def get_image_pair_loader(args, hparams, out_dir):
"""
TODO
"""
from mpunet.image import ImagePairLoader, ImagePair
if not args.f:
# No single file was specified with -f flag, load the desired dataset
dataset = args.dataset.replace("_data", "") + "_data"
image_pair_loader = ImagePairLoader(predict_mode=args.no_eval,
**hparams[dataset])
else:
predict_mode = not bool(args.l)
image_pair_loader = ImagePairLoader(predict_mode=predict_mode,
initialize_empty=True)
image_pair_loader.add_image(ImagePair(args.f, args.l))
# Put image pairs into a dict and remove from image_pair_loader to gain
# more control with garbage collection
image_pair_dict = {
image.identifier: image
for image in image_pair_loader.images
}
if vars(args)["continue"]:
# Remove images that were already predicted
image_pair_dict = remove_already_predicted(image_pair_dict, out_dir)
return image_pair_loader, image_pair_dict
def entry_func(args=None):
# Project base path
args = vars(get_argparser().parse_args(args))
basedir = os.path.abspath(args["project_dir"])
overwrite = args["overwrite"]
continue_training = args["continue_training"]
eval_prob = args["eval_prob"]
await_PID = args["wait_for"]
dice_weight = args["dice_weight"]
print("Fitting fusion model for project-folder: %s" % basedir)
# Minimum images in validation set before also using training images
min_val_images = 15
# Fusion model training params
epochs = args['epochs']
fm_batch_size = args["batch_size"]
# Early stopping params
early_stopping = args["early_stopping"]
# Wait for PID?
if await_PID:
from mpunet.utils import await_PIDs
await_PIDs(await_PID)
# Fetch GPU(s)
num_GPUs = args["num_GPUs"]
force_gpu = args["force_GPU"]
# Wait for free GPU
if not force_gpu:
await_and_set_free_gpu(N=num_GPUs, sleep_seconds=120)
else:
set_gpu(force_gpu)
# Get logger
logger = Logger(base_path=basedir,
active_file="train_fusion",
overwrite_existing=overwrite)
# Get YAML hyperparameters
hparams = YAMLHParams(os.path.join(basedir, "train_hparams.yaml"))
# Get some key settings
n_classes = hparams["build"]["n_classes"]
if hparams["build"]["out_activation"] == "linear":
# Trained with logit targets?
hparams["build"][
"out_activation"] = "softmax" if n_classes > 1 else "sigmoid"
# Get views
views = np.load("%s/views.npz" % basedir)["arr_0"]
del hparams["fit"]["views"]
# Get weights and set fusion (output) path
weights = get_best_model("%s/model" % basedir)
weights_name = os.path.splitext(os.path.split(weights)[-1])[0]
fusion_weights = "%s/model/fusion_weights/" \
"%s_fusion_weights.h5" % (basedir, weights_name)
create_folders(os.path.split(fusion_weights)[0])
# Log a few things
log(logger, hparams, views, weights, fusion_weights)
# Check if exists already...
if not overwrite and os.path.exists(fusion_weights):
from sys import exit
print("\n[*] A fusion weights file already exists at '%s'."
"\n Use the --overwrite flag to overwrite." % fusion_weights)
exit(0)
# Load validation data
images = ImagePairLoader(**hparams["val_data"], logger=logger)
is_validation = {m.identifier: True for m in images}
# Define random sets of images to train on simul. (cant be all due
# to memory constraints)
image_IDs = [m.identifier for m in images]
if len(images) < min_val_images:
# Pick N random training images
diff = min_val_images - len(images)
logger("Adding %i training images to set" % diff)
# Load the training data and pick diff images
train = ImagePairLoader(**hparams["train_data"], logger=logger)
indx = np.random.choice(np.arange(len(train)),
diff,
replace=diff > len(train))
# Add the images to the image set set
train_add = [train[i] for i in indx]
for m in train_add:
is_validation[m.identifier] = False
image_IDs.append(m.identifier)
images.add_images(train_add)
# Append to length % sub_size == 0
sub_size = args["images_per_round"]
rest = int(sub_size * np.ceil(len(image_IDs) / sub_size)) - len(image_IDs)
if rest:
image_IDs += list(np.random.choice(image_IDs, rest, replace=False))
# Shuffle and split
random.shuffle(image_IDs)
sets = [
set(s) for s in np.array_split(image_IDs,
len(image_IDs) / sub_size)
]
assert (contains_all_images(sets, image_IDs))
# Define fusion model (named 'org' to store reference to orgiginal model if
# multi gpu model is created below)
fusion_model = FusionModel(n_inputs=len(views),
n_classes=n_classes,
weight=dice_weight,
logger=logger,
verbose=False)
if continue_training:
fusion_model.load_weights(fusion_weights)
print("\n[OBS] CONTINUED TRAINING FROM:\n", fusion_weights)
import tensorflow as tf
with tf.distribute.MirroredStrategy().scope():
# Define model
unet = init_model(hparams["build"], logger)
print("\n[*] Loading weights: %s\n" % weights)
unet.load_weights(weights, by_name=True)
# Compile the model
logger("Compiling...")
metrics = [
"sparse_categorical_accuracy", sparse_fg_precision, sparse_fg_recall
]
fusion_model.compile(optimizer=Adam(lr=1e-3),
loss=fusion_model.loss,
metrics=metrics)
fusion_model._log()
try:
_run_fusion_training(sets, logger, hparams, min_val_images,
is_validation, views, n_classes, unet,
fusion_model, early_stopping, fm_batch_size,
epochs, eval_prob, fusion_weights)
except KeyboardInterrupt:
pass
finally:
if not os.path.exists(os.path.split(fusion_weights)[0]):
os.mkdir(os.path.split(fusion_weights)[0])
# Save fusion model weights
# OBS: Must be original model if multi-gpu is performed!
fusion_model.save_weights(fusion_weights)
def _run_fusion_training(sets, logger, hparams, min_val_images, is_validation,
views, n_classes, unet, fusion_model, early_stopping,
fm_batch_size, epochs, eval_prob,
fusion_weights_path):
"""
TODO
"""
for _round, _set in enumerate(sets):
s = "Set %i/%i:\n%s" % (_round + 1, len(sets), _set)
logger("\n%s" % highlighted(s))
# Reload data
images = ImagePairLoader(**hparams["val_data"])
if len(images) < min_val_images:
images.add_images(ImagePairLoader(**hparams["train_data"]))
# Get list of ImagePair objects to run on
image_set_dict = {
m.identifier: m
for m in images if m.identifier in _set
}
# Set scaler and bg values
images.set_scaler_and_bg_values(
bg_value=hparams.get_from_anywhere('bg_value'),
scaler=hparams.get_from_anywhere('scaler'),
compute_now=False)
# Init LazyQueue and get its sequencer
from mpunet.sequences.utils import get_sequence
seq = get_sequence(data_queue=images,
is_validation=True,
views=views,
**hparams["fit"],
**hparams["build"])
# Fetch points from the set images
points_collection = []
targets_collection = []
N_im = len(image_set_dict)
for num_im, image_id in enumerate(list(image_set_dict.keys())):
logger("")
logger(
highlighted("(%i/%i) Running on %s (%s)" %
(num_im + 1, N_im, image_id,
"val" if is_validation[image_id] else "train")))
with seq.image_pair_queue.get_image_by_id(image_id) as image:
# Get voxel grid in real space
voxel_grid_real_space = get_voxel_grid_real_space(image)
# Get array to store predictions across all views
targets = image.labels.reshape(-1, 1)
points = np.empty(shape=(len(targets), len(views), n_classes),
dtype=np.float32)
points.fill(np.nan)
# Predict on all views
for k, v in enumerate(views):
print("\n%s" % "View: %s" % v)
points[:, k, :] = predict_and_map(
model=unet,
seq=seq,
image=image,
view=v,
voxel_grid_real_space=voxel_grid_real_space,
n_planes='same+20',
targets=targets,
eval_prob=eval_prob).reshape(-1, n_classes)
# add to collections
points_collection.append(points)
targets_collection.append(targets)
print(image.is_loaded)
# Stack points into one matrix
logger("Stacking points...")
X, y = stack_collections(points_collection, targets_collection)
# Shuffle train
print("Shuffling points...")
X, y = shuffle(X, y)
print("Getting validation set...")
val_ind = int(0.20 * X.shape[0])
X_val, y_val = X[:val_ind], y[:val_ind]
X, y = X[val_ind:], y[val_ind:]
# Prepare dice score callback for validation data
val_cb = ValDiceScores((X_val, y_val), n_classes, 50000, logger)
# Callbacks
cbs = [
val_cb,
CSVLogger(filename="logs/fusion_training.csv",
separator=",",
append=True),
PrintLayerWeights(fusion_model.layers[-1],
every=1,
first=1000,
per_epoch=True,
logger=logger)
]
es = EarlyStopping(monitor='val_dice',
min_delta=0.0,
patience=early_stopping,
verbose=1,
mode='max')
cbs.append(es)
# Start training
try:
fusion_model.fit(X,
y,
batch_size=fm_batch_size,
epochs=epochs,
callbacks=cbs,
verbose=1)
except KeyboardInterrupt:
pass
fusion_model.save_weights(fusion_weights_path)
def predict_single(image, model, hparams, verbose=1):
"""
A generic prediction function that sets up a ImagePairLoader object for the
given image, prepares the image and predicts.
Note that this function should only be used for convinience in scripts that
work on single images at a time anyway, as batch-preparing the entire
ImagePairLoader object prior to prediction is faster.
NOTE: Only works with iso_live intrp modes at this time
"""
mode = hparams["fit"]["intrp_style"].lower()
assert mode in ("iso_live", "iso_live_3d")
# Prepare image for prediction
kwargs = hparams["fit"]
kwargs.update(hparams["build"])
# Set verbose memory
verb_mem = kwargs["verbose"]
kwargs["verbose"] = verbose
# Create a ImagePairLoader with only the given file
from mpunet.image import ImagePairLoader
image_pair_loader = ImagePairLoader(predict_mode=True,
initialize_empty=True,
no_log=bool(verbose))
image_pair_loader.add_image(image)
# Get N classes
n_classes = kwargs["n_classes"]
if mode == "iso_live":
# Add views if SMMV model
kwargs["views"] = np.load(hparams.project_path + "/views.npz")["arr_0"]
# Get sequence object
sequence = image_pair_loader.get_sequencer(**kwargs)
# Get voxel grid in real space
voxel_grid_real_space = get_voxel_grid_real_space(image)
# Prepare tensor to store combined prediction
d = image.image.shape
predicted = np.empty(shape=(len(kwargs["views"]), d[0], d[1], d[2],
n_classes),
dtype=np.float32)
print("Predicting on brain hyper-volume of shape:", predicted.shape)
for n_view, v in enumerate(kwargs["views"]):
print("\nView %i/%i: %s" % (n_view + 1, len(kwargs["views"]), v))
# Sample the volume along the view
X, y, grid, inv_basis = sequence.get_view_from(image.id,
v,
n_planes="same+20")
# Predict on volume using model
pred = predict_volume(model, X, axis=2)
# Map the real space coordiante predictions to nearest
# real space coordinates defined on voxel grid
predicted[n_view] = map_real_space_pred(pred,
grid,
inv_basis,
voxel_grid_real_space,
method="nearest")
else:
predicted = pred_3D_iso(
model=model,
sequence=image_pair_loader.get_sequencer(**kwargs),
image=image,
extra_boxes="3x",
min_coverage=None)
# Revert verbose mem
kwargs["verbose"] = verb_mem
return predicted
def entry_func(args=None):
# Get command line arguments
args = vars(get_argparser().parse_args(args))
base_dir = os.path.abspath(args["project_dir"])
_file = args["f"]
label = args["l"]
N_extra = args["extra"]
try:
N_extra = int(N_extra)
except ValueError:
pass
# Get settings from YAML file
from mpunet.hyperparameters import YAMLHParams
hparams = YAMLHParams(os.path.join(base_dir, "train_hparams.yaml"))
# Set strides
hparams["fit"]["strides"] = args["strides"]
if not _file:
try:
# Data specified from command line?
data_dir = os.path.abspath(args["data_dir"])
# Set with default sub dirs
hparams["test_data"] = {
"base_dir": data_dir,
"img_subdir": "images",
"label_subdir": "labels"
}
except (AttributeError, TypeError):
data_dir = hparams["test_data"]["base_dir"]
else:
data_dir = False
out_dir = os.path.abspath(args["out_dir"])
overwrite = args["overwrite"]
predict_mode = args["no_eval"]
save_only_pred = args["save_only_pred"]
# Check if valid dir structures
validate_folders(base_dir, data_dir, out_dir, overwrite)
# Import all needed modules (folder is valid at this point)
import numpy as np
from mpunet.image import ImagePairLoader, ImagePair
from mpunet.utils import get_best_model, create_folders, \
pred_to_class, await_and_set_free_gpu, set_gpu
from mpunet.utils.fusion import predict_3D_patches, predict_3D_patches_binary, pred_3D_iso
from mpunet.logging import init_result_dict_3D, save_all_3D
from mpunet.evaluate import dice_all
from mpunet.bin.predict import save_nii_files
# Fetch GPU(s)
num_GPUs = args["num_GPUs"]
force_gpu = args["force_GPU"]
# Wait for free GPU
if force_gpu == -1:
await_and_set_free_gpu(N=num_GPUs, sleep_seconds=240)
else:
set_gpu(force_gpu)
# Read settings from the project hyperparameter file
dim = hparams["build"]["dim"]
n_classes = hparams["build"]["n_classes"]
mode = hparams["fit"]["intrp_style"]
# Set ImagePairLoader object
if not _file:
image_pair_loader = ImagePairLoader(predict_mode=predict_mode,
**hparams["test_data"])
else:
predict_mode = not bool(label)
image_pair_loader = ImagePairLoader(predict_mode=predict_mode,
initialize_empty=True)
image_pair_loader.add_image(ImagePair(_file, label))
all_images = {
image.identifier: image
for image in image_pair_loader.images
}
# Set scaler and bg values
image_pair_loader.set_scaler_and_bg_values(
bg_value=hparams.get_from_anywhere('bg_value'),
scaler=hparams.get_from_anywhere('scaler'),
compute_now=False)
# Init LazyQueue and get its sequencer
from mpunet.sequences.utils import get_sequence
seq = get_sequence(data_queue=image_pair_loader,
is_validation=True,
**hparams["fit"],
**hparams["build"])
""" Define UNet model """
from mpunet.models import model_initializer
hparams["build"]["batch_size"] = 1
unet = model_initializer(hparams, False, base_dir)
model_path = get_best_model(base_dir + "/model")
unet.load_weights(model_path)
# Evaluate?
if not predict_mode:
# Prepare dictionary to store results in pd df
results, detailed_res = init_result_dict_3D(all_images, n_classes)
# Save to check correct format
save_all_3D(results, detailed_res, out_dir)
# Define result paths
nii_res_dir = os.path.join(out_dir, "nii_files")
create_folders(nii_res_dir)
image_ids = sorted(all_images)
for n_image, image_id in enumerate(image_ids):
print("\n[*] Running on: %s" % image_id)
with seq.image_pair_queue.get_image_by_id(image_id) as image_pair:
if mode.lower() == "iso_live_3d":
pred = pred_3D_iso(model=unet,
sequence=seq,
image=image_pair,
extra_boxes=N_extra,
min_coverage=None)
else:
# Predict on volume using model
if n_classes > 1:
pred = predict_3D_patches(model=unet,
patches=seq,
image=image_pair,
N_extra=N_extra)
else:
pred = predict_3D_patches_binary(model=unet,
patches=seq,
image_id=image_id,
N_extra=N_extra)
if not predict_mode:
# Get patches for the current image
y = image_pair.labels
# Calculate dice score
print("Mean dice: ", end="", flush=True)
p = pred_to_class(pred, img_dims=3, has_batch_dim=False)
dices = dice_all(y, p, n_classes=n_classes, ignore_zero=True)
mean_dice = dices[~np.isnan(dices)].mean()
print("Dices: ", dices)
print("%s (n=%i)" % (mean_dice, len(dices)))
# Add to results
results[image_id] = [mean_dice]
detailed_res[image_id] = dices
# Overwrite with so-far results
save_all_3D(results, detailed_res, out_dir)
# Save results
save_nii_files(p, image_pair, nii_res_dir, save_only_pred)
if not predict_mode:
# Write final results
save_all_3D(results, detailed_res, out_dir)
def predict_single(image, model, hparams, verbose=1):
"""
A generic prediction function that sets up a ImagePairLoader object for the
given image, prepares the image and predicts.
Note that this function should only be used for convinience in scripts that
work on single images at a time anyway, as batch-preparing the entire
ImagePairLoader object prior to prediction is faster.
NOTE: Only works with iso_live intrp modes at this time
"""
mode = hparams["fit"]["intrp_style"].lower()
assert mode in ("iso_live", "iso_live_3d")
# Create a ImagePairLoader with only the given file
from mpunet.image import ImagePairLoader
image_pair_loader = ImagePairLoader(predict_mode=True,
initialize_empty=True,
no_log=bool(verbose))
image_pair_loader.add_image(image)
# Set scaler and bg values
image_pair_loader.set_scaler_and_bg_values(
bg_value=hparams.get_from_anywhere('bg_value'),
scaler=hparams.get_from_anywhere('scaler'),
compute_now=False)
if mode == "iso_live":
# Init LazyQueue and get its sequencer
seq = get_sequence(data_queue=image_pair_loader,
views=np.load(hparams.project_path +
"/views.npz")["arr_0"],
is_validation=True,
**hparams["fit"],
**hparams["build"])
with seq.image_pair_queue.get_image_by_id(image.identifier) as image:
# Get voxel grid in real space
voxel_grid_real_space = get_voxel_grid_real_space(image)
# Prepare tensor to store combined prediction
d = image.image.shape
predicted = np.empty(shape=(len(seq.views), d[0], d[1], d[2],
seq.n_classes),
dtype=np.float32)
print("Predicting on brain hyper-volume of shape:",
predicted.shape)
for n_view, v in enumerate(seq.views):
print("\nView %i/%i: %s" % (n_view + 1, len(seq.views), v))
# Sample the volume along the view
X, y, grid, inv_basis = seq.get_view_from(image,
v,
n_planes="same+20")
# Predict on volume using model
pred = predict_volume(model, X, axis=2)
# Map the real space coordiante predictions to nearest
# real space coordinates defined on voxel grid
predicted[n_view] = map_real_space_pred(pred,
grid,
inv_basis,
voxel_grid_real_space,
method="nearest")
else:
# Init LazyQueue and get its sequencer
seq = get_sequence(data_queue=image_pair_loader,
is_validation=True,
**hparams["fit"],
**hparams["build"])
predicted = pred_3D_iso(model=model,
sequence=seq,
image=image,
extra_boxes="3x",
min_coverage=None)
return predicted
def _base_loader_func(hparams, just_one, no_val, logger, mtype):
"""
Base loader function used for all models. This function performs a series
of actions:
1) Loads train, val and test data according to hparams
2) Performs a hparam audit on the training + validation images
3) If any audited parameters were not manually specified, updates the
hparams dict with the audited values and updates the YAML file on disk
4) If just_one, discards all but the first training and validation images
5) Initializes a ImageQueue object on the training and validation data
if needed.
Args:
hparams: A mpunet.train.YAMLHParams object
just_one: A bool specifying whether to keep only the first train and
validation samples (for quick testing purposes)
no_val: A bool specifying whether to omit validation data entirely
Note: This setting applies even if validation data is
specified in the YAMLHparams object
logger: A mpunet.logger object
mtype: A string identifier for the dimensionality of the model,
currently either '2d', '3d'
(upper/lower ignored)
Returns:
train_data: An ImagePairLoader object storing the training images
val_data: An ImagePairLoader object storing the validation images, or
an 'empty' ImagePairLoader storing no images if no_val=True
logger: The passed logger object or a ScreenLogger object
auditor: An auditor object storing statistics on the training data
"""
# Get basic ScreenLogger if no logger is passed
logger = logger or ScreenLogger()
logger("Looking for images...")
# Get data loaders
train_data = ImagePairLoader(logger=logger, **hparams["train_data"])
val_data = ImagePairLoader(logger=logger, **hparams["val_data"])
# Audit
lab_paths = train_data.label_paths + val_data.label_paths
auditor = Auditor(train_data.image_paths + val_data.image_paths,
nii_lab_paths=lab_paths, logger=logger,
dim_3d=hparams.get_from_anywhere("dim") or 64,
hparams=hparams)
# Fill hparams with audited values, if not specified manually
auditor.fill(hparams, mtype)
# Add augmented data?
if hparams.get("aug_data"):
aug_data = hparams["aug_data"]
if "include" not in aug_data:
logger.warn("Found 'aug_data' group, but the group does not "
"contain the key 'include', which is required in "
"version 2.0 and above. OBS: Not including aug data!")
elif aug_data["include"]:
logger("\n[*] Adding augmented data with weight ", aug_data["sample_weight"])
train_data.add_images(ImagePairLoader(logger=logger, **aug_data))
if just_one:
# For testing purposes, run only on one train and one val image?
logger("[**NOTTICE**] Only running on first train & val samples.")
train_data.images = [train_data.images[0]]
val_data.images = [val_data.images[0]]
if no_val:
# Run without performing validation (even if specified in param file)
val_data.images = []
# Set queue object if necessary
train_data.set_queue(hparams["train_data"].get("max_load"))
val_data.set_queue(hparams["val_data"].get("max_load"))
return train_data, val_data, logger, auditor