=False):
"""
This function creates a list of image data generators corresponds to each model
used to ensemble using the path to directories with different types of augmentation
(224x224 or 331x331 in size, cropped or uncopped).
Parameters:
image_path_224/331_crop/NONE (string): the path to images with different
augmentations.
tta (boolean): a flag to check if test-time augmentation should be applied
or not.
Returns:
datagenlist (list): a list of image data generators corresponds to each model
used to ensemble.
"""
img_util_224 = imgUtils(224)
test_datagen_224_tta, test_datagen_224_notta = img_util_224.dataGen(
rotation=15, h_shift=0.05, w_shift=0.05)
img_util_331 = imgUtils(331)
test_datagen_331_tta, test_datagen_331_notta = img_util_331.dataGen(
rotation=15, h_shift=0.05, w_shift=0.05)
if tta:
test_datagen_224 = test_datagen_224_tta
test_datagen_331 = test_datagen_331_tta
else:
test_datagen_224 = test_datagen_224_notta
test_datagen_331 = test_datagen_331_notta
images_224 = glob.glob(os.path.join(image_path_224, '*'))
images_224 = pd.DataFrame.from_dict({'filename': images_224})
images_224_crop = glob.glob(os.path.join(image_path_224_crop, '*'))
images_224_crop = pd.DataFrame.from_dict({'filename': images_224_crop})
images_331 = glob.glob(os.path.join(image_path_331, '*'))
images_331 = pd.DataFrame.from_dict({'filename': images_331})
images_331_crop = glob.glob(os.path.join(image_path_331_crop, '*'))
images_331_crop = pd.DataFrame.from_dict({'filename': images_331_crop})
batch_size = 100
test_generator_224 = img_util_224.testgenerator_from_dataframe(
batch_size=batch_size,
test=test_datagen_224,
dataframe=images_224,
class_mode=None)
test_generator_224_crop = img_util_224.testgenerator_from_dataframe(
batch_size=batch_size,
test=test_datagen_224,
dataframe=images_224_crop,
class_mode=None)
test_generator_331 = img_util_331.testgenerator_from_dataframe(
batch_size=batch_size,
test=test_datagen_331,
dataframe=images_331,
class_mode=None)
test_generator_331_crop = img_util_331.testgenerator_from_dataframe(
batch_size=batch_size,
test=test_datagen_331,
dataframe=images_331_crop,
class_mode=None)
datagenlist = [
test_generator_224, test_generator_224_crop, test_generator_331,
test_generator_331_crop
]
datagenlist = datagenlist * 6
return datagenlist
def generate_gradCAM(input_path, img_size, models, output_path=os.getcwd()):
"""
This function generates a Grad-CAM image for a given image. The Grad-CAM image
is overlayed on the original image. The image will be demonstrated and saved
in a given path.
Parameters:
input_path (string): the path to the image.
img_size (int): the size of the input image.
models (list): a list of models used to generate the Grad-CAM image.
output_path (string): the path to save the output image.
"""
img_list = []
img_array_list = []
img_name_list = []
# current_path, input_filename = os.path.split(input_path)
result_base = os.path.join(output_path, 'gradCAM_img')
img_proc = imgUtils(img_size)
# Create new folder to save the resulting image
if not os.path.exists(result_base):
os.mkdir(result_base)
# Check if input path is a directory or file
if os.path.isdir(input_path):
for filename in os.listdir(input_path):
img_path = os.path.join(input_path, filename)
input_img = image.load_img(img_path,
target_size=(img_size, img_size),
color_mode='rgb',
interpolation='lanczos')
img_array = np.asarray(input_img, dtype='float64')
img_array_list.append(img_array)
img_preproc = img_proc.preprocess(img_array)
img = np.expand_dims(img_preproc, axis=0)
img_list.append(img)
# Split filename for result image name
base_name = os.path.basename(filename)
img_name = (os.path.splitext(base_name))[0]
result_name = img_name + '_gradCAM.jpg'
result_path = os.path.join(result_base, result_name)
img_name_list.append(result_path)
else:
input_img = image.load_img(input_path,
target_size=(img_size, img_size),
color_mode='rgb',
interpolation='lanczos')
img_array = np.asarray(input_img, dtype='float64')
img_array_list.append(img_array)
img_preproc = img_proc.preprocess(img_array)
img = np.expand_dims(img_preproc, axis=0)
img_list.append(img)
# Split filename for result image name
base_name = os.path.basename(input_path)
img_name, ext = os.path.splitext(base_name)
result_name = img_name + '_gradCAM.png'
result_path = os.path.join(result_base, result_name)
img_name_list.append(result_path)
i = 0
for img in img_list:
result_path = img_name_list[i]
img_array = img_array_list[i]
visualization = img_proc.gradCAM(img, models)
plt.rcParams['figure.figsize'] = (18, 6)
# Matplotlib preparations
fig, axes = plt.subplots(1, 3)
axes[0].imshow(img_array[..., 0], cmap='gray')
axes[0].set_title('Input')
axes[1].imshow(visualization)
axes[1].set_title('Grad-CAM')
heatmap = np.uint8(cm.jet(visualization)[..., :3] * 255)
original = np.uint8(cm.gray(img_array[..., 0])[..., :3] * 255)
axes[2].imshow(overlay(heatmap, original))
axes[2].set_title('Overlay')
plt.savefig(result_path)
i += 1