def get_model():
return unet_model(
N_CLASSES,
PATCH_SZ,
n_channels=N_BANDS,
upconv=UPCONV,
class_weights=CLASS_WEIGHTS,
)
def run_model(learning_rate):
print("Training with learning rate: " + str(learning_rate))
train_name = datetime.now().strftime("%Y%m%d-%H%M%S") + "_lr" + str(
learning_rate).split('.')[1]
logdir = "logs/" + train_name
tensorboard_callback = keras.callbacks.TensorBoard(log_dir=logdir)
model = unet_model((512, 512, 1))
opt = keras.optimizers.Adam(
learning_rate=learning_rate) # default is 0.001
model.compile(optimizer=opt,
loss="binary_crossentropy",
metrics=["accuracy"])
# create_generators
seed = 1
batchsize = 2
train_generator, val_generator = create_generators(seed=seed,
batchsize=batchsize)
# train_generator, val_generator = create_generators("tranformed_val", seed=seed, batchsize=batchsize)
epochs = 50
model.fit(
train_generator,
steps_per_epoch=24 /
batchsize, # TODO: fix this #TODO: check that mult by 10 reduces number of epochs to get to same thing
validation_data=val_generator,
validation_steps=6 / batchsize, # TODO: fix this
epochs=epochs,
verbose=1,
callbacks=[tensorboard_callback])
os.makedirs("saved_models/" + train_name + "/")
model.save("saved_models/" + train_name + "/")
print()
parser.add_argument('--labels',
help='labels file',
default='../stage1_labels.csv')
parser.add_argument('--model', help='model', required=True)
parser.add_argument('--augment',
help='augment images',
dest='deterministic',
action='store_false')
parser.set_defaults(deterministic=True)
args = parser.parse_args()
opts = vars(args)
lmap = read_labels(args.labels)
model = unet_model(True, False)
print "loading existing weights into model from ", args.model
model.load_weights(args.model)
patients = os.listdir(args.input_folder)
patients.sort()
print "deterministic -> ", args.deterministic
t1 = int(round(time.time() * 1000))
fnum = 0
for filename in patients:
fnum += 1
print "processing file ", filename, fnum, "/", len(patients)
key = filename.split(".")[0]
def get_model():
if MODEL == 'U':
model = unet_model(N_CLASSES, PATCH_SZ, n_channels=N_BANDS)
elif MODEL == 'W':
model = wnet_model(N_CLASSES, PATCH_SZ, n_channels=N_BANDS)
return model
parser.add_argument('--embedding',
help='train embedding',
dest='embed',
action='store_true')
parser.set_defaults(embed=False)
args = parser.parse_args()
opts = vars(args)
train_folder = args.luna16_folder + "/1_1_1mm_slices_lung/subset[0-8]"
test_folder = args.luna16_folder + "/1_1_1mm_slices_lung/subset9"
# embedding params
if args.embed:
print "Training an embedding model"
model = unet_model(False, True)
else:
print "Training a unet model from scratch"
model = unet_model(True, False)
if not args.checkpoint is None:
print "loading existing weights into model from ", args.checkpoint
model.load_weights(args.checkpoint, by_name=True)
file_glob = train_folder + "/*.pkl.gz"
print file_glob
train_filenames = glob.glob(file_glob)
shuffle(train_filenames)
file_glob = test_folder + "/*.pkl.gz"
print file_glob
def get_model():
return unet_model(N_CLASSES,
PATCH_SZ,
n_channels=N_BANDS,
upconv=UPCONV,
class_weights=[0.5, 0.1, 0.4])
def get_model():
return unet_model(outputClasses,
dimpatchsz,
n_channels=numbands,
upconv=True,
class_weights=outputClassWeight)
