def test_gan_custom_layer_graph():
z_shape = (1, 8, 8)
z = Input(shape=z_shape, name='z')
gen_cond = Input(shape=(1, 8, 8), name='gen_cond')
inputs = [z, gen_cond]
gen_input = merge(inputs, mode='concat', concat_axis=1)
gen_output = Convolution2D(1, 2, 2, activation='relu',
name='g1',
border_mode='same')(gen_input)
generator = Container(inputs, gen_output)
f, r = Input(z_shape, name='fake'), Input(z_shape, name='real')
inputs = [f, r]
dis_input = merge(inputs, mode='concat', concat_axis=0)
dis_conv = Convolution2D(5, 2, 2, name='d1', activation='relu')(dis_input)
dis_flatten = Flatten()(dis_conv)
dis = Dense(1, activation='sigmoid')(dis_flatten)
discriminator = Container(inputs, gan_outputs(dis))
gan = GAN(generator, discriminator, z_shape=z_shape, real_shape=z_shape)
gan.build('adam', 'adam', gan_binary_crossentropy)
fn = gan.compile_custom_layers(['g1', 'd1'])
z = np.random.uniform(-1, 1, (64,) + z_shape)
real = np.random.uniform(-1, 1, (64,) + z_shape)
cond = np.random.uniform(-1, 1, (64,) + z_shape)
print(z.shape)
print(real.shape)
print(cond.shape)
fn({'z': z, 'gen_cond': cond, 'real': real})
def test_gan_utility_funcs(simple_gan: GAN):
simple_gan.build('adam', 'adam', gan_binary_crossentropy)
simple_gan.compile()
xy_shp = simple_gan_z_shape[1:]
x = np.zeros(xy_shp, dtype=np.float32)
y = np.zeros(xy_shp, dtype=np.float32)
simple_gan.interpolate(x, y)
z_point = simple_gan.random_z_point()
neighbors = simple_gan.neighborhood(z_point, std=0.05)
diff = np.stack([neighbors[0]]*len(neighbors)) - neighbors
assert np.abs(diff).mean() < 0.1
def mogan(self, gan: GAN, loss_fn, d_optimizer, name="mogan",
gan_objective=binary_crossentropy, gan_regulizer=None,
cond_true_ndim=4):
assert len(gan.conditionals) >= 1
g_dummy_opt = SGD()
d_optimizer = d_optimizer
v = gan.build(g_dummy_opt, d_optimizer, gan_objective)
del v['g_updates']
cond_true = K.placeholder(ndim=cond_true_ndim)
inputs = copy(gan.graph.inputs)
inputs['cond_true'] = cond_true
cond_loss = loss_fn(cond_true, v.g_outmap)
metrics = {
"cond_loss": cond_loss.mean(),
"d_loss": v.d_loss,
"g_loss": v.g_loss,
}
params = flatten([n.trainable_weights
for n in gan.get_generator_nodes().values()])
return MultipleObjectives(
name, inputs, metrics=metrics, params=params,
objectives={'g_loss': v['g_loss'], 'cond_loss': cond_loss},
additional_updates=v['d_updates'] + gan.updates)
def test_gan_graph():
z_shape = (1, 8, 8)
z = Input(shape=z_shape, name='z')
gen_cond = Input(shape=(1, 8, 8), name='gen_cond')
inputs = [z, gen_cond]
gen_input = merge(inputs, mode='concat', concat_axis=1)
gen_output = Convolution2D(10, 2, 2, activation='relu',
border_mode='same')(gen_input)
generator = Container(inputs, gen_output)
f, r = Input(z_shape, name='f'), Input(z_shape, name='r')
inputs = [f, r]
dis_input = merge(inputs, mode='concat', concat_axis=1)
dis_conv = Convolution2D(5, 2, 2, activation='relu')(dis_input)
dis_flatten = Flatten()(dis_conv)
dis = Dense(1, activation='sigmoid')(dis_flatten)
discriminator = Container(inputs, gan_outputs(dis))
gan = GAN(generator, discriminator, z_shape=z_shape, real_shape=z_shape)
gan.build('adam', 'adam', gan_binary_crossentropy)
gan.compile()
gan.generate({'gen_cond': np.zeros((64,) + z_shape)}, nb_samples=64)
def mogan(self,
gan: GAN,
loss_fn,
d_optimizer,
name="mogan",
gan_objective=binary_crossentropy,
gan_regulizer=None,
cond_true_ndim=4):
assert len(gan.conditionals) >= 1
g_dummy_opt = SGD()
d_optimizer = d_optimizer
v = gan.build(g_dummy_opt, d_optimizer, gan_objective)
del v['g_updates']
cond_true = K.placeholder(ndim=cond_true_ndim)
inputs = copy(gan.graph.inputs)
inputs['cond_true'] = cond_true
cond_loss = loss_fn(cond_true, v.g_outmap)
metrics = {
"cond_loss": cond_loss.mean(),
"d_loss": v.d_loss,
"g_loss": v.g_loss,
}
params = flatten(
[n.trainable_weights for n in gan.get_generator_nodes().values()])
return MultipleObjectives(name,
inputs,
metrics=metrics,
params=params,
objectives={
'g_loss': v['g_loss'],
'cond_loss': cond_loss
},
additional_updates=v['d_updates'] + gan.updates)
def test_mask_blending_generator():
nb_driver = 20
def driver(z):
return Dense(nb_driver)(z)
def mask_generator(x):
return sequential([
Dense(16),
Reshape((1, 4, 4)),
UpSampling2D((16, 16))
])(x)
def merge_mask(subsample):
def call(x):
if subsample:
x = MaxPooling2D(subsample)(x)
return Convolution2D(1, 3, 3, border_mode='same')(x)
return call
def light_generator(ins):
seq = sequential([
Convolution2D(1, 3, 3, border_mode='same')
])(concat(ins))
return UpSampling2D((4, 4))(seq), UpSampling2D((4, 4))(seq),
def offset_front(x):
return sequential([
Dense(16),
Reshape((1, 4, 4)),
UpSampling2D((4, 4))
])(concat(x))
def offset_middle(x):
return UpSampling2D()(concat(x))
def offset_back(x):
feature_map = sequential([
UpSampling2D(),
])(concat(x))
return feature_map, Convolution2D(1, 3, 3,
border_mode='same')(feature_map)
def mask_post(x):
return sequential([
Convolution2D(1, 3, 3, border_mode='same')
])(concat(x))
def mask_weight_blending(x):
return sequential([
Flatten(),
Dense(1),
])(x)
def discriminator(x):
return gan_outputs(sequential([
Flatten(),
Dense(1),
])(concat(x)), fake_for_gen=(0, 10), fake_for_dis=(0, 10),
real=(10, 20))
gen = mask_blending_generator(
mask_driver=driver,
mask_generator=mask_generator,
light_merge_mask16=merge_mask(None),
offset_merge_light16=merge_mask((4, 4)),
offset_merge_mask16=merge_mask(None),
offset_merge_mask32=merge_mask(None),
lighting_generator=light_generator,
offset_front=offset_front,
offset_middle=offset_middle,
offset_back=offset_back,
mask_weight_blending32=mask_weight_blending,
mask_weight_blending64=mask_weight_blending,
mask_postprocess=mask_post,
z_for_driver=(0, 10),
z_for_offset=(10, 20),
z_for_bits=(20, 32),
)
z_shape = (32, )
real_shape = (1, 64, 64)
gan = GAN(gen, discriminator, z_shape, real_shape)
gan.build(Adam(), Adam(), gan_binary_crossentropy)
for l in gan.gen_layers:
print("{}: {}, {}".format(
l.name, l.output_shape, getattr(l, 'regularizers', [])))
bs = 10
z_in = np.random.sample((bs,) + z_shape)
gan.compile_generate()
gan.generate({'z': z_in})
def test_mask_blending_generator():
nb_driver = 20
def driver(z):
return Dense(nb_driver)(z)
def mask_generator(x):
return sequential(
[Dense(16), Reshape((1, 4, 4)),
UpSampling2D((16, 16))])(x)
def merge_mask(subsample):
def call(x):
if subsample:
x = MaxPooling2D(subsample)(x)
return Convolution2D(1, 3, 3, border_mode='same')(x)
return call
def light_generator(ins):
seq = sequential([Convolution2D(1, 3, 3,
border_mode='same')])(concat(ins))
return UpSampling2D((4, 4))(seq), UpSampling2D((4, 4))(seq),
def offset_front(x):
return sequential(
[Dense(16), Reshape((1, 4, 4)),
UpSampling2D((4, 4))])(concat(x))
def offset_middle(x):
return UpSampling2D()(concat(x))
def offset_back(x):
feature_map = sequential([
UpSampling2D(),
])(concat(x))
return feature_map, Convolution2D(1, 3, 3,
border_mode='same')(feature_map)
def mask_post(x):
return sequential([Convolution2D(1, 3, 3,
border_mode='same')])(concat(x))
def mask_weight_blending(x):
return sequential([
Flatten(),
Dense(1),
])(x)
def discriminator(x):
return gan_outputs(sequential([
Flatten(),
Dense(1),
])(concat(x)),
fake_for_gen=(0, 10),
fake_for_dis=(0, 10),
real=(10, 20))
gen = mask_blending_generator(
mask_driver=driver,
mask_generator=mask_generator,
light_merge_mask16=merge_mask(None),
offset_merge_light16=merge_mask((4, 4)),
offset_merge_mask16=merge_mask(None),
offset_merge_mask32=merge_mask(None),
lighting_generator=light_generator,
offset_front=offset_front,
offset_middle=offset_middle,
offset_back=offset_back,
mask_weight_blending32=mask_weight_blending,
mask_weight_blending64=mask_weight_blending,
mask_postprocess=mask_post,
z_for_driver=(0, 10),
z_for_offset=(10, 20),
z_for_bits=(20, 32),
)
z_shape = (32, )
real_shape = (1, 64, 64)
gan = GAN(gen, discriminator, z_shape, real_shape)
gan.build(Adam(), Adam(), gan_binary_crossentropy)
for l in gan.gen_layers:
print("{}: {}, {}".format(l.name, l.output_shape,
getattr(l, 'regularizers', [])))
bs = 10
z_in = np.random.sample((bs, ) + z_shape)
gan.compile_generate()
gan.generate({'z': z_in})