|
|
|
@ -4,7 +4,7 @@ Hacked together by / Copyright 2020 Ross Wightman
|
|
|
|
|
"""
|
|
|
|
|
from .model_ema import ModelEma
|
|
|
|
|
import torch
|
|
|
|
|
|
|
|
|
|
import fnmatch
|
|
|
|
|
|
|
|
|
|
def unwrap_model(model):
|
|
|
|
|
if isinstance(model, ModelEma):
|
|
|
|
@ -23,32 +23,37 @@ def avg_sq_ch_mean(model, input, output):
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
def avg_ch_var(model, input, output):
|
|
|
|
|
"calculate average channel variance of output activations"
|
|
|
|
|
return torch.mean(output.var(axis=[0,2,3])).item()\
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
def avg_ch_var_residual(model, input, output):
|
|
|
|
|
"calculate average channel variance of output activations"
|
|
|
|
|
return torch.mean(output.var(axis=[0,2,3])).item()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
class ActivationStatsHook:
|
|
|
|
|
"""Iterates through each of `model`'s modules and if module's class name
|
|
|
|
|
is present in `layer_names` then registers `hook_fns` inside that module
|
|
|
|
|
and stores activation stats inside `self.stats`.
|
|
|
|
|
"""Iterates through each of `model`'s modules and matches modules using unix pattern
|
|
|
|
|
matching based on `layer_name` and `layer_type`. If there is match, this class adds
|
|
|
|
|
creates a hook using `hook_fn` and adds it to the module.
|
|
|
|
|
|
|
|
|
|
Arguments:
|
|
|
|
|
model (nn.Module): model from which we will extract the activation stats
|
|
|
|
|
layer_names (List[str]): The layer name to look for to register forward
|
|
|
|
|
hook. Example, `BasicBlock`, `Bottleneck`
|
|
|
|
|
layer_names (str): The layer name to look for to register forward
|
|
|
|
|
hook. Example, 'stem', 'stages'
|
|
|
|
|
hook_fns (List[Callable]): List of hook functions to be registered at every
|
|
|
|
|
module in `layer_names`.
|
|
|
|
|
|
|
|
|
|
Inspiration from https://docs.fast.ai/callback.hook.html.
|
|
|
|
|
"""
|
|
|
|
|
|
|
|
|
|
def __init__(self, model, layer_names, hook_fns=[avg_sq_ch_mean, avg_ch_var]):
|
|
|
|
|
def __init__(self, model, hook_fn_locs, hook_fns):
|
|
|
|
|
self.model = model
|
|
|
|
|
self.layer_names = layer_names
|
|
|
|
|
self.hook_fn_locs = hook_fn_locs
|
|
|
|
|
self.hook_fns = hook_fns
|
|
|
|
|
self.stats = dict((hook_fn.__name__, []) for hook_fn in hook_fns)
|
|
|
|
|
for hook_fn in hook_fns:
|
|
|
|
|
self.register_hook(layer_names, hook_fn)
|
|
|
|
|
for hook_fn_loc, hook_fn in zip(hook_fn_locs, hook_fns):
|
|
|
|
|
self.register_hook(hook_fn_loc, hook_fn)
|
|
|
|
|
|
|
|
|
|
def _create_hook(self, hook_fn):
|
|
|
|
|
def append_activation_stats(module, input, output):
|
|
|
|
@ -56,17 +61,16 @@ class ActivationStatsHook:
|
|
|
|
|
self.stats[hook_fn.__name__].append(out)
|
|
|
|
|
return append_activation_stats
|
|
|
|
|
|
|
|
|
|
def register_hook(self, layer_names, hook_fn):
|
|
|
|
|
for layer in self.model.modules():
|
|
|
|
|
layer_name = layer.__class__.__name__
|
|
|
|
|
if layer_name not in layer_names:
|
|
|
|
|
def register_hook(self, hook_fn_loc, hook_fn):
|
|
|
|
|
for name, module in self.model.named_modules():
|
|
|
|
|
if not fnmatch.fnmatch(name, hook_fn_loc):
|
|
|
|
|
continue
|
|
|
|
|
layer.register_forward_hook(self._create_hook(hook_fn))
|
|
|
|
|
module.register_forward_hook(self._create_hook(hook_fn))
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
def extract_spp_stats(model,
|
|
|
|
|
layer_names,
|
|
|
|
|
hook_fns=[avg_sq_ch_mean, avg_ch_var],
|
|
|
|
|
hook_fn_locs,
|
|
|
|
|
hook_fns,
|
|
|
|
|
input_shape=[8, 3, 224, 224]):
|
|
|
|
|
"""Extract average square channel mean and variance of activations during
|
|
|
|
|
forward pass to plot Signal Propogation Plots (SPP).
|
|
|
|
@ -74,7 +78,7 @@ def extract_spp_stats(model,
|
|
|
|
|
Paper: https://arxiv.org/abs/2101.08692
|
|
|
|
|
"""
|
|
|
|
|
x = torch.normal(0., 1., input_shape)
|
|
|
|
|
hook = ActivationStatsHook(model, layer_names, hook_fns)
|
|
|
|
|
hook = ActivationStatsHook(model, hook_fn_locs=hook_fn_locs, hook_fns=hook_fns)
|
|
|
|
|
_ = model(x)
|
|
|
|
|
return hook.stats
|
|
|
|
|
|