def _transform_and_predict(model,
x,
block_shape,
rotation,
translation=[0, 0, 0],
verbose=False):
"""Predict on rigidly transformed features.
The rigid transformation is applied to the volumes prior to prediction, and
the prediced labels are transformed with the inverse warp, so that they are
in the same space.
Parameters
----------
model: `tf.keras.Model`, model used for prediction.
x: 3D array, volume of features.
block_shape: tuple of length 3, shape of non-overlapping blocks to take
from the features. This also corresponds to the input of the model, not
including the batch or channel dimensions.
rotation: tuple of length 3, rotation angle in radians in each dimension.
translation: tuple of length 3, units of translation in each dimension.
verbose: bool, whether to print progress bar.
Returns
-------
Array of predictions with the same shape and in the same space as the
original input features.
"""
x = np.asarray(x).astype(np.float32)
affine = get_affine(x.shape, rotation=rotation, translation=translation)
inverse_affine = tf.linalg.inv(affine)
x_warped = warp(x, affine, order=1)
x_warped_blocks = to_blocks_numpy(x_warped, block_shape)
x_warped_blocks = x_warped_blocks[..., np.newaxis] # add grayscale channel
x_warped_blocks = standardize_numpy(x_warped_blocks)
y = model.predict(x_warped_blocks, batch_size=1, verbose=verbose)
n_classes = y.shape[-1]
if n_classes == 1:
y = y.squeeze(-1)
else:
# Usually, the argmax would be taken to get the class membership of
# each voxel, but if we get hard values, then we cannot average
# multiple predictions.
raise ValueError(
"This function is not compatible with multi-class predictions.")
y = from_blocks_numpy(y, x.shape)
y = warp(y, inverse_affine, order=0).numpy()
return y
def predict_from_array(inputs,
model,
block_shape,
batch_size=1,
normalizer=None,
n_samples=1,
return_variance=False,
return_entropy=False):
"""Return a prediction given a filepath and an ndarray of features.
Parameters
----------
inputs: ndarray, array of features.
model: `tf.keras.Model`, trained model.
block_shape: tuple of length 3, shape of sub-volumes on which to
predict.
batch_size: int, number of sub-volumes per batch for predictions.
normalizer: callable, function that accepts an ndarray and returns an
ndarray. Called before separating volume into blocks.
n_samples: The number of sampling. If set as 1, it will just return the
single prediction value. The default value is 1
return_variance: Boolean. If set True, it returns the running population
variance along with mean. Note, if the n_samples is smaller or equal to 1,
the variance will not be returned; instead it will return None
return_entropy: Boolean. If set True, it returns the running entropy.
along with mean.
Returns
-------
ndarray of predictions.
"""
if normalizer:
features = normalizer(inputs)
else:
features = inputs
if block_shape is not None:
features = to_blocks_numpy(features, block_shape=block_shape)
else:
features = features[None] # Add batch dimension.
# Add a dimension for single channel.
features = features[..., None]
# Predict per block to reduce memory consumption.
n_blocks = features.shape[0]
n_batches = math.ceil(n_blocks / batch_size)
if not return_variance and not return_entropy and n_samples == 1:
outputs = model.predict(features, batch_size=1, verbose=0)
if outputs.shape[-1] == 1:
# Binarize according to threshold.
outputs = outputs > 0.3
outputs = outputs.squeeze(-1)
# Nibabel doesn't like saving boolean arrays as Nifti.
outputs = outputs.astype(np.uint8)
else:
# Hard classes for multi-class segmentation.
outputs = np.argmax(outputs, -1)
outputs = from_blocks_numpy(outputs, output_shape=inputs.shape)
return outputs
raise NotImplementedError(
"Predicting from Bayesian nets is not implemented yet.")
means = np.zeros_like(features)
variances = np.zeros_like(features)
entropies = np.zeros_like(features)
progbar = tf.keras.utils.Progbar(n_batches)
progbar.update(0)
for j in range(0, n_blocks, batch_size):
this_x = features[j:j + batch_size]
new_prediction = model.predict(this_x, batch_size=1, verbose=0)
prev_mean = np.zeros_like(new_prediction['probabilities'])
curr_mean = new_prediction['probabilities']
M = np.zeros_like(new_prediction['probabilities'])
for n in range(1, n_samples):
new_prediction = model.predict(this_x)
prev_mean = curr_mean
curr_mean = prev_mean + (new_prediction['probabilities'] -
prev_mean) / float(n + 1)
M = M + np.multiply(prev_mean - new_prediction['probabilities'],
curr_mean - new_prediction['probabilities'])
means[j:j + batch_size] = np.argmax(curr_mean, axis=-1) # max mean
variances[j:j + batch_size] = np.sum(M / n_samples, axis=-1)
entropies[j:j + batch_size] = -np.sum(np.multiply(
np.log(curr_mean + 0.001), curr_mean),
axis=-1) # entropy
progbar.add(1)
total_means = from_blocks_numpy(means, output_shape=inputs.shape)
total_variance = from_blocks_numpy(variances, output_shape=inputs.shape)
total_entropy = from_blocks_numpy(entropies, output_shape=inputs.shape)
mean_var_voxels = np.mean(total_variance)
std_var_voxels = np.std(total_variance)
include_variance = ((n_samples > 1) and (return_variance))
if include_variance:
if return_entropy:
return total_means, total_variance, total_entropy
else:
return total_means, total_variance
else:
if return_entropy:
return total_means, total_entropy
else:
return total_means,