naswot/models/shape_searchs/SearchImagenetResNet.py
2020-06-03 12:59:01 +01:00

483 lines
22 KiB
Python

import math, torch
from collections import OrderedDict
from bisect import bisect_right
import torch.nn as nn
from ..initialization import initialize_resnet
from ..SharedUtils import additive_func
from .SoftSelect import select2withP, ChannelWiseInter
from .SoftSelect import linear_forward
from .SoftSelect import get_width_choices
def get_depth_choices(layers):
min_depth = min(layers)
info = {'num': min_depth}
for i, depth in enumerate(layers):
choices = []
for j in range(1, min_depth+1):
choices.append( int( float(depth)*j/min_depth ) )
info[i] = choices
return info
def conv_forward(inputs, conv, choices):
iC = conv.in_channels
fill_size = list(inputs.size())
fill_size[1] = iC - fill_size[1]
filled = torch.zeros(fill_size, device=inputs.device)
xinputs = torch.cat((inputs, filled), dim=1)
outputs = conv(xinputs)
selecteds = [outputs[:,:oC] for oC in choices]
return selecteds
class ConvBNReLU(nn.Module):
num_conv = 1
def __init__(self, nIn, nOut, kernel, stride, padding, bias, has_avg, has_bn, has_relu, last_max_pool=False):
super(ConvBNReLU, self).__init__()
self.InShape = None
self.OutShape = None
self.choices = get_width_choices(nOut)
self.register_buffer('choices_tensor', torch.Tensor( self.choices ))
if has_avg : self.avg = nn.AvgPool2d(kernel_size=2, stride=2, padding=0)
else : self.avg = None
self.conv = nn.Conv2d(nIn, nOut, kernel_size=kernel, stride=stride, padding=padding, dilation=1, groups=1, bias=bias)
#if has_bn : self.bn = nn.BatchNorm2d(nOut)
#else : self.bn = None
self.has_bn = has_bn
self.BNs = nn.ModuleList()
for i, _out in enumerate(self.choices):
self.BNs.append(nn.BatchNorm2d(_out))
if has_relu: self.relu = nn.ReLU(inplace=True)
else : self.relu = None
if last_max_pool: self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
else : self.maxpool = None
self.in_dim = nIn
self.out_dim = nOut
self.search_mode = 'basic'
def get_flops(self, channels, check_range=True, divide=1):
iC, oC = channels
if check_range: assert iC <= self.conv.in_channels and oC <= self.conv.out_channels, '{:} vs {:} | {:} vs {:}'.format(iC, self.conv.in_channels, oC, self.conv.out_channels)
assert isinstance(self.InShape, tuple) and len(self.InShape) == 2, 'invalid in-shape : {:}'.format(self.InShape)
assert isinstance(self.OutShape, tuple) and len(self.OutShape) == 2, 'invalid out-shape : {:}'.format(self.OutShape)
#conv_per_position_flops = self.conv.kernel_size[0] * self.conv.kernel_size[1] * iC * oC / self.conv.groups
conv_per_position_flops = (self.conv.kernel_size[0] * self.conv.kernel_size[1] * 1.0 / self.conv.groups)
all_positions = self.OutShape[0] * self.OutShape[1]
flops = (conv_per_position_flops * all_positions / divide) * iC * oC
if self.conv.bias is not None: flops += all_positions / divide
return flops
def get_range(self):
return [self.choices]
def forward(self, inputs):
if self.search_mode == 'basic':
return self.basic_forward(inputs)
elif self.search_mode == 'search':
return self.search_forward(inputs)
else:
raise ValueError('invalid search_mode = {:}'.format(self.search_mode))
def search_forward(self, tuple_inputs):
assert isinstance(tuple_inputs, tuple) and len(tuple_inputs) == 5, 'invalid type input : {:}'.format( type(tuple_inputs) )
inputs, expected_inC, probability, index, prob = tuple_inputs
index, prob = torch.squeeze(index).tolist(), torch.squeeze(prob)
probability = torch.squeeze(probability)
assert len(index) == 2, 'invalid length : {:}'.format(index)
# compute expected flop
#coordinates = torch.arange(self.x_range[0], self.x_range[1]+1).type_as(probability)
expected_outC = (self.choices_tensor * probability).sum()
expected_flop = self.get_flops([expected_inC, expected_outC], False, 1e6)
if self.avg : out = self.avg( inputs )
else : out = inputs
# convolutional layer
out_convs = conv_forward(out, self.conv, [self.choices[i] for i in index])
out_bns = [self.BNs[idx](out_conv) for idx, out_conv in zip(index, out_convs)]
# merge
out_channel = max([x.size(1) for x in out_bns])
outA = ChannelWiseInter(out_bns[0], out_channel)
outB = ChannelWiseInter(out_bns[1], out_channel)
out = outA * prob[0] + outB * prob[1]
#out = additive_func(out_bns[0]*prob[0], out_bns[1]*prob[1])
if self.relu : out = self.relu( out )
if self.maxpool: out = self.maxpool(out)
return out, expected_outC, expected_flop
def basic_forward(self, inputs):
if self.avg : out = self.avg( inputs )
else : out = inputs
conv = self.conv( out )
if self.has_bn:out= self.BNs[-1]( conv )
else : out = conv
if self.relu: out = self.relu( out )
else : out = out
if self.InShape is None:
self.InShape = (inputs.size(-2), inputs.size(-1))
self.OutShape = (out.size(-2) , out.size(-1))
if self.maxpool: out = self.maxpool(out)
return out
class ResNetBasicblock(nn.Module):
expansion = 1
num_conv = 2
def __init__(self, inplanes, planes, stride):
super(ResNetBasicblock, self).__init__()
assert stride == 1 or stride == 2, 'invalid stride {:}'.format(stride)
self.conv_a = ConvBNReLU(inplanes, planes, 3, stride, 1, False, has_avg=False, has_bn=True, has_relu=True)
self.conv_b = ConvBNReLU( planes, planes, 3, 1, 1, False, has_avg=False, has_bn=True, has_relu=False)
if stride == 2:
self.downsample = ConvBNReLU(inplanes, planes, 1, 1, 0, False, has_avg=True, has_bn=True, has_relu=False)
elif inplanes != planes:
self.downsample = ConvBNReLU(inplanes, planes, 1, 1, 0, False, has_avg=False,has_bn=True, has_relu=False)
else:
self.downsample = None
self.out_dim = planes
self.search_mode = 'basic'
def get_range(self):
return self.conv_a.get_range() + self.conv_b.get_range()
def get_flops(self, channels):
assert len(channels) == 3, 'invalid channels : {:}'.format(channels)
flop_A = self.conv_a.get_flops([channels[0], channels[1]])
flop_B = self.conv_b.get_flops([channels[1], channels[2]])
if hasattr(self.downsample, 'get_flops'):
flop_C = self.downsample.get_flops([channels[0], channels[-1]])
else:
flop_C = 0
if channels[0] != channels[-1] and self.downsample is None: # this short-cut will be added during the infer-train
flop_C = channels[0] * channels[-1] * self.conv_b.OutShape[0] * self.conv_b.OutShape[1]
return flop_A + flop_B + flop_C
def forward(self, inputs):
if self.search_mode == 'basic' : return self.basic_forward(inputs)
elif self.search_mode == 'search': return self.search_forward(inputs)
else: raise ValueError('invalid search_mode = {:}'.format(self.search_mode))
def search_forward(self, tuple_inputs):
assert isinstance(tuple_inputs, tuple) and len(tuple_inputs) == 5, 'invalid type input : {:}'.format( type(tuple_inputs) )
inputs, expected_inC, probability, indexes, probs = tuple_inputs
assert indexes.size(0) == 2 and probs.size(0) == 2 and probability.size(0) == 2
#import pdb; pdb.set_trace()
out_a, expected_inC_a, expected_flop_a = self.conv_a( (inputs, expected_inC , probability[0], indexes[0], probs[0]) )
out_b, expected_inC_b, expected_flop_b = self.conv_b( (out_a , expected_inC_a, probability[1], indexes[1], probs[1]) )
if self.downsample is not None:
residual, _, expected_flop_c = self.downsample( (inputs, expected_inC , probability[1], indexes[1], probs[1]) )
else:
residual, expected_flop_c = inputs, 0
out = additive_func(residual, out_b)
return nn.functional.relu(out, inplace=True), expected_inC_b, sum([expected_flop_a, expected_flop_b, expected_flop_c])
def basic_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
out = additive_func(residual, basicblock)
return nn.functional.relu(out, inplace=True)
class ResNetBottleneck(nn.Module):
expansion = 4
num_conv = 3
def __init__(self, inplanes, planes, stride):
super(ResNetBottleneck, self).__init__()
assert stride == 1 or stride == 2, 'invalid stride {:}'.format(stride)
self.conv_1x1 = ConvBNReLU(inplanes, planes, 1, 1, 0, False, has_avg=False, has_bn=True, has_relu=True)
self.conv_3x3 = ConvBNReLU( planes, planes, 3, stride, 1, False, has_avg=False, has_bn=True, has_relu=True)
self.conv_1x4 = ConvBNReLU(planes, planes*self.expansion, 1, 1, 0, False, has_avg=False, has_bn=True, has_relu=False)
if stride == 2:
self.downsample = ConvBNReLU(inplanes, planes*self.expansion, 1, 1, 0, False, has_avg=True, has_bn=True, has_relu=False)
elif inplanes != planes*self.expansion:
self.downsample = ConvBNReLU(inplanes, planes*self.expansion, 1, 1, 0, False, has_avg=False,has_bn=True, has_relu=False)
else:
self.downsample = None
self.out_dim = planes * self.expansion
self.search_mode = 'basic'
def get_range(self):
return self.conv_1x1.get_range() + self.conv_3x3.get_range() + self.conv_1x4.get_range()
def get_flops(self, channels):
assert len(channels) == 4, 'invalid channels : {:}'.format(channels)
flop_A = self.conv_1x1.get_flops([channels[0], channels[1]])
flop_B = self.conv_3x3.get_flops([channels[1], channels[2]])
flop_C = self.conv_1x4.get_flops([channels[2], channels[3]])
if hasattr(self.downsample, 'get_flops'):
flop_D = self.downsample.get_flops([channels[0], channels[-1]])
else:
flop_D = 0
if channels[0] != channels[-1] and self.downsample is None: # this short-cut will be added during the infer-train
flop_D = channels[0] * channels[-1] * self.conv_1x4.OutShape[0] * self.conv_1x4.OutShape[1]
return flop_A + flop_B + flop_C + flop_D
def forward(self, inputs):
if self.search_mode == 'basic' : return self.basic_forward(inputs)
elif self.search_mode == 'search': return self.search_forward(inputs)
else: raise ValueError('invalid search_mode = {:}'.format(self.search_mode))
def basic_forward(self, inputs):
bottleneck = self.conv_1x1(inputs)
bottleneck = self.conv_3x3(bottleneck)
bottleneck = self.conv_1x4(bottleneck)
if self.downsample is not None: residual = self.downsample(inputs)
else : residual = inputs
out = additive_func(residual, bottleneck)
return nn.functional.relu(out, inplace=True)
def search_forward(self, tuple_inputs):
assert isinstance(tuple_inputs, tuple) and len(tuple_inputs) == 5, 'invalid type input : {:}'.format( type(tuple_inputs) )
inputs, expected_inC, probability, indexes, probs = tuple_inputs
assert indexes.size(0) == 3 and probs.size(0) == 3 and probability.size(0) == 3
out_1x1, expected_inC_1x1, expected_flop_1x1 = self.conv_1x1( (inputs, expected_inC , probability[0], indexes[0], probs[0]) )
out_3x3, expected_inC_3x3, expected_flop_3x3 = self.conv_3x3( (out_1x1,expected_inC_1x1, probability[1], indexes[1], probs[1]) )
out_1x4, expected_inC_1x4, expected_flop_1x4 = self.conv_1x4( (out_3x3,expected_inC_3x3, probability[2], indexes[2], probs[2]) )
if self.downsample is not None:
residual, _, expected_flop_c = self.downsample( (inputs, expected_inC , probability[2], indexes[2], probs[2]) )
else:
residual, expected_flop_c = inputs, 0
out = additive_func(residual, out_1x4)
return nn.functional.relu(out, inplace=True), expected_inC_1x4, sum([expected_flop_1x1, expected_flop_3x3, expected_flop_1x4, expected_flop_c])
class SearchShapeImagenetResNet(nn.Module):
def __init__(self, block_name, layers, deep_stem, num_classes):
super(SearchShapeImagenetResNet, self).__init__()
#Model type specifies number of layers for CIFAR-10 and CIFAR-100 model
if block_name == 'BasicBlock':
block = ResNetBasicblock
elif block_name == 'Bottleneck':
block = ResNetBottleneck
else:
raise ValueError('invalid block : {:}'.format(block_name))
self.message = 'SearchShapeCifarResNet : Depth : {:} , Layers for each block : {:}'.format(sum(layers)*block.num_conv, layers)
self.num_classes = num_classes
if not deep_stem:
self.layers = nn.ModuleList( [ ConvBNReLU(3, 64, 7, 2, 3, False, has_avg=False, has_bn=True, has_relu=True, last_max_pool=True) ] )
self.channels = [64]
else:
self.layers = nn.ModuleList( [ ConvBNReLU(3, 32, 3, 2, 1, False, has_avg=False, has_bn=True, has_relu=True)
,ConvBNReLU(32,64, 3, 1, 1, False, has_avg=False, has_bn=True, has_relu=True, last_max_pool=True) ] )
self.channels = [32, 64]
meta_depth_info = get_depth_choices(layers)
self.InShape = None
self.depth_info = OrderedDict()
self.depth_at_i = OrderedDict()
for stage, layer_blocks in enumerate(layers):
cur_block_choices = meta_depth_info[stage]
assert cur_block_choices[-1] == layer_blocks, 'stage={:}, {:} vs {:}'.format(stage, cur_block_choices, layer_blocks)
block_choices, xstart = [], len(self.layers)
for iL in range(layer_blocks):
iC = self.channels[-1]
planes = 64 * (2**stage)
stride = 2 if stage > 0 and iL == 0 else 1
module = block(iC, planes, stride)
self.channels.append( module.out_dim )
self.layers.append ( module )
self.message += "\nstage={:}, ilayer={:02d}/{:02d}, block={:03d}, iC={:3d}, oC={:3d}, stride={:}".format(stage, iL, layer_blocks, len(self.layers)-1, iC, module.out_dim, stride)
# added for depth
layer_index = len(self.layers) - 1
if iL + 1 in cur_block_choices: block_choices.append( layer_index )
if iL + 1 == layer_blocks:
self.depth_info[layer_index] = {'choices': block_choices,
'stage' : stage,
'xstart' : xstart}
self.depth_info_list = []
for xend, info in self.depth_info.items():
self.depth_info_list.append( (xend, info) )
xstart, xstage = info['xstart'], info['stage']
for ilayer in range(xstart, xend+1):
idx = bisect_right(info['choices'], ilayer-1)
self.depth_at_i[ilayer] = (xstage, idx)
self.avgpool = nn.AdaptiveAvgPool2d((1,1))
self.classifier = nn.Linear(module.out_dim, num_classes)
self.InShape = None
self.tau = -1
self.search_mode = 'basic'
#assert sum(x.num_conv for x in self.layers) + 1 == depth, 'invalid depth check {:} vs {:}'.format(sum(x.num_conv for x in self.layers)+1, depth)
# parameters for width
self.Ranges = []
self.layer2indexRange = []
for i, layer in enumerate(self.layers):
start_index = len(self.Ranges)
self.Ranges += layer.get_range()
self.layer2indexRange.append( (start_index, len(self.Ranges)) )
self.register_parameter('width_attentions', nn.Parameter(torch.Tensor(len(self.Ranges), get_width_choices(None))))
self.register_parameter('depth_attentions', nn.Parameter(torch.Tensor(len(layers), meta_depth_info['num'])))
nn.init.normal_(self.width_attentions, 0, 0.01)
nn.init.normal_(self.depth_attentions, 0, 0.01)
self.apply(initialize_resnet)
def arch_parameters(self, LR=None):
if LR is None:
return [self.width_attentions, self.depth_attentions]
else:
return [
{"params": self.width_attentions, "lr": LR},
{"params": self.depth_attentions, "lr": LR},
]
def base_parameters(self):
return list(self.layers.parameters()) + list(self.avgpool.parameters()) + list(self.classifier.parameters())
def get_flop(self, mode, config_dict, extra_info):
if config_dict is not None: config_dict = config_dict.copy()
# select channels
channels = [3]
for i, weight in enumerate(self.width_attentions):
if mode == 'genotype':
with torch.no_grad():
probe = nn.functional.softmax(weight, dim=0)
C = self.Ranges[i][ torch.argmax(probe).item() ]
else:
raise ValueError('invalid mode : {:}'.format(mode))
channels.append( C )
# select depth
if mode == 'genotype':
with torch.no_grad():
depth_probs = nn.functional.softmax(self.depth_attentions, dim=1)
choices = torch.argmax(depth_probs, dim=1).cpu().tolist()
else:
raise ValueError('invalid mode : {:}'.format(mode))
selected_layers = []
for choice, xvalue in zip(choices, self.depth_info_list):
xtemp = xvalue[1]['choices'][choice] - xvalue[1]['xstart'] + 1
selected_layers.append(xtemp)
flop = 0
for i, layer in enumerate(self.layers):
s, e = self.layer2indexRange[i]
xchl = tuple( channels[s:e+1] )
if i in self.depth_at_i:
xstagei, xatti = self.depth_at_i[i]
if xatti <= choices[xstagei]: # leave this depth
flop+= layer.get_flops(xchl)
else:
flop+= 0 # do not use this layer
else:
flop+= layer.get_flops(xchl)
# the last fc layer
flop += channels[-1] * self.classifier.out_features
if config_dict is None:
return flop / 1e6
else:
config_dict['xchannels'] = channels
config_dict['xblocks'] = selected_layers
config_dict['super_type'] = 'infer-shape'
config_dict['estimated_FLOP'] = flop / 1e6
return flop / 1e6, config_dict
def get_arch_info(self):
string = "for depth and width, there are {:} + {:} attention probabilities.".format(len(self.depth_attentions), len(self.width_attentions))
string+= '\n{:}'.format(self.depth_info)
discrepancy = []
with torch.no_grad():
for i, att in enumerate(self.depth_attentions):
prob = nn.functional.softmax(att, dim=0)
prob = prob.cpu() ; selc = prob.argmax().item() ; prob = prob.tolist()
prob = ['{:.3f}'.format(x) for x in prob]
xstring = '{:03d}/{:03d}-th : {:}'.format(i, len(self.depth_attentions), ' '.join(prob))
logt = ['{:.4f}'.format(x) for x in att.cpu().tolist()]
xstring += ' || {:17s}'.format(' '.join(logt))
prob = sorted( [float(x) for x in prob] )
disc = prob[-1] - prob[-2]
xstring += ' || discrepancy={:.2f} || select={:}/{:}'.format(disc, selc, len(prob))
discrepancy.append( disc )
string += '\n{:}'.format(xstring)
string += '\n-----------------------------------------------'
for i, att in enumerate(self.width_attentions):
prob = nn.functional.softmax(att, dim=0)
prob = prob.cpu() ; selc = prob.argmax().item() ; prob = prob.tolist()
prob = ['{:.3f}'.format(x) for x in prob]
xstring = '{:03d}/{:03d}-th : {:}'.format(i, len(self.width_attentions), ' '.join(prob))
logt = ['{:.3f}'.format(x) for x in att.cpu().tolist()]
xstring += ' || {:52s}'.format(' '.join(logt))
prob = sorted( [float(x) for x in prob] )
disc = prob[-1] - prob[-2]
xstring += ' || dis={:.2f} || select={:}/{:}'.format(disc, selc, len(prob))
discrepancy.append( disc )
string += '\n{:}'.format(xstring)
return string, discrepancy
def set_tau(self, tau_max, tau_min, epoch_ratio):
assert epoch_ratio >= 0 and epoch_ratio <= 1, 'invalid epoch-ratio : {:}'.format(epoch_ratio)
tau = tau_min + (tau_max-tau_min) * (1 + math.cos(math.pi * epoch_ratio)) / 2
self.tau = tau
def get_message(self):
return self.message
def forward(self, inputs):
if self.search_mode == 'basic':
return self.basic_forward(inputs)
elif self.search_mode == 'search':
return self.search_forward(inputs)
else:
raise ValueError('invalid search_mode = {:}'.format(self.search_mode))
def search_forward(self, inputs):
flop_width_probs = nn.functional.softmax(self.width_attentions, dim=1)
flop_depth_probs = nn.functional.softmax(self.depth_attentions, dim=1)
flop_depth_probs = torch.flip( torch.cumsum( torch.flip(flop_depth_probs, [1]), 1 ), [1] )
selected_widths, selected_width_probs = select2withP(self.width_attentions, self.tau)
selected_depth_probs = select2withP(self.depth_attentions, self.tau, True)
with torch.no_grad():
selected_widths = selected_widths.cpu()
x, last_channel_idx, expected_inC, flops = inputs, 0, 3, []
feature_maps = []
for i, layer in enumerate(self.layers):
selected_w_index = selected_widths [last_channel_idx: last_channel_idx+layer.num_conv]
selected_w_probs = selected_width_probs[last_channel_idx: last_channel_idx+layer.num_conv]
layer_prob = flop_width_probs [last_channel_idx: last_channel_idx+layer.num_conv]
x, expected_inC, expected_flop = layer( (x, expected_inC, layer_prob, selected_w_index, selected_w_probs) )
feature_maps.append( x )
last_channel_idx += layer.num_conv
if i in self.depth_info: # aggregate the information
choices = self.depth_info[i]['choices']
xstagei = self.depth_info[i]['stage']
#print ('iL={:}, choices={:}, stage={:}, probs={:}'.format(i, choices, xstagei, selected_depth_probs[xstagei].cpu().tolist()))
#for A, W in zip(choices, selected_depth_probs[xstagei]):
# print('Size = {:}, W = {:}'.format(feature_maps[A].size(), W))
possible_tensors = []
max_C = max( feature_maps[A].size(1) for A in choices )
for tempi, A in enumerate(choices):
xtensor = ChannelWiseInter(feature_maps[A], max_C)
possible_tensors.append( xtensor )
weighted_sum = sum( xtensor * W for xtensor, W in zip(possible_tensors, selected_depth_probs[xstagei]) )
x = weighted_sum
if i in self.depth_at_i:
xstagei, xatti = self.depth_at_i[i]
x_expected_flop = flop_depth_probs[xstagei, xatti] * expected_flop
else:
x_expected_flop = expected_flop
flops.append( x_expected_flop )
flops.append( expected_inC * (self.classifier.out_features*1.0/1e6) )
features = self.avgpool(x)
features = features.view(features.size(0), -1)
logits = linear_forward(features, self.classifier)
return logits, torch.stack( [sum(flops)] )
def basic_forward(self, inputs):
if self.InShape is None: self.InShape = (inputs.size(-2), inputs.size(-1))
x = inputs
for i, layer in enumerate(self.layers):
x = layer( x )
features = self.avgpool(x)
features = features.view(features.size(0), -1)
logits = self.classifier(features)
return features, logits