mirrors/pytorch-image-models
1
0
Fork 0
mirror of https://github.com/huggingface/pytorch-image-models.git synced 2025-06-03 15:01:08 +08:00
Code Issues Projects Releases Wiki Activity

Fixup Resnext, remove alternate shortcut types

Browse Source
This commit is contained in:
Ross Wightman 2019-04-01 11:03:37 -07:00
parent 45cde6f0c7
commit 5cb1a35c6b
2 changed files with 13 additions and 32 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
data/random_erasing.py
View File

@ -115,9 +115,9 @@ class RandomErasingTorch:
h = int(round(math.sqrt(target_area * aspect_ratio))) h = int(round(math.sqrt(target_area * aspect_ratio)))
w = int(round(math.sqrt(target_area / aspect_ratio))) w = int(round(math.sqrt(target_area / aspect_ratio)))
if self.rand_color: if self.rand_color:
c = torch.empty(chan, dtype=batch.dtype, device=self.device).normal_() c = torch.empty((chan, 1, 1), dtype=batch.dtype, device=self.device).normal_()
elif not self.per_pixel: elif not self.per_pixel:
c = torch.zeros(chan, dtype=batch.dtype, device=self.device) c = torch.zeros((chan, 1, 1), dtype=batch.dtype, device=self.device)
if w < img_w and h < img_h: if w < img_w and h < img_h:
top = random.randint(0, img_h - h) top = random.randint(0, img_h - h)
left = random.randint(0, img_w - w) left = random.randint(0, img_w - w)

41
models/resnext.py
View File

@ -19,7 +19,7 @@ class ResNeXtBottleneckC(nn.Module):
def __init__(self, inplanes, planes, stride=1, downsample=None, cardinality=32, base_width=4): def __init__(self, inplanes, planes, stride=1, downsample=None, cardinality=32, base_width=4):
super(ResNeXtBottleneckC, self).__init__() super(ResNeXtBottleneckC, self).__init__()
width = math.floor(planes / 64 * cardinality * base_width) width = math.floor(planes * (base_width / 64)) * cardinality
self.conv1 = nn.Conv2d(inplanes, width, kernel_size=1, bias=False) self.conv1 = nn.Conv2d(inplanes, width, kernel_size=1, bias=False)
self.bn1 = nn.BatchNorm2d(width) self.bn1 = nn.BatchNorm2d(width)
@ -57,13 +57,12 @@ class ResNeXtBottleneckC(nn.Module):
class ResNeXt(nn.Module): class ResNeXt(nn.Module):
def __init__(self, block, layers, num_classes=1000, cardinality=32, base_width=4, shortcut='C', def __init__(self, block, layers, num_classes=1000, cardinality=32, base_width=4,
drop_rate=0., global_pool='avg'): drop_rate=0., global_pool='avg'):
self.num_classes = num_classes self.num_classes = num_classes
self.inplanes = 64 self.inplanes = 64
self.cardinality = cardinality self.cardinality = cardinality
self.base_width = base_width self.base_width = base_width
self.shortcut = shortcut
self.drop_rate = drop_rate self.drop_rate = drop_rate
super(ResNeXt, self).__init__() super(ResNeXt, self).__init__()
self.conv1 = nn.Conv2d(3, 64, kernel_size=7, stride=2, padding=3, bias=False) self.conv1 = nn.Conv2d(3, 64, kernel_size=7, stride=2, padding=3, bias=False)
@ -80,39 +79,24 @@ class ResNeXt(nn.Module):
for m in self.modules(): for m in self.modules():
if isinstance(m, nn.Conv2d): if isinstance(m, nn.Conv2d):
n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
m.weight.data.normal_(0, math.sqrt(2. / n))
elif isinstance(m, nn.BatchNorm2d): elif isinstance(m, nn.BatchNorm2d):
m.weight.data.fill_(1) m.weight.data.fill_(1)
m.bias.data.zero_() m.bias.data.zero_()
def _make_layer(self, block, planes, blocks, stride=1): def _make_layer(self, block, planes, blocks, stride=1):
downsample = None downsample = None
reshape = stride != 1 or self.inplanes != planes * block.expansion if stride != 1 or self.inplanes != planes * block.expansion:
use_conv = (self.shortcut == 'C') or (self.shortcut == 'B' and reshape)
if use_conv:
downsample = nn.Sequential( downsample = nn.Sequential(
nn.Conv2d( nn.Conv2d(
self.inplanes, planes * block.expansion, kernel_size=1, stride=stride, bias=False), self.inplanes, planes * block.expansion, kernel_size=1, stride=stride, bias=False),
nn.BatchNorm2d(planes * block.expansion), nn.BatchNorm2d(planes * block.expansion),
) )
elif reshape:
downsample = nn.AvgPool2d(3, stride=stride)
layers = [block(self.inplanes, planes, stride, downsample, self.cardinality, self.base_width)] layers = [block(self.inplanes, planes, stride, downsample, self.cardinality, self.base_width)]
self.inplanes = planes * block.expansion self.inplanes = planes * block.expansion
if self.shortcut == 'C':
shortcut = nn.Sequential(
nn.Conv2d(
self.inplanes, planes * block.expansion, kernel_size=1, stride=1, bias=False),
nn.BatchNorm2d(planes * block.expansion),
)
else:
shortcut = None
for i in range(1, blocks): for i in range(1, blocks):
layers.append(block(self.inplanes, planes, 1, shortcut, self.cardinality, self.base_width)) layers.append(block(self.inplanes, planes, 1, None, self.cardinality, self.base_width))
return nn.Sequential(*layers) return nn.Sequential(*layers)
@ -152,7 +136,7 @@ class ResNeXt(nn.Module):
return x return x
def resnext50(cardinality=32, base_width=4, shortcut='C', pretrained=False, **kwargs): def resnext50(cardinality=32, base_width=4, pretrained=False, **kwargs):
"""Constructs a ResNeXt-50 model. """Constructs a ResNeXt-50 model.
Args: Args:
@ -161,12 +145,11 @@ def resnext50(cardinality=32, base_width=4, shortcut='C', pretrained=False, **kw
shortcut ('A'|'B'|'C'): 'B' use 1x1 conv to downsample, 'C' use 1x1 conv on every residual connection shortcut ('A'|'B'|'C'): 'B' use 1x1 conv to downsample, 'C' use 1x1 conv on every residual connection
""" """
model = ResNeXt( model = ResNeXt(
ResNeXtBottleneckC, [3, 4, 6, 3], cardinality=cardinality, ResNeXtBottleneckC, [3, 4, 6, 3], cardinality=cardinality, base_width=base_width, **kwargs)
base_width=base_width, shortcut=shortcut, **kwargs)
return model return model
def resnext101(cardinality=32, base_width=4, shortcut='C', pretrained=False, **kwargs): def resnext101(cardinality=32, base_width=4, pretrained=False, **kwargs):
"""Constructs a ResNeXt-101 model. """Constructs a ResNeXt-101 model.
Args: Args:
@ -175,12 +158,11 @@ def resnext101(cardinality=32, base_width=4, shortcut='C', pretrained=False, **k
shortcut ('A'|'B'|'C'): 'B' use 1x1 conv to downsample, 'C' use 1x1 conv on every residual connection shortcut ('A'|'B'|'C'): 'B' use 1x1 conv to downsample, 'C' use 1x1 conv on every residual connection
""" """
model = ResNeXt( model = ResNeXt(
ResNeXtBottleneckC, [3, 4, 23, 3], cardinality=cardinality, ResNeXtBottleneckC, [3, 4, 23, 3], cardinality=cardinality, base_width=base_width, **kwargs)
base_width=base_width, shortcut=shortcut, **kwargs)
return model return model
def resnext152(cardinality=32, base_width=4, shortcut='C', pretrained=False, **kwargs): def resnext152(cardinality=32, base_width=4, pretrained=False, **kwargs):
"""Constructs a ResNeXt-152 model. """Constructs a ResNeXt-152 model.
Args: Args:
@ -189,6 +171,5 @@ def resnext152(cardinality=32, base_width=4, shortcut='C', pretrained=False, **k
shortcut ('A'|'B'|'C'): 'B' use 1x1 conv to downsample, 'C' use 1x1 conv on every residual connection shortcut ('A'|'B'|'C'): 'B' use 1x1 conv to downsample, 'C' use 1x1 conv on every residual connection
""" """
model = ResNeXt( model = ResNeXt(
ResNeXtBottleneckC, [3, 8, 36, 3], cardinality=cardinality, ResNeXtBottleneckC, [3, 8, 36, 3], cardinality=cardinality, base_width=base_width, **kwargs)
base_width=base_width, shortcut=shortcut, **kwargs)
return model return model