update GDAS (TO-FINISH)

This commit is contained in:
D-X-Y 2019-10-16 16:29:57 +11:00
parent 6814816d5f
commit d28826793d
6 changed files with 429 additions and 4 deletions

View File

@ -48,7 +48,7 @@ CUDA_VISIBLE_DEVICES=0,1 bash ./scripts-search/search-cifar.sh cifar10 ResNet56
args: `cifar10` indicates the dataset name, `ResNet56` indicates the basemodel name, `CIFARX` indicates the searching hyper-parameters, `0.47/0.57` indicates the expected FLOP ratio, `-1` indicates the random seed.
## One-Shot Neural Architecture Search via Self-Evaluated Template Network
## [One-Shot Neural Architecture Search via Self-Evaluated Template Network](https://arxiv.org/abs/1910.05733)
<img align="right" src="https://d-x-y.github.com/resources/paper-icon/ICCV-2019-SETN.png" width="450">
@ -67,7 +67,7 @@ CUDA_VISIBLE_DEVICES=0,1,2,3 bash ./scripts/nas-infer-train.sh imagenet-1k SETN
Searching codes come soon!
## [Searching for A Robust Neural Architecture in Four GPU Hours](http://openaccess.thecvf.com/content_CVPR_2019/papers/Dong_Searching_for_a_Robust_Neural_Architecture_in_Four_GPU_Hours_CVPR_2019_paper.pdf)
## [Searching for A Robust Neural Architecture in Four GPU Hours](https://arxiv.org/abs/1910.04465)
<img align="right" src="https://d-x-y.github.com/resources/paper-icon/CVPR-2019-GDAS.png" width="300">

View File

View File

@ -0,0 +1,196 @@
import math, torch
import torch.nn as nn
import torch.nn.functional as F
from copy import deepcopy
from .operations import OPS, ReLUConvBN
class SearchCell(nn.Module):
def __init__(self, C_in, C_out, stride, max_nodes, op_names):
super(SearchCell, self).__init__()
self.op_names = deepcopy(op_names)
self.edges = nn.ModuleDict()
self.max_nodes = max_nodes
self.in_dim = C_in
self.out_dim = C_out
for i in range(1, max_nodes):
for j in range(i):
node_str = '{:}<-{:}'.format(i, j)
if j == 0:
xlists = [OPS[op_name](C_in , C_out, stride) for op_name in op_names]
else:
xlists = [OPS[op_name](C_in , C_out, 1) for op_name in op_names]
self.edges[ node_str ] = nn.ModuleList( xlists )
self.edge_keys = sorted(list(self.edges.keys()))
self.edge2index = {key:i for i, key in enumerate(self.edge_keys)}
self.num_edges = len(self.edges)
def extra_repr(self):
string = 'info :: {max_nodes} nodes, inC={in_dim}, outC={out_dim}'.format(**self.__dict__)
return string
def forward(self, inputs, weightss):
nodes = [inputs]
for i in range(1, self.max_nodes):
inter_nodes = []
for j in range(i):
node_str = '{:}<-{:}'.format(i, j)
weights = weightss[ self.edge2index[node_str] ]
inter_nodes.append( sum( layer(nodes[j]) * w for layer, w in zip(self.edges[node_str], weights) ) )
nodes.append( sum(inter_nodes) )
return nodes[-1]
# GDAS
def forward_acc(self, inputs, weightss, indexess):
nodes = [inputs]
for i in range(1, self.max_nodes):
inter_nodes = []
for j in range(i):
node_str = '{:}<-{:}'.format(i, j)
weights = weightss[ self.edge2index[node_str] ]
indexes = indexess[ self.edge2index[node_str] ].item()
import pdb; pdb.set_trace() # to-do
#inter_nodes.append( self.edges[node_str][indexes](nodes[j]) * weights[indexes] )
nodes.append( sum(inter_nodes) )
return nodes[-1]
# joint
def forward_joint(self, inputs, weightss):
nodes = [inputs]
for i in range(1, self.max_nodes):
inter_nodes = []
for j in range(i):
node_str = '{:}<-{:}'.format(i, j)
weights = weightss[ self.edge2index[node_str] ]
aggregation = sum( layer(nodes[j]) * w for layer, w in zip(self.edges[node_str], weights) ) / weights.numel()
inter_nodes.append( aggregation )
nodes.append( sum(inter_nodes) )
return nodes[-1]
# uniform random sampling per iteration
def forward_urs(self, inputs):
nodes = [inputs]
for i in range(1, self.max_nodes):
while True: # to avoid select zero for all ops
sops, has_non_zero = [], False
for j in range(i):
node_str = '{:}<-{:}'.format(i, j)
candidates = self.edges[node_str]
select_op = random.choice(candidates)
sops.append( select_op )
if not hasattr(select_op, 'is_zero') or select_op.is_zero == False: has_non_zero=True
if has_non_zero: break
inter_nodes = []
for j, select_op in enumerate(sops):
inter_nodes.append( select_op(nodes[j]) )
nodes.append( sum(inter_nodes) )
return nodes[-1]
# select the argmax
def forward_select(self, inputs, weightss):
nodes = [inputs]
for i in range(1, self.max_nodes):
inter_nodes = []
for j in range(i):
node_str = '{:}<-{:}'.format(i, j)
weights = weightss[ self.edge2index[node_str] ]
inter_nodes.append( self.edges[node_str][ weights.argmax().item() ]( nodes[j] ) )
#inter_nodes.append( sum( layer(nodes[j]) * w for layer, w in zip(self.edges[node_str], weights) ) )
nodes.append( sum(inter_nodes) )
return nodes[-1]
# select the argmax
def forward_dynamic(self, inputs, structure):
nodes = [inputs]
for i in range(1, self.max_nodes):
cur_op_node = structure.nodes[i-1]
inter_nodes = []
for op_name, j in cur_op_node:
node_str = '{:}<-{:}'.format(i, j)
op_index = self.op_names.index( op_name )
inter_nodes.append( self.edges[node_str][op_index]( nodes[j] ) )
nodes.append( sum(inter_nodes) )
return nodes[-1]
class InferCell(nn.Module):
def __init__(self, genotype, C_in, C_out, stride):
super(InferCell, self).__init__()
self.layers = nn.ModuleList()
self.node_IN = []
self.node_IX = []
self.genotype = deepcopy(genotype)
for i in range(1, len(genotype)):
node_info = genotype[i-1]
cur_index = []
cur_innod = []
for (op_name, op_in) in node_info:
if op_in == 0:
layer = OPS[op_name](C_in , C_out, stride)
else:
layer = OPS[op_name](C_out, C_out, 1)
cur_index.append( len(self.layers) )
cur_innod.append( op_in )
self.layers.append( layer )
self.node_IX.append( cur_index )
self.node_IN.append( cur_innod )
self.nodes = len(genotype)
self.in_dim = C_in
self.out_dim = C_out
def extra_repr(self):
string = 'info :: nodes={nodes}, inC={in_dim}, outC={out_dim}'.format(**self.__dict__)
laystr = []
for i, (node_layers, node_innods) in enumerate(zip(self.node_IX,self.node_IN)):
y = ['I{:}-L{:}'.format(_ii, _il) for _il, _ii in zip(node_layers, node_innods)]
x = '{:}<-({:})'.format(i+1, ','.join(y))
laystr.append( x )
return string + ', [{:}]'.format( ' | '.join(laystr) ) + ', {:}'.format(self.genotype.tostr())
def forward(self, inputs):
nodes = [inputs]
for i, (node_layers, node_innods) in enumerate(zip(self.node_IX,self.node_IN)):
node_feature = sum( self.layers[_il](nodes[_ii]) for _il, _ii in zip(node_layers, node_innods) )
nodes.append( node_feature )
return nodes[-1]
class ResNetBasicblock(nn.Module):
def __init__(self, inplanes, planes, stride):
super(ResNetBasicblock, self).__init__()
assert stride == 1 or stride == 2, 'invalid stride {:}'.format(stride)
self.conv_a = ReLUConvBN(inplanes, planes, 3, stride, 1, 1)
self.conv_b = ReLUConvBN( planes, planes, 3, 1, 1, 1)
if stride == 2:
self.downsample = nn.Sequential(
nn.AvgPool2d(kernel_size=2, stride=2, padding=0),
nn.Conv2d(inplanes, planes, kernel_size=1, stride=1, padding=0, bias=False))
elif inplanes != planes:
self.downsample = ReLUConvBN(inplanes, planes, 1, 1, 0, 1)
else:
self.downsample = None
self.in_dim = inplanes
self.out_dim = planes
self.stride = stride
self.num_conv = 2
def extra_repr(self):
string = '{name}(inC={in_dim}, outC={out_dim}, stride={stride})'.format(name=self.__class__.__name__, **self.__dict__)
return string
def forward(self, inputs):
basicblock = self.conv_a(inputs)
basicblock = self.conv_b(basicblock)
if self.downsample is not None:
residual = self.downsample(inputs)
else:
residual = inputs
return residual + basicblock

View File

@ -0,0 +1,113 @@
import torch
import torch.nn as nn
__all__ = ['OPS', 'ReLUConvBN', 'SearchSpaceNames']
OPS = {
'none' : lambda C_in, C_out, stride: Zero(C_in, C_out, stride),
'avg_pool_3x3' : lambda C_in, C_out, stride: POOLING(C_in, C_out, stride, 'avg'),
'max_pool_3x3' : lambda C_in, C_out, stride: POOLING(C_in, C_out, stride, 'max'),
'nor_conv_7x7' : lambda C_in, C_out, stride: ReLUConvBN(C_in, C_out, (7,7), (stride,stride), (3,3), (1,1)),
'nor_conv_3x3' : lambda C_in, C_out, stride: ReLUConvBN(C_in, C_out, (3,3), (stride,stride), (1,1), (1,1)),
'nor_conv_1x1' : lambda C_in, C_out, stride: ReLUConvBN(C_in, C_out, (1,1), (stride,stride), (0,0), (1,1)),
'skip_connect' : lambda C_in, C_out, stride: Identity() if stride == 1 and C_in == C_out else FactorizedReduce(C_in, C_out, stride),
}
CONNECT_NAS_BENCHMARK = ['none', 'skip_connect', 'nor_conv_3x3']
SearchSpaceNames = {'connect-nas' : CONNECT_NAS_BENCHMARK}
class POOLING(nn.Module):
def __init__(self, C_in, C_out, stride, mode):
super(POOLING, self).__init__()
if C_in == C_out:
self.preprocess = None
else:
self.preprocess = ReLUConvBN(C_in, C_out, 1, 1, 0)
if mode == 'avg' : self.op = nn.AvgPool2d(3, stride=stride, padding=1, count_include_pad=False)
elif mode == 'max': self.op = nn.MaxPool2d(3, stride=stride, padding=1)
else : raise ValueError('Invalid mode={:} in POOLING'.format(mode))
def forward(self, inputs):
if self.preprocess: x = self.preprocess(inputs)
else : x = inputs
return self.op(x)
class ReLUConvBN(nn.Module):
def __init__(self, C_in, C_out, kernel_size, stride, padding, dilation):
super(ReLUConvBN, self).__init__()
self.op = nn.Sequential(
nn.ReLU(inplace=False),
nn.Conv2d(C_in, C_out, kernel_size, stride=stride, padding=padding, dilation=dilation, bias=False),
nn.BatchNorm2d(C_out)
)
def forward(self, x):
return self.op(x)
class Identity(nn.Module):
def __init__(self):
super(Identity, self).__init__()
def forward(self, x):
return x
class Zero(nn.Module):
def __init__(self, C_in, C_out, stride):
super(Zero, self).__init__()
self.C_in = C_in
self.C_out = C_out
self.stride = stride
self.is_zero = True
def forward(self, x):
if self.C_in == self.C_out:
if self.stride == 1: return x.mul(0.)
else : return x[:,:,::self.stride,::self.stride].mul(0.)
else:
shape = list(x.shape)
shape[1] = self.C_out
zeros = x.new_zeros(shape, dtype=x.dtype, device=x.device)
return zeros
def extra_repr(self):
return 'C_in={C_in}, C_out={C_out}, stride={stride}'.format(**self.__dict__)
class FactorizedReduce(nn.Module):
def __init__(self, C_in, C_out, stride):
super(FactorizedReduce, self).__init__()
self.stride = stride
self.C_in = C_in
self.C_out = C_out
self.relu = nn.ReLU(inplace=False)
if stride == 2:
#assert C_out % 2 == 0, 'C_out : {:}'.format(C_out)
C_outs = [C_out // 2, C_out - C_out // 2]
self.convs = nn.ModuleList()
for i in range(2):
self.convs.append( nn.Conv2d(C_in, C_outs[i], 1, stride=stride, padding=0, bias=False) )
self.pad = nn.ConstantPad2d((0, 1, 0, 1), 0)
else:
raise ValueError('Invalid stride : {:}'.format(stride))
self.bn = nn.BatchNorm2d(C_out)
def forward(self, x):
x = self.relu(x)
y = self.pad(x)
out = torch.cat([self.convs[0](x), self.convs[1](y[:,:,1:,1:])], dim=1)
out = self.bn(out)
return out
def extra_repr(self):
return 'C_in={C_in}, C_out={C_out}, stride={stride}'.format(**self.__dict__)

View File

@ -0,0 +1,117 @@
##################################################
# Copyright (c) Xuanyi Dong [GitHub D-X-Y], 2019 #
###########################################################################
# Searching for A Robust Neural Architecture in Four GPU Hours, CVPR 2019 #
###########################################################################
import torch
import torch.nn as nn
from copy import deepcopy
from .infer_cells import ResNetBasicblock
from .search_cells import SearchCell
from .genotypes import Structure
class TinyNetworkGDAS(nn.Module):
def __init__(self, C, N, max_nodes, num_classes, search_space):
super(TinyNetworkGDAS, self).__init__()
self._C = C
self._layerN = N
self.max_nodes = max_nodes
self.stem = nn.Sequential(
nn.Conv2d(3, C, kernel_size=3, padding=1, bias=False),
nn.BatchNorm2d(C))
layer_channels = [C ] * N + [C*2 ] + [C*2 ] * N + [C*4 ] + [C*4 ] * N
layer_reductions = [False] * N + [True] + [False] * N + [True] + [False] * N
C_prev, num_edge, edge2index = C, None, None
self.cells = nn.ModuleList()
for index, (C_curr, reduction) in enumerate(zip(layer_channels, layer_reductions)):
if reduction:
cell = ResNetBasicblock(C_prev, C_curr, 2)
else:
cell = SearchCell(C_prev, C_curr, 1, max_nodes, search_space)
if num_edge is None: num_edge, edge2index = cell.num_edges, cell.edge2index
else: assert num_edge == cell.num_edges and edge2index == cell.edge2index, 'invalid {:} vs. {:}.'.format(num_edge, cell.num_edges)
self.cells.append( cell )
C_prev = cell.out_dim
self.op_names = deepcopy( search_space )
self._Layer = len(self.cells)
self.edge2index = edge2index
self.lastact = nn.Sequential(nn.BatchNorm2d(C_prev), nn.ReLU(inplace=True))
self.global_pooling = nn.AdaptiveAvgPool2d(1)
self.classifier = nn.Linear(C_prev, num_classes)
self.arch_parameters = nn.Parameter( 1e-3*torch.randn(num_edge, len(search_space)) )
self.tau = 10
self.nan_count = 0
def get_weights(self):
xlist = list( self.stem.parameters() ) + list( self.cells.parameters() )
xlist+= list( self.lastact.parameters() ) + list( self.global_pooling.parameters() )
xlist+= list( self.classifier.parameters() )
return xlist
def set_tau(self, tau, _nan_count=0):
self.tau = tau
self.nan_count = _nan_count
def get_tau(self):
return self.tau
def get_alphas(self):
return [self.arch_parameters]
def get_message(self):
string = self.extra_repr()
for i, cell in enumerate(self.cells):
string += '\n {:02d}/{:02d} :: {:}'.format(i, len(self.cells), cell.extra_repr())
return string
def extra_repr(self):
return ('{name}(C={_C}, Max-Nodes={max_nodes}, N={_layerN}, L={_Layer})'.format(name=self.__class__.__name__, **self.__dict__))
def genotype(self):
genotypes = []
for i in range(1, self.max_nodes):
xlist = []
for j in range(i):
node_str = '{:}<-{:}'.format(i, j)
with torch.no_grad():
weights = self.arch_parameters[ self.edge2index[node_str] ]
op_name = self.op_names[ weights.argmax().item() ]
xlist.append((op_name, j))
genotypes.append( tuple(xlist) )
return Structure( genotypes )
def forward(self, inputs):
def gumbel_softmax(_logits, _tau):
while True: # a trick to avoid the gumbels bug
gumbels = -torch.empty_like(_logits).exponential_().log()
new_logits = (_logits.log_softmax(dim=1) + gumbels) / _tau
probs = nn.functional.softmax(new_logits, dim=1)
index = probs.max(-1, keepdim=True)[1]
if index[0].item() == self.op_names.index('none') and index[3].item() == self.op_names.index('none') and index[5].item() == self.op_names.index('none'): continue
if index[1].item() == self.op_names.index('none') and index[2].item() == self.op_names.index('none') and index[3].item() == self.op_names.index('none') and index[4].item() == self.op_names.index('none'): continue
if index[3].item() == self.op_names.index('none') and index[4].item() == self.op_names.index('none') and index[5].item() == self.op_names.index('none'): continue
if index[3].item() == self.op_names.index('none') and index[0].item() == self.op_names.index('none') and index[1].item() == self.op_names.index('none'): continue
one_h = torch.zeros_like(_logits).scatter_(-1, index, 1.0)
xres = one_h - probs.detach() + probs
if (not torch.isinf(gumbels).any()) and (not torch.isinf(probs).any()) and (not torch.isnan(probs).any()): break
self.nan_count += 1
return xres, index
feature = self.stem(inputs)
for i, cell in enumerate(self.cells):
if isinstance(cell, SearchCell):
alphas, IDX = gumbel_softmax(self.arch_parameters, self.tau)
feature = cell.forward_gdas(feature, alphas, IDX.cpu())
else:
feature = cell(feature)
out = self.lastact(feature)
out = self.global_pooling( out )
out = out.view(out.size(0), -1)
logits = self.classifier(out)
return out, logits

View File

@ -9,8 +9,7 @@ def select2withP(logits, tau, just_prob=False, num=2, eps=1e-7):
else :
while True: # a trick to avoid the gumbels bug
gumbels = -torch.empty_like(logits).exponential_().log()
new_logits = (logits + gumbels) / tau
#new_logits = (logits.log_softmax(dim=1) + gumbels) / tau
new_logits = (logits.log_softmax(dim=1) + gumbels) / tau
probs = nn.functional.softmax(new_logits, dim=1)
if (not torch.isinf(gumbels).any()) and (not torch.isinf(probs).any()) and (not torch.isnan(probs).any()): break