Merge branch 'develop' into fix/value-init

CONTRIBUTING.md

@@ -170,12 +170,29 @@ Or:
 ]
 ```
 
-To only hide the input 
+To only hide the input and
+
+```yaml
+"tags": [
+  "hide-output"
+]
+```
+for hiding the output only.
 
 It is important to leave a warning at the top of the document to avoid confusion.
 Examples for hidden imports and plots are available in the notebooks, e.g. in
-[Shapley for data valuation](https://aai-institute.github.io/pyDVL/examples/shapley_basic_spotify.ipynb).
+*notebooks/shapley_basic_spotify.ipynb*.
 
+### Plots in Notebooks
+If you add a plot to a notebook, which should also render nicely in browser
+dark mode, add the tag *invertible-output*, i.e.
+
+```yaml
+"tags": [
+  "invertible-output"
+]
+```
+This applies a simple CSS-filter to the output image of the cell.
 
 ## Documentation
 

docs/css/extra.css

@@ -107,14 +107,17 @@ a.autorefs-external:hover::after {
   color: black;
 }
 
-
 body[data-md-color-scheme="default"] .invertible img {
 }
 
 body[data-md-color-scheme="slate"] .invertible img {
     filter: invert(100%) hue-rotate(180deg);
 }
 
+body[data-md-color-scheme="slate"] .celltag_invertible-output img {
+    filter: invert(100%) hue-rotate(180deg);
+}
+
 /* Rendered dataframe from jupyter */
 table.dataframe {
     display: block;

mkdocs.yml

@@ -53,6 +53,8 @@ plugins:
           - hide
         remove_input_tags:
           - hide-input
+        remove_all_outputs_tags:
+          - hide-output
       binder: true
       binder_service_name: "gh"
       binder_branch: "develop"

notebooks/shapley_utility_learning.ipynb

@@ -374,8 +374,13 @@
    "cell_type": "code",
    "execution_count": 12,
    "metadata": {
+    "editable": true,
+    "slideshow": {
+     "slide_type": ""
+    },
     "tags": [
-     "hide-input"
+     "hide-input",
+     "invertible-output"
     ]
    },
    "outputs": [
@@ -436,8 +441,13 @@
    "cell_type": "code",
    "execution_count": 14,
    "metadata": {
+    "editable": true,
+    "slideshow": {
+     "slide_type": ""
+    },
     "tags": [
-     "hide-input"
+     "hide-input",
+     "invertible-output"
     ]
    },
    "outputs": [
@@ -477,8 +487,13 @@
    "cell_type": "code",
    "execution_count": 16,
    "metadata": {
+    "editable": true,
+    "slideshow": {
+     "slide_type": ""
+    },
     "tags": [
-     "hide-input"
+     "hide-input",
+     "invertible-output"
     ]
    },
    "outputs": [
@@ -628,8 +643,13 @@
    "cell_type": "code",
    "execution_count": 21,
    "metadata": {
+    "editable": true,
+    "slideshow": {
+     "slide_type": ""
+    },
     "tags": [
-     "hide-input"
+     "hide-input",
+     "invertible-output"
     ]
    },
    "outputs": [

src/pydvl/influence/general.py

@@ -104,8 +104,9 @@ def test_grads() -> Generator[TensorType, None, None]:
             )  # type:ignore
 
     try:
-        # if provided input_data implements __len__, pre-allocate the result tensor to reduce memory consumption
-        resulting_shape = (len(test_data), model.num_params)  # type:ignore
+        # in case input_data is a torch DataLoader created from a Dataset,
+        # we can pre-allocate the result tensor to reduce memory consumption
+        resulting_shape = (len(test_data.dataset), model.num_params)  # type:ignore
         rhs = cat_gen(
             test_grads(), resulting_shape, model  # type:ignore
         )  # type:ignore
@@ -174,8 +175,9 @@ def train_grads() -> Generator[TensorType, None, None]:
             )  # type:ignore
 
     try:
-        # if provided input_data implements __len__, pre-allocate the result tensor to reduce memory consumption
-        resulting_shape = (len(input_data), model.num_params)  # type:ignore
+        # in case input_data is a torch DataLoader created from a Dataset,
+        # we can pre-allocate the result tensor to reduce memory consumption
+        resulting_shape = (len(input_data.dataset), model.num_params)  # type:ignore
         train_grad_tensor = cat_gen(
             train_grads(), resulting_shape, model  # type:ignore
         )  # type:ignore

src/pydvl/influence/torch/functional.py

@@ -91,7 +91,7 @@ def batch_hvp_gen(
 
     for inputs, targets in iter(data_loader):
         batch_loss = batch_loss_function(model, loss, inputs, targets)
-        model_params = dict(model.named_parameters())
+        model_params = {k: p for k, p in model.named_parameters() if p.requires_grad}
 
         def batch_hvp(vec: torch.Tensor):
             return flatten_tensors_to_vector(
@@ -166,9 +166,7 @@ def batch_loss_function(
     """
 
     def batch_loss(params: Dict[str, torch.Tensor]):
-        outputs = functional_call(
-            model, params, (to_model_device(x, model),), strict=True
-        )
+        outputs = functional_call(model, params, (to_model_device(x, model),))
         return loss(outputs, y)
 
     return batch_loss
@@ -209,7 +207,9 @@ def get_hvp_function(
     """
 
     params = {
-        k: p if track_gradients else p.detach() for k, p in model.named_parameters()
+        k: p if track_gradients else p.detach()
+        for k, p in model.named_parameters()
+        if p.requires_grad
     }
 
     def hvp_function(vec: torch.Tensor) -> torch.Tensor:

src/pydvl/influence/torch/torch_differentiable.py

@@ -149,7 +149,6 @@ def model_func(param):
                     param,
                 ),
                 (x.to(self.device),),
-                strict=True,
             )
             return self.loss(outputs, y.to(self.device))