108 lines
4.0 KiB
Python
108 lines
4.0 KiB
Python
# Copyright 2021 Samsung Electronics Co., Ltd.
|
|
#
|
|
# Licensed under the Apache License, Version 2.0 (the "License");
|
|
# you may not use this file except in compliance with the License.
|
|
# You may obtain a copy of the License at
|
|
|
|
# http://www.apache.org/licenses/LICENSE-2.0
|
|
|
|
# Unless required by applicable law or agreed to in writing, software
|
|
# distributed under the License is distributed on an "AS IS" BASIS,
|
|
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
# See the License for the specific language governing permissions and
|
|
# limitations under the License.
|
|
# =============================================================================
|
|
|
|
import torch
|
|
import torch.nn as nn
|
|
import torch.nn.functional as F
|
|
|
|
import types
|
|
|
|
from . import measure
|
|
from ..p_utils import get_layer_metric_array, reshape_elements
|
|
|
|
|
|
def fisher_forward_conv2d(self, x):
|
|
x = F.conv2d(x, self.weight, self.bias, self.stride,
|
|
self.padding, self.dilation, self.groups)
|
|
#intercept and store the activations after passing through 'hooked' identity op
|
|
self.act = self.dummy(x)
|
|
return self.act
|
|
|
|
def fisher_forward_linear(self, x):
|
|
x = F.linear(x, self.weight, self.bias)
|
|
self.act = self.dummy(x)
|
|
return self.act
|
|
|
|
@measure('fisher', bn=True, mode='channel')
|
|
def compute_fisher_per_weight(net, inputs, targets, loss_fn, mode, split_data=1):
|
|
|
|
device = inputs.device
|
|
|
|
if mode == 'param':
|
|
raise ValueError('Fisher pruning does not support parameter pruning.')
|
|
|
|
net.train()
|
|
all_hooks = []
|
|
for layer in net.modules():
|
|
if isinstance(layer, nn.Conv2d) or isinstance(layer, nn.Linear):
|
|
#variables/op needed for fisher computation
|
|
layer.fisher = None
|
|
layer.act = 0.
|
|
layer.dummy = nn.Identity()
|
|
|
|
#replace forward method of conv/linear
|
|
if isinstance(layer, nn.Conv2d):
|
|
layer.forward = types.MethodType(fisher_forward_conv2d, layer)
|
|
if isinstance(layer, nn.Linear):
|
|
layer.forward = types.MethodType(fisher_forward_linear, layer)
|
|
|
|
#function to call during backward pass (hooked on identity op at output of layer)
|
|
def hook_factory(layer):
|
|
def hook(module, grad_input, grad_output):
|
|
act = layer.act.detach()
|
|
grad = grad_output[0].detach()
|
|
if len(act.shape) > 2:
|
|
g_nk = torch.sum((act * grad), list(range(2,len(act.shape))))
|
|
else:
|
|
g_nk = act * grad
|
|
del_k = g_nk.pow(2).mean(0).mul(0.5)
|
|
if layer.fisher is None:
|
|
layer.fisher = del_k
|
|
else:
|
|
layer.fisher += del_k
|
|
del layer.act #without deleting this, a nasty memory leak occurs! related: https://discuss.pytorch.org/t/memory-leak-when-using-forward-hook-and-backward-hook-simultaneously/27555
|
|
return hook
|
|
|
|
#register backward hook on identity fcn to compute fisher info
|
|
layer.dummy.register_backward_hook(hook_factory(layer))
|
|
|
|
N = inputs.shape[0]
|
|
for sp in range(split_data):
|
|
st=sp*N//split_data
|
|
en=(sp+1)*N//split_data
|
|
|
|
net.zero_grad()
|
|
outputs = net(inputs[st:en])
|
|
loss = loss_fn(outputs, targets[st:en])
|
|
loss.backward()
|
|
|
|
# retrieve fisher info
|
|
def fisher(layer):
|
|
if layer.fisher is not None:
|
|
return torch.abs(layer.fisher.detach())
|
|
else:
|
|
return torch.zeros(layer.weight.shape[0]) #size=ch
|
|
|
|
grads_abs_ch = get_layer_metric_array(net, fisher, mode)
|
|
|
|
#broadcast channel value here to all parameters in that channel
|
|
#to be compatible with stuff downstream (which expects per-parameter metrics)
|
|
#TODO cleanup on the selectors/apply_prune_mask side (?)
|
|
shapes = get_layer_metric_array(net, lambda l : l.weight.shape[1:], mode)
|
|
|
|
grads_abs = reshape_elements(grads_abs_ch, shapes, device)
|
|
|
|
return grads_abs
|