add test loss (pytorch#1073)

Summary:
Pull Request resolved: pytorch#1073

- For all `TracInCPBase` implementations, this adds an additional `test_loss_fn` initialization argument, which is the loss function to apply to test examples when computing the influence of a training example on a test example.  With this change,the influence score is a sum over terms for each checkpoint, where each term is the gradient of `loss_fn` for a given training example, multiplied with the gradient of `test_loss_fn` for a given test example. Before, `test_loss_fn` was assumed to be the same as `loss_fn`.
- checks regarding the reduction type of both `loss_fn` and `test_loss_fn` are now handled by helper functions `_check_tracincp_loss_fn` and `_check_tracincp_fast_loss_fn`.
- documentation is updated.  one detail: for `TracInCP`, we assume that `sample_wise_grads_per_batch` is applied to both `loss_fn` and `test_loss_fn` (if provided), and this is mentioned in the documentation.
- `test_tracin_regression.test_tracin_regression` is slightly modified - `DataInfluenceConstructor` now can explicitly pass in the same loss function for both `loss_fn` and `test_loss_fn` (done when `duplicate_loss_fn=True`). Doing so would have the same effect as not passing in `test_loss_fn`, so the original tests are also applied to the case when `duplicate_loss_fn=True`, as the expected behavior should be the same as before.
- a new test, `test_tracin_regression.test_tracin_constant_test_loss_fn` is added. For all implementations of `TracInCPBase`, it checks that if `test_loss_fn` is a constant loss function, the influence scores are all 0's. This should be the case, because if `test_loss_fn` is constant, its gradients would all be 0's, so that training examples have 0 influence on test examples.

Differential Revision: https://internalfb.com/D41202866

fbshipit-source-id: 68898a2a57f9a3599a06902c4eeb9d7c748f53f8

Author: 99warriors

captum/_utils/gradient.py

@@ -849,18 +849,21 @@ def _compute_jacobian_wrt_params_with_sample_wise_trick(
             if labels is not None and loss_fn is not None:
                 loss = loss_fn(out, labels)
                 # TODO: allow loss_fn to be Callable
-                if isinstance(loss_fn, Module) and hasattr(loss_fn, "reduction"):
+                if (isinstance(loss_fn, Module) or callable(loss_fn)) and hasattr(
+                    loss_fn, "reduction"
+                ):
+                    reduction = loss_fn.reduction  # type: ignore
                     msg0 = (
                         "Please ensure that loss_fn.reduction is set to `sum` or `mean`"
                     )
 
-                    assert loss_fn.reduction != "none", msg0
+                    assert reduction != "none", msg0
                     msg1 = (
-                        f"loss_fn.reduction ({loss_fn.reduction}) does not match"
+                        f"loss_fn.reduction ({reduction}) does not match"
                         f"reduction type ({reduction_type}). Please ensure they are"
                         " matching."
                     )
-                    assert loss_fn.reduction == reduction_type, msg1
+                    assert reduction == reduction_type, msg1
                 msg2 = (
                     "Please ensure custom loss function is applying either a "
                     "sum or mean reduction."

captum/influence/_core/tracincp.py

@@ -26,6 +26,7 @@
 from captum._utils.progress import NullProgress, progress
 from captum.influence._core.influence import DataInfluence
 from captum.influence._utils.common import (
+    _check_loss_fn,
     _format_inputs_dataset,
     _get_k_most_influential_helper,
     _gradient_dot_product,
@@ -102,6 +103,7 @@ def __init__(
         checkpoints_load_func: Callable = _load_flexible_state_dict,
         loss_fn: Optional[Union[Module, Callable]] = None,
         batch_size: Union[int, None] = 1,
+        test_loss_fn: Optional[Union[Module, Callable]] = None,
     ) -> None:
         r"""
         Args:
@@ -152,6 +154,19 @@ def __init__(
                     `train_dataset` is a Dataset. If `train_dataset`
                     is a DataLoader, then `batch_size` is ignored as an argument.
                     Default: 1
+            test_loss_fn (Callable, optional): In some cases, one may want to use a
+                    separate loss functions for training examples, i.e. those in
+                    `train_dataset`, and for test examples, i.e. those
+                    represented by the `inputs` and `targets` arguments to the
+                    `influence` method. For example, if one wants to calculate the
+                    influence score of a training example on a test example's
+                    prediction for a fixed class, `test_loss_fn` could map from the
+                    logits for all classes to the logits for a fixed class.
+                    `test_loss_fn` needs to satisfy the same constraints as `loss_fn`.
+                    If not provided, the loss function for test examples is assumed to
+                    be the same as the loss function for training examples, i.e.
+                    `loss_fn`.
+                    Default: None
         """
 
         self.model = model
@@ -167,6 +182,8 @@ def __init__(
 
         self.checkpoints_load_func = checkpoints_load_func
         self.loss_fn = loss_fn
+        # If test_loss_fn not provided, it's assumed to be same as loss_fn
+        self.test_loss_fn = loss_fn if test_loss_fn is None else test_loss_fn
         self.batch_size = batch_size
 
         if not isinstance(train_dataset, DataLoader):
@@ -489,6 +506,7 @@ def __init__(
         layers: Optional[List[str]] = None,
         loss_fn: Optional[Union[Module, Callable]] = None,
         batch_size: Union[int, None] = 1,
+        test_loss_fn: Optional[Union[Module, Callable]] = None,
         sample_wise_grads_per_batch: bool = False,
     ) -> None:
         r"""
@@ -561,6 +579,24 @@ def __init__(
                     `train_dataset` is a Dataset. If `train_dataset`
                     is a DataLoader, then `batch_size` is ignored as an argument.
                     Default: 1
+            test_loss_fn (Callable, optional): In some cases, one may want to use a
+                    separate loss functions for training examples, i.e. those in
+                    `train_dataset`, and for test examples, i.e. those
+                    represented by the `inputs` and `targets` arguments to the
+                    `influence` method. For example, if one wants to calculate the
+                    influence score of a training example on a test example's
+                    prediction for a fixed class, `test_loss_fn` could map from the
+                    logits for all classes to the logits for a fixed class.
+                    `test_loss_fn` needs satisfy the same constraints as `loss_fn`.
+                    Thus, the same checks that we apply to `loss_fn` are also applied
+                    to `test_loss_fn`, if the latter is provided. Note that the
+                    constraints on both `loss_fn` and `test_loss_fn` both depend on
+                    `sample_wise_grads_per_batch`. This means `loss_fn` and
+                    `test_loss_fn` must either both be "per-example"  loss functions,
+                    or both be "reduction" loss functions. If not provided, the loss
+                    function for test examples is assumed to be the same as the loss
+                    function for training examples, i.e. `loss_fn`.
+                    Default: None
             sample_wise_grads_per_batch (bool, optional): PyTorch's native gradient
                     computations w.r.t. model parameters aggregates the results for a
                     batch and does not allow to access sample-wise gradients w.r.t.
@@ -590,51 +626,23 @@ def __init__(
             checkpoints_load_func,
             loss_fn,
             batch_size,
+            test_loss_fn,
         )
 
         self.sample_wise_grads_per_batch = sample_wise_grads_per_batch
 
-        # If we are able to access the reduction used by `loss_fn`, we check whether
-        # the reduction is compatible with `sample_wise_grads_per_batch`
-        if isinstance(loss_fn, Module) and hasattr(
-            loss_fn, "reduction"
-        ):  # TODO: allow loss_fn to be Callable
-            if self.sample_wise_grads_per_batch:
-                assert loss_fn.reduction in ["sum", "mean"], (
-                    'reduction for `loss_fn` must be "sum" or "mean" when '
-                    "`sample_wise_grads_per_batch` is True"
-                )
-                self.reduction_type = str(loss_fn.reduction)
-            else:
-                assert loss_fn.reduction == "none", (
-                    'reduction for `loss_fn` must be "none" when '
-                    "`sample_wise_grads_per_batch` is False"
-                )
-        else:
-            # if we are unable to access the reduction used by `loss_fn`, we warn
-            # the user about the assumptions we are making regarding the reduction
-            # used by `loss_fn`
-            if self.sample_wise_grads_per_batch:
-                warnings.warn(
-                    'Since `loss_fn` has no "reduction" attribute, and '
-                    "`sample_wise_grads_per_batch` is True, the implementation assumes "
-                    'that `loss_fn` is a "reduction" loss function that reduces the '
-                    "per-example losses by taking their *sum*. If `loss_fn` "
-                    "instead reduces the per-example losses by taking their mean, "
-                    'please set the reduction attribute of `loss_fn` to "mean", i.e. '
-                    '`loss_fn.reduction = "mean"`. Note that if '
-                    "`sample_wise_grads_per_batch` is True, the implementation "
-                    "assumes the reduction is either a sum or mean reduction."
-                )
-                self.reduction_type = "sum"
-            else:
-                warnings.warn(
-                    'Since `loss_fn` has no "reduction" attribute, and '
-                    "`sample_wise_grads_per_batch` is False, the implementation "
-                    'assumes that `loss_fn` is a "per-example" loss function (see '
-                    "documentation for `loss_fn` for details).  Please ensure that "
-                    "this is the case."
-                )
+        # check `loss_fn`
+        self.reduction_type = _check_loss_fn(
+            self, loss_fn, "loss_fn", sample_wise_grads_per_batch
+        )
+        # check `test_loss_fn` if it was provided
+        self.test_reduction_type = (
+            self.reduction_type
+            if test_loss_fn is None
+            else _check_loss_fn(
+                self, test_loss_fn, "test_loss_fn", sample_wise_grads_per_batch
+            )
+        )
 
         r"""
         TODO: Either restore model state after done (would have to place functionality
@@ -790,11 +798,15 @@ def get_checkpoint_contribution(checkpoint):
             input_jacobians = self._basic_computation_tracincp(
                 inputs,
                 targets,
+                self.test_loss_fn,
+                self.test_reduction_type,
             )
             return (
                 _gradient_dot_product(
                     input_jacobians,
-                    self._basic_computation_tracincp(batch[0:-1], batch[-1]),
+                    self._basic_computation_tracincp(
+                        batch[0:-1], batch[-1], self.loss_fn, self.reduction_type
+                    ),
                 )
                 * learning_rate
             )
@@ -1042,7 +1054,10 @@ def get_checkpoint_contribution(checkpoint):
             for batch in _inputs_dataset:
 
                 layer_jacobians = self._basic_computation_tracincp(
-                    batch[0:-1], batch[-1]
+                    batch[0:-1],
+                    batch[-1],
+                    self.loss_fn,
+                    self.reduction_type,
                 )
 
                 # Note that all variables in this function are for an entire batch.
@@ -1179,11 +1194,14 @@ def _basic_computation_tracincp(
         self,
         inputs: Tuple[Any, ...],
         targets: Optional[Tensor] = None,
+        loss_fn: Optional[Union[Module, Callable]] = None,
+        reduction_type: Optional[str] = None,
     ) -> Tuple[Tensor, ...]:
         """
         For instances of TracInCP, computation of influence scores or self influence
         scores repeatedly calls this function for different checkpoints
-        and batches.
+        and batches. In particular, this function computes the jacobian of a loss
+        function w.r.t. parameters in the `layers` initialization argument.
 
         Args:
 
@@ -1193,20 +1211,26 @@ def _basic_computation_tracincp(
                     that `model(*inputs)` produces the predictions for the batch.
             targets (tensor or None): If computing influence scores on a loss function,
                     these are the labels corresponding to the batch `inputs`.
+                    Default: none
+            loss_fn (Callable, optional): The loss function to use when computing the
+                    jacobian.
+            reduction_type (str, optional): The reduction type of `loss_fn`. This
+                    argument is only used if `sample_wise_grads_per_batch` was true in
+                    initialization.
         """
         if self.sample_wise_grads_per_batch:
             return _compute_jacobian_wrt_params_with_sample_wise_trick(
                 self.model,
                 inputs,
                 targets,
-                self.loss_fn,
-                self.reduction_type,
+                loss_fn,
+                reduction_type,
                 self.layer_modules,
             )
         return _compute_jacobian_wrt_params(
             self.model,
             inputs,
             targets,
-            self.loss_fn,
+            loss_fn,
             self.layer_modules,
         )