def __call__(self, image, style_name):
"""
Deep Dream Main Algorithm.
Passes the given image through the model and uses the
Gradient Ascent Method to update the image.
"""
# prepare style
style = self.cfg.STYLES_CFG[style_name]
logger.info(f'Running deep dream algorithm using style: {style_name}')
# prepare model
model = DeepGoogLeNet(loi=style.loi)
model.eval()
# prepare input image
try:
original_size = image.size
image = image.resize((style.size, style.size))
target = utils.preprocess(image).to(self.device)
except Exception:
raise PreProcessingError()
start = time()
for e in range(style.epochs):
# reset gradient
if target.grad is not None:
target.grad.zero_()
# loss backward
loss = self.deep_dream_loss(model, target)
loss.backward(retain_graph=True)
# gradient ascent step (standarizing the gradient)
grad = target.grad.data / (torch.std(target.grad.data) + 1e-8)
target.data = target.data + grad * style.learning_rate
# clip pixel values
target.data = utils.clip(target.data)
logger.debug(f'Epoch {e}/{style.epochs} '
f'took: {time() - start:.2f}')
logger.info('Deep Dream with style: '
f'{style_name} took: {time() - start:.2f}')
try:
dream = target.cpu().clone().detach().squeeze(0)
dream = utils.postprocess(dream).resize(original_size)
except Exception:
raise PostProcessingError()
return dream
def test_feature_extraction_count(self):
model = DeepGoogLeNet((2, 4))
inp = torch.randn(1, 3, 256, 256)
_ = model(inp)
assert len(model.features) == 2, 'Incorrect feature count'
def test_init_loi_interpretation(self):
model = DeepGoogLeNet((2, 4))
assert len(model.hooks) == 2, 'Incorrect hook count'
assert len(model.layers) == 4, 'Incorrect layer count'
def test_init_over_b(self):
self.assertRaises(ModelInitializationError,
lambda: DeepGoogLeNet((1, 17)))
def test_init_neg_a(self):
self.assertRaises(ModelInitializationError,
lambda: DeepGoogLeNet((-1, 9)))
def test_init(self):
self.assertNotRaise(lambda: DeepGoogLeNet((0, 4)),
'Model Initalization failed')