MeCo/zero-cost-nas/foresight/pruners/measures/zen.py
HamsterMimi 993d55076e update
2023-05-14 10:57:08 +08:00

111 lines
3.5 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
from torch import nn
import numpy as np
from . import measure
def network_weight_gaussian_init(net: nn.Module):
with torch.no_grad():
for n, m in net.named_modules():
if isinstance(m, nn.Conv2d):
nn.init.normal_(m.weight)
if hasattr(m, 'bias') and m.bias is not None:
nn.init.zeros_(m.bias)
elif isinstance(m, nn.BatchNorm2d):
try:
nn.init.ones_(m.weight)
nn.init.zeros_(m.bias)
except:
pass
elif isinstance(m, nn.Linear):
nn.init.normal_(m.weight)
if hasattr(m, 'bias') and m.bias is not None:
nn.init.zeros_(m.bias)
else:
continue
return net
def get_zen(gpu, model, mixup_gamma=1e-2, resolution=32, batch_size=64, repeat=32,
fp16=False):
info = {}
nas_score_list = []
if gpu is not None:
device = torch.device(gpu)
else:
device = torch.device('cpu')
if fp16:
dtype = torch.half
else:
dtype = torch.float32
with torch.no_grad():
for repeat_count in range(repeat):
network_weight_gaussian_init(model)
input = torch.randn(size=[batch_size, 3, resolution, resolution], device=device, dtype=dtype)
input2 = torch.randn(size=[batch_size, 3, resolution, resolution], device=device, dtype=dtype)
mixup_input = input + mixup_gamma * input2
output = model.forward_pre_GAP(input)
mixup_output = model.forward_pre_GAP(mixup_input)
nas_score = torch.sum(torch.abs(output - mixup_output), dim=[1, 2, 3])
nas_score = torch.mean(nas_score)
# compute BN scaling
log_bn_scaling_factor = 0.0
for m in model.modules():
if isinstance(m, nn.BatchNorm2d):
try:
bn_scaling_factor = torch.sqrt(torch.mean(m.running_var))
log_bn_scaling_factor += torch.log(bn_scaling_factor)
except:
pass
pass
pass
nas_score = torch.log(nas_score) + log_bn_scaling_factor
nas_score_list.append(float(nas_score))
std_nas_score = np.std(nas_score_list)
avg_precision = 1.96 * std_nas_score / np.sqrt(len(nas_score_list))
avg_nas_score = np.mean(nas_score_list)
info = float(avg_nas_score)
return info
@measure('zen', bn=True)
def compute_zen(net, inputs, targets, split_data=1, loss_fn=None):
device = inputs.device
# Compute gradients (but don't apply them)
net.zero_grad()
try:
zen = get_zen(device,net)
except Exception as e:
print(e)
zen= np.nan
return zen