58 lines
1.7 KiB
Python
58 lines
1.7 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 numpy as np
|
|
|
|
from . import measure
|
|
|
|
|
|
def get_batch_jacobian(net, x, target, device, split_data):
|
|
x.requires_grad_(True)
|
|
|
|
N = x.shape[0]
|
|
for sp in range(split_data):
|
|
st=sp*N//split_data
|
|
en=(sp+1)*N//split_data
|
|
y = net(x[st:en])
|
|
y.backward(torch.ones_like(y))
|
|
|
|
jacob = x.grad.detach()
|
|
x.requires_grad_(False)
|
|
return jacob, target.detach()
|
|
|
|
def eval_score(jacob, labels=None):
|
|
corrs = np.corrcoef(jacob)
|
|
v, _ = np.linalg.eig(corrs)
|
|
k = 1e-5
|
|
return -np.sum(np.log(v + k) + 1./(v + k))
|
|
|
|
@measure('jacob_cov', bn=True)
|
|
def compute_jacob_cov(net, inputs, targets, split_data=1, loss_fn=None):
|
|
device = inputs.device
|
|
# Compute gradients (but don't apply them)
|
|
net.zero_grad()
|
|
|
|
jacobs, labels = get_batch_jacobian(net, inputs, targets, device, split_data=split_data)
|
|
jacobs = jacobs.reshape(jacobs.size(0), -1).cpu().numpy()
|
|
|
|
try:
|
|
jc = eval_score(jacobs, labels)
|
|
except Exception as e:
|
|
print(e)
|
|
jc = np.nan
|
|
|
|
return jc
|