def own_complex_fit(epochs=10):
tf.random.set_seed(1)
init = tf.keras.initializers.GlorotUniform(seed=117)
model = models.Sequential()
model.add(complex_layers.ComplexConv2D(32, (3, 3), activation='cart_relu', input_shape=(32, 32, 3),
kernel_initializer=init, use_bias=False, init_technique='zero_imag'))
model.add(complex_layers.ComplexMaxPooling2D((2, 2)))
model.add(complex_layers.ComplexConv2D(64, (3, 3), activation='cart_relu', kernel_initializer=init,
use_bias=False, init_technique='zero_imag'))
model.add(complex_layers.ComplexMaxPooling2D((2, 2)))
model.add(complex_layers.ComplexConv2D(64, (3, 3), activation='cart_relu', kernel_initializer=init,
use_bias=False, init_technique='zero_imag'))
model.add(complex_layers.ComplexFlatten())
model.add(complex_layers.ComplexDense(64, activation='cart_relu', kernel_initializer=init,
use_bias=False, init_technique='zero_imag'))
model.add(complex_layers.ComplexDense(10, activation='cast_to_real', kernel_initializer=init,
use_bias=False, init_technique='zero_imag'))
# model.summary()
model.compile(optimizer='sgd',
loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True),
metrics=['accuracy'])
weigths = model.get_weights()
with tf.GradientTape() as tape:
loss = model.compiled_loss(y_true=tf.convert_to_tensor(test_labels), y_pred=model(test_images))
gradients = tape.gradient(loss, model.trainable_weights) # back-propagation
history = model.fit(train_images, train_labels, epochs=epochs, validation_data=(test_images, test_labels))
test_loss, test_acc = model.evaluate(test_images, test_labels, verbose=2)
logs = {
'weights_at_init': weigths,
'loss': loss,
'gradients': gradients,
'weights_at_end': model.get_weights()
}
return history, logs
def cifar10_test():
dtype_1 = np.complex64
(train_images, train_labels), (test_images,
test_labels) = datasets.cifar10.load_data()
# Normalize pixel values to be between 0 and 1
train_images, test_images = train_images / 255.0, test_images / 255.0
train_images = train_images.astype(dtype_1)
test_images = test_images.astype(dtype_1)
train_labels = train_labels.astype(dtype_1)
test_labels = test_labels.astype(dtype_1)
model = models.Sequential()
model.add(layers.ComplexInput(input_shape=(32, 32, 3),
dtype=dtype_1)) # Never forget this!!!
model.add(layers.ComplexConv2D(32, (3, 3), activation='cart_relu'))
model.add(layers.ComplexMaxPooling2D(
(2, 2))) # TODO: This is changing the dtype!
model.add(layers.ComplexConv2D(64, (3, 3), activation='cart_relu'))
model.add(layers.ComplexAvgPooling2D((2, 2)))
model.add(layers.ComplexConv2D(64, (3, 3), activation='cart_relu'))
model.add(layers.ComplexFlatten())
model.add(layers.ComplexDense(64, activation='cart_relu'))
model.add(layers.ComplexDense(10, activation='convert_to_real_with_abs'))
model.summary()
model.compile(
optimizer='adam',
loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True),
metrics=['accuracy'])
history = model.fit(train_images,
train_labels,
epochs=2,
validation_data=(test_images, test_labels))
def cifar10_test_model_2(train_images,
train_labels,
test_images,
test_labels,
dtype_1='complex64'):
x = layers.complex_input(shape=(32, 32, 3), dtype=dtype_1)
conv1 = layers.ComplexConv2D(32, (3, 3), activation='cart_relu')(x)
pool1 = layers.ComplexMaxPooling2D((2, 2))(conv1)
conv2 = layers.ComplexConv2D(64, (3, 3), activation='cart_relu')(pool1)
pool2 = layers.ComplexAvgPooling2D((2, 2))(conv2)
conv3 = layers.ComplexConv2D(64, (3, 3), activation='cart_relu')(pool2)
flat = layers.ComplexFlatten()(conv3)
dense1 = layers.ComplexDense(64, activation='cart_relu')(flat)
y = layers.ComplexDense(10, activation='convert_to_real_with_abs')(dense1)
model = models.Model(inputs=[x], outputs=[y])
model.compile(
optimizer='adam',
loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True),
metrics=['accuracy'])
model.summary()
return model.fit(train_images,
train_labels,
epochs=2,
validation_data=(test_images, test_labels),
shuffle=False)
def random_dataset():
x_train = np.complex64(tf.complex(tf.random.uniform([640, 65, 82, 1]), tf.random.uniform([640, 65, 82, 1])))
x_test = np.complex64(tf.complex(tf.random.uniform([200, 65, 82, 1]), tf.random.uniform([200, 65, 82, 1])))
y_train = np.uint8(np.random.randint(5, size=(640, 1)))
y_test = np.uint8(np.random.randint(5, size=(200, 1)))
model = tf.keras.models.Sequential()
model.add(complex_layers.ComplexInput(input_shape=(65, 82, 1))) # Always use ComplexInput at the start
model.add(complex_layers.ComplexConv2D(8, (5, 5), activation='cart_relu'))
model.add(complex_layers.ComplexMaxPooling2D((2, 2)))
model.add(complex_layers.ComplexConv2D(16, (5, 5), activation='cart_relu'))
model.add(complex_layers.ComplexFlatten())
model.add(complex_layers.ComplexDense(256, activation='cart_relu'))
model.add(complex_layers.ComplexDropout(0.1))
model.add(complex_layers.ComplexDense(64, activation='cart_relu'))
model.add(complex_layers.ComplexDropout(0.1))
model.add(complex_layers.ComplexDense(5, activation='convert_to_real_with_abs'))
# An activation that casts to real must be used at the last layer.
# The loss function cannot minimize a complex number
# Compile it
model.compile(optimizer='adam',
loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True),
metrics=['accuracy'],
# run_eagerly=Trutest_regressione
)
model.summary()
# Train and evaluate
history = model.fit(x_train, y_train, epochs=2, validation_data=(x_test, y_test))
test_loss, test_acc = model.evaluate(x_test, y_test, verbose=2)
def cifar10_test_model_1(train_images,
train_labels,
test_images,
test_labels,
dtype_1='complex64'):
model = models.Sequential()
model.add(layers.ComplexInput(input_shape=(32, 32, 3),
dtype=dtype_1)) # Never forget this!!!
model.add(layers.ComplexConv2D(32, (3, 3), activation='cart_relu'))
model.add(layers.ComplexMaxPooling2D((2, 2)))
model.add(layers.ComplexConv2D(64, (3, 3), activation='cart_relu'))
model.add(layers.ComplexAvgPooling2D((2, 2)))
model.add(layers.ComplexConv2D(64, (3, 3), activation='cart_relu'))
model.add(layers.ComplexFlatten())
model.add(layers.ComplexDense(64, activation='cart_relu'))
model.add(layers.ComplexDense(10, activation='convert_to_real_with_abs'))
model.summary()
model.compile(
optimizer='adam',
loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True),
metrics=['accuracy'])
return model.fit(train_images,
train_labels,
epochs=2,
validation_data=(test_images, test_labels),
shuffle=False)
def get_cvnn_model(dtype=tf.float32):
tf.random.set_seed(1)
inputs = layers.complex_input(shape=INPUT_SIZE + (3, ), dtype=dtype)
# inputs = tf.keras.layers.InputLayer(input_shape=INPUT_SIZE + (3,), dtype=dtype)
# inputs = tf.keras.layers.Input(shape=INPUT_SIZE + (3,))
c0, c1, c2 = _downsample_cvnn(inputs, 64, dtype)
c3, c4, c5 = _downsample_cvnn(c2, 128, dtype)
c6, c7, c8 = _downsample_cvnn(c5, 256, dtype)
c9, c10, c11 = _downsample_cvnn(c8, 512, dtype)
c12 = layers.ComplexConv2D(1024,
activation='relu',
kernel_size=3,
dtype=dtype)(c11)
c13 = layers.ComplexConv2D(1024,
activation='relu',
kernel_size=3,
padding='valid',
dtype=dtype)(c12)
c14, c15 = _upsample_cvnn(c13, c10, 512, 4, dtype)
c16, c17 = _upsample_cvnn(c15, c7, 256, 16, dtype)
c18, c19 = _upsample_cvnn(c17, c4, 128, 40, dtype)
c20, c21 = _upsample_cvnn(c19, c1, 64, 88, dtype)
outputs = layers.ComplexConv2D(4, kernel_size=1, dtype=dtype)(c21)
model = tf.keras.Model(inputs=inputs, outputs=outputs, name="u-net-cvnn")
model.compile(loss=tf.keras.losses.SparseCategoricalCrossentropy(),
optimizer="adam",
metrics=["accuracy"])
return model
def all_layers_model():
"""
Creates a model using all possible layers to assert no layer changes the dtype to real.
"""
input_shape = (4, 28, 28, 3)
x = tf.cast(tf.random.normal(input_shape), tf.complex64)
model = tf.keras.models.Sequential()
model.add(complex_layers.ComplexInput(
input_shape=input_shape[1:])) # Always use ComplexInput at the start
model.add(complex_layers.ComplexConv2D(32, (3, 3), activation='cart_relu'))
model.add(complex_layers.ComplexAvgPooling2D((2, 2)))
model.add(
complex_layers.ComplexConv2D(64, (3, 3), activation='cart_sigmoid'))
model.add(complex_layers.ComplexDropout(0.5))
model.add(complex_layers.ComplexMaxPooling2D((2, 2)))
model.add(complex_layers.ComplexConv2DTranspose(32, (2, 2)))
model.add(complex_layers.ComplexFlatten())
model.add(complex_layers.ComplexDense(64, activation='cart_tanh'))
model.compile(loss=tf.keras.losses.MeanAbsoluteError(),
optimizer='adam',
metrics=['accuracy'])
y = model(x)
assert y.dtype == np.complex64
return model
def test_functional_api():
inputs = complex_layers.complex_input(shape=(128, 128, 3))
c0 = complex_layers.ComplexConv2D(32, activation='cart_relu', kernel_size=3)(inputs)
c1 = complex_layers.ComplexConv2D(32, activation='cart_relu', kernel_size=3)(c0)
c2 = complex_layers.ComplexMaxPooling2D(pool_size=(2, 2), strides=(2, 2), padding='valid')(c1)
t01 = complex_layers.ComplexConv2DTranspose(5, kernel_size=2, strides=(2, 2), activation='cart_relu')(c2)
concat01 = tf.keras.layers.concatenate([t01, c1], axis=-1)
c3 = complex_layers.ComplexConv2D(4, activation='cart_relu', kernel_size=3)(concat01)
out = complex_layers.ComplexConv2D(4, activation='cart_relu', kernel_size=3)(c3)
model = tf.keras.Model(inputs, out)
def _downsample_cvnn(inputs, units, dtype=tf.float32):
c0 = layers.ComplexConv2D(units,
activation='cart_relu',
kernel_size=3,
dtype=dtype)(inputs)
c1 = layers.ComplexConv2D(units,
activation='cart_relu',
kernel_size=3,
dtype=dtype)(c0)
c2 = layers.ComplexMaxPooling2D(pool_size=(2, 2),
strides=(2, 2),
padding='valid',
dtype=dtype)(c1)
return c0, c1, c2
def test_cifar():
(train_images, train_labels), (test_images, test_labels) = get_dataset()
# Create your model
model = tf.keras.models.Sequential()
model.add(complex_layers.ComplexInput(input_shape=(32, 32, 3))) # Always use ComplexInput at the start
model.add(complex_layers.ComplexConv2D(32, (3, 3), activation='cart_relu'))
model.add(complex_layers.ComplexAvgPooling2D((2, 2)))
model.add(complex_layers.ComplexConv2D(64, (3, 3), activation='cart_relu'))
model.add(complex_layers.ComplexFlatten())
model.add(complex_layers.ComplexDense(64, activation='cart_relu'))
model.add(complex_layers.ComplexDense(10, activation='convert_to_real_with_abs'))
model.compile(optimizer='adam',
loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True),
metrics=['accuracy'])
# model.summary()
history = model.fit(train_images, train_labels, epochs=1, validation_data=(test_images, test_labels))
test_loss, test_acc = model.evaluate(test_images, test_labels, verbose=2)
def own_fit(epochs=10):
tf.random.set_seed(1)
init = tf.keras.initializers.GlorotUniform(seed=117)
model = models.Sequential()
model.add(
complex_layers.ComplexConv2D(32, (3, 3),
activation='cart_relu',
input_shape=(32, 32, 3),
dtype=np.float32,
kernel_initializer=init))
model.add(complex_layers.ComplexMaxPooling2D((2, 2), dtype=np.float32))
model.add(
complex_layers.ComplexConv2D(64, (3, 3),
activation='cart_relu',
dtype=np.float32,
kernel_initializer=init))
model.add(complex_layers.ComplexMaxPooling2D((2, 2), dtype=np.float32))
model.add(
complex_layers.ComplexConv2D(64, (3, 3),
activation='cart_relu',
dtype=np.float32,
kernel_initializer=init))
model.add(complex_layers.ComplexFlatten())
model.add(
complex_layers.ComplexDense(64,
activation='cart_relu',
dtype=np.float32,
kernel_initializer=init))
model.add(
complex_layers.ComplexDense(10,
dtype=np.float32,
kernel_initializer=init))
# model.summary()
model.compile(
optimizer='adam',
loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True),
metrics=['accuracy'])
history = model.fit(train_images,
train_labels,
epochs=epochs,
validation_data=(test_images, test_labels))
test_loss, test_acc = model.evaluate(test_images, test_labels, verbose=2)
return history
def _upsample_cvnn(in1, in2, units, crop, dtype=tf.float32):
t01 = layers.ComplexConv2DTranspose(units,
kernel_size=2,
strides=(2, 2),
activation='relu',
dtype=dtype)(in1)
crop01 = tf.keras.layers.Cropping2D(cropping=(crop, crop))(in2)
concat01 = tf.keras.layers.concatenate([t01, crop01], axis=-1)
out1 = layers.ComplexConv2D(units,
activation='relu',
kernel_size=3,
dtype=dtype)(concat01)
out2 = layers.ComplexConv2D(units,
activation='relu',
kernel_size=3,
dtype=dtype)(out1)
return out1, out2
def test_complex_conv2d(self):
conv2d = layers.ComplexConv2D(10, kernel_size=(3, 3))
cimg = np.random.randn(1, 10, 10,
3) + 1j * np.random.randn(1, 10, 10, 3)
_ = conv2d(cimg.astype('complex64'))
