373 lines
14 KiB
Python
373 lines
14 KiB
Python
# Taken and modified from https://github.com/lucifer443/SpineNet-Pytorch/blob/master/mmdet/models/backbones/spinenet.py
|
|
|
|
import torch
|
|
import torch.nn as nn
|
|
import torch.nn.functional as F
|
|
from torch.nn.init import kaiming_normal
|
|
|
|
from torchvision.models.resnet import BasicBlock, Bottleneck
|
|
from models.arch_util import ConvGnSilu, ConvBnSilu, ConvBnRelu
|
|
from trainer.networks import register_model
|
|
|
|
|
|
def constant_init(module, val, bias=0):
|
|
if hasattr(module, 'weight') and module.weight is not None:
|
|
nn.init.constant_(module.weight, val)
|
|
if hasattr(module, 'bias') and module.bias is not None:
|
|
nn.init.constant_(module.bias, bias)
|
|
|
|
def kaiming_init(module,
|
|
a=0,
|
|
mode='fan_out',
|
|
nonlinearity='relu',
|
|
bias=0,
|
|
distribution='normal'):
|
|
assert distribution in ['uniform', 'normal']
|
|
if distribution == 'uniform':
|
|
nn.init.kaiming_uniform_(
|
|
module.weight, a=a, mode=mode, nonlinearity=nonlinearity)
|
|
else:
|
|
nn.init.kaiming_normal_(
|
|
module.weight, a=a, mode=mode, nonlinearity=nonlinearity)
|
|
if hasattr(module, 'bias') and module.bias is not None:
|
|
nn.init.constant_(module.bias, bias)
|
|
|
|
FILTER_SIZE_MAP = {
|
|
1: 32,
|
|
2: 64,
|
|
3: 128,
|
|
4: 256,
|
|
5: 256,
|
|
6: 256,
|
|
7: 256,
|
|
}
|
|
|
|
def make_res_layer(block,
|
|
inplanes,
|
|
planes,
|
|
blocks,
|
|
stride=1,
|
|
dilation=1):
|
|
downsample = None
|
|
if stride != 1 or inplanes != planes * block.expansion:
|
|
downsample = nn.Sequential(
|
|
nn.Conv2d(
|
|
inplanes,
|
|
planes * block.expansion,
|
|
kernel_size=1,
|
|
stride=stride,
|
|
bias=False),
|
|
nn.BatchNorm2d(planes * block.expansion),
|
|
)
|
|
|
|
layers = []
|
|
layers.append(
|
|
block(
|
|
inplanes=inplanes,
|
|
planes=planes,
|
|
stride=stride,
|
|
dilation=dilation,
|
|
downsample=downsample))
|
|
inplanes = planes * block.expansion
|
|
for i in range(1, blocks):
|
|
layers.append(
|
|
block(
|
|
inplanes=inplanes,
|
|
planes=planes,
|
|
stride=1,
|
|
dilation=dilation))
|
|
|
|
return nn.Sequential(*layers)
|
|
|
|
# The fixed SpineNet architecture discovered by NAS.
|
|
# Each element represents a specification of a building block:
|
|
# (block_level, block_fn, (input_offset0, input_offset1), is_output).
|
|
SPINENET_BLOCK_SPECS = [
|
|
(2, Bottleneck, (None, None), False), # init block
|
|
(2, Bottleneck, (None, None), False), # init block
|
|
(2, Bottleneck, (0, 1), False),
|
|
(4, BasicBlock, (0, 1), False),
|
|
(3, Bottleneck, (2, 3), False),
|
|
(4, Bottleneck, (2, 4), False),
|
|
(6, BasicBlock, (3, 5), False),
|
|
(4, Bottleneck, (3, 5), False),
|
|
(5, BasicBlock, (6, 7), False),
|
|
(7, BasicBlock, (6, 8), False),
|
|
(5, Bottleneck, (8, 9), False),
|
|
(4, Bottleneck, (5, 10), True),
|
|
(3, Bottleneck, (4, 10), True),
|
|
]
|
|
|
|
SCALING_MAP = {
|
|
'49S': {
|
|
'endpoints_num_filters': 128,
|
|
'filter_size_scale': 0.65,
|
|
'resample_alpha': 0.5,
|
|
'block_repeats': 1,
|
|
},
|
|
'49': {
|
|
'endpoints_num_filters': 256,
|
|
'filter_size_scale': 1.0,
|
|
'resample_alpha': 0.5,
|
|
'block_repeats': 1,
|
|
},
|
|
'96': {
|
|
'endpoints_num_filters': 256,
|
|
'filter_size_scale': 1.0,
|
|
'resample_alpha': 0.5,
|
|
'block_repeats': 2,
|
|
},
|
|
'143': {
|
|
'endpoints_num_filters': 256,
|
|
'filter_size_scale': 1.0,
|
|
'resample_alpha': 1.0,
|
|
'block_repeats': 3,
|
|
},
|
|
'190': {
|
|
'endpoints_num_filters': 512,
|
|
'filter_size_scale': 1.3,
|
|
'resample_alpha': 1.0,
|
|
'block_repeats': 4,
|
|
},
|
|
}
|
|
|
|
|
|
class BlockSpec(object):
|
|
"""A container class that specifies the block configuration for SpineNet."""
|
|
|
|
def __init__(self, level, block_fn, input_offsets, is_output):
|
|
self.level = level
|
|
self.block_fn = block_fn
|
|
self.input_offsets = input_offsets
|
|
self.is_output = is_output
|
|
|
|
|
|
def build_block_specs(block_specs=None):
|
|
"""Builds the list of BlockSpec objects for SpineNet."""
|
|
if not block_specs:
|
|
block_specs = SPINENET_BLOCK_SPECS
|
|
return [BlockSpec(*b) for b in block_specs]
|
|
|
|
|
|
class Resample(nn.Module):
|
|
def __init__(self, in_channels, out_channels, scale, block_type, alpha=1.0):
|
|
super(Resample, self).__init__()
|
|
self.scale = scale
|
|
new_in_channels = int(in_channels * alpha)
|
|
if block_type == Bottleneck:
|
|
in_channels *= 4
|
|
self.squeeze_conv = ConvGnSilu(in_channels, new_in_channels, kernel_size=1)
|
|
if scale < 1:
|
|
self.downsample_conv = ConvGnSilu(new_in_channels, new_in_channels, kernel_size=3, stride=2)
|
|
self.expand_conv = ConvGnSilu(new_in_channels, out_channels, kernel_size=1, activation=False)
|
|
|
|
def _resize(self, x):
|
|
if self.scale == 1:
|
|
return x
|
|
elif self.scale > 1:
|
|
return F.interpolate(x, scale_factor=self.scale, mode='nearest')
|
|
else:
|
|
x = self.downsample_conv(x)
|
|
if self.scale < 0.5:
|
|
new_kernel_size = 3 if self.scale >= 0.25 else 5
|
|
x = F.max_pool2d(x, kernel_size=new_kernel_size, stride=int(0.5/self.scale), padding=new_kernel_size//2)
|
|
return x
|
|
|
|
def forward(self, inputs):
|
|
feat = self.squeeze_conv(inputs)
|
|
feat = self._resize(feat)
|
|
feat = self.expand_conv(feat)
|
|
return feat
|
|
|
|
|
|
class Merge(nn.Module):
|
|
"""Merge two input tensors"""
|
|
def __init__(self, block_spec, alpha, filter_size_scale):
|
|
super(Merge, self).__init__()
|
|
out_channels = int(FILTER_SIZE_MAP[block_spec.level] * filter_size_scale)
|
|
if block_spec.block_fn == Bottleneck:
|
|
out_channels *= 4
|
|
self.block = block_spec.block_fn
|
|
self.resample_ops = nn.ModuleList()
|
|
for spec_idx in block_spec.input_offsets:
|
|
spec = BlockSpec(*SPINENET_BLOCK_SPECS[spec_idx])
|
|
in_channels = int(FILTER_SIZE_MAP[spec.level] * filter_size_scale)
|
|
scale = 2**(spec.level - block_spec.level)
|
|
self.resample_ops.append(
|
|
Resample(in_channels, out_channels, scale, spec.block_fn, alpha)
|
|
)
|
|
|
|
def forward(self, inputs):
|
|
assert len(inputs) == len(self.resample_ops)
|
|
parent0_feat = self.resample_ops[0](inputs[0])
|
|
parent1_feat = self.resample_ops[1](inputs[1])
|
|
target_feat = parent0_feat + parent1_feat
|
|
return target_feat
|
|
|
|
|
|
class SpineNet(nn.Module):
|
|
"""Class to build SpineNet backbone"""
|
|
def __init__(self,
|
|
arch,
|
|
in_channels=3,
|
|
output_level=[3, 4],
|
|
conv_cfg=None,
|
|
norm_cfg=dict(type='BN', requires_grad=True),
|
|
zero_init_residual=True,
|
|
activation='relu',
|
|
use_input_norm=False,
|
|
double_reduce_early=True):
|
|
super(SpineNet, self).__init__()
|
|
self._block_specs = build_block_specs()[2:]
|
|
self._endpoints_num_filters = SCALING_MAP[arch]['endpoints_num_filters']
|
|
self._resample_alpha = SCALING_MAP[arch]['resample_alpha']
|
|
self._block_repeats = SCALING_MAP[arch]['block_repeats']
|
|
self._filter_size_scale = SCALING_MAP[arch]['filter_size_scale']
|
|
self._init_block_fn = Bottleneck
|
|
self._num_init_blocks = 2
|
|
self._early_double_reduce = double_reduce_early
|
|
self.zero_init_residual = zero_init_residual
|
|
assert min(output_level) > 2 and max(output_level) < 8, "Output level out of range"
|
|
self.output_level = output_level
|
|
self.use_input_norm = use_input_norm
|
|
|
|
self._make_stem_layer(in_channels)
|
|
self._make_scale_permuted_network()
|
|
self._make_endpoints()
|
|
|
|
def _make_stem_layer(self, in_channels):
|
|
"""Build the stem network."""
|
|
# Build the first conv and maxpooling layers.
|
|
if self._early_double_reduce:
|
|
self.conv1 = ConvGnSilu(
|
|
in_channels,
|
|
64,
|
|
kernel_size=7,
|
|
stride=2)
|
|
self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
|
|
else:
|
|
self.conv1 = None
|
|
|
|
# Build the initial level 2 blocks.
|
|
self.init_block1 = make_res_layer(
|
|
self._init_block_fn,
|
|
64,
|
|
int(FILTER_SIZE_MAP[2] * self._filter_size_scale),
|
|
self._block_repeats)
|
|
self.init_block2 = make_res_layer(
|
|
self._init_block_fn,
|
|
int(FILTER_SIZE_MAP[2] * self._filter_size_scale) * 4,
|
|
int(FILTER_SIZE_MAP[2] * self._filter_size_scale),
|
|
self._block_repeats)
|
|
|
|
def _make_endpoints(self):
|
|
self.endpoint_convs = nn.ModuleDict()
|
|
for block_spec in self._block_specs:
|
|
if block_spec.is_output:
|
|
in_channels = int(FILTER_SIZE_MAP[block_spec.level]*self._filter_size_scale) * 4
|
|
self.endpoint_convs[str(block_spec.level)] = ConvGnSilu(in_channels,
|
|
self._endpoints_num_filters,
|
|
kernel_size=1,
|
|
activation=False)
|
|
|
|
def _make_scale_permuted_network(self):
|
|
self.merge_ops = nn.ModuleList()
|
|
self.scale_permuted_blocks = nn.ModuleList()
|
|
for spec in self._block_specs:
|
|
self.merge_ops.append(
|
|
Merge(spec, self._resample_alpha, self._filter_size_scale)
|
|
)
|
|
channels = int(FILTER_SIZE_MAP[spec.level] * self._filter_size_scale)
|
|
in_channels = channels * 4 if spec.block_fn == Bottleneck else channels
|
|
self.scale_permuted_blocks.append(
|
|
make_res_layer(spec.block_fn,
|
|
in_channels,
|
|
channels,
|
|
self._block_repeats)
|
|
)
|
|
|
|
def init_weights(self, pretrained=None):
|
|
for m in self.modules():
|
|
if isinstance(m, nn.Conv2d):
|
|
kaiming_init(m)
|
|
elif isinstance(m, (_BatchNorm, nn.GroupNorm)):
|
|
constant_init(m, 1)
|
|
if self.zero_init_residual:
|
|
for m in self.modules():
|
|
if isinstance(m, Bottleneck):
|
|
constant_init(m.bn3, 0)
|
|
elif isinstance(m, BasicBlock):
|
|
constant_init(m.bn2, 0)
|
|
|
|
def forward(self, input):
|
|
if self.conv1 is not None:
|
|
if self.use_input_norm:
|
|
mean = torch.Tensor([0.485, 0.456, 0.406]).view(1, 3, 1, 1).to(input.device)
|
|
std = torch.Tensor([0.229, 0.224, 0.225]).view(1, 3, 1, 1).to(input.device)
|
|
input = (input - mean) / std
|
|
feat = self.conv1(input)
|
|
feat = self.maxpool(feat)
|
|
else:
|
|
feat = input
|
|
feat1 = self.init_block1(feat)
|
|
feat2 = self.init_block2(feat1)
|
|
block_feats = [feat1, feat2]
|
|
output_feat = {}
|
|
num_outgoing_connections = [0, 0]
|
|
|
|
for i, spec in enumerate(self._block_specs):
|
|
target_feat = self.merge_ops[i]([block_feats[feat_idx] for feat_idx in spec.input_offsets])
|
|
# Connect intermediate blocks with outdegree 0 to the output block.
|
|
if spec.is_output:
|
|
for j, (j_feat, j_connections) in enumerate(
|
|
zip(block_feats, num_outgoing_connections)):
|
|
if j_connections == 0 and j_feat.shape == target_feat.shape:
|
|
target_feat += j_feat
|
|
num_outgoing_connections[j] += 1
|
|
target_feat = F.relu(target_feat, inplace=True)
|
|
target_feat = self.scale_permuted_blocks[i](target_feat)
|
|
block_feats.append(target_feat)
|
|
num_outgoing_connections.append(0)
|
|
for feat_idx in spec.input_offsets:
|
|
num_outgoing_connections[feat_idx] += 1
|
|
if spec.is_output:
|
|
output_feat[spec.level] = target_feat
|
|
|
|
return tuple([self.endpoint_convs[str(level)](output_feat[level]) for level in self.output_level])
|
|
|
|
|
|
# Attachs a simple 1x1 conv prediction head to a Spinenet.
|
|
class SpinenetWithLogits(SpineNet):
|
|
def __init__(self,
|
|
arch,
|
|
output_to_attach,
|
|
num_labels,
|
|
in_channels=3,
|
|
output_level=[3, 4],
|
|
conv_cfg=None,
|
|
norm_cfg=dict(type='BN', requires_grad=True),
|
|
zero_init_residual=True,
|
|
activation='relu',
|
|
use_input_norm=False,
|
|
double_reduce_early=True):
|
|
super().__init__(arch, in_channels, output_level, conv_cfg, norm_cfg, zero_init_residual, activation, use_input_norm, double_reduce_early)
|
|
self.output_to_attach = output_to_attach
|
|
self.tail = nn.Sequential(ConvBnRelu(256, 128, kernel_size=1, activation=True, norm=True, bias=False),
|
|
ConvBnRelu(128, 64, kernel_size=1, activation=True, norm=True, bias=False),
|
|
ConvBnRelu(64, num_labels, kernel_size=1, activation=False, norm=False, bias=True),
|
|
nn.Softmax(dim=1))
|
|
|
|
def forward(self, x):
|
|
fea = super().forward(x)[self.output_to_attach]
|
|
return self.tail(fea)
|
|
|
|
@register_model
|
|
def register_spinenet(opt_net, opt):
|
|
return SpineNet(str(opt_net['arch']), in_channels=3, use_input_norm=opt_net['use_input_norm'])
|
|
|
|
|
|
@register_model
|
|
def register_spinenet_with_logits(opt_net, opt):
|
|
return SpinenetWithLogits(str(opt_net['arch']), opt_net['output_to_attach'], opt_net['num_labels'],
|
|
in_channels=3, use_input_norm=opt_net['use_input_norm'])
|