LukasMut
commented
1 year ago Hi there! This is because the keys for the features_per_layer dictionary that is used to store the activations for each module in the extract_all_layers convenience function (see below) correspond to custom names and not to the original module names. You can replace features_per_layer[f'layer_{l:02d}'] with features_per_layer[f'{module_name}'] for using original module names. See my comments in the below function.
def extract_all_layers(
model_name: str,
extractor: Any,
image_path: str,
out_path: str,
batch_size: int,
flatten_activations: bool,
apply_center_crop: bool,
layer: Any=nn.Linear,
file_format: str = "npy",
class_names: Optional[List[str]]=None,
file_names: Optional[List[str]]=None,
) -> Dict[str, Union[np.ndarray, torch.Tensor]]:
"""Extract features for all selected layers and save them to disk."""
features_per_layer = {}
for l, (module_name, module) in enumerate(extractor.model.named_modules(), start=1):
if isinstance(module, layer):
# extract features for layer "module_name"
features = extract_features(
extractor=extractor,
module_name=module_name,
image_path=image_path,
out_path=out_path,
batch_size=batch_size,
flatten_activations=flatten_activations,
apply_center_crop=apply_center_crop,
class_names=class_names,
file_names=file_names,
)
# NOTE: for original module names use features_per_layer[f'{module_name}'] = features
# replace with e.g., [f'conv_{l:02d}'] or [f'fc_{l:02d}']
features_per_layer[f'layer_{l:02d}'] = features
# save the features to disk
save_features(features, out_path=f'{out_path}/features_{model_name}_{module_name}', file_format=file_format)
return features_per_layer
spaladin
Hi!
I am using the Collab notebook for PyTorch. I tried Alexnet – by changing some parameters in the notebook examples (VGG-16 with batch norm pretrained on ImageNet) I wanted to extract activations from all the convolutional layers, so I used the following code modification:
The number of extracted layers (5) is correct, however, their names are wrong. Also, they don’t seem to match activations if I extract them layer by layer.