def predict_3D_patches(model, patches, image, N_extra=0, logger=None):
# Get box dim and image dim
d = patches.dim
i1, i2, i3 = image.shape[:3]
# Prepare reconstruction volume. Predictions will be summed in this volume.
recon = np.zeros(shape=(i1, i2, i3, model.n_classes), dtype=np.float32)
# Predict on base patches + N extra randomly
# sampled patches from the volume
for patch, (i, k, v), status in patches.get_patches_from(image, N_extra):
# Log the status of the generator
print(status, end="\r", flush=True)
# Predict on patch
pred = model.predict(reshape_add_axis(patch, im_dims=3))
# Add prediction to reconstructed volume
recon[i:i+d, k:k+d, v:v+d] += pred.squeeze()
print("")
# Normalize
recon /= np.sum(recon, axis=-1, keepdims=True)
return recon
def predict_3D_patches_binary(model, patches, image_id, N_extra=0, logger=None):
# Get box dim and image dim
d = patches.dim
i1, i2, i3 = patches.im_dim
# Prepare reconstruction volume. Predictions will be summed in this volume.
recon = np.zeros(shape=(i1, i2, i3, 2), dtype=np.uint32)
# Predict on base patches + N extra randomly
# sampled patches from the volume
for patch, (i, k, v), status in patches.get_patches_from(image_id, N_extra):
# Log the status of the generator
print(status, end="\r", flush=True)
# Predict on patch
pred = model.predict(reshape_add_axis(patch, im_dims=3)).squeeze()
mask = pred > 0.5
# Add prediction to reconstructed volume
recon[i:i+d, k:k+d, v:v+d, 0] += ~mask
recon[i:i+d, k:k+d, v:v+d, 1] += mask
print("")
total = np.sum(recon, axis=-1)
return (recon[..., 1] > (0.20 * total)).astype(np.uint8)