def test_custom_image_param_set_param(self):
"""
checks if custom_image_param.set_param correctly
loads the image without discrepancies with an absolute tolerance of 1e-5 element-wise
"""
model = models.inception_v3(pretrained=True)
dreamy_boi = dreamer(model=model, device='cpu', quiet=False)
param = custom_image_param(image='images/sample_small.jpg',
device='cpu')
image_param = dreamy_boi.render(image_parameter=param,
layers=[model.Mixed_6a],
iters=5,
lr=2e-4,
grad_clip=0.1,
weight_decay=1e-1)
image_tensor = image_param.to_nchw_tensor()
image_param.set_param(tensor=image_tensor)
# print(torch.abs((image_tensor - image_param.to_nchw_tensor())).mean())
self.assertTrue(
torch.allclose(image_tensor,
image_param.to_nchw_tensor(),
atol=1e-5))
def test_custom_image_param(self):
model = models.inception_v3(pretrained=True)
dreamy_boi = dreamer(model=model, device='cpu', quiet=False)
param = custom_image_param(image='images/sample_small.jpg',
device='cpu')
image_param = dreamy_boi.render(image_parameter=param,
layers=[model.Mixed_6a],
iters=5,
lr=2e-4,
grad_clip=0.1,
weight_decay=1e-1)
image_param.save(filename='test_custom_image_param.jpg')
self.assertTrue(os.path.exists('test_custom_image_param.jpg'))
self.assertTrue(isinstance(image_param, custom_image_param),
'should be an instance of auto_image_param')
self.assertTrue(isinstance(image_param.__array__(), np.ndarray))
self.assertTrue(isinstance(image_param.to_hwc_tensor(), torch.Tensor),
'should be a torch.Tensor')
self.assertTrue(isinstance(image_param.to_chw_tensor(), torch.Tensor),
'should be a torch.Tensor')
os.remove('test_custom_image_param.jpg')
def run(config, img, title=None):
if not dreamy_boi or not model:
setup_torch_dreams()
# get config
assert 'objective' in config, 'No objective to optimize in config'
c = EasyDict(config)
width, height = c.size if 'size' in c else (256, 256)
iters = c.iters if 'iters' in c else 150
lr = c.lr if 'lr' in c else 9e-3
rotate_degrees = c.rotate_degrees if 'rotate_degrees' in c else 15
scale_max = c.scale_max if 'scale_max' in c else 1.2
scale_min = c.scale_min if 'scale_min' in c else 0.5
translate = c.translate if 'translate' in c else (0.2, 0.2)
weight_decay = c.weight_decay if 'weight_decay' in c else 1e-2
grad_clip = c.grad_clip if 'grad_clip' in c else 1
layer = get_layer(c.objective['layer'])
channel = c.objective['channel']
# set target
layers_to_use = [layer]
my_custom_func = make_custom_func(layer_number=0, channel_number=channel)
if img is None:
input_param = auto_image_param(height=height,
width=width,
device='cuda',
standard_deviation=0.01)
else:
if isinstance(img, str):
img = image.load_image(img)
img = image.resize(img, (width, height))
img = torch.tensor(np.array(img) / 255).permute(-1, 0, 1).unsqueeze(0)
input_param = custom_image_param(image=img, device='cuda')
# run torch_dreams
output_param = dreamy_boi.render(image_parameter=input_param,
layers=layers_to_use,
custom_func=my_custom_func,
width=width,
height=height,
iters=iters,
lr=lr,
rotate_degrees=rotate_degrees,
scale_max=scale_max,
scale_min=scale_min,
translate_x=translate[0],
translate_y=translate[1],
weight_decay=weight_decay,
grad_clip=grad_clip)
output_image = (255 * np.clip(np.array(output_param), 0, 1)).astype(
np.uint8)
return output_image
def test_caricature(self):
model = models.resnet18(pretrained=True)
dreamy_boi = dreamer(model, device='cpu')
param = custom_image_param(image='images/sample_small.jpg',
device='cpu')
image_tensor = param.to_nchw_tensor()
param = dreamy_boi.caricature(input_tensor=image_tensor,
layers=[model.layer3],
power=1.0,
iters=5)
self.assertTrue(isinstance(param, auto_image_param),
'should be an auto_image_param')