Does Lightning disable PyTorch registered forward hooks when using trainer.validate?

[jonathanking]

I've registered some forward hooks in my model in order to extract intermediate layer activations. I've done this via this method:

model = MyModel()
activations = {"input": [], "output": []}

def hook(module, _input, _output):
    activations["input"].append( _input)
    activations["output"].append( _output)

_hook = model.my_layer.register_forward_hook(hook)

# ... model training, validation, etc. 

The activations dictionary gets populated when using trainer.fit but not trainer.validate. Why is this? How can I keep this behavior during trainer.validate?