def fit(self,
x,
y,
validation_data=None,
batch_size=128,
epochs=1,
iterations_to_validate=None,
patience=0,
verbose=1,
**kwargs):
print('\nTraining Model')
self.verbose = verbose
if validation_data is not None:
return self.fit_tf_val(x, y, validation_data, batch_size, epochs,
iterations_to_validate, patience, verbose)
train_ds = tf.data.Dataset.from_tensor_slices(
(x, y)).shuffle(10000).batch(batch_size)
for epoch in range(epochs):
start_time = time.time()
for images, labels in train_ds:
self.train_step(images, labels)
if self.verbose:
self.eval_step(images, labels)
template = 'Epoch {}, Loss: {}, Acc: {}, Time: {}'
print(
template.format(epoch + 1, self.eval_loss.result(),
self.eval_accuracy.result() * 100,
delta_timer(time.time() - start_time)))
self.eval_loss.reset_states()
self.eval_accuracy.reset_states()
self.eval_tf(x, y, verbose=verbose)
def print_scores_times_to_file(file_path, header, scores_list, times_list):
with open(file_path, "a+") as text_file:
text_file.write(header + '\n')
scores_grouped = group_scores(scores_list)
metric_names = ['diri', 'simple', 'entropy', 'xH']
for i in range(len(scores_grouped)):
text_file.write('%s: %.5f+/-%.5f\n' %
(metric_names[i], np.mean(scores_grouped[i]),
scipy.stats.sem(scores_grouped[i])))
text_file.write("Time: %s\n\n" %
utils.delta_timer(np.mean(times_list)))
def fit_tf_val(self,
x,
y,
validation_data=None,
batch_size=128,
epochs=1,
iterations_to_validate=None,
patience=0,
verbose=1):
self.verbose = verbose
n_iterations_in_epoch = len(y) // batch_size
# check if validate at end of epoch
if iterations_to_validate is None:
iterations_to_validate = n_iterations_in_epoch
self.best_model_so_far = {
general_keys.ITERATION: 0,
general_keys.LOSS: 1e100,
general_keys.NOT_IMPROVED_COUNTER: 0,
}
train_ds = tf.data.Dataset.from_tensor_slices(
(x, y)).shuffle(10000).batch(batch_size)
val_ds = tf.data.Dataset.from_tensor_slices(
(validation_data[0], validation_data[1])).batch(1024)
self.check_best_model_save(iteration=0)
for epoch in range(epochs):
start_time = time.time()
for it_i, (images, labels) in enumerate(train_ds):
self.train_step(images, labels)
if it_i % iterations_to_validate == 0 and it_i != 0:
if self.check_early_stopping(patience):
return
for test_images, test_labels in val_ds:
self.eval_step(test_images, test_labels)
# TODO: check train printer
if self.verbose:
template = 'Epoch {}, Loss: {}, Acc: {}, Val loss: {}, Val acc: {}, Time: {}'
print(
template.format(
epoch + 1, self.train_loss.result(),
self.train_accuracy.result() * 100,
self.eval_loss.result(),
self.eval_accuracy.result() * 100,
delta_timer(time.time() - start_time)))
self.check_best_model_save(it_i + (
(epoch + 1) * n_iterations_in_epoch))
self.eval_loss.reset_states()
self.eval_accuracy.reset_states()
self.train_loss.reset_states()
self.train_accuracy.reset_states()
self.load_weights(self.best_model_weights_path).expect_partial()
def eval_tf(self, x, y, batch_size=1024, verbose=1):
self.verbose = verbose
self.eval_loss.reset_states()
self.eval_accuracy.reset_states()
dataset = tf.data.Dataset.from_tensor_slices((x, y)).batch(batch_size)
start_time = time.time()
for images, labels in dataset:
self.eval_step(images, labels)
if self.verbose:
template = 'Loss: {}, Acc: {}, Time: {}'
print(
template.format(self.eval_loss.result(),
self.eval_accuracy.result() * 100,
delta_timer(time.time() - start_time)))
results_dict = {
general_keys.LOSS: self.eval_loss.result(),
general_keys.ACCURACY: self.eval_accuracy.result()
}
self.eval_loss.reset_states()
self.eval_accuracy.reset_states()
return results_dict