def test_image_feature_generates_correct_ouput(self):
attribution = torch.zeros(1, 3, 4, 4)
data = torch.ones(1, 3, 4, 4)
contribution = 1.0
name = "photo"
orig_fig = Figure(figsize=(4, 4))
attr_fig = Figure(figsize=(4, 4))
def mock_viz_attr(*args, **kwargs):
if kwargs["method"] == "original_image":
return orig_fig, None
else:
return attr_fig, None
feature = ImageFeature(
name=name,
baseline_transforms=None,
input_transforms=None,
visualization_transform=None,
)
with patch(
"captum.attr._utils.visualization.visualize_image_attr", mock_viz_attr
):
feature_output = feature.visualize(attribution, data, contribution)
expected_feature_output = FeatureOutput(
name=name,
base=_convert_figure_base64(orig_fig),
modified=_convert_figure_base64(attr_fig),
type="image",
contribution=contribution,
)
self.assertEqual(expected_feature_output, feature_output)
def test_one_feature(self):
batch_size = 2
classes = _get_classes()
dataset = list(
_labelled_img_data(num_labels=len(classes),
num_samples=batch_size))
# NOTE: using DataLoader to batch the inputs
# since AttributionVisualizer requires the input to be of size `B x ...`
data_loader = torch.utils.data.DataLoader(list(dataset),
batch_size=batch_size,
shuffle=False,
num_workers=0)
visualizer = AttributionVisualizer(
models=[_get_cnn()],
classes=classes,
features=[
ImageFeature(
"Photo",
input_transforms=[lambda x: x],
baseline_transforms=[lambda x: x * 0],
)
],
dataset=to_iter(data_loader),
score_func=None,
)
visualizer._config = FilterConfig(attribution_arguments={"n_steps": 2})
outputs = visualizer.visualize()
for output in outputs:
total_contrib = sum(
abs(f.contribution) for f in output.feature_outputs)
self.assertAlmostEqual(total_contrib, 1.0, places=6)
def insights(x: CaptumInterpretation, inp_data, debug=True):
_baseline_func = lambda o: o * 0
_get_vocab = lambda vocab: list(map(str, vocab)) if isinstance(
vocab[0], bool) else vocab
dl = x.dls.test_dl(L(inp_data), with_labels=True, bs=4)
normalize_func = next(
(func for func in dl.after_batch if type(func) == Normalize), noop)
# captum v0.3 expects tensors without the batch dimension.
if hasattr(normalize_func, 'mean'):
if normalize_func.mean.ndim == 4: normalize_func.mean.squeeze_(0)
if hasattr(normalize_func, 'std'):
if normalize_func.std.ndim == 4: normalize_func.std.squeeze_(0)
visualizer = AttributionVisualizer(
models=[x.model],
score_func=lambda o: torch.nn.functional.softmax(o, 1),
classes=_get_vocab(dl.vocab),
features=[
ImageFeature(
"Image",
baseline_transforms=[_baseline_func],
input_transforms=[normalize_func],
)
],
dataset=x._formatted_data_iter(dl, normalize_func))
visualizer.render(debug=debug)
def main():
normalize = transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
model = get_pretrained_model()
visualizer = AttributionVisualizer(
models=[model],
score_func=lambda o: torch.nn.functional.softmax(o, 1),
classes=get_classes(),
features=[
ImageFeature(
"Photo",
baseline_transforms=[baseline_func],
input_transforms=[normalize],
)
],
dataset=formatted_data_iter(),
)
visualizer.serve(debug=True)
def __init__(self,
path,
bs,
learn,
application,
baseline_func_default=False,
baseline_token=None,
baseline_fn=None):
self.path = path if path else None
self.bs = bs if bs else 8
self.data = learn.data
self.model = learn.model.eval()
self.application = application
self.baseline_func_default = baseline_func_default
if self.application == 'text':
self.vocab = learn.data.x.vocab
self.baseline_token = (baseline_token if baseline_token else
'.' if self.application == 'text' else 0)
self.baseline_fn = baseline_fn
self.features = ([
TextFeature(
"Sentence",
baseline_transforms=(
[self.baseline_fn] if self.baseline_fn else None
if self.baseline_func_default else [self.baseline_func]),
visualization_transform=self.itos,
input_transforms=[])
] if self.application == 'text' else [
ImageFeature(
'Image',
baseline_transforms=(
[self.baseline_fn] if self.baseline_fn else None if self.
baseline_func_default else [self.baseline_func]),
input_transforms=[])
] if self.application == 'vision' else [
TabularFeature(
'Table',
baseline_transforms=(
[self.baseline_fn] if self.baseline_fn else None if self.
baseline_func_default else [self.baseline_func]))
])
i = transform(i)
i = transform_normalize(i)
i = i.unsqueeze(0)
return i
input_imgs = torch.cat(list(map(lambda i: full_img_transform(i), imgs)), 0)
visualizer = AttributionVisualizer(
models=[model],
score_func=lambda o: torch.nn.functional.softmax(o, 1),
classes=list(map(lambda k: idx_to_labels[k][1], idx_to_labels.keys())),
features=[
ImageFeature(
"Photo",
baseline_transforms=[baseline_func],
input_transforms=[],
)
],
dataset=[Batch(input_imgs, labels=[282, 849, 69])])
#########################################################################
# Note that running the cell above didn’t take much time at all, unlike
# our attributions above. That’s because Captum Insights lets you
# configure different attribution algorithms in a visual widget, after
# which it will compute and display the attributions. *That* process will
# take a few minutes.
#
# Running the cell below will render the Captum Insights widget. You can
# then choose attributions methods and their arguments, filter model
# responses based on predicted class or prediction correctness, see the
