def test_split(self):
options = Options()
_, _, x_test, _ = get_dataset(options)
model = train_network(get_model(x_test, options), options)
part1, part2 = split_network(model, 3)
model_output = model(x_test[0:5])
split_output = part2(part1(x_test[0:5]))
equality = tf.math.reduce_all(tf.equal(model_output, split_output))
self.assertEqual(equality, True)
def test_compose_rotation(self):
"""Test whether rotation -> convolution == convolution -> rotation
It isn't. This has implications for our project.
This is not as much a test as a demonstration of fact"""
options = Options()
_, _, x_test, _ = get_dataset(options)
model = train_network(get_model(x_test, options), options)
part1, _ = split_network(model, 8)
image = tf.expand_dims(x_test[0], 0)
rotated_image = tfa.image.rotate(image, math.pi)
out = tfa.image.rotate(part1(image), math.pi)
rotated_out = part1(rotated_image)
equality = tf.math.reduce_all(
tf.equal(tfa.image.rotate(out, math.pi), rotated_out))
self.assertEqual(equality, False)
def basic(iterations, profile):
options = Options()
options.serial = True
#options.accperclass = True
options.convlayers = 0
options.step = 20
options.combine = True
options.post_init()
x_train, y_train, x_test, y_test = get_dataset(options) # pylint: disable=unused-variable
model = train_network(get_model(x_test, options), options)
only_convolutional, _ = split_network(model, options.convlayers) # pylint: disable=unused-variable
cmd_string = """run_experiment(iterations, options, x_train, y_train, x_test, y_test,
model, only_convolutional, EXPERIMENT_DIR + '/basic.csv')"""
if profile:
cProfile.run(cmd_string, PROFILING_DIR + '/basic')
else:
exec(cmd_string)
def rot_first(iterations, profile):
options = Options()
options.serial = True
options.convlayers = 5
options.step = 20
options.combine = False
options.representatives = False
options.rotate_first = True
options.post_init()
x_train, y_train, x_test, y_test = get_dataset(options) # pylint: disable=unused-variable
model = train_network(get_model(x_test, options), options)
options.convlayers = get_last_conv_layer(model) + 1
only_convolutional, _ = split_network(model, options.convlayers) # pylint: disable=unused-variable
cmd_string = """run_experiment(iterations, options, x_train, y_train, x_test, y_test,
model, only_convolutional, EXPERIMENT_DIR + '/rotate_first.csv')"""
if profile:
cProfile.run(cmd_string, PROFILING_DIR + '/rotate_first')
else:
exec(cmd_string)
import sys
import os
from src.runner import check_some
from src.options import Options
from src.training_sample_finder import get_training_samples
from src.network import train_network, get_dataset, get_model, split_network
sys.path.append(os.path.join(os.path.dirname(__file__), 'src'))
if __name__ == "__main__":
# Initialize the options class.
options = Options()
# Parse arguments
options.parse_args(10000)
# Get the MNIST dataset
x_train, y_train, x_test, y_test = get_dataset(options)
# Gets the model and the conv part of model, the function also trains the
# network, if there doesn't exist a saved network
model = train_network(get_model(x_test, options), options)
only_convolutional, _ = split_network(model, options.convlayers)
model.summary()
if only_convolutional is not None:
only_convolutional.summary()
training_samples = get_training_samples(only_convolutional, x_train,
y_train, options)
# Run the checksome function, which tests a chosen algorithm to find if it
# can properly rotate pictures back
check_some(only_convolutional, x_test, y_test, model, training_samples,
options)