def test_bs_PCA_fit():
img_path_placeholder, ending_points, SESS \
= _get_cate_seed0_ending_points_sess(
64,
'/mnt/fs4/chengxuz/brainscore_model_caches/cate_aug/res18/exp_seed0_bn_wd/checkpoint-505505')
identifier = 'test-cate-bn-wd-pca'
activations_model = TensorflowSlimWrapper(
identifier=identifier, labels_offset=0,
endpoints=ending_points, inputs=img_path_placeholder,
session=SESS)
LayerPCA.hook(activations_model, n_components=1000)
#benchmark = bench.load('dicarlo.Majaj2015.IT-pls')
benchmark = bench.load('dicarlo.Majaj2015.V4-pls')
all_layers = ['encode_%i' % i for i in range(1, 10)]
model_scores = LayerScores(
identifier,
activations_model=activations_model,
)
score = model_scores(
benchmark=benchmark,
layers=all_layers,
)
print(score)
def test_from_stimulus_set(model_ctr, layers, pca_components):
image_names = [
'rgb.jpg', 'grayscale.png', 'grayscale2.jpg', 'grayscale_alpha.png'
]
stimulus_set = StimulusSet([{
'image_id': image_name,
'some_meta': image_name[::-1]
} for image_name in image_names])
stimulus_set.image_paths = {
image_name: os.path.join(os.path.dirname(__file__), image_name)
for image_name in image_names
}
activations_extractor = model_ctr()
if pca_components:
LayerPCA.hook(activations_extractor, pca_components)
activations = activations_extractor.from_stimulus_set(
stimulus_set, layers=layers, stimuli_identifier=False)
assert activations is not None
assert set(activations['image_id'].values) == set(image_names)
assert all(activations['some_meta'].values ==
[image_name[::-1] for image_name in image_names])
assert len(np.unique(activations['layer'])) == len(layers)
if pca_components is not None:
assert len(activations['neuroid']) == pca_components * len(layers)
def test_base_model_pool_reload(self):
activations_model = base_model_pool['alexnet']
LayerPCA.hook(activations_model, n_components=1000)
assert len(activations_model._extractor._batch_activations_hooks) == 1
activations_model = base_model_pool['alexnet']
assert len(activations_model._extractor._stimulus_set_hooks) == 0
assert len(activations_model._extractor._batch_activations_hooks) == 0
LayerPCA.hook(activations_model, n_components=1000)
assert len(activations_model._extractor._batch_activations_hooks) == 1
def load(model_name=model_name, layer=layer, region=region, pca_components=pca_components):
activations_model = base_model_pool[model_name]
if pca_components:
LayerPCA.hook(activations_model, n_components=pca_components)
activations_model.identifier += "-pca_1000"
model = LayerMappedModel(f"{model_name}-{layer}", activations_model=activations_model, visual_degrees=8)
model.commit(region, layer)
model = TemporalIgnore(model)
return model
def test_brain_translated_pool_reload(self, model_identifier):
activations_model = brain_translated_pool[model_identifier].content.activations_model
LayerPCA.hook(activations_model, n_components=1000)
assert len(activations_model._extractor._batch_activations_hooks) == 1
activations_model = brain_translated_pool[model_identifier].content.activations_model
assert len(activations_model._extractor._stimulus_set_hooks) == 0
assert len(activations_model._extractor._batch_activations_hooks) == 0
LayerPCA.hook(activations_model, n_components=1000)
assert len(activations_model._extractor._batch_activations_hooks) == 1
def test_MajajHong2015ITpls(self, model_identifier, expected_score, attach_hook):
model = brain_translated_pool[model_identifier]
if attach_hook:
activations_model = model.layer_model._layer_model.activations_model
LayerPCA.hook(activations_model, n_components=1000)
identifier = activations_model.identifier + "-pca_1000"
activations_model.identifier = identifier
score = score_model(model_identifier, 'dicarlo.MajajHong2015.IT-pls', model=model)
assert score.raw.sel(aggregation='center') == approx(expected_score, abs=0.005)
def test_from_stimulus_set(model_ctr, layers, pca_components):
image_names = ['rgb.jpg', 'grayscale.png', 'grayscale2.jpg', 'grayscale_alpha.png', 'palletized.png']
stimulus_set = _build_stimulus_set(image_names)
activations_extractor = model_ctr()
if pca_components:
LayerPCA.hook(activations_extractor, pca_components)
activations = activations_extractor.from_stimulus_set(stimulus_set, layers=layers, stimuli_identifier=False)
assert activations is not None
assert set(activations['image_id'].values) == set(image_names)
assert all(activations['some_meta'].values == [image_name[::-1] for image_name in image_names])
assert len(np.unique(activations['layer'])) == len(layers)
if pca_components is not None:
assert len(activations['neuroid']) == pca_components * len(layers)
def __call__(self, selection_identifier, benchmark):
# for layer-mapping, attach LayerPCA so that we can cache activations
model_identifier = self.model_identifier
pca_hooked = LayerPCA.is_hooked(self._layer_scoring._activations_model)
if not pca_hooked:
pca_handle = LayerPCA.hook(self._layer_scoring._activations_model, n_components=1000)
identifier = self._layer_scoring._activations_model.identifier
self._layer_scoring._activations_model.identifier = identifier + "-pca_1000"
model_identifier += "-pca_1000"
result = self._call(model_identifier=model_identifier, selection_identifier=selection_identifier,
benchmark=benchmark)
if not pca_hooked:
pca_handle.remove()
self._layer_scoring._activations_model.identifier = identifier
return result
def test_from_image_path(model_ctr, layers, image_name, pca_components, logits):
stimuli_paths = [os.path.join(os.path.dirname(__file__), image_name)]
activations_extractor = model_ctr()
if pca_components:
LayerPCA.hook(activations_extractor, pca_components)
activations = activations_extractor.from_paths(stimuli_paths=stimuli_paths,
layers=layers if not logits else None)
assert activations is not None
assert len(activations['stimulus_path']) == 1
assert len(np.unique(activations['layer'])) == len(layers) if not logits else 1
if logits and not pca_components:
assert len(activations['neuroid']) == 1000
elif pca_components is not None:
assert len(activations['neuroid']) == pca_components * len(layers)
import gc
gc.collect() # free some memory, we're piling up a lot of activations at this point
return activations