def main():
'''
for now, this function grabs a ndpi image, splits the image and the mask
and saves the splits in the split directory.
'''
# NOTE: Pasar de "data/test" a "data/train" para hacer el training set
preprocess_dir = "data/train"
os.chdir(preprocess_dir)
print(os.getcwd())
build_dirs(preprocess_dir)
file_list, _, _ = list_files_from_dir(extension=".ndpi")
for i, f in enumerate(file_list):
print("{}. ".format(i) + f)
width, height = 128, 128
print("Tile size: {}x{}".format(height, width))
for ndpi_file in file_list:
print(ndpi_file)
ndp_image, image_annotation_list = call_ndpi_ndpa(ndpi_file)
rectangle_split_ndpi_ndpa(ndp_image=ndp_image,
image_annotation_list=image_annotation_list,
split_height=height,
split_width=width,
tohsv=False,
path_ndpi="split/X",
n_bkgnd_tiles=5000)
data_augmentation()
def main(clean=False):
'''
for now, this function grabs a ndpi image, splits the image and the mask
and saves the splits in the split directory.
'''
os.chdir("data/test")
file_list, _, _ = list_files_from_dir(extension=".ndpi")
for i, f in enumerate(file_list):
print("{}. ".format(i) + f)
print(file_list)
for ndpi_file in file_list:
print(ndpi_file)
ndp_image, image_annotation_list = call_ndpi_ndpa(ndpi_file)
width = floor(ndp_image.width_lvl_0 / 4)
height = floor(ndp_image.height_lvl_0 / 4)
rectangle_split_ndpi(ndp_image=ndp_image,
split_width=width,
split_height=height,
tohsv=False,
path="grandes_RGB")
if clean:
clean_split_files()
def main():
# NOTE: remember, the model used must have been trained with the same tile side
# as the one being used here for testing
test_set = "128px_x20_RGB_Box5_newPreprocessing_modeRangePreprocessing_partialsAre0_CD8"
test_set_directory = "data/test/split/{}/X".format(test_set)
file_list, dir_list, counts = list_files_from_dir(directory=test_set_directory,
extension=".tif")
print("\n\nNumber of elements in file list: " + str(len(file_list)))
ild = {file_list[i]: dir_list[i] for i in range(len(dir_list))}
tile_side = 128
model_directory = "D:/felipe/epi_seg/src/models/"
# model_list, _, _ = list_files_from_dir(directory=model_directory[:-1], extension=".h5")
# NOTE: por si solo necesito usar un modelo, puedo usar la linea:
model_list = [
"20200117_InceptionV3_lossBCE_colorRGB_Box1-4_x20_modeRangePrep_partialsAre0_p16+CD8_optSGD_wDropout_128px_e40.h5",
"20200117_InceptionV3_lossBCE_colorRGB_Box1-4_x20_modeRangePrep_partialsAre0_p16+CD8_optSGD_wDropout_128px_e10.h5"
]
print("\nNumber models: " + str(len(model_list)) +"\n\n")
for m in model_list:
print("Model name: " + m)
model = tf.keras.models.load_model(model_directory + m,
custom_objects={'precision_m': precision_m,
'recall_m': recall_m,
'f1_m': f1_m,
'auc_m': auc_m})
test_generator = ImageTestGenerator(list_IDs=file_list,
image_label_directory=ild,
source_directory=test_set_directory,
tile_side=tile_side,
batch_size=64)
results = model.predict_generator(generator=test_generator,
workers=8,
use_multiprocessing=True,
verbose=1)
print("shape of results: {}".format(results.shape))
rows = zip(file_list, dir_list, results.reshape(results.shape[0],))
with open("data/test/split_results/" + test_set + "/" + m[:-3] + ".csv", "w", newline='') as f:
writer = csv.writer(f)
for row in rows:
writer.writerow(row)
K.clear_session()
def data_augmentation(directory="split/X/1", flip_imgs=True):
'''
generates extra images for eah image in directory: 3 addicional
images corresponding to 90, 180, 270 rotations.
if flip=True the flipped version of the 4 images (original + 3)
will be created as well.
'''
labels = {}
print("Data augmentation commencing for images labelled 'epithelium'.")
print("flip_imgs={}".format(flip_imgs))
file_list, _, _ = list_files_from_dir(directory=directory, extension=".tif")
for img_name in tqdm(file_list):
img = imread(directory + "/" + img_name)
img_90, labels["90_" + img_name] = flip(transpose(img), 1), 1
save_np_as_image(cvtColor(img_90, COLOR_BGR2RGB),
directory + "/90_" + img_name)
img_180, labels["180_" + img_name] = flip(img, -1), 1
save_np_as_image(cvtColor(img_180, COLOR_BGR2RGB),
directory + "/180_" + img_name)
img_270, labels["270_" + img_name] = flip(transpose(img), 0), 1
save_np_as_image(cvtColor(img_270, COLOR_BGR2RGB),
directory + "/270_" + img_name)
if flip_imgs:
img_f, labels["f_" + img_name] = flip(img, 1), 1
save_np_as_image(cvtColor(img_f, COLOR_BGR2RGB),
directory + "/f_" + img_name)
img_90f, labels["90f_" + img_name] = flip(img_90, 1), 1
save_np_as_image(cvtColor(img_90f, COLOR_BGR2RGB),
directory + "/90f_" + img_name)
img_180f, labels["180f_" + img_name] = flip(img_180, 1), 1
save_np_as_image(cvtColor(img_180f, COLOR_BGR2RGB),
directory + "/180f_" + img_name)
img_270f, labels["270f_" + img_name] = flip(img_270, 1), 1
save_np_as_image(cvtColor(img_270f, COLOR_BGR2RGB),
directory + "/270f_" + img_name)
json.dump(labels, open("split/X/augmentations.txt", "w"))
pass
def main():
'''
trains a model from a set of tiles previously preprocessed and separated into
folders according to the label
'''
### background_filters[0] --> ["stdDev", 0, "InceptionV3", True, "BCE", ]:
### "stdDev" --> stdDev o modeRange
### "0" --> "0" or "1" depending on "partialsAre"
### "InceptionV3" --> determines the base model. Can be one between "InceptionResNetV2", "InceptionV3", "ResNet50", "Xception", "basic" (which is default)
### True --> wether conv_base is trainable or not (if not, then transfer learning is applied)
### "BCE" --> loss function: can be "BCE" (binary crossentropy) or "ST" (stability training) (TODO: todavia no esta implementado este parametro, falta terminarlo)
### "RGB" --> color model, can be "RGB" or "HSV" (TODO: todavia no esta implementado este parametro, falta terminarlo)
parameter_sets = [
# ["modeRange", "1", "InceptionV3", True, "BCE", "RGB"],
# ["stdDev", "1", "InceptionV3", True, "BCE", "RGB"],
# ["modeRange", "0", "InceptionV3", True, "BCE", "RGB"],
# ["stdDev", "0", "InceptionV3", True, "BCE", "RGB"],
# ["modeRange", "0", "InceptionV3", False, "BCE", "RGB"],
# ["modeRange", "1", "InceptionV3", False, "BCE", "RGB"],
# ["modeRange", "1", "ResNet50", True, "BCE", "RGB"],
# ["stdDev", "1", "ResNet50", True, "BCE", "RGB"],
# ["modeRange", "0", "ResNet50", True, "BCE", "RGB"],
# ["stdDev", "0", "ResNet50", True, "BCE", "RGB"],
# ["modeRange", "1", "basic", True, "BCE", "RGB"], # <--- got to 39 eopchs!
["stdDev", "1", "basic", True, "BCE", "RGB"],
# ["modeRange", "0", "basic", True, "BCE", "RGB"],
# ["stdDev", "0", "basic", True, "BCE", "RGB"],
# ["modeRange", "1", "InceptionResNetV2", True, "BCE", "RGB"],
# ["stdDev", "1", "InceptionResNetV2", True, "BCE", "RGB"],
# ["modeRange", "0", "InceptionResNetV2", True, "BCE", "RGB"],
# ["stdDev", "0", "InceptionResNetV2", True, "BCE", "RGB"],
["modeRange", "1", "Xception", True, "BCE", "RGB"],
["stdDev", "1", "Xception", True, "BCE", "RGB"],
["modeRange", "0", "Xception", True, "BCE", "RGB"],
["stdDev", "0", "Xception", True, "BCE", "RGB"],
]
for p_set in parameter_sets:
name_base = "models/" + time.strftime("%Y%m%d") + "_" + p_set[2] + "_trainable" + \
str(p_set[3]) + "_loss" + p_set[4] + "_color" + p_set[5] + "_15pics_x20_" + \
p_set[0] + "Preprocessing_partialsAre" + p_set[1]
model_name = name_base + "_{}px_epoch_{}.h5"
bg_filter_dir = "128px_x20_RGB_15pics_{}Preprocessing_partialsAre{}/".format(
p_set[0], p_set[1])
full_dir = "data/train/split/{}X".format(bg_filter_dir)
file_list, dir_list, counts = list_files_from_dir(directory=full_dir,
extension=".tif")
print("\n\nImage folder: " + bg_filter_dir)
print("Image path: " + full_dir)
print("Number of elements in file list: " + str(len(file_list)))
print(str(counts) + "\n\n")
print("Base model: " + p_set[2])
print("Base model trainable: " + str(p_set[3]))
print("Loss function: " + str(p_set[4]))
print("Color model: " + str(p_set[5]))
train_list, val_list, _ = train_validation_test_partition(file_list,
prop=(0.6,
0.4,
0.0))
ild = {file_list[i]: dir_list[i] for i in range(len(dir_list))}
tile_side = 128
saver = CustomSaver(model_name=model_name, tile_side=tile_side)
training_generator = ImageGenerator2(list_IDs=train_list,
image_label_directory=ild,
source_directory=full_dir,
tile_side=tile_side,
batch_size=64)
validation_generator = ImageGenerator2(list_IDs=val_list,
image_label_directory=ild,
source_directory=full_dir,
tile_side=tile_side,
batch_size=64)
class_weight = {0: 1., 1: (counts["-1"] + counts["0"]) / counts["1"]}
# epochs = [40]
# for e in epochs:
model, history = basic_dl_model(
tile_side,
saver=saver,
model_name=model_name,
training_generator=training_generator,
validation_generator=validation_generator,
class_weight=class_weight,
epochs=40,
base=p_set[2],
trainable=p_set[3])
# model.save(model_name.format(tile_side, e))
history_name = name_base + "_" + str(tile_side) + "px.json"
with open(history_name, 'w') as jsonfile:
json.dump(history.history, jsonfile)
K.clear_session()
def main():
'''
trains a model from a set of tiles previously preprocessed and separated into
folders according to the label
'''
### ["modeRange", "1", "InceptionV3", True, "BCE", "RGB", "p16", "15pics", "SGD"]
### [ 0 1 2 3 4 5 6 7 8 ]
### 0: type of background filterinf --> "stdDev" or "modeRange"
### 1: labelling of partially annotated triles --> "0" (non-epithelium) or "1" (epithelium)
### 2: DL base architecture --> "InceptionResNetV2", "InceptionV3", "ResNet50", "Xception", "basic" (which is default)
### 3: Variable that determines if layers of base architecture are trainable (no transfer learning) or not (transfer learning) --> True, False
### 4: Loss function --> "BCE" (binary crossentropy) or "ST" (stability training) (TODO: todavia no esta implementado este parametro, falta terminarlo)
### 5: Color model --> "RGB" or "HSV"
### 6: IHC used --> "p16" or "p16+CD8"
### 7: images used --> "15pics" or "Box1-4"
### 8: optimizer --> "Adam" o "SGD" (TODO: todavia no esta implemenbtado este parametro)
parameter_sets = [
# ["modeRange", "1", "InceptionV3", True, "BCE", "HSV", "p16", "15pics", "Adam"],
# ["stdDev", "1", "InceptionV3", True, "BCE", "RGB", "p16", "15pics", "Adam"],
# ["modeRange", "0", "InceptionV3", True, "BCE", "RGB", "p16+CD8", "Box1-4", "SGD"],
[
"modeRange", "0", "InceptionV3", True, "ST", "RGB", "p16",
"15pics", "SGD"
],
# ["stdDev", "0", "InceptionV3", True, "BCE", "RGB", "p16", "15pics", "Adam"],
# ["modeRange", "0", "InceptionV3", False, "BCE", "RGB", "p16", "15pics", "Adam"],
# ["modeRange", "1", "InceptionV3", False, "BCE", "RGB", "p16", "15pics", "Adam"],
]
for p_set in parameter_sets:
name_base = "models/" + time.strftime("%Y%m%d") + "_" + p_set[2] + "_loss" + p_set[4] + \
"_color" + p_set[5] + "_" + p_set[7] + "_x20_" + p_set[0] + "Prep_partialsAre" + \
p_set[1] + "_" + p_set[6] + "_opt" + p_set[8] + "_wDropout"
model_name = name_base + "_{}px_e{}.h5"
bg_filter_dir = "128px_x20_RGB_{}_{}Preprocessing_partialsAre{}_{}/".format(
p_set[7], p_set[0], p_set[1], p_set[6])
full_dir = "data/train/split/{}X".format(bg_filter_dir)
file_list, dir_list, counts = list_files_from_dir(directory=full_dir,
extension=".tif")
tohsv = False
if p_set[5] == "HSV":
tohsv = True
print("\n\nImage folder: " + bg_filter_dir)
print("Image path: " + full_dir)
print("Number of elements in file list: " + str(len(file_list)))
print(str(counts) + "\n\n")
print("Base model: " + p_set[2])
print("Base model trainable: " + str(p_set[3]))
print("optimizer: " + p_set[8])
print("Loss function: " + str(p_set[4]))
print("Color model: " + str(p_set[5]) + "- tohsv=" + str(tohsv))
train_list, val_list, _ = train_validation_test_partition(file_list,
prop=(0.6,
0.4,
0.0))
ild = {file_list[i]: dir_list[i] for i in range(len(dir_list))}
tile_side = 128
saver = CustomSaver(model_name=model_name, tile_side=tile_side)
training_generator = ImageGenerator2(list_IDs=train_list,
image_label_directory=ild,
source_directory=full_dir,
tile_side=tile_side,
batch_size=64,
hsv=tohsv)
validation_generator = ImageGenerator2(list_IDs=val_list,
image_label_directory=ild,
source_directory=full_dir,
tile_side=tile_side,
batch_size=64,
hsv=tohsv)
class_weight = {0: 1., 1: (counts["-1"] + counts["0"]) / counts["1"]}
# epochs = [40]
# for e in epochs:
model, history = basic_dl_model(
tile_side,
saver=saver,
model_name=model_name,
training_generator=training_generator,
validation_generator=validation_generator,
class_weight=class_weight,
epochs=40,
base=p_set[2],
trainable=p_set[3])
# model.save(model_name.format(tile_side, e))
history_name = name_base + "_" + str(tile_side) + "px.json"
with open(history_name, 'w') as jsonfile:
json.dump(history.history, jsonfile)
K.clear_session()
# Evaluate model on data
# ndpi_file = "src/data/test/prueba2.ndpi"
# ndp_image, image_annotation_list = call_ndpi_ndpa(ndpi_file)
# val_X = np.array(ndp_image.read_region(location=(0, 0),
# level=0,
# size=(ndp_image.width_lvl_0,
# ndp_image.height_lvl_0)))
# val_X = cv2.cvtColor(val_X[:, :, :3], cv2.COLOR_RGB2BGR)
tile_size = 128
resolution = "x20"
image_dir = "src/data/test/grandes_RGB/{}".format(resolution)
file_list, _, _ = list_files_from_dir(directory=image_dir, extension=".tif")
# image_filename = "S04_2819_p16_RTU_ER1_20 - 2016-04-12 15.39.25_(30720,16384)_10240x8192.tif"
for image_filename in file_list:
print(image_filename)
val_X = cv2.imread(image_dir + "/" + image_filename)
val_X = normalize_image(val_X)
n_ver = floor(val_X.shape[0] / tile_size)
n_hor = floor(val_X.shape[1] / tile_size)
val_X = convert_image_to_stack_of_tiles(val_X, tile_size, tile_size)