shap--shap/shap/explainers/_deep/deep_tf.py

import warnings
from collections.abc import Callable

import numpy as np
import tensorflow as tf
from packaging import version
from tensorflow.python.eager import backprop as tf_backprop
from tensorflow.python.eager import execute as tf_execute
from tensorflow.python.framework import (
    ops as tf_ops,
)
from tensorflow.python.ops import (
    gradients_impl as tf_gradients_impl,
)

from ...utils._exceptions import DimensionError
from .._explainer import Explainer
from ..tf_utils import _get_graph, _get_model_inputs, _get_model_output, _get_session
from .deep_utils import _check_additivity

if not hasattr(tf_gradients_impl, "_IsBackpropagatable"):
    from tensorflow.python.ops import gradients_util as tf_gradients_impl


def custom_record_gradient(op_name, inputs, attrs, results):
    """This overrides tensorflow.python.eager.backprop._record_gradient.

    We need to override _record_gradient in order to get gradient backprop to
    get called for ResourceGather operations. In order to make this work we
    temporarily "lie" about the input type to prevent the node from getting
    pruned from the gradient backprop process. We then reset the type directly
    afterwards back to what it was (an integer type).
    """
    reset_input = False
    if op_name == "ResourceGather" and inputs[1].dtype == tf.int32:
        inputs[1].__dict__["_dtype"] = tf.float32
        reset_input = True
    try:
        out = tf_backprop._record_gradient("shap_" + op_name, inputs, attrs, results)
    except AttributeError:
        out = tf_backprop.record_gradient("shap_" + op_name, inputs, attrs, results)

    if reset_input:
        inputs[1].__dict__["_dtype"] = tf.int32

    return out


class TFDeep(Explainer):
    """Using tf.gradients to implement the backpropagation was
    inspired by the gradient-based implementation approach proposed by Ancona et al, ICLR 2018. Note
    that this package does not currently use the reveal-cancel rule for ReLu units proposed in DeepLIFT.
    """

    def __init__(self, model, data, session=None, learning_phase_flags=None):
        """An explainer object for a deep model using a given background dataset.

        Note that the complexity of the method scales linearly with the number of background data
        samples. Passing the entire training dataset as `data` will give very accurate expected
        values, but will be computationally expensive. The variance of the expectation estimates scales by
        roughly 1/sqrt(N) for N background data samples. So 100 samples will give a good estimate,
        and 1000 samples a very good estimate of the expected values.

        Parameters
        ----------
        model : tf.keras.Model or (input : [tf.Operation], output : tf.Operation)
            A keras model object or a pair of TensorFlow operations (or a list and an op) that
            specifies the input and output of the model to be explained. Note that SHAP values
            are specific to a single output value, so you get an explanation for each element of
            the output tensor (which must be a flat rank one vector).

        data : [numpy.array] or [pandas.DataFrame] or function
            The background dataset to use for integrating out features. DeepExplainer integrates
            over all these samples for each explanation. The data passed here must match the input
            operations given to the model. If a function is supplied, it must be a function that
            takes a particular input example and generates the background dataset for that example
        session : None or tensorflow.Session
            The TensorFlow session that has the model we are explaining. If None is passed then
            we do our best to find the right session, first looking for a keras session, then
            falling back to the default TensorFlow session.

        learning_phase_flags : None or list of tensors
            If you have your own custom learning phase flags pass them here. When explaining a prediction
            we need to ensure we are not in training mode, since this changes the behavior of ops like
            batch norm or dropout. If None is passed then we look for tensors in the graph that look like
            learning phase flags (this works for Keras models). Note that we assume all the flags should
            have a value of False during predictions (and hence explanations).

        """
        if version.parse(tf.__version__) < version.parse("1.4.0"):
            warnings.warn("Your TensorFlow version is older than 1.4.0 and not supported.")

        if version.parse(tf.__version__) >= version.parse("2.4.0"):
            warnings.warn(
                "Your TensorFlow version is newer than 2.4.0 and so graph support has been removed in eager mode and some static graphs may not be supported. See PR #1483 for discussion."
            )

        # determine the model inputs and outputs
        self.model_inputs = _get_model_inputs(model)
        self.model_output = _get_model_output(model)
        assert not isinstance(self.model_output, list), "The model output to be explained must be a single tensor!"
        assert len(self.model_output.shape) < 3, "The model output must be a vector or a single value!"
        self.multi_output = True
        if len(self.model_output.shape) == 1:
            self.multi_output = False

        if tf.executing_eagerly():
            if isinstance(model, (list, tuple)):
                assert len(model) == 2, "When a tuple is passed it must be of the form (inputs, outputs)"
                from tensorflow import keras

                self.model = keras.Model(model[0], model[1])
            else:
                self.model = model

        # check if we have multiple inputs
        self.multi_input = True
        if not isinstance(self.model_inputs, list) or len(self.model_inputs) == 1:
            self.multi_input = False
            if not isinstance(self.model_inputs, list):
                self.model_inputs = [self.model_inputs]
        if not isinstance(data, list) and (hasattr(data, "__call__") is False):
            data = [data]
        self.data = data

        self._vinputs = {}  # used to track what op inputs depends on the model inputs
        self.orig_grads = {}

        if not tf.executing_eagerly():
            self.session = _get_session(session)

        self.graph = _get_graph(self)

        # if no learning phase flags were given we go looking for them
        # ...this will catch the one that keras uses
        # we need to find them since we want to make sure learning phase flags are set to False
        if learning_phase_flags is None:
            self.learning_phase_ops = []
            for op in self.graph.get_operations():
                if "learning_phase" in op.name and op.type == "Const" and len(op.outputs[0].shape) == 0:
                    if op.outputs[0].dtype == tf.bool:
                        self.learning_phase_ops.append(op)
            self.learning_phase_flags = [op.outputs[0] for op in self.learning_phase_ops]
        else:
            self.learning_phase_ops = [t.op for t in learning_phase_flags]

        # save the expected output of the model
        # if self.data is a function, set self.expected_value to None
        if hasattr(self.data, "__call__"):
            self.expected_value = None
        else:
            if self.data[0].shape[0] > 5000:
                warnings.warn(
                    "You have provided over 5k background samples! For better performance consider using smaller random sample."
                )
            if not tf.executing_eagerly():
                self.expected_value = self.run(self.model_output, self.model_inputs, self.data).mean(0)
            else:
                # if type(self.model)is tuple:
                #    self.fModel(cnn.inputs, cnn.get_layer(theNameYouWant).outputs)
                self.expected_value = tf.reduce_mean(self.model(self.data), 0)

        if not tf.executing_eagerly():
            self._init_between_tensors(self.model_output.op, self.model_inputs)

        # make a blank array that will get lazily filled in with the SHAP value computation
        # graphs for each output. Lazy is important since if there are 1000 outputs and we
        # only explain the top 5 it would be a waste to build graphs for the other 995
        if not self.multi_output:
            self.phi_symbolics = [None]
        else:
            model_output_shape = self.model_output.shape
            if isinstance(model_output_shape, tuple):
                noutputs = model_output_shape[1]
            else:
                noutputs = model_output_shape.as_list()[1]
            if noutputs is not None:
                self.phi_symbolics = [None for _ in range(noutputs)]
            else:
                raise DimensionError(
                    "The model output tensor to be explained cannot have a static shape in dim 1 of None!"
                )

    def _get_model_output(self, model):
        if len(model.layers[-1]._inbound_nodes) == 0:
            if len(model.outputs) > 1:
                warnings.warn("Only one model output supported.")
            return model.outputs[0]
        else:
            return model.layers[-1].output

    def _init_between_tensors(self, out_op, model_inputs):
        # find all the operations in the graph between our inputs and outputs
        tensor_blacklist = tensors_blocked_by_false(self.learning_phase_ops)  # don't follow learning phase branches
        dependence_breakers = [k for k in op_handlers if op_handlers[k] == break_dependence]
        back_ops = backward_walk_ops([out_op], tensor_blacklist, dependence_breakers)
        start_ops = []
        for minput in model_inputs:
            for op in minput.consumers():
                start_ops.append(op)
        self.between_ops = forward_walk_ops(start_ops, tensor_blacklist, dependence_breakers, within_ops=back_ops)

        # note all the tensors that are on the path between the inputs and the output
        self.between_tensors = {}
        for op in self.between_ops:
            for t in op.outputs:
                self.between_tensors[t.name] = True
        for t in model_inputs:
            self.between_tensors[t.name] = True

        # save what types are being used
        self.used_types = {}
        for op in self.between_ops:
            self.used_types[op.type] = True

    def _variable_inputs(self, op):
        """Return which inputs of this operation are variable (i.e. depend on the model inputs)."""
        if op not in self._vinputs:
            out = np.zeros(len(op.inputs), dtype=bool)
            for i, t in enumerate(op.inputs):
                out[i] = t.name in self.between_tensors
            self._vinputs[op] = out
        return self._vinputs[op]

    def phi_symbolic(self, i):
        """Get the SHAP value computation graph for a given model output."""
        if self.phi_symbolics[i] is None:
            if not tf.executing_eagerly():

                def anon():
                    out = self.model_output[:, i] if self.multi_output else self.model_output
                    return tf.gradients(out, self.model_inputs)

                self.phi_symbolics[i] = self.execute_with_overridden_gradients(anon)
            else:
                if version.parse(tf.__version__) < version.parse("2.16.0"):
                    # TODO: set a deprecation warning for this
                    @tf.function
                    def grad_graph(shap_rAnD):
                        phase = tf.keras.backend.learning_phase()
                        tf.keras.backend.set_learning_phase(0)

                        with tf.GradientTape(watch_accessed_variables=False) as tape:
                            tape.watch(shap_rAnD)
                            out = self.model(shap_rAnD)
                            if self.multi_output:
                                out = out[:, i]

                        self._init_between_tensors(out.op, shap_rAnD)
                        x_grad = tape.gradient(out, shap_rAnD)
                        tf.keras.backend.set_learning_phase(phase)
                        return x_grad

                else:

                    @tf.function
                    def grad_graph(shap_rAnD):
                        with tf.GradientTape(watch_accessed_variables=False) as tape:
                            tape.watch(shap_rAnD)
                            out = self.model(shap_rAnD, training=False)
                            if self.multi_output:
                                out = out[:, i]

                        self._init_between_tensors(out.op, shap_rAnD)
                        x_grad = tape.gradient(out, shap_rAnD)
                        return x_grad

                self.phi_symbolics[i] = grad_graph

        return self.phi_symbolics[i]

    def shap_values(self, X, ranked_outputs=None, output_rank_order="max", check_additivity=True):
        # check if we have multiple inputs
        if not self.multi_input:
            if isinstance(X, list) and len(X) != 1:
                raise ValueError("Expected a single tensor as model input!")
            elif not isinstance(X, list):
                X = [X]
        else:
            if not isinstance(X, list):
                raise TypeError("Expected a list of model inputs!")
        if len(self.model_inputs) != len(X):
            raise ValueError(
                f"Number of model inputs ({len(self.model_inputs)}) does not match the number given ({len(X)})!"
            )

        # rank and determine the model outputs that we will explain
        if ranked_outputs is not None and self.multi_output:
            if not tf.executing_eagerly():
                model_output_values = self.run(self.model_output, self.model_inputs, X)
            else:
                model_output_values = self.model(X)

            if output_rank_order == "max":
                model_output_ranks = np.argsort(-model_output_values)
            elif output_rank_order == "min":
                model_output_ranks = np.argsort(model_output_values)
            elif output_rank_order == "max_abs":
                model_output_ranks = np.argsort(np.abs(model_output_values))
            else:
                emsg = "output_rank_order must be max, min, or max_abs!"
                raise ValueError(emsg)
            model_output_ranks = model_output_ranks[:, :ranked_outputs]
        else:
            model_output_ranks = np.tile(np.arange(len(self.phi_symbolics)), (X[0].shape[0], 1))

        # compute the attributions
        output_phis = []
        for i in range(model_output_ranks.shape[1]):
            phis = []
            for k in range(len(X)):
                phis.append(np.zeros(X[k].shape))
            for j in range(X[0].shape[0]):
                if hasattr(self.data, "__call__"):
                    bg_data = self.data([X[t][j] for t in range(len(X))])
                    if not isinstance(bg_data, list):
                        bg_data = [bg_data]
                else:
                    bg_data = self.data

                # tile the inputs to line up with the background data samples
                tiled_X = [
                    np.tile(X[t][j : j + 1], (bg_data[t].shape[0],) + tuple([1 for k in range(len(X[t].shape) - 1)]))
                    for t in range(len(X))
                ]

                # we use the first sample for the current sample and the rest for the references
                joint_input = [np.concatenate([tiled_X[t], bg_data[t]], 0) for t in range(len(X))]

                # run attribution computation graph
                feature_ind = model_output_ranks[j, i]
                sample_phis = self.run(self.phi_symbolic(feature_ind), self.model_inputs, joint_input)

                # assign the attributions to the right part of the output arrays
                for t in range(len(X)):
                    phis[t][j] = (sample_phis[t][bg_data[t].shape[0] :] * (X[t][j] - bg_data[t])).mean(0)

            output_phis.append(phis[0] if not self.multi_input else phis)

        # check that the SHAP values sum up to the model output
        if check_additivity:
            if not tf.executing_eagerly():
                model_output = self.run(self.model_output, self.model_inputs, X)
            else:
                model_output = self.model(X)

            _check_additivity(self, model_output, output_phis)

        if isinstance(output_phis, list):
            # in this case we have multiple inputs and potentially multiple outputs
            if isinstance(output_phis[0], list):
                output_phis = [np.stack([phi[i] for phi in output_phis], axis=-1) for i in range(len(output_phis[0]))]
            # multiple outputs case
            else:
                output_phis = np.stack(output_phis, axis=-1)
        if ranked_outputs is not None:
            return output_phis, model_output_ranks
        else:
            return output_phis

    def run(self, out, model_inputs, X):
        """Runs the model while also setting the learning phase flags to False."""
        if not tf.executing_eagerly():
            feed_dict = dict(zip(model_inputs, X))
            for t in self.learning_phase_flags:
                feed_dict[t] = False
            return self.session.run(out, feed_dict)
        else:

            def anon():
                tf_execute.record_gradient = custom_record_gradient

                # build inputs that are correctly shaped, typed, and tf-wrapped
                inputs = []
                for i in range(len(X)):
                    shape = list(self.model_inputs[i].shape)
                    shape[0] = -1
                    data = X[i].reshape(shape)
                    v = tf.constant(data, dtype=self.model_inputs[i].dtype)
                    inputs.append(v)
                final_out = out(inputs)
                try:
                    tf_execute.record_gradient = tf_backprop._record_gradient
                except AttributeError:
                    tf_execute.record_gradient = tf_backprop.record_gradient

                return final_out

            return self.execute_with_overridden_gradients(anon)

    def custom_grad(self, op, *grads):
        """Passes a gradient op creation request to the correct handler."""
        type_name = op.type[5:] if op.type.startswith("shap_") else op.type
        out = op_handlers[type_name](self, op, *grads)  # we cut off the shap_ prefix before the lookup
        return out

    def execute_with_overridden_gradients(self, f):
        # replace the gradients for all the non-linear activations
        # we do this by hacking our way into the registry (TODO: find a public API for this if it exists)
        reg = tf_ops._gradient_registry._registry
        ops_not_in_registry = ["TensorListReserve"]
        # NOTE: location_tag taken from tensorflow source for None type ops
        location_tag = ("UNKNOWN", "UNKNOWN", "UNKNOWN", "UNKNOWN", "UNKNOWN")
        # TODO: unclear why some ops are not in the registry with TF 2.0 like TensorListReserve
        for non_reg_ops in ops_not_in_registry:
            reg[non_reg_ops] = {"type": None, "location": location_tag}
        for n in op_handlers:
            if n in reg:
                self.orig_grads[n] = reg[n]["type"]
                reg["shap_" + n] = {"type": self.custom_grad, "location": reg[n]["location"]}
                reg[n]["type"] = self.custom_grad

        # In TensorFlow 1.10 they started pruning out nodes that they think can't be backpropped
        # unfortunately that includes the index of embedding layers so we disable that check here
        if hasattr(tf_gradients_impl, "_IsBackpropagatable"):
            orig_IsBackpropagatable = tf_gradients_impl._IsBackpropagatable
            tf_gradients_impl._IsBackpropagatable = lambda tensor: True

        # define the computation graph for the attribution values using a custom gradient-like computation
        try:
            out = f()
        finally:
            # reinstate the backpropagatable check
            if hasattr(tf_gradients_impl, "_IsBackpropagatable"):
                tf_gradients_impl._IsBackpropagatable = orig_IsBackpropagatable

            # restore the original gradient definitions
            for n in op_handlers:
                if n in reg:
                    del reg["shap_" + n]
                    reg[n]["type"] = self.orig_grads[n]
            for non_reg_ops in ops_not_in_registry:
                del reg[non_reg_ops]
        if not tf.executing_eagerly():
            return out
        else:
            return [v.numpy() for v in out]


def tensors_blocked_by_false(ops):
    """Follows a set of ops assuming their value is False and find blocked Switch paths.

    This is used to prune away parts of the model graph that are only used during the training
    phase (like dropout, batch norm, etc.).
    """
    blocked = []

    def recurse(op):
        if op.type == "Switch":
            blocked.append(op.outputs[1])  # the true path is blocked since we assume the ops we trace are False
        else:
            for out in op.outputs:
                for c in out.consumers():
                    recurse(c)

    for op in ops:
        recurse(op)

    return blocked


def backward_walk_ops(start_ops, tensor_blacklist, op_type_blacklist):
    found_ops = []
    op_stack = [op for op in start_ops]
    while len(op_stack) > 0:
        op = op_stack.pop()
        if op.type not in op_type_blacklist and op not in found_ops:
            found_ops.append(op)
            for input in op.inputs:
                if input not in tensor_blacklist:
                    op_stack.append(input.op)
    return found_ops


def forward_walk_ops(start_ops, tensor_blacklist, op_type_blacklist, within_ops):
    found_ops = []
    op_stack = [op for op in start_ops]
    while len(op_stack) > 0:
        op = op_stack.pop()
        if op.type not in op_type_blacklist and op in within_ops and op not in found_ops:
            found_ops.append(op)
            for out in op.outputs:
                if out not in tensor_blacklist:
                    for c in out.consumers():
                        op_stack.append(c)
    return found_ops


def softmax(explainer, op, *grads):
    """Just decompose softmax into its components and recurse, we can handle all of them :)

    We assume the 'axis' is the last dimension because the TF codebase swaps the 'axis' to
    the last dimension before the softmax op if 'axis' is not already the last dimension.
    We also don't subtract the max before tf.exp for numerical stability since that might
    mess up the attributions and it seems like TensorFlow doesn't define softmax that way
    (according to the docs)
    """
    in0 = op.inputs[0]
    in0_max = tf.reduce_max(in0, axis=-1, keepdims=True, name="in0_max")
    in0_centered = in0 - in0_max
    evals = tf.exp(in0_centered, name="custom_exp")
    rsum = tf.reduce_sum(evals, axis=-1, keepdims=True)
    div = evals / rsum

    # mark these as in-between the inputs and outputs
    for op in [evals.op, rsum.op, div.op, in0_centered.op]:
        for t in op.outputs:
            if t.name not in explainer.between_tensors:
                explainer.between_tensors[t.name] = False

    out = tf.gradients(div, in0_centered, grad_ys=grads[0])[0]

    # remove the names we just added
    for op in [evals.op, rsum.op, div.op, in0_centered.op]:
        for t in op.outputs:
            if explainer.between_tensors[t.name] is False:
                del explainer.between_tensors[t.name]

    # rescale to account for our shift by in0_max (which we did for numerical stability)
    xin0, rin0 = tf.split(in0, 2)
    xin0_centered, rin0_centered = tf.split(in0_centered, 2)
    delta_in0 = xin0 - rin0
    dup0 = [2] + [1 for i in delta_in0.shape[1:]]
    return tf.where(
        tf.tile(tf.abs(delta_in0), dup0) < 1e-6, out, out * tf.tile((xin0_centered - rin0_centered) / delta_in0, dup0)
    )


def maxpool(explainer, op, *grads):
    xin0, rin0 = tf.split(op.inputs[0], 2)
    xout, rout = tf.split(op.outputs[0], 2)
    delta_in0 = xin0 - rin0
    dup0 = [2] + [1 for i in delta_in0.shape[1:]]
    cross_max = tf.maximum(xout, rout)
    diffs = tf.concat([cross_max - rout, xout - cross_max], 0)
    if op.type.startswith("shap_"):
        op.type = op.type[5:]
    xmax_pos, rmax_pos = tf.split(explainer.orig_grads[op.type](op, grads[0] * diffs), 2)
    return tf.tile(
        tf.where(tf.abs(delta_in0) < 1e-7, tf.zeros_like(delta_in0), (xmax_pos + rmax_pos) / delta_in0), dup0
    )


def gather(explainer, op, *grads):
    # params = op.inputs[0]
    indices = op.inputs[1]
    # axis = op.inputs[2]
    var = explainer._variable_inputs(op)
    if var[1] and not var[0]:
        assert len(indices.shape) == 2, "Only scalar indices supported right now in GatherV2!"

        xin1, rin1 = tf.split(tf.cast(op.inputs[1], tf.float32), 2)
        xout, rout = tf.split(op.outputs[0], 2)
        dup_in1 = [2] + [1 for i in xin1.shape[1:]]
        dup_out = [2] + [1 for i in xout.shape[1:]]
        delta_in1_t = tf.tile(xin1 - rin1, dup_in1)
        out_sum = tf.reduce_sum(
            grads[0] * tf.tile(xout - rout, dup_out), list(range(len(indices.shape), len(grads[0].shape)))
        )
        if op.type == "ResourceGather":
            return [None, tf.where(tf.abs(delta_in1_t) < 1e-6, tf.zeros_like(delta_in1_t), out_sum / delta_in1_t)]
        return [None, tf.where(tf.abs(delta_in1_t) < 1e-6, tf.zeros_like(delta_in1_t), out_sum / delta_in1_t), None]
    elif var[0] and not var[1]:
        if op.type.startswith("shap_"):
            op.type = op.type[5:]
        return [explainer.orig_grads[op.type](op, grads[0]), None]  # linear in this case
    else:
        raise ValueError("Axis not yet supported to be varying for gather op!")


def linearity_1d_nonlinearity_2d(input_ind0, input_ind1, op_func):
    def handler(explainer, op, *grads):
        var = explainer._variable_inputs(op)
        if var[input_ind0] and not var[input_ind1]:
            return linearity_1d_handler(input_ind0, explainer, op, *grads)
        elif var[input_ind1] and not var[input_ind0]:
            return linearity_1d_handler(input_ind1, explainer, op, *grads)
        elif var[input_ind0] and var[input_ind1]:
            return nonlinearity_2d_handler(input_ind0, input_ind1, op_func, explainer, op, *grads)
        else:
            return [None for _ in op.inputs]  # no inputs vary, we must be hidden by a switch function

    return handler


def nonlinearity_1d_nonlinearity_2d(input_ind0: int, input_ind1: int, op_func: Callable) -> Callable:
    def handler(explainer, op, *grads):
        var = explainer._variable_inputs(op)
        if var[input_ind0] and not var[input_ind1]:
            return nonlinearity_1d_handler(input_ind0, explainer, op, *grads)
        elif var[input_ind1] and not var[input_ind0]:
            return nonlinearity_1d_handler(input_ind1, explainer, op, *grads)
        elif var[input_ind0] and var[input_ind1]:
            return nonlinearity_2d_handler(input_ind0, input_ind1, op_func, explainer, op, *grads)
        else:
            return [None for _ in op.inputs]  # no inputs vary, we must be hidden by a switch function

    return handler


def nonlinearity_1d(input_ind):
    def handler(explainer, op, *grads):
        return nonlinearity_1d_handler(input_ind, explainer, op, *grads)

    return handler


def nonlinearity_1d_handler(input_ind, explainer, op, *grads):
    # make sure only the given input varies
    op_inputs = op.inputs
    if op_inputs is None:
        op_inputs = op.outputs[0].op.inputs

    for i in range(len(op_inputs)):
        if i != input_ind:
            assert not explainer._variable_inputs(op)[i], str(i) + "th input to " + op.name + " cannot vary!"

    xin0, rin0 = tf.split(op_inputs[input_ind], 2)
    xout, rout = tf.split(op.outputs[input_ind], 2)
    delta_in0 = xin0 - rin0
    if delta_in0.shape is None:
        dup0 = [2, 1]
    else:
        dup0 = [2] + [1 for i in delta_in0.shape[1:]]
    out = [None for _ in op_inputs]
    if op.type.startswith("shap_"):
        op.type = op.type[5:]
    orig_grad = explainer.orig_grads[op.type](op, grads[0])
    out[input_ind] = tf.where(
        tf.tile(tf.abs(delta_in0), dup0) < 1e-6,
        orig_grad[input_ind] if len(op_inputs) > 1 else orig_grad,
        grads[0] * tf.tile((xout - rout) / delta_in0, dup0),
    )
    return out


def nonlinearity_2d_handler(input_ind0, input_ind1, op_func, explainer, op, *grads):
    if not (input_ind0 == 0 and input_ind1 == 1):
        emsg = "TODO: Can't yet handle double inputs that are not first!"
        raise Exception(emsg)
    xout, rout = tf.split(op.outputs[0], 2)
    in0 = op.inputs[input_ind0]
    in1 = op.inputs[input_ind1]
    xin0, rin0 = tf.split(in0, 2)
    xin1, rin1 = tf.split(in1, 2)
    delta_in0 = xin0 - rin0
    delta_in1 = xin1 - rin1
    dup0 = [2] + [1 for i in delta_in0.shape[1:]]
    out10 = op_func(xin0, rin1)
    out01 = op_func(rin0, xin1)
    out11, out00 = xout, rout
    out0 = 0.5 * (out11 - out01 + out10 - out00)
    out0 = grads[0] * tf.tile(out0 / delta_in0, dup0)
    out1 = 0.5 * (out11 - out10 + out01 - out00)
    out1 = grads[0] * tf.tile(out1 / delta_in1, dup0)

    # Avoid divide by zero nans
    out0 = tf.where(tf.abs(tf.tile(delta_in0, dup0)) < 1e-7, tf.zeros_like(out0), out0)
    out1 = tf.where(tf.abs(tf.tile(delta_in1, dup0)) < 1e-7, tf.zeros_like(out1), out1)

    # see if due to broadcasting our gradient shapes don't match our input shapes
    if np.any(np.array(out1.shape) != np.array(in1.shape)):
        broadcast_index = np.where(np.array(out1.shape) != np.array(in1.shape))[0][0]
        out1 = tf.reduce_sum(out1, axis=broadcast_index, keepdims=True)
    elif np.any(np.array(out0.shape) != np.array(in0.shape)):
        broadcast_index = np.where(np.array(out0.shape) != np.array(in0.shape))[0][0]
        out0 = tf.reduce_sum(out0, axis=broadcast_index, keepdims=True)

    return [out0, out1]


def linearity_1d(input_ind):
    def handler(explainer, op, *grads):
        return linearity_1d_handler(input_ind, explainer, op, *grads)

    return handler


def linearity_1d_handler(input_ind, explainer, op, *grads):
    # make sure only the given input varies (negative means only that input cannot vary, and is measured from the end of the list)
    for i in range(len(op.inputs)):
        if i != input_ind:
            assert not explainer._variable_inputs(op)[i], str(i) + "th input to " + op.name + " cannot vary!"
    if op.type.startswith("shap_"):
        op.type = op.type[5:]
    return explainer.orig_grads[op.type](op, *grads)


def linearity_with_excluded(input_inds):
    def handler(explainer, op, *grads):
        return linearity_with_excluded_handler(input_inds, explainer, op, *grads)

    return handler


def linearity_with_excluded_handler(input_inds, explainer, op, *grads):
    # make sure the given inputs don't vary (negative is measured from the end of the list)
    for i in range(len(op.inputs)):
        if i in input_inds or i - len(op.inputs) in input_inds:
            assert not explainer._variable_inputs(op)[i], str(i) + "th input to " + op.name + " cannot vary!"
    if op.type.startswith("shap_"):
        op.type = op.type[5:]
    return explainer.orig_grads[op.type](op, *grads)


def passthrough(explainer, op, *grads):
    if op.type.startswith("shap_"):
        op.type = op.type[5:]
    return explainer.orig_grads[op.type](op, *grads)


def break_dependence(explainer, op, *grads):
    """This function name is used to break attribution dependence in the graph traversal.

    These operation types may be connected above input data values in the graph but their outputs
    don't depend on the input values (for example they just depend on the shape).
    """
    return [None for _ in op.inputs]


op_handlers: dict[str, Callable] = {}

# ops that are always linear
op_handlers["Identity"] = passthrough
op_handlers["StridedSlice"] = passthrough
op_handlers["Squeeze"] = passthrough
op_handlers["ExpandDims"] = passthrough
op_handlers["Pack"] = passthrough
op_handlers["BiasAdd"] = passthrough
op_handlers["Unpack"] = passthrough
op_handlers["Add"] = passthrough
op_handlers["AddV2"] = passthrough
op_handlers["Sub"] = passthrough
op_handlers["Merge"] = passthrough
op_handlers["Sum"] = passthrough
op_handlers["Mean"] = passthrough
op_handlers["Cast"] = passthrough
op_handlers["Transpose"] = passthrough
op_handlers["Enter"] = passthrough
op_handlers["Exit"] = passthrough
op_handlers["NextIteration"] = passthrough
op_handlers["Tile"] = passthrough
op_handlers["TensorArrayScatterV3"] = passthrough
op_handlers["TensorArrayReadV3"] = passthrough
op_handlers["TensorArrayWriteV3"] = passthrough


# ops that don't pass any attributions to their inputs
op_handlers["Shape"] = break_dependence
op_handlers["RandomUniform"] = break_dependence
op_handlers["ZerosLike"] = break_dependence
# op_handlers["StopGradient"] = break_dependence # this allows us to stop attributions when we want to (like softmax re-centering)

# ops that are linear and only allow a single input to vary
op_handlers["Reshape"] = linearity_1d(0)
op_handlers["Pad"] = linearity_1d(0)
op_handlers["ReverseV2"] = linearity_1d(0)
op_handlers["ConcatV2"] = linearity_with_excluded([-1])
op_handlers["Conv2D"] = linearity_1d(0)
op_handlers["Switch"] = linearity_1d(0)
op_handlers["AvgPool"] = linearity_1d(0)
op_handlers["FusedBatchNorm"] = linearity_1d(0)

# ops that are nonlinear and only allow a single input to vary
op_handlers["Relu"] = nonlinearity_1d(0)
op_handlers["Selu"] = nonlinearity_1d(0)
op_handlers["Elu"] = nonlinearity_1d(0)
op_handlers["Sigmoid"] = nonlinearity_1d(0)
op_handlers["Tanh"] = nonlinearity_1d(0)
op_handlers["Softplus"] = nonlinearity_1d(0)
op_handlers["Exp"] = nonlinearity_1d(0)
op_handlers["ClipByValue"] = nonlinearity_1d(0)
op_handlers["Rsqrt"] = nonlinearity_1d(0)
op_handlers["Square"] = nonlinearity_1d(0)
op_handlers["Max"] = nonlinearity_1d(0)

# ops that are nonlinear and allow two inputs to vary
op_handlers["SquaredDifference"] = nonlinearity_1d_nonlinearity_2d(0, 1, lambda x, y: (x - y) * (x - y))
op_handlers["Minimum"] = nonlinearity_1d_nonlinearity_2d(0, 1, lambda x, y: tf.minimum(x, y))
op_handlers["Maximum"] = nonlinearity_1d_nonlinearity_2d(0, 1, lambda x, y: tf.maximum(x, y))

# ops that allow up to two inputs to vary are are linear when only one input varies
op_handlers["Mul"] = linearity_1d_nonlinearity_2d(0, 1, lambda x, y: x * y)
op_handlers["RealDiv"] = linearity_1d_nonlinearity_2d(0, 1, lambda x, y: x / y)
op_handlers["MatMul"] = linearity_1d_nonlinearity_2d(0, 1, lambda x, y: tf.matmul(x, y))

# ops that need their own custom attribution functions
op_handlers["GatherV2"] = gather
op_handlers["ResourceGather"] = gather
op_handlers["MaxPool"] = maxpool
op_handlers["Softmax"] = softmax


# TODO items
# TensorArrayGatherV3
# Max
# TensorArraySizeV3
# Range