def register_slot_builder(builder,
_reg_id=None,
args_fmt=None,
kwargs_fmt=None):
"""Register a builder that converts slot to module."""
_reg_id = _reg_id or builder.__qualname__
register(get_slot_builder(builder, args_fmt, kwargs_fmt), _reg_id)
return builder
def resnext101_32x8d(resnet_cls, **kwargs):
"""Construct a ResNeXt-50 32x8d model."""
return resnet_cls(block=Bottleneck,
layers=[3, 4, 23, 3],
groups=32,
width_per_group=8,
**kwargs)
def resnet(resnet_cls, bottleneck=False, **kwargs):
"""Construct a ResNet model."""
block = Bottleneck if bottleneck else BasicBlock
return resnet_cls(block=block, **kwargs)
for net_cls in [CIFARResNet, ImageNetResNet]:
name = 'CIFAR-' if net_cls == CIFARResNet else 'ImageNet-'
register(partial(resnet10, net_cls), name + 'ResNet-10')
register(partial(resnet18, net_cls), name + 'ResNet-18')
register(partial(resnet32, net_cls), name + 'ResNet-32')
register(partial(resnet34, net_cls), name + 'ResNet-34')
register(partial(resnet50, net_cls), name + 'ResNet-50')
register(partial(resnet56, net_cls), name + 'ResNet-56')
register(partial(resnet101, net_cls), name + 'ResNet-101')
register(partial(resnet152, net_cls), name + 'ResNet-152')
register(partial(resnext50_32x4d, net_cls), name + 'ResNeXt-50')
register(partial(resnext101_32x8d, net_cls), name + 'ResNeXt-101')
register(partial(resnet, net_cls), name + 'ResNet')
[6, 80, 4, 2],
[6, 96, 4, 1],
[6, 192, 4, 2],
[6, 320, 1, 1]
]
def mobilenetv2(cfgs=None, cifar=False, **kwargs):
"""Return MobileNetV2 model."""
if cfgs is None:
cfgs = _mbv2_ori_cfgs
if cifar:
cfgs[0][3] = 1
cfgs[6][3] = 1
return MobileNetV2(cfgs=cfgs, **kwargs)
for cifar in [True, False]:
img = 'CIFAR' if cifar else 'ImageNet'
register(partial(mobilenetv2, cifar=cifar), '{}_MobileNetV2'.format(img))
register(partial(mobilenetv2, cfgs=_mbv2_gpu_cfgs, cifar=cifar), '{}_MobileNetV2_GPU'.format(img))
kernel_sizes = [3, 5, 7, 9]
expand_ratios = [1, 3, 6, 9]
for k in kernel_sizes:
for e in expand_ratios:
p = get_same_padding(k)
builder = partial(MobileInvertedConv, expansion=e, kernel_size=k, padding=p)
register_slot_builder(builder, 'MB{}E{}'.format(k, e), 'i1o1s2')
def __init__(self, inplace=True):
super(HardSigmoid, self).__init__()
self.inplace = inplace
def forward(self, x):
"""Return module output."""
return F.relu6(x + 3., inplace=self.inplace) / 6.
class HardSwish(nn.Module):
"""Hard Swish activation function."""
def __init__(self, inplace=True):
super(HardSwish, self).__init__()
self.inplace = inplace
def forward(self, x):
"""Return module output."""
return x * F.relu6(x + 3., inplace=self.inplace) / 6.
class Swish(nn.Module):
"""Swish activation function."""
def forward(self, x):
"""Return module output."""
return x * F.sigmoid(x)
register(HardSigmoid)
register(HardSwish)
register(Swish)
"""Torch activation functions."""
import torch.nn
from modnas.registry.arch_space import register
modules = [
'ELU',
'Hardshrink',
'Hardtanh',
'LeakyReLU',
'LogSigmoid',
'PReLU',
'ReLU',
'ReLU6',
'RReLU',
'SELU',
'CELU',
'Sigmoid',
'Softplus',
'Softshrink',
'Softsign',
'Tanh',
'Tanhshrink',
'Threshold',
]
for name in modules:
attr = getattr(torch.nn, name, None)
if attr is not None:
register(attr)
"""Return architecture built from config."""
n_nodes = 4
if 'nodes' in kwargs:
n_nodes = kwargs.pop('nodes')
darts_kwargs = {
'cell_cls': DAGLayer,
'cell_kwargs': {
'chn_in': None,
'chn_out': None,
'stride': None,
'n_nodes': n_nodes,
'allocator': 'ReplicateAllocator',
'merger_state': 'SumMerger',
'merger_out': 'ConcatMerger',
'enumerator': 'CombinationEnumerator',
'preproc': None,
'edge_cls': Slot,
'edge_kwargs': {
'_chn_in': None,
'_chn_out': None,
'_stride': None,
},
},
}
darts_kwargs.update(kwargs)
return darts_cls(**darts_kwargs)
register(build_from_config, 'DARTS')
register(partial(build_from_config, ImageNetDARTSLikeNet), 'ImageNet-DARTS')
# -*- coding:utf-8 -*- # Copyright (C) 2020. Huawei Technologies Co., Ltd. All rights reserved. # This program is free software; you can redistribute it and/or modify # it under the terms of the MIT License. # This program is distributed in the hope that it will be useful, # but WITHOUT ANY WARRANTY; without even the implied warranty of # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the # MIT License for more details. """Torch activation functions.""" import torch.nn as nn from modnas.registry.arch_space import register # torch activations register(nn.ELU) register(nn.Hardshrink) register(nn.Hardtanh) register(nn.LeakyReLU) register(nn.LogSigmoid) # register(torch.nn.MultiheadAttention) register(nn.PReLU) register(nn.ReLU) register(nn.ReLU6) register(nn.RReLU) register(nn.SELU) register(nn.CELU) register(nn.Sigmoid) register(nn.Softplus) register(nn.Softshrink) register(nn.Softsign)