Module models.CNN_Base
+
+
++Expand source code +
+import os
+
+import glob
+import datetime
+
+import skimage.io
+import numpy as np
+
+import tensorflow as tf
+
+import keras
+from keras import backend as K
+from keras.models import Model, load_model
+from keras.callbacks import EarlyStopping, ReduceLROnPlateau, ModelCheckpoint, TensorBoard, ProgbarLogger
+
+from .internals.image_functions import Image_Functions
+from .internals.network_config import Network_Config
+from .internals.dataset import Dataset
+
+class CNN_Base(Dataset, Image_Functions):
+ def __init__(self, model_dir = None, config_filepath = None, **kwargs):
+ """Creates the base neural network class with basic functions
+
+ Parameters
+ ----------
+ model_dir : `str`, optional
+ [Default: None] Folder where the model is stored
+ config_filepath : `str`, optional
+ [Default: None] Filepath to the config file
+ **kwargs
+ Parameters that are passed to :class:`network_config.Network_Config`
+
+ Attributes
+ ----------
+ config : :class:`network_config.Network_Config`
+ Network_config object containing the config and necessary functions
+ """
+
+ super().__init__()
+
+ self.config = Network_Config(model_dir = model_dir, config_filepath = config_filepath, **kwargs)
+
+ self.config.update_parameter(["general", "now"], datetime.datetime.now())
+
+ if self.config.get_parameter("use_cpu") is True:
+ self.initialize_cpu()
+ else:
+ self.initialize_gpu()
+
+ #######################
+ # Logging functions
+ #######################
+ def init_logs(self):
+ """Initiates the parameters required for the log file
+ """
+ # Directory for training logs
+ print(self.config.get_parameter("name"), self.config.get_parameter("now"))
+ self.log_dir = os.path.join(self.config.get_parameter("model_dir"), "{}-{:%Y%m%dT%H%M}".format(self.config.get_parameter("name"), self.config.get_parameter("now")))
+
+ # Path to save after each epoch. Include placeholders that get filled by Keras.
+ self.checkpoint_path = os.path.join(self.log_dir, "{}-{:%Y%m%dT%H%M}_*epoch*.h5".format(self.config.get_parameter("name"), self.config.get_parameter("now")))
+ self.checkpoint_path = self.checkpoint_path.replace("*epoch*", "{epoch:04d}")
+
+ def write_logs(self):
+ """Writes the log file
+ """
+ # Create log_dir if it does not exist
+ if os.path.exists(self.log_dir) is False:
+ os.makedirs(self.log_dir)
+
+ # save the parameters used in current run to logs dir
+ self.config.write_config(os.path.join(self.log_dir, "{}-{:%Y%m%dT%H%M}-config.yml".format(self.config.get_parameter("name"), self.config.get_parameter("now"))))
+
+ #######################
+ # Initialization functions
+ #######################
+ def summary(self):
+ """Summary of the layers in the model
+ """
+ self.model.summary()
+
+ def compile_model(self, optimizer, loss):
+ """Compiles model
+
+ Parameters
+ ----------
+ optimizer
+ Gradient optimizer used in during the training of the network
+ loss
+ Loss function of the network
+ """
+ self.model.compile(optimizer, loss = loss, metrics = self.config.get_parameter("metrics"))
+
+ def initialize_model(self):
+ """Initializes the logs, builds the model, and chooses the correct initialization function
+ """
+ # write parameters to yaml file
+ self.init_logs()
+ if self.config.get_parameter("for_prediction") is False:
+ self.write_logs()
+
+ # build model
+ self.model = self.build_model(self.config.get_parameter("input_size"))
+
+ # save model to yaml file
+ if self.config.get_parameter("for_prediction") is False:
+ self.config.write_model(self.model, os.path.join(self.log_dir, "{}-{:%Y%m%dT%H%M}-model.yml".format(self.config.get_parameter("name"), self.config.get_parameter("now"))))
+
+ print("{} using single GPU or CPU..".format("Predicting" if self.config.get_parameter("for_prediction") else "Training"))
+ self.initialize_model_normal()
+
+ def initialize_cpu(self):
+ """Sets the session to only use the CPU
+ """
+ config = tf.ConfigProto(
+ device_count = {'CPU' : 1,
+ 'GPU' : 0}
+ )
+ session = tf.Session(config=config)
+ K.set_session(session)
+
+ def initialize_gpu(self):
+ """Sets the seesion to use the gpu specified in config file
+ """
+ os.environ["CUDA_DEVICE_ORDER"]="PCI_BUS_ID" # see issue #152
+ os.environ['CUDA_VISIBLE_DEVICES'] = str(self.config.get_parameter("visible_gpu")) # needs to be a string
+
+ config = tf.ConfigProto()
+ config.gpu_options.allow_growth = True
+ sess = tf.Session(config=config)
+ K.tensorflow_backend.set_session(sess)
+
+ def initialize_model_normal(self):
+ """Initializes the optimizer and any specified callback functions
+ """
+ opt = self.optimizer_function()
+ self.compile_model(optimizer = opt, loss = self.loss_function(self.config.get_parameter("loss")))
+
+ if self.config.get_parameter("for_prediction") == False:
+ self.callbacks = [self.model_checkpoint_call(verbose = True)]
+
+ if self.config.get_parameter("use_tensorboard") is True:
+ self.callbacks.append(self.tensorboard_call())
+
+ if self.config.get_parameter("reduce_LR_on_plateau") is True:
+ self.callbacks.append(ReduceLROnPlateau(monitor=self.config.get_parameter("reduce_LR_monitor"),
+ factor = self.config.get_parameter("reduce_LR_factor"),
+ patience = self.config.get_parameter("reduce_LR_patience"),
+ min_lr = self.config.get_parameter("reduce_LR_min_lr"),
+ verbose = True))
+
+ if self.config.get_parameter("early_stopping") is True:
+ self.callbacks.append(EarlyStopping(monitor=self.config.get_parameter("early_stopping_monitor"),
+ patience = self.config.get_parameter("early_stopping_patience"),
+ min_delta = self.config.get_parameter("early_stopping_min_delta"),
+ verbose = True))
+
+ #######################
+ # Optimizer/Loss functions
+ #######################
+ def optimizer_function(self, learning_rate = None):
+ """Initialize optimizer function
+
+ Parameters
+ ----------
+ learning_rate : `int`
+ Learning rate of the descent algorithm
+
+ Returns
+ ----------
+ optimizer
+ Function to call the optimizer
+ """
+ if learning_rate is None:
+ learning_rate = self.config.get_parameter("learning_rate")
+ if self.config.get_parameter("optimizer_function") == 'sgd':
+ return keras.optimizers.SGD(lr = learning_rate,
+ decay = self.config.get_parameter("decay"),
+ momentum = self.config.get_parameter("momentum"),
+ nesterov = self.config.get_parameter("nesterov"))
+ elif self.config.get_parameter("optimizer_function") == 'rmsprop':
+ return keras.optimizers.RMSprop(lr = learning_rate,
+ decay = self.config.get_parameter("decay"))
+ elif self.config.get_parameter("optimizer_function") == 'adam':
+ return keras.optimizers.Adam(lr = learning_rate,
+ decay = self.config.get_parameter("decay"))
+
+ def loss_function(self, loss):
+ """Initialize loss function
+
+ Parameters
+ ----------
+ loss : `str`
+ Name of the loss function
+
+ Returns
+ ----------
+ loss
+ Function to call loss function
+ """
+ if loss == "binary_crossentropy":
+ print("Using binary crossentropy")
+ return loss
+ elif loss == "jaccard_distance_loss":
+ print("Using jaccard distance loss")
+ from .internals.losses import jaccard_distance_loss
+ return jaccard_distance_loss
+ elif loss == "lovasz_hinge":
+ print("Using Lovasz-hinge loss")
+ from .internals.losses import lovasz_loss
+ return lovasz_loss
+ elif loss == "dice_loss":
+ print("Using Dice loss")
+ from .internals.losses import dice_coef_loss
+ return dice_coef_loss
+ elif loss == "bce_dice_loss":
+ print("Using 1 - Dice + BCE loss")
+ from .internals.losses import bce_dice_loss
+ return bce_dice_loss
+ elif loss == "ssim_loss":
+ print("Using DSSIM loss")
+ from .internals.losses import DSSIM_loss
+ return DSSIM_loss
+ elif loss == "bce_ssim_loss":
+ print("Using BCE + DSSIM loss")
+ from .internals.losses import bce_ssim_loss
+ return bce_ssim_loss
+ elif loss == "mean_squared_error":
+ return keras.losses.mean_squared_error
+ elif loss == "mean_absolute_error":
+ return keras.losses.mean_absolute_error
+ elif loss == "ssim_mae_loss":
+ print("Using DSSIM + MAE loss")
+ from .internals.losses import dssim_mae_loss
+ return dssim_mae_loss
+ else:
+ print("Using {}".format(loss))
+ return loss
+
+ #######################
+ # Callbacks
+ #######################
+ def tensorboard_call(self):
+ """Initialize tensorboard call
+ """
+ return TensorBoard(log_dir=self.log_dir,
+ batch_size = self.config.get_parameter("batch_size_per_GPU"),
+ write_graph=self.config.get_parameter("write_graph"),
+ write_images=self.config.get_parameter("write_images"),
+ write_grads=self.config.get_parameter("write_grads"),
+ update_freq='epoch',
+ histogram_freq=self.config.get_parameter("histogram_freq"))
+
+ def model_checkpoint_call(self, verbose = 0):
+ """Initialize model checkpoint call
+ """
+ return ModelCheckpoint(self.checkpoint_path, save_weights_only=True, verbose=verbose)
+
+ #######################
+ # Clear memory once training is done
+ #######################
+ def end_training(self):
+ """Deletes model and releases gpu memory held by tensorflow
+ """
+ # del reference to model
+ del self.model
+
+ # clear memory
+ tf.reset_default_graph()
+ K.clear_session()
+
+ # take hold of cuda device to shut it down
+ from numba import cuda
+ cuda.select_device(0)
+ cuda.close()
+
+ #######################
+ # Train Model
+ #######################
+ def train_model(self, verbose = True):
+ """Trains model
+
+ Parameters
+ ----------
+ verbose : `int`, optional
+ [Default: True] Verbose output
+ """
+ history = self.model.fit(self.aug_images, self.aug_ground_truth, validation_split = self.config.get_parameter("val_split"),
+ batch_size = self.config.get_parameter("batch_size"), epochs = self.config.get_parameter("num_epochs"), shuffle = True,
+ callbacks=self.callbacks, verbose=verbose)
+
+ self.end_training()
+
+ #######################
+ # Predict using loaded model weights
+ #######################
+ # TODO: change to load model from yaml file
+ def load_model(self, model_dir = None):
+ """Loads model from h5 file
+
+ Parameters
+ ----------
+ model_dir : `str`, optional
+ [Default: None] Directory containing the model file
+ """
+ # TODO: rewrite to load model from yaml file
+ if model_dir is None:
+ model_dir = self.config.get_parameter("model_dir")
+
+ if os.path.isdir(model_dir) is True:
+ list_weights_files = glob.glob(os.path.join(model_dir,'*.h5'))
+ list_weights_files.sort() # To ensure that [-1] gives the last file
+
+ model_dir = os.path.join(model_dir,list_weights_files[-1])
+
+ self.model.load_model(model_dir)
+ print("Loaded model from: " + model_dir)
+
+ def load_weights(self, model_dir = None, weights_index = -1):
+ """Loads weights from h5 file
+
+ Parameters
+ ----------
+ model_dir : `str`, optional
+ [Default: None] Directory containing the weights file
+ weights_index : `int`, optional
+ [Default: -1]
+ """
+ if model_dir is None:
+ model_dir = self.config.get_parameter("model_dir")
+
+ if os.path.isdir(model_dir) is True:
+ list_weights_files = glob.glob(os.path.join(model_dir,'*.h5'))
+ list_weights_files.sort() # To ensure that [-1] gives the last file
+ self.weights_path = list_weights_files[weights_index]
+ model_dir = os.path.join(model_dir, self.weights_path)
+ else:
+ self.weights_path = model_dir
+
+ self.model.load_weights(model_dir)
+ print("Loaded weights from: " + model_dir)
+
+ def predict_images(self, image_dir):
+ """Perform prediction on images found in ``image_dir``
+
+ Parameters
+ ----------
+ image_dir : `str`
+ Directory containing the images to perform prediction on
+
+ Returns
+ ----------
+ image : `array_like`
+ Last image that prediction was perfromed on
+ """
+ # load image list
+ image_list = self.list_images(image_dir)
+
+ for image_path in image_list:
+ image = self.load_image(image_path = image_path)
+
+ # percentile normalization
+ if self.config.get_parameter("percentile_normalization"):
+ image, _, _ = self.percentile_normalization(image, in_bound = self.config.get_parameter("percentile"))
+
+ if self.config.get_parameter("tile_overlap_size") == [0,0]:
+ padding = None
+ if image.shape[0] < self.config.get_parameter("tile_size")[0] or image.shape[1] < self.config.get_parameter("tile_size")[1]:
+ image, padding = self.pad_image(image, image_size = self.config.get_parameter("tile_size"))
+ input_image = image[np.newaxis,:,:,np.newaxis]
+
+ output_image = self.model.predict(input_image, verbose=1)
+
+ if padding is not None:
+ h, w = output_image.shape[1:3]
+ output_image = np.reshape(output_image, (h, w))
+ output_image = self.remove_pad_image(output_image, padding = padding)
+ else:
+ tile_image_list, num_rows, num_cols, padding = self.tile_image(image, self.config.get_parameter("tile_size"), self.config.get_parameter("tile_overlap_size"))
+
+ pred_train_list = []
+ for tile in tile_image_list:
+
+ # reshape image to correct dimensions for unet
+ h, w = tile.shape[:2]
+
+ tile = np.reshape(tile, (1, h, w, 1))
+
+ pred_train_list.extend(self.model.predict(tile, verbose=1))
+
+ output_image = self.untile_image(pred_train_list, self.config.get_parameter("tile_size"), self.config.get_parameter("tile_overlap_size"),
+ num_rows, num_cols, padding = padding)
+
+ self.save_image(output_image, image_path)
+
+ return output_image
+
+ def save_image(self, image, image_path, subfolder = 'Masks', suffix = '-preds'):
+ """Saves image to image_path
+
+ Final location of image is as follows:
+ - image_path
+ - subfolder
+ - model/weights file name
+
+ Parameters
+ ----------
+ image : `array_like`
+ Image to be saved
+ image_path : `str`
+ Location to save the image in
+ subfolder : `str`
+ [Default: 'Masks'] Subfolder in which the image is to be saved in
+ suffix : `str`
+ [Default: '-preds'] Suffix to append to the filename of the predicted image
+ """
+ image_dir = os.path.dirname(image_path)
+
+ output_dir = os.path.join(image_dir, subfolder)
+ if not os.path.exists(output_dir):
+ os.makedirs(output_dir)
+
+ basename, _ = os.path.splitext(os.path.basename(self.weights_path))
+
+ output_dir = os.path.join(output_dir, basename)
+ if not os.path.exists(output_dir):
+ os.makedirs(output_dir)
+
+ filename, _ = os.path.splitext(os.path.basename(image_path))
+ output_path = os.path.join(output_dir, "{}{}.tif".format(filename, suffix))
+
+ skimage.io.imsave(output_path, image)Classes
+-
+
+class CNN_Base +(model_dir=None, config_filepath=None, **kwargs) +
+-
+
Creates the base neural network class with basic functions
+Parameters
+-
+
model_dir:str, optional
+- [Default: None] Folder where the model is stored +
config_filepath:str, optional
+- [Default: None] Filepath to the config file +
**kwargs
+- Parameters that are passed to :class:
network_config.Network_Config
+
Attributes
+-
+
config: :class:network_config.Network_Config
+- Network_config object containing the config and necessary functions +
+++Expand source code +
+
+class CNN_Base(Dataset, Image_Functions): + def __init__(self, model_dir = None, config_filepath = None, **kwargs): + """Creates the base neural network class with basic functions + + Parameters + ---------- + model_dir : `str`, optional + [Default: None] Folder where the model is stored + config_filepath : `str`, optional + [Default: None] Filepath to the config file + **kwargs + Parameters that are passed to :class:`network_config.Network_Config` + + Attributes + ---------- + config : :class:`network_config.Network_Config` + Network_config object containing the config and necessary functions + """ + + super().__init__() + + self.config = Network_Config(model_dir = model_dir, config_filepath = config_filepath, **kwargs) + + self.config.update_parameter(["general", "now"], datetime.datetime.now()) + + if self.config.get_parameter("use_cpu") is True: + self.initialize_cpu() + else: + self.initialize_gpu() + + ####################### + # Logging functions + ####################### + def init_logs(self): + """Initiates the parameters required for the log file + """ + # Directory for training logs + print(self.config.get_parameter("name"), self.config.get_parameter("now")) + self.log_dir = os.path.join(self.config.get_parameter("model_dir"), "{}-{:%Y%m%dT%H%M}".format(self.config.get_parameter("name"), self.config.get_parameter("now"))) + + # Path to save after each epoch. Include placeholders that get filled by Keras. + self.checkpoint_path = os.path.join(self.log_dir, "{}-{:%Y%m%dT%H%M}_*epoch*.h5".format(self.config.get_parameter("name"), self.config.get_parameter("now"))) + self.checkpoint_path = self.checkpoint_path.replace("*epoch*", "{epoch:04d}") + + def write_logs(self): + """Writes the log file + """ + # Create log_dir if it does not exist + if os.path.exists(self.log_dir) is False: + os.makedirs(self.log_dir) + + # save the parameters used in current run to logs dir + self.config.write_config(os.path.join(self.log_dir, "{}-{:%Y%m%dT%H%M}-config.yml".format(self.config.get_parameter("name"), self.config.get_parameter("now")))) + + ####################### + # Initialization functions + ####################### + def summary(self): + """Summary of the layers in the model + """ + self.model.summary() + + def compile_model(self, optimizer, loss): + """Compiles model + + Parameters + ---------- + optimizer + Gradient optimizer used in during the training of the network + loss + Loss function of the network + """ + self.model.compile(optimizer, loss = loss, metrics = self.config.get_parameter("metrics")) + + def initialize_model(self): + """Initializes the logs, builds the model, and chooses the correct initialization function + """ + # write parameters to yaml file + self.init_logs() + if self.config.get_parameter("for_prediction") is False: + self.write_logs() + + # build model + self.model = self.build_model(self.config.get_parameter("input_size")) + + # save model to yaml file + if self.config.get_parameter("for_prediction") is False: + self.config.write_model(self.model, os.path.join(self.log_dir, "{}-{:%Y%m%dT%H%M}-model.yml".format(self.config.get_parameter("name"), self.config.get_parameter("now")))) + + print("{} using single GPU or CPU..".format("Predicting" if self.config.get_parameter("for_prediction") else "Training")) + self.initialize_model_normal() + + def initialize_cpu(self): + """Sets the session to only use the CPU + """ + config = tf.ConfigProto( + device_count = {'CPU' : 1, + 'GPU' : 0} + ) + session = tf.Session(config=config) + K.set_session(session) + + def initialize_gpu(self): + """Sets the seesion to use the gpu specified in config file + """ + os.environ["CUDA_DEVICE_ORDER"]="PCI_BUS_ID" # see issue #152 + os.environ['CUDA_VISIBLE_DEVICES'] = str(self.config.get_parameter("visible_gpu")) # needs to be a string + + config = tf.ConfigProto() + config.gpu_options.allow_growth = True + sess = tf.Session(config=config) + K.tensorflow_backend.set_session(sess) + + def initialize_model_normal(self): + """Initializes the optimizer and any specified callback functions + """ + opt = self.optimizer_function() + self.compile_model(optimizer = opt, loss = self.loss_function(self.config.get_parameter("loss"))) + + if self.config.get_parameter("for_prediction") == False: + self.callbacks = [self.model_checkpoint_call(verbose = True)] + + if self.config.get_parameter("use_tensorboard") is True: + self.callbacks.append(self.tensorboard_call()) + + if self.config.get_parameter("reduce_LR_on_plateau") is True: + self.callbacks.append(ReduceLROnPlateau(monitor=self.config.get_parameter("reduce_LR_monitor"), + factor = self.config.get_parameter("reduce_LR_factor"), + patience = self.config.get_parameter("reduce_LR_patience"), + min_lr = self.config.get_parameter("reduce_LR_min_lr"), + verbose = True)) + + if self.config.get_parameter("early_stopping") is True: + self.callbacks.append(EarlyStopping(monitor=self.config.get_parameter("early_stopping_monitor"), + patience = self.config.get_parameter("early_stopping_patience"), + min_delta = self.config.get_parameter("early_stopping_min_delta"), + verbose = True)) + + ####################### + # Optimizer/Loss functions + ####################### + def optimizer_function(self, learning_rate = None): + """Initialize optimizer function + + Parameters + ---------- + learning_rate : `int` + Learning rate of the descent algorithm + + Returns + ---------- + optimizer + Function to call the optimizer + """ + if learning_rate is None: + learning_rate = self.config.get_parameter("learning_rate") + if self.config.get_parameter("optimizer_function") == 'sgd': + return keras.optimizers.SGD(lr = learning_rate, + decay = self.config.get_parameter("decay"), + momentum = self.config.get_parameter("momentum"), + nesterov = self.config.get_parameter("nesterov")) + elif self.config.get_parameter("optimizer_function") == 'rmsprop': + return keras.optimizers.RMSprop(lr = learning_rate, + decay = self.config.get_parameter("decay")) + elif self.config.get_parameter("optimizer_function") == 'adam': + return keras.optimizers.Adam(lr = learning_rate, + decay = self.config.get_parameter("decay")) + + def loss_function(self, loss): + """Initialize loss function + + Parameters + ---------- + loss : `str` + Name of the loss function + + Returns + ---------- + loss + Function to call loss function + """ + if loss == "binary_crossentropy": + print("Using binary crossentropy") + return loss + elif loss == "jaccard_distance_loss": + print("Using jaccard distance loss") + from .internals.losses import jaccard_distance_loss + return jaccard_distance_loss + elif loss == "lovasz_hinge": + print("Using Lovasz-hinge loss") + from .internals.losses import lovasz_loss + return lovasz_loss + elif loss == "dice_loss": + print("Using Dice loss") + from .internals.losses import dice_coef_loss + return dice_coef_loss + elif loss == "bce_dice_loss": + print("Using 1 - Dice + BCE loss") + from .internals.losses import bce_dice_loss + return bce_dice_loss + elif loss == "ssim_loss": + print("Using DSSIM loss") + from .internals.losses import DSSIM_loss + return DSSIM_loss + elif loss == "bce_ssim_loss": + print("Using BCE + DSSIM loss") + from .internals.losses import bce_ssim_loss + return bce_ssim_loss + elif loss == "mean_squared_error": + return keras.losses.mean_squared_error + elif loss == "mean_absolute_error": + return keras.losses.mean_absolute_error + elif loss == "ssim_mae_loss": + print("Using DSSIM + MAE loss") + from .internals.losses import dssim_mae_loss + return dssim_mae_loss + else: + print("Using {}".format(loss)) + return loss + + ####################### + # Callbacks + ####################### + def tensorboard_call(self): + """Initialize tensorboard call + """ + return TensorBoard(log_dir=self.log_dir, + batch_size = self.config.get_parameter("batch_size_per_GPU"), + write_graph=self.config.get_parameter("write_graph"), + write_images=self.config.get_parameter("write_images"), + write_grads=self.config.get_parameter("write_grads"), + update_freq='epoch', + histogram_freq=self.config.get_parameter("histogram_freq")) + + def model_checkpoint_call(self, verbose = 0): + """Initialize model checkpoint call + """ + return ModelCheckpoint(self.checkpoint_path, save_weights_only=True, verbose=verbose) + + ####################### + # Clear memory once training is done + ####################### + def end_training(self): + """Deletes model and releases gpu memory held by tensorflow + """ + # del reference to model + del self.model + + # clear memory + tf.reset_default_graph() + K.clear_session() + + # take hold of cuda device to shut it down + from numba import cuda + cuda.select_device(0) + cuda.close() + + ####################### + # Train Model + ####################### + def train_model(self, verbose = True): + """Trains model + + Parameters + ---------- + verbose : `int`, optional + [Default: True] Verbose output + """ + history = self.model.fit(self.aug_images, self.aug_ground_truth, validation_split = self.config.get_parameter("val_split"), + batch_size = self.config.get_parameter("batch_size"), epochs = self.config.get_parameter("num_epochs"), shuffle = True, + callbacks=self.callbacks, verbose=verbose) + + self.end_training() + + ####################### + # Predict using loaded model weights + ####################### + # TODO: change to load model from yaml file + def load_model(self, model_dir = None): + """Loads model from h5 file + + Parameters + ---------- + model_dir : `str`, optional + [Default: None] Directory containing the model file + """ + # TODO: rewrite to load model from yaml file + if model_dir is None: + model_dir = self.config.get_parameter("model_dir") + + if os.path.isdir(model_dir) is True: + list_weights_files = glob.glob(os.path.join(model_dir,'*.h5')) + list_weights_files.sort() # To ensure that [-1] gives the last file + + model_dir = os.path.join(model_dir,list_weights_files[-1]) + + self.model.load_model(model_dir) + print("Loaded model from: " + model_dir) + + def load_weights(self, model_dir = None, weights_index = -1): + """Loads weights from h5 file + + Parameters + ---------- + model_dir : `str`, optional + [Default: None] Directory containing the weights file + weights_index : `int`, optional + [Default: -1] + """ + if model_dir is None: + model_dir = self.config.get_parameter("model_dir") + + if os.path.isdir(model_dir) is True: + list_weights_files = glob.glob(os.path.join(model_dir,'*.h5')) + list_weights_files.sort() # To ensure that [-1] gives the last file + self.weights_path = list_weights_files[weights_index] + model_dir = os.path.join(model_dir, self.weights_path) + else: + self.weights_path = model_dir + + self.model.load_weights(model_dir) + print("Loaded weights from: " + model_dir) + + def predict_images(self, image_dir): + """Perform prediction on images found in ``image_dir`` + + Parameters + ---------- + image_dir : `str` + Directory containing the images to perform prediction on + + Returns + ---------- + image : `array_like` + Last image that prediction was perfromed on + """ + # load image list + image_list = self.list_images(image_dir) + + for image_path in image_list: + image = self.load_image(image_path = image_path) + + # percentile normalization + if self.config.get_parameter("percentile_normalization"): + image, _, _ = self.percentile_normalization(image, in_bound = self.config.get_parameter("percentile")) + + if self.config.get_parameter("tile_overlap_size") == [0,0]: + padding = None + if image.shape[0] < self.config.get_parameter("tile_size")[0] or image.shape[1] < self.config.get_parameter("tile_size")[1]: + image, padding = self.pad_image(image, image_size = self.config.get_parameter("tile_size")) + input_image = image[np.newaxis,:,:,np.newaxis] + + output_image = self.model.predict(input_image, verbose=1) + + if padding is not None: + h, w = output_image.shape[1:3] + output_image = np.reshape(output_image, (h, w)) + output_image = self.remove_pad_image(output_image, padding = padding) + else: + tile_image_list, num_rows, num_cols, padding = self.tile_image(image, self.config.get_parameter("tile_size"), self.config.get_parameter("tile_overlap_size")) + + pred_train_list = [] + for tile in tile_image_list: + + # reshape image to correct dimensions for unet + h, w = tile.shape[:2] + + tile = np.reshape(tile, (1, h, w, 1)) + + pred_train_list.extend(self.model.predict(tile, verbose=1)) + + output_image = self.untile_image(pred_train_list, self.config.get_parameter("tile_size"), self.config.get_parameter("tile_overlap_size"), + num_rows, num_cols, padding = padding) + + self.save_image(output_image, image_path) + + return output_image + + def save_image(self, image, image_path, subfolder = 'Masks', suffix = '-preds'): + """Saves image to image_path + + Final location of image is as follows: + - image_path + - subfolder + - model/weights file name + + Parameters + ---------- + image : `array_like` + Image to be saved + image_path : `str` + Location to save the image in + subfolder : `str` + [Default: 'Masks'] Subfolder in which the image is to be saved in + suffix : `str` + [Default: '-preds'] Suffix to append to the filename of the predicted image + """ + image_dir = os.path.dirname(image_path) + + output_dir = os.path.join(image_dir, subfolder) + if not os.path.exists(output_dir): + os.makedirs(output_dir) + + basename, _ = os.path.splitext(os.path.basename(self.weights_path)) + + output_dir = os.path.join(output_dir, basename) + if not os.path.exists(output_dir): + os.makedirs(output_dir) + + filename, _ = os.path.splitext(os.path.basename(image_path)) + output_path = os.path.join(output_dir, "{}{}.tif".format(filename, suffix)) + + skimage.io.imsave(output_path, image)Ancestors
+ +Subclasses
+-
+
- Unet +
- Unet_Resnet +
Methods
+-
+
+def compile_model(self, optimizer, loss) +
+-
+
Compiles model
+Parameters
+-
+
optimizer
+- Gradient optimizer used in during the training of the network +
loss
+- Loss function of the network +
+++Expand source code +
+
+def compile_model(self, optimizer, loss): + """Compiles model + + Parameters + ---------- + optimizer + Gradient optimizer used in during the training of the network + loss + Loss function of the network + """ + self.model.compile(optimizer, loss = loss, metrics = self.config.get_parameter("metrics"))
+ +def end_training(self) +
+-
+
Deletes model and releases gpu memory held by tensorflow
+++Expand source code +
+
+def end_training(self): + """Deletes model and releases gpu memory held by tensorflow + """ + # del reference to model + del self.model + + # clear memory + tf.reset_default_graph() + K.clear_session() + + # take hold of cuda device to shut it down + from numba import cuda + cuda.select_device(0) + cuda.close()
+ +def init_logs(self) +
+-
+
Initiates the parameters required for the log file
+++Expand source code +
+
+def init_logs(self): + """Initiates the parameters required for the log file + """ + # Directory for training logs + print(self.config.get_parameter("name"), self.config.get_parameter("now")) + self.log_dir = os.path.join(self.config.get_parameter("model_dir"), "{}-{:%Y%m%dT%H%M}".format(self.config.get_parameter("name"), self.config.get_parameter("now"))) + + # Path to save after each epoch. Include placeholders that get filled by Keras. + self.checkpoint_path = os.path.join(self.log_dir, "{}-{:%Y%m%dT%H%M}_*epoch*.h5".format(self.config.get_parameter("name"), self.config.get_parameter("now"))) + self.checkpoint_path = self.checkpoint_path.replace("*epoch*", "{epoch:04d}")
+ +def initialize_cpu(self) +
+-
+
Sets the session to only use the CPU
+++Expand source code +
+
+def initialize_cpu(self): + """Sets the session to only use the CPU + """ + config = tf.ConfigProto( + device_count = {'CPU' : 1, + 'GPU' : 0} + ) + session = tf.Session(config=config) + K.set_session(session)
+ +def initialize_gpu(self) +
+-
+
Sets the seesion to use the gpu specified in config file
+++Expand source code +
+
+def initialize_gpu(self): + """Sets the seesion to use the gpu specified in config file + """ + os.environ["CUDA_DEVICE_ORDER"]="PCI_BUS_ID" # see issue #152 + os.environ['CUDA_VISIBLE_DEVICES'] = str(self.config.get_parameter("visible_gpu")) # needs to be a string + + config = tf.ConfigProto() + config.gpu_options.allow_growth = True + sess = tf.Session(config=config) + K.tensorflow_backend.set_session(sess)
+ +def initialize_model(self) +
+-
+
Initializes the logs, builds the model, and chooses the correct initialization function
+++Expand source code +
+
+def initialize_model(self): + """Initializes the logs, builds the model, and chooses the correct initialization function + """ + # write parameters to yaml file + self.init_logs() + if self.config.get_parameter("for_prediction") is False: + self.write_logs() + + # build model + self.model = self.build_model(self.config.get_parameter("input_size")) + + # save model to yaml file + if self.config.get_parameter("for_prediction") is False: + self.config.write_model(self.model, os.path.join(self.log_dir, "{}-{:%Y%m%dT%H%M}-model.yml".format(self.config.get_parameter("name"), self.config.get_parameter("now")))) + + print("{} using single GPU or CPU..".format("Predicting" if self.config.get_parameter("for_prediction") else "Training")) + self.initialize_model_normal()
+ +def initialize_model_normal(self) +
+-
+
Initializes the optimizer and any specified callback functions
+++Expand source code +
+
+def initialize_model_normal(self): + """Initializes the optimizer and any specified callback functions + """ + opt = self.optimizer_function() + self.compile_model(optimizer = opt, loss = self.loss_function(self.config.get_parameter("loss"))) + + if self.config.get_parameter("for_prediction") == False: + self.callbacks = [self.model_checkpoint_call(verbose = True)] + + if self.config.get_parameter("use_tensorboard") is True: + self.callbacks.append(self.tensorboard_call()) + + if self.config.get_parameter("reduce_LR_on_plateau") is True: + self.callbacks.append(ReduceLROnPlateau(monitor=self.config.get_parameter("reduce_LR_monitor"), + factor = self.config.get_parameter("reduce_LR_factor"), + patience = self.config.get_parameter("reduce_LR_patience"), + min_lr = self.config.get_parameter("reduce_LR_min_lr"), + verbose = True)) + + if self.config.get_parameter("early_stopping") is True: + self.callbacks.append(EarlyStopping(monitor=self.config.get_parameter("early_stopping_monitor"), + patience = self.config.get_parameter("early_stopping_patience"), + min_delta = self.config.get_parameter("early_stopping_min_delta"), + verbose = True))
+ +def load_model(self, model_dir=None) +
+-
+
Loads model from h5 file
+Parameters
+-
+
model_dir:str, optional
+- [Default: None] Directory containing the model file +
+++Expand source code +
+
+def load_model(self, model_dir = None): + """Loads model from h5 file + + Parameters + ---------- + model_dir : `str`, optional + [Default: None] Directory containing the model file + """ + # TODO: rewrite to load model from yaml file + if model_dir is None: + model_dir = self.config.get_parameter("model_dir") + + if os.path.isdir(model_dir) is True: + list_weights_files = glob.glob(os.path.join(model_dir,'*.h5')) + list_weights_files.sort() # To ensure that [-1] gives the last file + + model_dir = os.path.join(model_dir,list_weights_files[-1]) + + self.model.load_model(model_dir) + print("Loaded model from: " + model_dir)
+ +def load_weights(self, model_dir=None, weights_index=-1) +
+-
+
Loads weights from h5 file
+Parameters
+-
+
model_dir:str, optional
+- [Default: None] Directory containing the weights file +
weights_index:int, optional
+- [Default: -1] +
+++Expand source code +
+
+def load_weights(self, model_dir = None, weights_index = -1): + """Loads weights from h5 file + + Parameters + ---------- + model_dir : `str`, optional + [Default: None] Directory containing the weights file + weights_index : `int`, optional + [Default: -1] + """ + if model_dir is None: + model_dir = self.config.get_parameter("model_dir") + + if os.path.isdir(model_dir) is True: + list_weights_files = glob.glob(os.path.join(model_dir,'*.h5')) + list_weights_files.sort() # To ensure that [-1] gives the last file + self.weights_path = list_weights_files[weights_index] + model_dir = os.path.join(model_dir, self.weights_path) + else: + self.weights_path = model_dir + + self.model.load_weights(model_dir) + print("Loaded weights from: " + model_dir)
+ +def loss_function(self, loss) +
+-
+
Initialize loss function
+Parameters
+-
+
loss:str
+- Name of the loss function +
Returns
+-
+
loss
+- Function to call loss function +
+++Expand source code +
+
+def loss_function(self, loss): + """Initialize loss function + + Parameters + ---------- + loss : `str` + Name of the loss function + + Returns + ---------- + loss + Function to call loss function + """ + if loss == "binary_crossentropy": + print("Using binary crossentropy") + return loss + elif loss == "jaccard_distance_loss": + print("Using jaccard distance loss") + from .internals.losses import jaccard_distance_loss + return jaccard_distance_loss + elif loss == "lovasz_hinge": + print("Using Lovasz-hinge loss") + from .internals.losses import lovasz_loss + return lovasz_loss + elif loss == "dice_loss": + print("Using Dice loss") + from .internals.losses import dice_coef_loss + return dice_coef_loss + elif loss == "bce_dice_loss": + print("Using 1 - Dice + BCE loss") + from .internals.losses import bce_dice_loss + return bce_dice_loss + elif loss == "ssim_loss": + print("Using DSSIM loss") + from .internals.losses import DSSIM_loss + return DSSIM_loss + elif loss == "bce_ssim_loss": + print("Using BCE + DSSIM loss") + from .internals.losses import bce_ssim_loss + return bce_ssim_loss + elif loss == "mean_squared_error": + return keras.losses.mean_squared_error + elif loss == "mean_absolute_error": + return keras.losses.mean_absolute_error + elif loss == "ssim_mae_loss": + print("Using DSSIM + MAE loss") + from .internals.losses import dssim_mae_loss + return dssim_mae_loss + else: + print("Using {}".format(loss)) + return loss
+ +def model_checkpoint_call(self, verbose=0) +
+-
+
Initialize model checkpoint call
+++Expand source code +
+
+def model_checkpoint_call(self, verbose = 0): + """Initialize model checkpoint call + """ + return ModelCheckpoint(self.checkpoint_path, save_weights_only=True, verbose=verbose)
+ +def optimizer_function(self, learning_rate=None) +
+-
+
Initialize optimizer function
+Parameters
+-
+
learning_rate:int
+- Learning rate of the descent algorithm +
Returns
+-
+
optimizer
+- Function to call the optimizer +
+++Expand source code +
+
+def optimizer_function(self, learning_rate = None): + """Initialize optimizer function + + Parameters + ---------- + learning_rate : `int` + Learning rate of the descent algorithm + + Returns + ---------- + optimizer + Function to call the optimizer + """ + if learning_rate is None: + learning_rate = self.config.get_parameter("learning_rate") + if self.config.get_parameter("optimizer_function") == 'sgd': + return keras.optimizers.SGD(lr = learning_rate, + decay = self.config.get_parameter("decay"), + momentum = self.config.get_parameter("momentum"), + nesterov = self.config.get_parameter("nesterov")) + elif self.config.get_parameter("optimizer_function") == 'rmsprop': + return keras.optimizers.RMSprop(lr = learning_rate, + decay = self.config.get_parameter("decay")) + elif self.config.get_parameter("optimizer_function") == 'adam': + return keras.optimizers.Adam(lr = learning_rate, + decay = self.config.get_parameter("decay"))
+ +def predict_images(self, image_dir) +
+-
+
Perform prediction on images found in
+image_dirParameters
+-
+
image_dir:str
+- Directory containing the images to perform prediction on +
Returns
+-
+
image:array_like
+- Last image that prediction was perfromed on +
+++Expand source code +
+
+def predict_images(self, image_dir): + """Perform prediction on images found in ``image_dir`` + + Parameters + ---------- + image_dir : `str` + Directory containing the images to perform prediction on + + Returns + ---------- + image : `array_like` + Last image that prediction was perfromed on + """ + # load image list + image_list = self.list_images(image_dir) + + for image_path in image_list: + image = self.load_image(image_path = image_path) + + # percentile normalization + if self.config.get_parameter("percentile_normalization"): + image, _, _ = self.percentile_normalization(image, in_bound = self.config.get_parameter("percentile")) + + if self.config.get_parameter("tile_overlap_size") == [0,0]: + padding = None + if image.shape[0] < self.config.get_parameter("tile_size")[0] or image.shape[1] < self.config.get_parameter("tile_size")[1]: + image, padding = self.pad_image(image, image_size = self.config.get_parameter("tile_size")) + input_image = image[np.newaxis,:,:,np.newaxis] + + output_image = self.model.predict(input_image, verbose=1) + + if padding is not None: + h, w = output_image.shape[1:3] + output_image = np.reshape(output_image, (h, w)) + output_image = self.remove_pad_image(output_image, padding = padding) + else: + tile_image_list, num_rows, num_cols, padding = self.tile_image(image, self.config.get_parameter("tile_size"), self.config.get_parameter("tile_overlap_size")) + + pred_train_list = [] + for tile in tile_image_list: + + # reshape image to correct dimensions for unet + h, w = tile.shape[:2] + + tile = np.reshape(tile, (1, h, w, 1)) + + pred_train_list.extend(self.model.predict(tile, verbose=1)) + + output_image = self.untile_image(pred_train_list, self.config.get_parameter("tile_size"), self.config.get_parameter("tile_overlap_size"), + num_rows, num_cols, padding = padding) + + self.save_image(output_image, image_path) + + return output_image
+ +def save_image(self, image, image_path, subfolder='Masks', suffix='-preds') +
+-
+
Saves image to image_path
+Final location of image is as follows: +- image_path +- subfolder +- model/weights file name
+Parameters
+-
+
image:array_like
+- Image to be saved +
image_path:str
+- Location to save the image in +
subfolder:str
+- [Default: 'Masks'] Subfolder in which the image is to be saved in +
suffix:str
+- [Default: '-preds'] Suffix to append to the filename of the predicted image +
+++Expand source code +
+
+def save_image(self, image, image_path, subfolder = 'Masks', suffix = '-preds'): + """Saves image to image_path + + Final location of image is as follows: + - image_path + - subfolder + - model/weights file name + + Parameters + ---------- + image : `array_like` + Image to be saved + image_path : `str` + Location to save the image in + subfolder : `str` + [Default: 'Masks'] Subfolder in which the image is to be saved in + suffix : `str` + [Default: '-preds'] Suffix to append to the filename of the predicted image + """ + image_dir = os.path.dirname(image_path) + + output_dir = os.path.join(image_dir, subfolder) + if not os.path.exists(output_dir): + os.makedirs(output_dir) + + basename, _ = os.path.splitext(os.path.basename(self.weights_path)) + + output_dir = os.path.join(output_dir, basename) + if not os.path.exists(output_dir): + os.makedirs(output_dir) + + filename, _ = os.path.splitext(os.path.basename(image_path)) + output_path = os.path.join(output_dir, "{}{}.tif".format(filename, suffix)) + + skimage.io.imsave(output_path, image)
+ +def summary(self) +
+-
+
Summary of the layers in the model
+++Expand source code +
+
+def summary(self): + """Summary of the layers in the model + """ + self.model.summary()
+ +def tensorboard_call(self) +
+-
+
Initialize tensorboard call
+++Expand source code +
+
+def tensorboard_call(self): + """Initialize tensorboard call + """ + return TensorBoard(log_dir=self.log_dir, + batch_size = self.config.get_parameter("batch_size_per_GPU"), + write_graph=self.config.get_parameter("write_graph"), + write_images=self.config.get_parameter("write_images"), + write_grads=self.config.get_parameter("write_grads"), + update_freq='epoch', + histogram_freq=self.config.get_parameter("histogram_freq"))
+ +def train_model(self, verbose=True) +
+-
+
Trains model
+Parameters
+-
+
verbose:int, optional
+- [Default: True] Verbose output +
+++Expand source code +
+
+def train_model(self, verbose = True): + """Trains model + + Parameters + ---------- + verbose : `int`, optional + [Default: True] Verbose output + """ + history = self.model.fit(self.aug_images, self.aug_ground_truth, validation_split = self.config.get_parameter("val_split"), + batch_size = self.config.get_parameter("batch_size"), epochs = self.config.get_parameter("num_epochs"), shuffle = True, + callbacks=self.callbacks, verbose=verbose) + + self.end_training()
+ +def write_logs(self) +
+-
+
Writes the log file
+++Expand source code +
+
+def write_logs(self): + """Writes the log file + """ + # Create log_dir if it does not exist + if os.path.exists(self.log_dir) is False: + os.makedirs(self.log_dir) + + # save the parameters used in current run to logs dir + self.config.write_config(os.path.join(self.log_dir, "{}-{:%Y%m%dT%H%M}-config.yml".format(self.config.get_parameter("name"), self.config.get_parameter("now"))))
+
Inherited members
+ +
+