def evaluate_and_save_performance_power(seed, lr_power):
# Setup
quaternion_predicted = np.load("data/predicted_quaternions2.npy")
angles_predicted = quaternion2euler(quaternion_predicted)
angles_true = np.load("data/angles_true.npy")
NUM_PROJECTIONS = len(angles_true)
kwargs = {
'm': [1.0, 1.0, 1.0, 1.0],
'steps': 1000,
'batch_size': 256,
'projection_idx': range(NUM_PROJECTIONS),
'angles_true': angles_true,
'angles_predicted': angles_predicted
}
optimizer = Ftrl(learning_rate_power=lr_power, learning_rate=0.1)
opt_name = optimizer.__class__.__name__
# Optimization algorithm
rotation, loss, collect_data, trajectory = training_angle_alignment(
optimizer=optimizer,
seed=seed,
save_id=f"{lr_power}_power_{seed}",
**kwargs)
# Save loss
loss_file_path = f'results/alignment/loss_lr_power.csv'
if not os.path.exists(loss_file_path):
df = pd.DataFrame(columns=['lr', 'final_loss'])
else:
df = pd.read_csv(loss_file_path)
df = df.append({'lr': lr_power, 'final_loss': loss}, ignore_index=True)
df.to_csv(loss_file_path, index=False)
def test_initialize(self):
self._compare_initialize_values(Adam(), 4, "m",
init_ops.constant_initializer(0.0))
self._compare_initialize_values(
Ftrl(initial_accumulator_value=0.5),
4,
"accumulator",
init_ops.constant_initializer(0.5),
)
self._compare_initialize_values(
Adagrad(initial_accumulator_value=0.5),
4,
"accumulator",
init_ops.constant_initializer(0.5),
)
def test_allowed_slot_names(self):
opt_and_slots_pairs = [
(SGD(), []),
(SGD(momentum=0.2), ["momentum"]),
(Adam(), ["m", "v"]),
(Adam(amsgrad=True), ["m", "v", "vhat"]),
(Adamax(), ["m", "v"]),
(Nadam(), ["m", "v"]),
(Adadelta(), ["accum_grad", "accum_var"]),
(Adagrad(), ["accumulator"]),
(Ftrl(), ["accumulator", "linear"]),
(RMSprop(), ["rms"]),
(RMSprop(momentum=0.2), ["rms", "momentum"]),
(RMSprop(centered=True), ["rms", "mg"]),
(RMSprop(momentum=0.2, centered=True), ["rms", "momentum", "mg"]),
]
for opt, expected_slots in opt_and_slots_pairs:
self._compare_slot_names(opt, expected_slots)
def _get_optimizer(opt_type, learning_rate):
if opt_type == 'adam':
return Adam(learning_rate=learning_rate)
elif opt_type == 'nadam':
return Nadam(learning_rate=learning_rate)
elif opt_type == 'sgd':
return SGD(learning_rate=learning_rate)
elif opt_type == 'adadelta':
return Adadelta(learning_rate=learning_rate)
elif opt_type == 'adagrad':
return Adagrad(learning_rate=learning_rate)
elif opt_type == 'adamax':
return Adamax(learning_rate=learning_rate)
elif opt_type == 'ftrl':
return Ftrl(learning_rate=learning_rate)
elif opt_type == 'rmsprop':
return RMSprop(learning_rate=learning_rate)
else:
raise NotImplementedError('Optimizer type %s is not implemented.' %
opt_type)
# store confusion matrix
with open(result_folder + '/' + model_name + '_cm_' + str(num_classes),
'wb') as file_pi:
pickle.dump(confusion_matrix(labels, predictions), file_pi)
del model, history, predictions, labels, result, test_labels, result_fix
optimizers = {
'sgd': SGD(lr=0.001, momentum=0.9),
'adam': Adam(),
'adamax': Adamax(),
'adadelta': Adadelta(),
'adagrad': Adagrad(),
'ftrl': Ftrl(),
'nadam': Nadam(),
'rmsprop': RMSprop()
}
initializers = [
'constant', 'glorot_normal', 'glorot_uniform', 'he_normal', 'he_uniform',
'lecun_normal', 'lecun_uniform', 'random_normal', 'truncated_normal'
]
# Try lots of different small models, see what works best
fit_evaluate_model(mymodels.get_simple_cnn(IMG_DIM, num_classes), 'mymodel',
'baseline')
fit_evaluate_model(mymodels.vgg1(IMG_DIM, num_classes), 'mymodel', 'vgg1')
fit_evaluate_model(mymodels.vgg2(IMG_DIM, num_classes), 'mymodel', 'vgg2')
fit_evaluate_model(mymodels.vgg3(IMG_DIM, num_classes), 'mymodel', 'vgg3')