def test_optimizer_step():
batch_size = 64
width = 14
height = 28
hyper = {
'width_height': (width, height, 1),
'model_type': 'GS',
'batch_size': batch_size,
'learning_rate': 0.0003,
'units_per_layer': 240,
'temp': tf.constant(0.1)
}
shape = (batch_size, width, height, 1)
x_upper, x_lower = create_upper_and_lower_dummy_data(shape=shape)
sop = SOP(hyper=hyper)
optimizer = tf.keras.optimizers.Adam(learning_rate=hyper['learning_rate'],
beta_1=0.9,
beta_2=0.999,
decay=1.e-3)
sop_opt = SOPOptimizer(model=sop, optimizer=optimizer)
gradients, loss = sop_opt.compute_gradients_and_loss(x_upper=x_upper,
x_lower=x_lower)
sop_opt.apply_gradients(gradients=gradients)
print('\nTEST: Gradient step from optimizer')
assert gradients is not None
def setup_sop_optimizer(hyper):
optimizer = tf.keras.optimizers.Adam(learning_rate=hyper['learning_rate'],
beta_1=0.9,
beta_2=0.999)
model = SOP(hyper=hyper)
sop_optimizer = SOPOptimizer(model=model, optimizer=optimizer)
return sop_optimizer
def test_count_of_network_parameters():
batch_size, width, height, rgb = 64, 14, 28, 1
units_per_layer = 240
hyper = {
'width_height': (width, height, rgb),
'model_type': 'GS',
'batch_size': batch_size,
'units_per_layer': units_per_layer,
'temp': tf.constant(0.1)
}
shape = (batch_size, width, height, rgb)
sop = SOP(hyper=hyper)
sop.build(input_shape=shape)
print('\nTEST: Number of parameters in the network')
assert sop.h1_dense.count_params() == (392 + 1) * units_per_layer
assert sop.h2_dense.count_params() == (units_per_layer +
1) * units_per_layer
assert sop.out_dense.count_params() == (units_per_layer + 1) * 392
def test_fwd_pass_connections_and_gradient(self):
batch_size = 64
width = 14
height = 28
hyper = {
'width_height': (width, height, 1),
'model_type': 'GS',
'batch_size': batch_size,
'temp': tf.constant(0.1)
}
shape = (batch_size, width, height, 1)
x_upper, x_lower = create_upper_and_lower(shape=shape)
sop = SOP(hyper=hyper)
with tf.GradientTape() as tape:
logits = sop.call(x_upper=x_upper)
loss = tf.nn.sigmoid_cross_entropy_with_logits(labels=x_lower,
logits=logits)
grad = tape.gradient(sources=sop.trainable_variables, target=loss)
self.assertTrue(grad is not None)
def test_fwd_pass_connections_and_gradient():
batch_size, width, height, rgb, sample_size = 64, 14, 28, 1, 1
hyper = {
'width_height': (width, height, 1),
'model_type': 'GS',
'batch_size': batch_size,
'units_per_layer': 240,
'temp': tf.constant(0.1)
}
shape = (batch_size, width, height, rgb)
x_upper, x_lower = create_upper_and_lower_dummy_data(shape=shape)
sop = SOP(hyper=hyper)
with tf.GradientTape() as tape:
logits = sop.call(x_upper=x_upper)
x_lower = tf.reshape(x_lower, x_lower.shape + (sample_size, ))
loss = tf.nn.sigmoid_cross_entropy_with_logits(labels=x_lower,
logits=logits)
grad = tape.gradient(sources=sop.trainable_variables, target=loss)
print('\nTEST: Forward pass and gradient computation')
assert grad is not None
def test_optimizer_step(self):
batch_size = 64
width = 14
height = 28
hyper = {
'width_height': (width, height, 1),
'model_type': 'IGR',
'batch_size': batch_size,
'learning_rate': 0.001,
'temp': tf.constant(0.1)
}
shape = (batch_size, width, height, 1)
x_upper, x_lower = create_upper_and_lower(shape=shape)
sop = SOP(hyper=hyper)
optimizer = tf.keras.optimizers.Adam(
learning_rate=hyper['learning_rate'])
sop_opt = SOPOptimizer(model=sop, optimizer=optimizer)
gradients, loss = sop_opt.compute_gradients_and_loss(x_upper=x_upper,
x_lower=x_lower)
sop_opt.apply_gradients(gradients=gradients)
self.assertTrue(gradients is not None)
from Utils.load_data import load_mnist_sop_data
from Models.SOP import SOP
from Models.SOPOptimizer import viz_reconstruction
import tensorflow as tf
model_type = 'GS'
hyper = {
'width_height': (14, 28, 1),
'model_type': model_type,
'batch_size': 64,
'learning_rate': 0.0003,
'epochs': 100,
'iter_per_epoch': 937,
'temp': tf.constant(0.67)
}
data = load_mnist_sop_data(batch_n=hyper['batch_size'], epochs=hyper['epochs'])
train, test = data
model = SOP(hyper=hyper)
results_file = './Log/model_weights_GS.h5'
shape = (hyper['batch_size'], ) + hyper['width_height']
shape = (64, 14, 28, 1)
model.build(input_shape=shape)
model.load_weights(filepath=results_file)
for x_test in test.take(10):
images = x_test
viz_reconstruction(test_image=images, model=model)