def train_and_calculate_metric(params):
logging.info("params = " + str(params))
s = Model.Settings()
s.batch_size = FLAGS.batch_size
s.loss = params["loss"]
s.num_classes = len(params["class_weights"])
s.class_weights = params["class_weights"]
s.image_depth = FLAGS.image_depth
s.image_height = FLAGS.image_width
s.image_width = FLAGS.image_height
s.keep_prob = params["keep_prob"]
s.l2_reg = params["l2_reg"]
s.learning_rate = params["learning_rate"]
s.num_conv_blocks = 5
s.num_conv_channels = 40
s.num_dense_layers = 1
s.use_batch_norm = False
try:
trainer = Trainer(s)
trainer.train(FLAGS.num_steps)
except tf.errors.ResourceExhaustedError as e:
logging.info("Resource exhausted: %s", e.message)
trainer.clear()
return {'status': hyperopt.STATUS_FAIL}
finally:
trainer.clear()
logging.info(
"iou = " + str(trainer.val_iou_history[-1]) + ", params = " + str(params))
return {'loss': -trainer.val_iou_history[-1], 'status': hyperopt.STATUS_OK}
def make_best_settings():
s = Model.Settings()
s.loss = "softmax"
s.batch_size = FLAGS.batch_size
s.image_depth = FLAGS.image_depth
s.image_height = FLAGS.image_width
s.image_width = FLAGS.image_height
s.keep_prob = 0.5
s.learning_rate = 0.05
s.num_conv_blocks = 4
s.num_conv_channels = 64
s.l2_reg = 0.0
s.use_batch_norm = False
s.num_dense_layers = 0
if FLAGS.settings == "Abdomen":
s.num_classes = 13
s.class_weights = [1.0] + [5.0]*12
elif FLAGS.settings == "Cardiac":
s.num_classes = 2
s.class_weights = [1.0, 5.0]
elif FLAGS.settings == "LiTS":
s.num_classes = 3
s.class_weights = [1.0, 5.0, 5.0]
s.l2_reg = 1.0e-6
elif FLAGS.settings == "LCTSC":
s.learning_rate = 0.0005
s.num_conv_channels = 128
s.num_classes = 6
s.class_weights = [1.0] + [5.0]*5
s.keep_prob = 0.5
s.l2_reg = 1.0e-4
elif FLAGS.settings == "Tissue":
s.learning_rate = 0.0005
s.num_conv_channels = 128
s.num_classes = 6
s.class_weights = [1.0, 5.0, 1.0, 5.0, 5.0, 1.0]
s.keep_prob = 0.5
else:
raise ValueError("Unknown dataset")
return s