def test_extraction_batches(self):
values = [1] * 10
backend = 'pt'
dataset = helper.SimpleDataset(values, backend)
model = Model('vgg16',
pretrained=False,
device=helper.DEVICE,
backend=backend)
model.model = helper.pt_model
# no batch remainders -> 5 batches with 2 examples
# batch remainders -> 3 batches with 3 examples and 1 batch with 1 remainder
for batch_size in [2, 3]:
dl = DataLoader(
dataset,
batch_size=batch_size,
backend=backend,
)
features, targets = model.extract_features(
data_loader=dl,
module_name='relu',
flatten_acts=False,
return_probabilities=False)
self.assertEqual(features.shape[0], len(dataset))
self.assertEqual(targets.shape[0], len(dataset))
def test_load_custom_user_model(self):
model_name = 'VGG16bn_ecoset'
model = Model(model_name, True, 'cpu')
self.assertTrue(model.__class__.__name__, 'vgg')
model_name = 'Resnet50_ecoset'
model = Model(model_name, True, 'cpu')
self.assertTrue(model.__class__.__name__, 'resnet')
model_name = 'Alexnet_ecoset'
model = Model(model_name, True, 'cpu')
self.assertTrue(model.__class__.__name__, 'alexnet')
def extract_features_across_models_datasets_and_modules(
out_path: str,
model_names: List[str],
img_paths: List[str],
module_names: List[str],
clip: List[str],
pretrained: bool,
batch_size: int,
backend: str,
flatten_acts: bool,
f_format: str = 'txt') -> None:
device = 'cuda' if torch.cuda.is_available() else 'cpu'
for i, model_name in enumerate(model_names):
model = Model(
model_name=model_name,
pretrained=pretrained,
device=device,
model_path=None,
backend=backend,
)
transforms = model.get_transformations()
modules = get_module_names(model, module_names[i])
for img_path in img_paths:
for module_name in modules:
PATH = os.path.join(out_path, img_path, model_name,
module_name, 'features')
dl = load_dl(
root=img_path,
out_path=out_path,
backend=backend,
batch_size=batch_size,
transforms=transforms,
)
features, _ = model.extract_features(
data_loader=dl,
module_name=module_name,
flatten_acts=flatten_acts,
clip=clip[i],
)
save_features(features, PATH, f_format)
def create_model_and_dl(model_name, backend):
"""Iterate through all backends and models and create model, dataset and data loader."""
model = Model(model_name=model_name,
pretrained=True,
device=DEVICE,
backend=backend)
dataset = ImageDataset(root=TEST_PATH,
out_path=OUT_PATH,
backend=backend,
imagenet_train=None,
imagenet_val=None,
things=None,
things_behavior=None,
add_ref_imgs=None,
transforms=model.get_transformations())
dl = DataLoader(
dataset,
batch_size=BATCH_SIZE,
backend=backend,
)
return model, dataset, dl
def test_extract_features(self):
layer_name = 'relu'
values = [2, -10]
backends = [['pt', helper.pt_model, 'vgg16'],
['tf', helper.tf_model, 'VGG16']]
batch_size = 1
for backend, custom_model, vgg_model in backends:
dataset = helper.SimpleDataset(values, backend)
dl = DataLoader(
dataset,
batch_size=batch_size,
backend=backend,
)
model = Model(vgg_model,
pretrained=False,
device=helper.DEVICE,
backend=backend)
model.model = custom_model
expected_features = np.array([[2, 2], [0, 0]])
expected_targets = np.array([0, 0])
features, targets = model.extract_features(
data_loader=dl,
module_name=layer_name,
flatten_acts=False,
return_probabilities=False)
np.testing.assert_allclose(features, expected_features)
np.testing.assert_allclose(targets, expected_targets)
expected_probs = np.array([[0.5, 0.5], [0.5, 0.5]])
features, targets, probs = model.extract_features(
data_loader=dl,
module_name=layer_name,
flatten_acts=False,
return_probabilities=True)
np.testing.assert_allclose(features, expected_features)
np.testing.assert_allclose(targets, expected_targets)
np.testing.assert_allclose(probs, expected_probs)
def test_custom_torch_vs_tf_extraction(self):
layer_name = 'relu'
values = [2, -10]
backend = 'tf'
tf_dataset = helper.SimpleDataset(values, backend)
tf_dl = DataLoader(
tf_dataset,
batch_size=1,
backend=backend,
)
model = Model('VGG16',
pretrained=False,
device=helper.DEVICE,
backend=backend)
model.model = helper.tf_model
tf_features, _ = model.extract_features(
data_loader=tf_dl,
module_name=layer_name,
flatten_acts=False,
)
backend = 'pt'
pt_dataset = helper.SimpleDataset(values, backend)
pt_dl = DataLoader(
pt_dataset,
batch_size=1,
backend=backend,
)
model = Model('vgg16',
pretrained=False,
device=helper.DEVICE,
backend=backend)
model.model = helper.pt_model
pt_features, _ = model.extract_features(
data_loader=pt_dl,
module_name=layer_name,
flatten_acts=False,
)
np.testing.assert_allclose(tf_features, pt_features)
expected_features = np.array([[2, 2], [0, 0]])
np.testing.assert_allclose(pt_features, expected_features)
def get_features(
root: str,
out_path: str,
model_names: List[str],
module_names: List[str],
clip: List[bool],
pretrained: bool,
batch_size: int,
backend: str,
flatten_acts: bool,
) -> Dict[str, Dict[str, np.ndarray]]:
"""Extract features for a list of neural network models and corresponding modules.
Parameters
----------
root : str
Root directory. Directory where images are stored.
out_path : str
PATH where order of images features should be stored.
Files are alphabetically sorted and features are
extracted accordingly.
model_names : List[str]
List of neural network models for which features
should be extracted.
module_names : List[str]
List of neural network layers for which features
should be extracted. Modules must correspond to
models. This should be thought of as zipped lists.
clip : List[bool]
List of Booleans which indicates whether the
corresponding model in the <model_names> list
is a CLIP-based model or not (i.e., True if
CLIP, else False)
pretrained : bool
Whether pretrained or randomly initialized models
should be loaded into memory.
batch_size : int
Integer value that determines the number of images
within a single mini-batch (i.e., subsample
of the data).
flatten_acts : bool
Whether activation tensor (e.g., activations
from an early layer of the neural network model)
should be transformed into a feature vector.
Returns
-------
output : Dict[str, Dict[str, np.ndarray]]
Returns a dictionary of feature matrices
corresponding to the selected models and layers.
"""
device = 'cuda' if torch.cuda.is_available() else 'cpu'
model_features = defaultdict(dict)
for i, model_name in enumerate(model_names):
model = Model(
model_name=model_name,
pretrained=pretrained,
device=device,
model_path=None,
backend=backend,
)
transforms = model.get_transformations()
dl = load_dl(
root=root,
out_path=out_path,
backend=backend,
batch_size=batch_size,
transforms=transforms,
)
features, _ = model.extract_features(
data_loader=dl,
module_name=module_names[i],
flatten_acts=flatten_acts,
clip=clip[i],
)
model_features[model_name][module_names[i]] = features
return model_features