def _run_hoc(self, summary, sample_id, sample_input, prob):
"""
Run HOC search for a sample image, and then save the result to summary.
Args:
summary (SummaryRecord): The summary object to store the data.
sample_id (int): The sample ID.
sample_input (Union[Tensor, np.ndarray]): Sample image tensor in CHW or NCWH(N=1).
prob (Union[Tensor, np.ndarray]): List of sample's classification prediction output, HOC will run for
labels with prediction output strictly larger then HOC searcher's threshold(0.5 by default).
"""
if isinstance(sample_input, ms.Tensor):
sample_input = sample_input.asnumpy()
if len(sample_input.shape) == 3:
sample_input = np.expand_dims(sample_input, axis=0)
has_rec = False
explain = Explain()
explain.sample_id = sample_id
str_mask = hoc.auto_str_mask(sample_input)
compiled_mask = None
for label_idx, label_prob in enumerate(prob):
if label_prob > self._hoc_searcher.threshold:
if compiled_mask is None:
compiled_mask = hoc.compile_mask(str_mask, sample_input)
try:
edit_tree, layer_outputs = self._hoc_searcher.search(
sample_input, label_idx, compiled_mask)
except hoc.NoValidResultError:
log.warning(
f"No Hierarchical Occlusion result was found in sample#{sample_id} "
f"label:{self._labels[label_idx]}, skipped.")
continue
has_rec = True
hoc_rec = explain.hoc.add()
hoc_rec.label = label_idx
hoc_rec.mask = str_mask
layer_count = edit_tree.max_layer + 1
for layer in range(layer_count):
steps = edit_tree.get_layer_or_leaf_steps(layer)
layer_output = layer_outputs[layer]
hoc_layer = hoc_rec.layer.add()
hoc_layer.prob = layer_output
for step in steps:
hoc_layer.box.extend(list(step.box))
if has_rec:
summary.add_value("explainer", "hoc", explain)
summary.record(1)
self._manifest['hierarchical_occlusion'] = True
def _run_hoc(self, summary, sample_id, sample_input, prob):
"""
Run HOC search for a sample image, and then save the result to summary.
Args:
summary (SummaryRecord): The summary object to store the data.
sample_id (int): The sample ID.
sample_input (Union[Tensor, np.ndarray]): Sample image tensor in CHW or NCWH(N=1).
prob (Union[Tensor, np.ndarray]): List of sample's classification prediction output, HOC will run for
labels with prediction output strictly larger then HOC searcher's threshold(0.5 by default).
"""
if isinstance(sample_input, ms.Tensor):
sample_input = sample_input.asnumpy()
if len(sample_input.shape) == 3:
sample_input = np.expand_dims(sample_input, axis=0)
explain = None
str_mask = hoc.auto_str_mask(sample_input)
compiled_mask = None
for label_idx, label_prob in enumerate(prob):
if label_prob <= self._hoc_searcher.threshold:
continue
if compiled_mask is None:
compiled_mask = hoc.compile_mask(str_mask, sample_input)
try:
edit_tree, layer_outputs = self._hoc_searcher.search(
sample_input, label_idx, compiled_mask)
except hoc.NoValidResultError:
log.warning(
f"No Hierarchical Occlusion result was found in sample#{sample_id} "
f"label:{self._labels[label_idx]}, skipped.")
continue
if explain is None:
explain = Explain()
explain.sample_id = sample_id
self._add_hoc_result_to_explain(label_idx, str_mask, edit_tree,
layer_outputs, explain)
if explain is not None:
summary.add_value("explainer", "hoc", explain)
summary.record(1)
self._manifest['hierarchical_occlusion'] = True