def save(self):
logger.info(f"save: {self._fname}")
# Save JSON config to disk
with self._fname_json.open("w") as json_file:
json_file.write(self._model.to_json())
# Save weights to disk
self._model.save_weights(str(self._fname_weights))
def load(self):
logger.info(f"load: {self._fname}")
# Reload the model from the 2 files we saved
with self._fname_json.open("r") as json_file:
json_config = json_file.read()
self._model = model_from_json(json_config)
self._model.load_weights(str(self._fname_weights))
self.compile()
def fit(self,
x=None,
y=None,
batch_size=None,
epochs=1,
verbose=1,
callbacks=None,
validation_split=0.0,
validation_data=None,
shuffle=True,
class_weight=None,
sample_weight=None,
initial_epoch=0,
steps_per_epoch=None,
validation_steps=None,
validation_freq=1,
max_queue_size=10,
workers=1,
use_multiprocessing=False):
if self.model_exists():
logger.info(f"{self._fname} exists. skip fitting")
return
logger.info(f"start fitting")
self.history = self._model.fit(
x, y, batch_size, epochs, verbose, callbacks, validation_split,
validation_data, shuffle, class_weight, sample_weight,
initial_epoch, steps_per_epoch, validation_steps, validation_freq,
max_queue_size, workers, use_multiprocessing)
logger.info(f"finish fitting")
self.save()
def fit(self, x=None, y=None, batch_size=None, epochs=1, verbose=1, callbacks=None, validation_split=0.0, validation_data=None, shuffle=True, class_weight=None, sample_weight=None, initial_epoch=0, steps_per_epoch=None, validation_steps=None, validation_freq=1, max_queue_size=10, workers=1, use_multiprocessing=False):
teacher = ConvMinst("teacher.h5", self.temp())
student = ConvMinst("student.h5", self.temp())
logger.info(f"start fitting the teacher")
self.teacher_history = teacher.fit(x, y, batch_size, epochs, verbose, callbacks, validation_split, validation_data, shuffle, class_weight,
sample_weight, initial_epoch, steps_per_epoch, validation_steps, validation_freq, max_queue_size,
workers, use_multiprocessing)
logger.info("evaluate the labels at temperature t")
teacher_logit = teacher.get_logit(x)
teacher_predicted = tf.nn.softmax(teacher_logit / self._temp)
logger.info("train the student model at temperature t")
self.student_history = student.fit(x, teacher_predicted, batch_size, epochs, verbose, callbacks, validation_split,
validation_data, shuffle, class_weight,
sample_weight, initial_epoch, steps_per_epoch, validation_steps,
validation_freq, max_queue_size,
workers, use_multiprocessing)
self._model = student.get_model()
def attack_info(k, v):
logger.info(f"{prefix}: {k}: {v:.2f}")
result[k] = v