def predict_shape(sketch, iv3_model, mapping_model, gan_model):
sketch = np.tile(sketch, (1, 1, 3)).T
sketch = preprocess_input(sketch)
features = iv3_model.predict(sketch, batch_size=1, verbose=1)
latent_vec = mapping_model.predict(features, batch_size=1)[0]
latent_vec = torch.Tensor(latent_vec)
latent_vec = latent_vec.view(1, -1, 1, 1, 1)
fake = gan_model(latent_vec)
np_fake = fake.detach().numpy()
voxels = np.reshape(fake.detach().numpy(), (64, 64, 64))
voxels = downsample(voxels, 4, method='max')
return threshold(voxels)
def predict_nearest_shape(bgr_sketch,
NNmodel,
gan_model,
latent_vectors,
n_neighbors=5,
colors=[[255, 255, 255], [0, 255, 0], [255, 0, 0],
[0, 0, 255]]):
# load data
gan_features_path = r"data/samples_screenshot_BT.npy"
# process rgb sketch
contours = process_rgb_sketch(bgr_sketch,
colors=colors,
precision=0.00001,
smoothing=10,
resolution=200,
diff_threshold=30,
bspline_degree=1,
n_contours=1)
contours = process_contours(contours[:, :, 0, :])
contours[0, :, :] *= OUTLINE_IMPORTANCE
x = np.reshape(
contours,
(1, contours.shape[0] * contours.shape[1] * contours.shape[2]))
# predict nearest neighbors
closest = indices[0][0]
# generate model
latent_vec = latent_vectors[closest]
latent_vec = torch.Tensor(latent_vec)
latent_vec = latent_vec.view(1, -1, 1, 1, 1)
fake = gan_model(latent_vec)
np_fake = fake.detach().numpy()
voxels = np.reshape(fake.detach().numpy(), (64, 64, 64))
voxels = downsample(voxels, 2, method='mean')
return threshold(voxels, 0.4)
contours,
(1, contours.shape[0] * contours.shape[1] * contours.shape[2]))
# predict nearest neighbors
nneighbors = 1
distances, indices = nn_model.kneighbors(x)
closest = indices[0][:nneighbors]
distances = np.array([(1 / np.exp((x + 1)**2))
for x in range(nneighbors)]).reshape(1, -1)
print(closest, distances[0])
# generate model
latent_vecs = latent_vectors[closest]
latent_vec = np.sum(np.multiply(latent_vecs, (1 / distances).T),
axis=0) / np.sum(1 / distances)
latent_vec = torch.Tensor(latent_vec)
latent_vec = latent_vec.view(1, -1, 1, 1, 1)
fake = gan_model(latent_vec)
np_fake = fake.detach().numpy()
voxels = np.reshape(fake.detach().numpy(), (64, 64, 64))
voxels = downsample(voxels, 2, method='mean')
visualization(voxels,
0.3,
title=None,
uniform_size=0,
use_colormap=False,
angle=0.3)
cv2.destroyWindow('image')
# get sketch bottleneck features
iv3_input = (139, 139, 3)
images = np.empty((1, iv3_input[0], iv3_input[1], iv3_input[2]))
resized = cv2.resize(image.copy(), (iv3_input[0], iv3_input[1]))
resized = preprocess_input(resized)
images[0] = resized.copy()
# extract bottleneck features
features = iv3_model.predict(images, batch_size=1, verbose=1)
# predict latent vector
latent_vec = mapping_model.predict(features, batch_size=1)[0]
latent_vec = torch.Tensor(latent_vec)
latent_vec = latent_vec.view(1, -1, 1, 1, 1)
# generate shape
gan_model.eval()
fake = gan_model(latent_vec)
np_fake = fake.detach().numpy()
voxels = np.reshape(fake.detach().numpy(), (64, 64, 64))
voxels = downsample(voxels, 4, method='max')
visualization(voxels,
0.5,
title=None,
uniform_size=1,
use_colormap=False,
angle=0.3)