def assert_gradient_descent(objective, model):
params, image = param.image(224, batch=2)
optimizer = torch.optim.Adam(params, lr=0.05)
T = render.hook_model(model, image)
objective_f = objectives.as_objective(objective)
model(image())
start_value = objective_f(T)
for _ in range(NUM_STEPS):
optimizer.zero_grad()
model(image())
loss = objective_f(T)
loss.backward()
optimizer.step()
end_value = objective_f(T)
assert start_value > end_value
def test_cppn_fits_xor():
params, image = param.cppn(16)
optimizer = torch.optim.Adam(params, lr=0.01)
objective_f = objectives.as_objective(xor_loss)
for _ in range(200):
optimizer.zero_grad()
loss = objective_f(image())
loss.backward()
optimizer.step()
vis = image()[0]
close_enough = (vis[:, 0, 0].mean() > .99
and vis[:, -1, -1].mean() > .99
and vis[:, -1, 0].mean() < .01
and vis[:, 0, -1].mean() < .01)
if close_enough:
return
assert False, "fitting XOR took more than 200 steps, failing test"
def render_vis(
model,
objective_f,
param_f=None,
optimizer=None,
transforms=None,
thresholds=(512, ),
verbose=False,
preprocess=True,
progress=True,
show_image=True,
save_image=False,
image_name=None,
show_inline=False,
fixed_image_size=None,
):
if param_f is None:
param_f = lambda: param.image(128)
# param_f is a function that should return two things
# params - parameters to update, which we pass to the optimizer
# image_f - a function that returns an image as a tensor
params, image_f = param_f()
if optimizer is None:
optimizer = lambda params: torch.optim.Adam(params, lr=5e-2)
optimizer = optimizer(params)
if transforms is None:
transforms = transform.standard_transforms
transforms = transforms.copy()
if preprocess:
if model._get_name() == "InceptionV1":
# Original Tensorflow InceptionV1 takes input range [-117, 138]
transforms.append(transform.preprocess_inceptionv1())
else:
# Assume we use normalization for torchvision.models
# See https://pytorch.org/docs/stable/torchvision/models.html
transforms.append(transform.normalize())
# Upsample images smaller than 224
image_shape = image_f().shape
if fixed_image_size is not None:
new_size = fixed_image_size
elif image_shape[2] < 224 or image_shape[3] < 224:
new_size = 224
else:
new_size = None
if new_size:
transforms.append(
torch.nn.Upsample(size=new_size,
mode="bilinear",
align_corners=True))
transform_f = transform.compose(transforms)
hook = hook_model(model, image_f)
objective_f = objectives.as_objective(objective_f)
if verbose:
model(transform_f(image_f()))
print("Initial loss: {:.3f}".format(objective_f(hook)))
images = []
try:
for i in tqdm(range(1, max(thresholds) + 1), disable=(not progress)):
optimizer.zero_grad()
try:
model.encode_image(transform_f(image_f()))
except RuntimeError as ex:
if i == 1:
# Only display the warning message
# on the first iteration, no need to do that
# every iteration
warnings.warn(
"Some layers could not be computed because the size of the "
"image is not big enough. It is fine, as long as the non"
"computed layers are not used in the objective function"
f"(exception details: '{ex}')")
loss = objective_f(hook)
loss.backward()
optimizer.step()
if i in thresholds:
image = tensor_to_img_array(image_f())
if verbose:
print("Loss at step {}: {:.3f}".format(
i, objective_f(hook)))
if show_inline:
show(image)
images.append(image)
except KeyboardInterrupt:
print("Interrupted optimization at step {:d}.".format(i))
if verbose:
print("Loss at step {}: {:.3f}".format(i, objective_f(hook)))
images.append(tensor_to_img_array(image_f()))
if save_image:
export(image_f(), image_name)
if show_inline:
show(tensor_to_img_array(image_f()))
elif show_image:
view(image_f())
return images
def render_vis(model,
objective_f,
param_f=None,
optimizer=None,
transforms=None,
thresholds=(512, ),
verbose=False,
preprocess=True,
progress=True,
show_image=True,
save_image=False,
image_name=None,
show_inline=False):
if param_f is None:
param_f = lambda: param.image(128)
# param_f is a function that should return two things
# params - parameters to update, which we pass to the optimizer
# image_f - a function that returns an image as a tensor
params, image_f = param_f()
if optimizer is None:
optimizer = lambda params: torch.optim.Adam(params, lr=5e-2)
optimizer = optimizer(params)
if transforms is None:
transforms = transform.standard_transforms.copy()
if preprocess:
if model._get_name() == "InceptionV1":
# Original Tensorflow InceptionV1 takes input range [-117, 138]
transforms.append(transform.preprocess_inceptionv1())
else:
# Assume we use normalization for torchvision.models
# See https://pytorch.org/docs/stable/torchvision/models.html
transforms.append(transform.normalize())
# Upsample images smaller than 224
image_shape = image_f().shape
if image_shape[2] < 224 or image_shape[3] < 224:
transforms.append(
torch.nn.Upsample(size=224, mode='bilinear', align_corners=True))
transform_f = transform.compose(transforms)
hook = hook_model(model, image_f)
objective_f = objectives.as_objective(objective_f)
if verbose:
model(transform_f(image_f()))
print("Initial loss: {:.3f}".format(objective_f(hook)))
images = []
try:
for i in tqdm(range(1, max(thresholds) + 1), disable=(not progress)):
optimizer.zero_grad()
model(transform_f(image_f()))
loss = objective_f(hook)
loss.backward()
optimizer.step()
if i in thresholds:
image = tensor_to_img_array(image_f())
if verbose:
print("Loss at step {}: {:.3f}".format(
i, objective_f(hook)))
if show_inline:
show(image)
images.append(image)
except KeyboardInterrupt:
print("Interrupted optimization at step {:d}.".format(i))
if verbose:
print("Loss at step {}: {:.3f}".format(i, objective_f(hook)))
images.append(tensor_to_img_array(image_f()))
if save_image:
export(image_f(), image_name)
if show_inline:
show(tensor_to_img_array(image_f()))
elif show_image:
view(image_f())
return images