stadv package

stadv.layers module

stadv.layers.flow_st(images, flows, data_format='NHWC')[source]

Flow-based spatial transformation of images. See Eq. (1) in Xiao et al. (arXiv:1801.02612).

Parameters:
  • images (tf.Tensor) – images of shape (B, H, W, C) or (B, C, H, W) depending on data_format.
  • flows (tf.Tensor) – flows of shape (B, 2, H, W), where the second dimension indicates the dimension on which the pixel shift is applied.
  • data_format (str) – 'NHWC' or 'NCHW' depending on the format of the input images and the desired output.
Returns:

tf.Tensor of the same shape and type as images.

stadv.losses module

stadv.losses.adv_loss(unscaled_logits, targets, kappa=None)[source]

Computes the adversarial loss. It was first suggested by Carlini and Wagner (arXiv:1608.04644). See also Eq. (3) in Xiao et al. (arXiv:1801.02612).

Parameters:
  • unscaled_logits (tf.Tensor) – logits of shape (B, K), where K is the number of input classes.
  • targets (tf.Tensor) – 1-D integer-encoded targets of length B with value corresponding to the class ID.
  • kappa (tf.Tensor) – confidence parameter, see Carlini and Wagner (arXiv:1608.04644). Defaults to 0.
Returns:

1-D tf.Tensor of length B of the same type as unscaled_logits.
stadv.losses.flow_loss(flows, padding_mode='SYMMETRIC', epsilon=1e-08)[source]

Computes the flow loss designed to “enforce the locally smooth spatial transformation perturbation”. See Eq. (4) in Xiao et al. (arXiv:1801.02612).

Parameters:
  • flows (tf.Tensor) – flows of shape (B, 2, H, W), where the second dimension indicates the dimension on which the pixel shift is applied.
  • padding_mode (str) –

    how to perform padding of the boundaries of the images. The value should be compatible with the mode argument of tf.pad. Expected values are:

    • 'SYMMETRIC': symmetric padding so as to not penalize a significant flow at the boundary of the images;
    • 'CONSTANT': 0-padding of the boundaries so as to enforce a small flow at the boundary of the images.
  • epsilon (float) – small value added to the argument of tf.sqrt to prevent NaN gradients when the argument is zero.
Returns:

1-D tf.Tensor of length B of the same type as flows.

stadv.optimization module

stadv.optimization.lbfgs(loss, flows, flows_x0, feed_dict=None, fmin_l_bfgs_b_extra_kwargs=None, sess=None)[source]

Optimize a given loss with (SciPy’s external) L-BFGS-B optimizer. It can be used to solve the optimization problem of Eq. (2) in Xiao et al. (arXiv:1801.02612). See the documentation on scipy.optimize.fmin_l_bfgs_b for reference on the optimizer.

Parameters:
  • loss (tf.Tensor) – loss (can be of any shape).
  • flows (tf.Tensor) – flows of shape (B, 2, H, W), where the second dimension indicates the dimension on which the pixel shift is applied.
  • flows_x0 (np.ndarray) – Initial guess for the flows. If the input is not of type np.ndarray, it will be converted as such if possible.
  • feed_dict (dict) – feed dictionary to the tf.run operation (for everything which might be needed to execute the graph beyond the input flows).
  • fmin_l_bfgs_b_extra_kwargs (dict) – extra arguments to scipy.optimize.fmin_l_bfgs_b (e.g. for modifying the stopping condition).
  • sess (tf.Session) – session within which the graph should be executed. If not provided a new session will be started.
Returns:

Dictionary with keys 'flows' (np.ndarray, estimated flows of the minimum), 'loss' (float, value of loss at the minimum), and 'info' (dict, information summary as returned by scipy.optimize.fmin_l_bfgs_b).