Forward Hooks Persist After Destroying FeatureExtractor

[ShunsukeOnoo]

Issue

When the FeatureExtractor is initialized in bdpy.recon.torch.icnn.reconstruct, it registers forward hooks on the encoder. However, the registered forward hooks are not erased after the reconstruct is returned and feature_extractor is destroyed. This leads to a problem that the remaining forward hooks keep accumulating features and occupy memory when the same encoder instance is used for reconstruct multiple times.

Here's a snippet that illustrates the behavior of FeatureExtractor

import open_clip
from bdpy.dl.torch import FeatureExtractor
from bdpy.dl.torch import models

model, _,  preprocess = open_clip.create_model_and_transforms('ViT-L-14', pretrained='openai')
layers = ['visual.transformer.resblocks.0']

def reconstruct_mock(model):
    feature_extractor = FeatureExtractor(model, layers, device='cpu', detach=False)
    
reconstruct_mock(model)

layer_object = models._parse_layer_name(model, layers[0])
print(layer_object._forward_hooks)
# OrderedDict([(0, <bdpy.dl.torch.torch.FeatureExtractorHandle at 0x7fc2f698bf40>)])

Yes, the example in the cookbook avoids this problem by initializing encoder for each image, but this is not effective when using larger models.

Suggestion

One possible fix is to add a destructor method that clears forward hooks from the layers:

class FeatureExtractor(object):
    
    ....

    def __del__(self):
        for layer in self.__layers:
            layer_object = models._parse_layer_name(self._encoder, layer)
            layer_object._forward_hooks = OrderedDict()