Migrations
/
pytorch-image-models
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Add HRNet feature extraction, fix senet type, lower feature testing res to 96x96

Browse Source
pull/175/head
Ross Wightman 4 years ago
parent 2ac663f340
commit c9d54bc1c3
3 changed files with 127 additions and 82 deletions
Show Stats Download Patch File Download Diff File

5
tests/test_models.py
Unescape View File

@ -109,12 +109,13 @@ def test_model_forward_torchscript(model_name, batch_size):
EXCLUDE_FEAT_FILTERS = [
'hrnet*', '*pruned*', # hopefully fix at some point
'*pruned*', # hopefully fix at some point
]
if 'GITHUB_ACTIONS' in os.environ and 'Linux' in platform.system():
# GitHub Linux runner is slower and hits memory limits sooner than MacOS, exclude bigger models
EXCLUDE_FEAT_FILTERS += ['*resnext101_32x32d']
@pytest.mark.timeout(120)
@pytest.mark.parametrize('model_name', list_models(exclude_filters=EXCLUDE_FILTERS + EXCLUDE_FEAT_FILTERS))
@pytest.mark.parametrize('batch_size', [1])
@ -124,7 +125,7 @@ def test_model_forward_features(model_name, batch_size):
model.eval()
expected_channels = model.feature_info.channels()
assert len(expected_channels) >= 4 # all models here should have at least 4 feature levels by default, some 5 or 6
input_size = (3, 128, 128) # jit compile is already a bit slow and we've tested normal res already...
input_size = (3, 96, 96) # jit compile is already a bit slow and we've tested normal res already...
outputs = model(torch.randn((batch_size, *input_size)))
assert len(expected_channels) == len(outputs)
for e, o in zip(expected_channels, outputs):

200
timm/models/hrnet.py
Unescape View File

@ -8,17 +8,15 @@ Original header:
Written by Bin Xiao (Bin.Xiao@microsoft.com)
Modified by Ke Sun (sunk@mail.ustc.edu.cn)
"""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import logging
from typing import List
import torch
import torch.nn as nn
import torch.nn.functional as F
from timm.data import IMAGENET_DEFAULT_MEAN, IMAGENET_DEFAULT_STD
from .features import FeatureInfo
from .helpers import build_model_with_cfg
from .layers import SelectAdaptivePool2d
from .registry import register_model
@ -403,32 +401,23 @@ class HighResolutionModule(nn.Module):
self.branches = self._make_branches(
num_branches, blocks, num_blocks, num_channels)
self.fuse_layers = self._make_fuse_layers()
self.relu = nn.ReLU(False)
self.fuse_act = nn.ReLU(False)
def _check_branches(self, num_branches, blocks, num_blocks, num_inchannels, num_channels):
error_msg = ''
if num_branches != len(num_blocks):
error_msg = 'NUM_BRANCHES({}) <> NUM_BLOCKS({})'.format(
num_branches, len(num_blocks))
logger.error(error_msg)
raise ValueError(error_msg)
if num_branches != len(num_channels):
error_msg = 'NUM_BRANCHES({}) <> NUM_CHANNELS({})'.format(
num_branches, len(num_channels))
error_msg = 'NUM_BRANCHES({}) <> NUM_BLOCKS({})'.format(num_branches, len(num_blocks))
elif num_branches != len(num_channels):
error_msg = 'NUM_BRANCHES({}) <> NUM_CHANNELS({})'.format(num_branches, len(num_channels))
elif num_branches != len(num_inchannels):
error_msg = 'NUM_BRANCHES({}) <> NUM_INCHANNELS({})'.format(num_branches, len(num_inchannels))
if error_msg:
logger.error(error_msg)
raise ValueError(error_msg)
if num_branches != len(num_inchannels):
error_msg = 'NUM_BRANCHES({}) <> NUM_INCHANNELS({})'.format(
num_branches, len(num_inchannels))
logger.error(error_msg)
raise ValueError(error_msg)
def _make_one_branch(self, branch_index, block, num_blocks, num_channels,
stride=1):
def _make_one_branch(self, branch_index, block, num_blocks, num_channels, stride=1):
downsample = None
if stride != 1 or \
self.num_inchannels[branch_index] != num_channels[branch_index] * block.expansion:
if stride != 1 or self.num_inchannels[branch_index] != num_channels[branch_index] * block.expansion:
downsample = nn.Sequential(
nn.Conv2d(
self.num_inchannels[branch_index], num_channels[branch_index] * block.expansion,
@ -489,22 +478,22 @@ class HighResolutionModule(nn.Module):
def get_num_inchannels(self):
return self.num_inchannels
def forward(self, x):
def forward(self, x: List[torch.Tensor]):
if self.num_branches == 1:
return [self.branches[0](x[0])]
for i in range(self.num_branches):
x[i] = self.branches[i](x[i])
for i, branch in enumerate(self.branches):
x[i] = branch(x[i])
x_fuse = []
for i in range(len(self.fuse_layers)):
y = x[0] if i == 0 else self.fuse_layers[i][0](x[0])
for i, fuse_outer in enumerate(self.fuse_layers):
y = x[0] if i == 0 else fuse_outer[0](x[0])
for j in range(1, self.num_branches):
if i == j:
y = y + x[j]
else:
y = y + self.fuse_layers[i][j](x[j])
x_fuse.append(self.relu(y))
y = y + fuse_outer[j](x[j])
x_fuse.append(self.fuse_act(y))
return x_fuse
@ -517,7 +506,7 @@ blocks_dict = {
class HighResolutionNet(nn.Module):
def __init__(self, cfg, in_chans=3, num_classes=1000, global_pool='avg', drop_rate=0.0):
def __init__(self, cfg, in_chans=3, num_classes=1000, global_pool='avg', drop_rate=0.0, head='classification'):
super(HighResolutionNet, self).__init__()
self.num_classes = num_classes
self.drop_rate = drop_rate
@ -525,9 +514,10 @@ class HighResolutionNet(nn.Module):
stem_width = cfg['STEM_WIDTH']
self.conv1 = nn.Conv2d(in_chans, stem_width, kernel_size=3, stride=2, padding=1, bias=False)
self.bn1 = nn.BatchNorm2d(stem_width, momentum=_BN_MOMENTUM)
self.act1 = nn.ReLU(inplace=True)
self.conv2 = nn.Conv2d(stem_width, 64, kernel_size=3, stride=2, padding=1, bias=False)
self.bn2 = nn.BatchNorm2d(64, momentum=_BN_MOMENTUM)
self.relu = nn.ReLU(inplace=True)
self.act2 = nn.ReLU(inplace=True)
self.stage1_cfg = cfg['STAGE1']
num_channels = self.stage1_cfg['NUM_CHANNELS'][0]
@ -557,31 +547,49 @@ class HighResolutionNet(nn.Module):
self.transition3 = self._make_transition_layer(pre_stage_channels, num_channels)
self.stage4, pre_stage_channels = self._make_stage(self.stage4_cfg, num_channels, multi_scale_output=True)
# Classification Head
self.num_features = 2048
self.incre_modules, self.downsamp_modules, self.final_layer = self._make_head(pre_stage_channels)
self.global_pool = SelectAdaptivePool2d(pool_type=global_pool)
self.classifier = nn.Linear(self.num_features * self.global_pool.feat_mult(), num_classes)
self.head = head
self.head_channels = None # set if _make_head called
if head == 'classification':
# Classification Head
self.num_features = 2048
self.incre_modules, self.downsamp_modules, self.final_layer = self._make_head(pre_stage_channels)
self.global_pool = SelectAdaptivePool2d(pool_type=global_pool)
self.classifier = nn.Linear(self.num_features * self.global_pool.feat_mult(), num_classes)
elif head == 'incre':
self.num_features = 2048
self.incre_modules, _, _ = self._make_head(pre_stage_channels, True)
else:
self.incre_modules = None
self.num_features = 256
curr_stride = 2
# module names aren't actually valid here, hook or FeatureNet based extraction would not work
self.feature_info = [dict(num_chs=64, reduction=curr_stride, module='stem')]
for i, c in enumerate(self.head_channels if self.head_channels else num_channels):
curr_stride *= 2
c = c * 4 if self.head_channels else c # head block expansion factor of 4
self.feature_info += [dict(num_chs=c, reduction=curr_stride, module=f'stage{i + 1}')]
self.init_weights()
def _make_head(self, pre_stage_channels):
def _make_head(self, pre_stage_channels, incre_only=False):
head_block = Bottleneck
head_channels = [32, 64, 128, 256]
self.head_channels = [32, 64, 128, 256]
# Increasing the #channels on each resolution
# from C, 2C, 4C, 8C to 128, 256, 512, 1024
incre_modules = []
for i, channels in enumerate(pre_stage_channels):
incre_modules.append(
self._make_layer(head_block, channels, head_channels[i], 1, stride=1))
incre_modules.append(self._make_layer(head_block, channels, self.head_channels[i], 1, stride=1))
incre_modules = nn.ModuleList(incre_modules)
if incre_only:
return incre_modules, None, None
# downsampling modules
downsamp_modules = []
for i in range(len(pre_stage_channels) - 1):
in_channels = head_channels[i] * head_block.expansion
out_channels = head_channels[i + 1] * head_block.expansion
in_channels = self.head_channels[i] * head_block.expansion
out_channels = self.head_channels[i + 1] * head_block.expansion
downsamp_module = nn.Sequential(
nn.Conv2d(
in_channels=in_channels, out_channels=out_channels, kernel_size=3, stride=2, padding=1),
@ -593,7 +601,7 @@ class HighResolutionNet(nn.Module):
final_layer = nn.Sequential(
nn.Conv2d(
in_channels=head_channels[3] * head_block.expansion,
in_channels=self.head_channels[3] * head_block.expansion,
out_channels=self.num_features, kernel_size=1, stride=1, padding=0
),
nn.BatchNorm2d(self.num_features, momentum=_BN_MOMENTUM),
@ -655,11 +663,7 @@ class HighResolutionNet(nn.Module):
modules = []
for i in range(num_modules):
# multi_scale_output is only used last module
if not multi_scale_output and i == num_modules - 1:
reset_multi_scale_output = False
else:
reset_multi_scale_output = True
reset_multi_scale_output = multi_scale_output or i < num_modules - 1
modules.append(HighResolutionModule(
num_branches, block, num_blocks, num_inchannels, num_channels, fuse_method, reset_multi_scale_output)
)
@ -688,40 +692,35 @@ class HighResolutionNet(nn.Module):
else:
self.classifier = nn.Identity()
def stages(self, x) -> List[torch.Tensor]:
x = self.layer1(x)
xl = [t(x) for i, t in enumerate(self.transition1)]
yl = self.stage2(xl)
xl = [t(yl[-1]) if not isinstance(t, nn.Identity) else yl[i] for i, t in enumerate(self.transition2)]
yl = self.stage3(xl)
xl = [t(yl[-1]) if not isinstance(t, nn.Identity) else yl[i] for i, t in enumerate(self.transition3)]
yl = self.stage4(xl)
return yl
def forward_features(self, x):
# Stem
x = self.conv1(x)
x = self.bn1(x)
x = self.relu(x)
x = self.act1(x)
x = self.conv2(x)
x = self.bn2(x)
x = self.relu(x)
x = self.layer1(x)
x_list = []
for i in range(len(self.transition1)):
x_list.append(self.transition1[i](x))
y_list = self.stage2(x_list)
x_list = []
for i in range(len(self.transition2)):
if not isinstance(self.transition2[i], nn.Identity):
x_list.append(self.transition2[i](y_list[-1]))
else:
x_list.append(y_list[i])
y_list = self.stage3(x_list)
x = self.act2(x)
x_list = []
for i in range(len(self.transition3)):
if not isinstance(self.transition3[i], nn.Identity):
x_list.append(self.transition3[i](y_list[-1]))
else:
x_list.append(y_list[i])
y_list = self.stage4(x_list)
# Stages
yl = self.stages(x)
# Classification Head
y = self.incre_modules[0](y_list[0])
for i in range(len(self.downsamp_modules)):
y = self.incre_modules[i + 1](y_list[i + 1]) + self.downsamp_modules[i](y)
y = self.incre_modules[0](yl[0])
for i, down in enumerate(self.downsamp_modules):
y = self.incre_modules[i + 1](yl[i + 1]) + down(y)
y = self.final_layer(y)
return y
@ -734,10 +733,55 @@ class HighResolutionNet(nn.Module):
return x
class HighResolutionNetFeatures(HighResolutionNet):
"""HighResolutionNet feature extraction
The design of HRNet makes it easy to grab feature maps, this class provides a simple wrapper to do so.
It would be more complicated to use the FeatureNet helpers.
The `feature_location=incre` allows grabbing increased channel count features using part of the
classification head. If `feature_location=''` the default HRNet features are returned. First stem
conv is used for stride 2 features.
"""
def __init__(self, cfg, in_chans=3, num_classes=1000, global_pool='avg', drop_rate=0.0,
feature_location='incre', out_indices=(0, 1, 2, 3, 4)):
assert feature_location in ('incre', '')
super(HighResolutionNetFeatures, self).__init__(
cfg, in_chans=in_chans, num_classes=num_classes, global_pool=global_pool,
drop_rate=drop_rate, head=feature_location)
self.feature_info = FeatureInfo(self.feature_info, out_indices)
self._out_idx = {i for i in out_indices}
def forward_features(self, x):
assert False, 'Not supported'
def forward(self, x) -> List[torch.tensor]:
out = []
x = self.conv1(x)
x = self.bn1(x)
x = self.act1(x)
if 0 in self._out_idx:
out.append(x)
x = self.conv2(x)
x = self.bn2(x)
x = self.act2(x)
x = self.stages(x)
if self.incre_modules is not None:
x = [incre(f) for f, incre in zip(x, self.incre_modules)]
for i, f in enumerate(x):
if i + 1 in self._out_idx:
out.append(f)
return out
def _create_hrnet(variant, pretrained, **model_kwargs):
assert not model_kwargs.pop('features_only', False) # feature extraction not figured out yet
model_cls = HighResolutionNet
if model_kwargs.pop('features_only', False):
model_cls = HighResolutionNetFeatures
return build_model_with_cfg(
HighResolutionNet, variant, pretrained, default_cfg=default_cfgs[variant],
model_cls, variant, pretrained, default_cfg=default_cfgs[variant],
model_cfg=cfg_cls[variant], **model_kwargs)

4
timm/models/senet.py
Unescape View File

@ -423,14 +423,14 @@ def legacy_seresnet34(pretrained=False, **kwargs):
@register_model
def legacy_seresnet50(pretrained=False, num_classes=1000, in_chans=3, **kwargs):
def legacy_seresnet50(pretrained=False, **kwargs):
model_args = dict(
block=SEResNetBottleneck, layers=[3, 4, 6, 3], groups=1, reduction=16, **kwargs)
return _create_senet('seresnet50', pretrained, **model_args)
@register_model
def legacy_seresnet101(pretrained=False, num_classes=1000, in_chans=3, **kwargs):
def legacy_seresnet101(pretrained=False, **kwargs):
model_args = dict(
block=SEResNetBottleneck, layers=[3, 4, 23, 3], groups=1, reduction=16, **kwargs)
return _create_senet('seresnet101', pretrained, **model_args)

Write Preview
Loading…
Cancel
Save