import typing
import torch
import torch.nn
import torch.utils.model_zoo
import thelper.nn
import thelper.nn.coordconv
[docs]def get_activation_layer(name: typing.AnyStr) -> torch.nn.Module:
# todo: support more prebuilt/custom layer types here, if needed...
assert name in ["relu", "leaky_relu"]
if name == "relu":
return torch.nn.ReLU(inplace=True)
elif name == "leaky_relu":
return torch.nn.LeakyReLU(inplace=True)
class Module(torch.nn.Module):
expansion = 1
def __init__(self, inplanes, planes, stride=1, downsample=None, coordconv=False, radius_channel=True):
super().__init__()
self.inplanes = inplanes
self.planes = planes
self.stride = stride
self.downsample = downsample
self.coordconv = coordconv
self.radius_channel = radius_channel
def _make_conv2d(self, *args, **kwargs):
if self.coordconv:
return thelper.nn.coordconv.CoordConv2d(*args, radius_channel=self.radius_channel, **kwargs)
else:
return torch.nn.Conv2d(*args, **kwargs)
class BasicBlock(Module):
def __init__(self, inplanes, planes, stride=1,
downsample=None, coordconv=False,
radius_channel=True, activation="relu"):
super().__init__(inplanes, planes, stride, downsample, coordconv, radius_channel)
self.conv1 = self._make_conv2d(inplanes, planes, kernel_size=3, stride=stride, padding=1, bias=False)
self.bn1 = torch.nn.BatchNorm2d(planes)
self.activ = get_activation_layer(activation)
self.conv2 = self._make_conv2d(planes, planes, kernel_size=3, stride=1, padding=1, bias=False)
self.bn2 = torch.nn.BatchNorm2d(planes)
def forward(self, x):
residual = x
out = self.conv1(x)
out = self.bn1(out)
out = self.activ(out)
out = self.conv2(out)
out = self.bn2(out)
if self.downsample is not None:
residual = self.downsample(x)
out += residual
out = self.activ(out)
return out
class Bottleneck(Module):
expansion = 4
def __init__(self, inplanes, planes, stride=1,
downsample=None, coordconv=False,
radius_channel=True, activation="relu"):
super().__init__(inplanes, planes, stride, downsample, coordconv, radius_channel)
self.conv1 = self._make_conv2d(inplanes, planes, kernel_size=1, bias=False)
self.bn1 = torch.nn.BatchNorm2d(planes)
self.conv2 = self._make_conv2d(planes, planes, kernel_size=3, stride=stride, padding=1, bias=False)
self.bn2 = torch.nn.BatchNorm2d(planes)
self.conv3 = self._make_conv2d(planes, planes * self.expansion, kernel_size=1, bias=False)
self.bn3 = torch.nn.BatchNorm2d(planes * self.expansion)
self.activ = get_activation_layer(activation)
def forward(self, x):
residual = x
out = self.conv1(x)
out = self.bn1(out)
out = self.activ(out)
out = self.conv2(out)
out = self.bn2(out)
out = self.activ(out)
out = self.conv3(out)
out = self.bn3(out)
if self.downsample is not None:
residual = self.downsample(x)
out += residual
out = self.activ(out)
return out
class SqueezeExcitationLayer(torch.nn.Module):
def __init__(self, channel, reduction=16, activation="relu"):
super().__init__()
self.pool = torch.nn.AdaptiveAvgPool2d(1)
self.fc = torch.nn.Sequential(
torch.nn.Linear(channel, channel // reduction),
get_activation_layer(activation),
torch.nn.Linear(channel // reduction, channel),
torch.nn.Sigmoid()
)
def forward(self, x):
b, c, _, _ = x.size()
y = self.pool(x).view(b, c)
y = self.fc(y).view(b, c, 1, 1)
return x * y
class SqueezeExcitationBlock(Module):
def __init__(self, inplanes, planes, stride=1,
downsample=None, reduction=16, coordconv=False,
radius_channel=True, activation="relu"):
super().__init__(inplanes, planes, stride, downsample, coordconv, radius_channel)
self.conv1 = self._make_conv2d(inplanes, planes, kernel_size=3, stride=stride, padding=1, bias=False)
self.bn1 = torch.nn.BatchNorm2d(planes)
self.activ = get_activation_layer(activation)
self.conv2 = self._make_conv2d(planes, planes, kernel_size=3, stride=1, padding=1, bias=False)
self.bn2 = torch.nn.BatchNorm2d(planes)
self.se = SqueezeExcitationLayer(planes, reduction, activation=activation)
def forward(self, x):
residual = x
out = self.conv1(x)
out = self.bn1(out)
out = self.activ(out)
out = self.conv2(out)
out = self.bn2(out)
out = self.se(out)
if self.downsample is not None:
residual = self.downsample(x)
out += residual
out = self.activ(out)
return out
class ResNet(thelper.nn.Module):
def __init__(self, task, block="thelper.nn.resnet.BasicBlock",
layers=[3, 4, 6, 3], strides=[1, 2, 2, 2], input_channels=3,
flexible_input_res=False, pool_size=7, head_type=None,
coordconv=False, radius_channel=True, activation="relu",
skip_max_pool=False, pretrained=False, conv1_config=[7, 2, 3]):
# note: must always forward args to base class to keep backup
super().__init__(task, **{k: v for k, v in vars().items() if k not in ["self", "task", "__class__"]})
if isinstance(block, str):
block = thelper.utils.import_class(block)
if not issubclass(block, Module):
raise AssertionError("block type must be subclass of thelper.nn.resnet.Module")
if isinstance(layers, str):
assert layers in ["18", "34"], "unknown basic block layer depth string postfix"
if layers == "18":
layers = [2, 2, 2, 2]
elif layers == "34":
layers = [3, 4, 6, 3]
if not isinstance(layers, list) or not isinstance(strides, list):
raise AssertionError("expected layers/strides to be provided as list of ints")
if len(layers) != len(strides):
raise AssertionError("layer/strides length mismatch")
# NOTE: conv1_config=[7,2,3] is the basic configuration of ResNet.
# other configuration more suitables for CIFAR for example can use conv1_config[3,1,1]
assert isinstance(conv1_config, list) and \
len(conv1_config) == 3 and all(isinstance(c, int) for c in conv1_config), \
"conv1 configuration must be a list of 3 parameters defining [kernel_size,stride,padding]"
self.input_channels = input_channels
self.flexible_input_res = flexible_input_res
self.pool_size = pool_size
self.head_type = head_type
self.coordconv = coordconv
self.radius_channel = radius_channel
self.pretrained = pretrained
self.inplanes = 64
self.conv1 = self._make_conv2d(
in_channels=input_channels, out_channels=self.inplanes,
kernel_size=conv1_config[0], stride=conv1_config[1],
padding=conv1_config[2], bias=False)
self.bn1 = torch.nn.BatchNorm2d(self.inplanes)
self.activ = get_activation_layer(activation)
if skip_max_pool:
self.maxpool = None
else:
self.maxpool = torch.nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
self.layer1 = self._make_layer(block, 64, layers[0], stride=strides[0], activation=activation)
self.layer2 = self._make_layer(block, 128, layers[1], stride=strides[1], activation=activation)
self.layer3 = self._make_layer(block, 256, layers[2], stride=strides[2], activation=activation)
self.layer4 = self._make_layer(block, 512, layers[3], stride=strides[3], activation=activation)
self.out_features = 512
self.layer5 = None
if len(layers) > 4:
self.layer5 = self._make_layer(block, 1024, layers[4], stride=strides[4])
self.out_features = 1024
if flexible_input_res:
self.avgpool = torch.nn.AdaptiveAvgPool2d(1)
else:
if pool_size < 1:
raise AssertionError("invalid avg pool size for non-flex resolution")
self.avgpool = torch.nn.AvgPool2d(pool_size, stride=1)
self.out_features *= block.expansion
self.fc = torch.nn.Linear(self.out_features, 1000) # output type/count will be specialized by task after init
self._init_weights(activation)
if pretrained:
# note: if using a non-default setup in the constructor, loading the pre-trained weights will most
# likely fail as the weights are downloaded from the pytorch model zoo for the regular resnet impls
import torchvision
default_weights_mapping = {
str([2, 2, 2, 2]) + str("BasicBlock"): "resnet18",
str([3, 4, 6, 3]) + str("BasicBlock"): "resnet34",
str([3, 4, 6, 3]) + str("Bottleneck"): "resnet50",
str([3, 4, 23, 3]) + str("Bottleneck"): "resnet101",
str([3, 8, 36, 3]) + str("Bottleneck"): "resnet152"
}
tag = str(layers) + block.__name__
assert tag in default_weights_mapping, "could not find corresponding weight url"
weights_url = torchvision.models.resnet.model_urls[default_weights_mapping[tag]]
state_dict = torchvision.models.utils.load_state_dict_from_url(weights_url)
self.load_state_dict(state_dict)
if isinstance(task, thelper.tasks.Segmentation):
# if base task is already associated with segmentation, add head attribute
if head_type is not None: # can also be manually defined (e.g. in autoencoder)
assert isinstance(head_type, str) and head_type in ["fcn", "deeplabv3"], \
f"unrecognized head type ('{head_type}') for segmentation resnet"
# note: head below will be fully instantiated when the task is assigned
self.fc = None
self.set_task(task)
def _init_weights(self, activation):
for m in self.modules():
if isinstance(m, torch.nn.Conv2d) or isinstance(m, torch.nn.ConvTranspose2d):
torch.nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity=activation)
elif isinstance(m, thelper.nn.coordconv.CoordConv2d) or isinstance(m, thelper.nn.coordconv.CoordConvTranspose2d):
torch.nn.init.kaiming_normal_(m.conv.weight, mode='fan_out', nonlinearity=activation)
elif isinstance(m, torch.nn.BatchNorm2d):
torch.nn.init.constant_(m.weight, 1)
torch.nn.init.constant_(m.bias, 0)
for m in self.modules():
if isinstance(m, Bottleneck):
torch.nn.init.constant_(m.bn3.weight, 0)
elif isinstance(m, BasicBlock) or isinstance(m, SqueezeExcitationBlock):
torch.nn.init.constant_(m.bn2.weight, 0)
def _make_conv2d(self, *args, **kwargs):
if self.coordconv:
return thelper.nn.coordconv.CoordConv2d(*args, radius_channel=self.radius_channel, **kwargs)
else:
return torch.nn.Conv2d(*args, **kwargs)
def _make_layer(self, block, planes, blocks, stride, activation):
downsample = None
if stride != 1 or self.inplanes != planes * block.expansion:
downsample = torch.nn.Sequential(
self._make_conv2d(self.inplanes, planes * block.expansion,
kernel_size=1, stride=stride, bias=False),
torch.nn.BatchNorm2d(planes * block.expansion),
)
layers = [block(self.inplanes, planes=planes, stride=stride,
downsample=downsample, activation=activation)]
self.inplanes = planes * block.expansion
for i in range(1, blocks):
layers.append(block(self.inplanes, planes))
return torch.nn.Sequential(*layers)
def get_embedding(self, x, pool=True):
x = self.conv1(x)
x = self.bn1(x)
x = self.activ(x)
if self.maxpool is not None:
x = self.maxpool(x)
x = self.layer1(x)
x = self.layer2(x)
x = self.layer3(x)
x = self.layer4(x)
if self.layer5 is not None:
x = self.layer5(x)
if pool:
x = self.avgpool(x)
x = x.view(x.size(0), -1)
return x
def forward(self, x):
if isinstance(self.task, thelper.tasks.Classification):
return self.fc(self.get_embedding(x, pool=True))
elif isinstance(self.task, thelper.tasks.Segmentation):
return self.fc(self.get_embedding(x), pool=False)
def set_task(self, task):
assert isinstance(task, (thelper.tasks.Classification, thelper.tasks.Segmentation)), \
"missing impl for non-classif task type"
num_classes = len(task.class_names)
if isinstance(task, thelper.tasks.Classification):
if self.fc.out_features != num_classes:
self.fc = torch.nn.Linear(self.out_features, num_classes)
elif isinstance(task, thelper.tasks.Segmentation):
import torchvision.models.segmentation
# note: heads below will be fully reinstantiated when the output class count changes
if self.fc is None or self.fc[len(self.fc) - 1].out_channels != num_classes:
if self.head_type == "fcn":
self.fc = torchvision.models.segmentation.fcn.FCNHead(self.out_features, num_classes)
elif self.head_type == "deeplabv3":
self.fc = torchvision.models.segmentation.deeplabv3.DeepLabHead(self.out_features, num_classes)
self.task = task
class ConvTailNet(torch.nn.Module):
"""DEPRECATED. Will be removed in a future version."""
def __init__(self, n_inputs, num_classes):
super(ConvTailNet, self).__init__()
self.conv1 = torch.nn.Conv2d(n_inputs, n_inputs, kernel_size=1, bias=False)
self.relu = torch.nn.ReLU(True)
self.conv2 = torch.nn.Conv2d(n_inputs, n_inputs, kernel_size=1, bias=False)
self.conv3 = torch.nn.Conv2d(n_inputs, num_classes, kernel_size=1, bias=False)
def forward(self, x):
x0 = x
x = self.conv1(x)
x = self.relu(x)
x = self.conv2(x)
x = self.relu(x)
x = torch.add(x0, x)
x = self.conv3(x)
return x
class ResNetFullyConv(ResNet):
"""DEPRECATED. Will be removed in a future version. Use the torchvision segmentation models or the ResNet above instead."""
def __init__(self, task, block="thelper.nn.resnet.BasicBlock",
layers=[3, 4, 6, 3], strides=[1, 2, 2, 2], input_channels=3,
flexible_input_res=False, pool_size=7, coordconv=False,
radius_channel=True, pretrained=False):
super().__init__(task=task, block=block, layers=layers, strides=strides, input_channels=input_channels,
flexible_input_res=flexible_input_res, pool_size=pool_size, coordconv=coordconv,
radius_channel=radius_channel, pretrained=pretrained)
self.set_task(task)
def forward(self, x):
x = self.get_embedding(x, pool=False)
x = self.avgpool(x)
x = self.fc(x)
x = torch.squeeze(x)
return x
def set_task(self, task):
assert isinstance(task, thelper.tasks.Classification), "missing impl for non-classif task type"
num_classes = len(task.class_names)
self.fc = ConvTailNet(self.out_features, num_classes)
self.task = task
class FCResNet(ResNet):
"""Fully Convolutional ResNet converter for pre-trained classification models."""
def __init__(self, task, ckptdata, map_location="cpu", avgpool_size=0):
if isinstance(ckptdata, str):
ckptdata = thelper.utils.load_checkpoint(ckptdata, map_location=map_location)
model_type = ckptdata["model_type"]
if model_type != "thelper.nn.resnet.ResNet":
raise AssertionError("cannot convert non-resnet model to fully conv with this impl")
model_params = ckptdata["model_params"]
if isinstance(ckptdata["task"], str):
old_model_task = thelper.tasks.create_task(ckptdata["task"])
else:
old_model_task = ckptdata["task"]
self.task = None
self.avgpool_size = avgpool_size
super().__init__(old_model_task, **model_params)
self.load_state_dict(ckptdata["model"], strict=False) # assumes model always stored as weight dict
self.finallayer = torch.nn.Conv2d(self.out_features, self.fc.out_features, kernel_size=1)
self.finallayer.weight = \
torch.nn.Parameter(self.fc.weight.view(self.fc.out_features, self.out_features, 1, 1))
self.finallayer.bias = torch.nn.Parameter(self.fc.bias)
self.set_task(task)
def forward(self, x):
x = self.get_embedding(x, pool=False)
if self.avgpool_size > 0:
x = torch.nn.functional.avg_pool2d(x, kernel_size=self.avgpool_size, stride=1)
x = self.finallayer(x)
return x
def set_task(self, task):
assert isinstance(task, (thelper.tasks.Segmentation, thelper.tasks.Classification)), \
"missing impl for non-segm/classif task type"
num_classes = len(task.class_names)
if self.fc.out_features != num_classes:
self.fc = torch.nn.Linear(self.out_features, num_classes)
self.finallayer = torch.nn.Conv2d(self.out_features, num_classes, kernel_size=1)
self.finallayer.weight = \
torch.nn.Parameter(self.fc.weight.view(self.fc.out_features, self.out_features, 1, 1))
self.finallayer.bias = torch.nn.Parameter(self.fc.bias)
self.task = task
class AutoEncoderResNet(ResNet):
"""Autoencoder-classifier architecture based on ResNet blocks+layers configurations."""
def __init__(self, task, output_pads=None, **kwargs):
assert isinstance(task, thelper.tasks.Classification)
super().__init__(task, activation="leaky_relu", **kwargs)
convt = thelper.nn.coordconv.CoordConvTranspose2d if self.coordconv else torch.nn.ConvTranspose2d
self.decoder_top = torch.nn.Sequential(
thelper.nn.coordconv.CoordConv2d(
self.out_features, self.out_features, kernel_size=1, stride=1, padding=0
),
torch.nn.BatchNorm2d(self.out_features),
torch.nn.LeakyReLU(),
)
self.decoder_depths = [self.out_features // 2 ** d for d in range(0, 5)]
self.output_pads = output_pads if not None else [1, 1, 1, 1, 1]
self.decoder_layers = torch.nn.ModuleList([
torch.nn.Sequential(
convt(depth, depth // 2, kernel_size=3, stride=2, padding=1, output_padding=out_pad),
torch.nn.BatchNorm2d(depth // 2),
torch.nn.LeakyReLU(),
)
for depth, out_pad in zip(self.decoder_depths, self.output_pads)
])
self.decoder_bottom = torch.nn.Sequential(
thelper.nn.coordconv.CoordConv2d(
self.decoder_depths[-1] // 2, self.input_channels,
kernel_size=3, stride=1, padding=1,
),
torch.nn.Tanh()
)
self._init_weights(activation="leaky_relu")
# note: cannot rely on pretrained imagenet weights since we reset just above
def forward(self, input):
featmap = self.get_embedding(input, pool=False)
embedding = self.avgpool(featmap)
embedding = embedding.view(embedding.size(0), -1)
class_logits = self.fc(embedding)
featmap = self.decoder_top(featmap)
for decoder_layer in self.decoder_layers:
featmap = decoder_layer(featmap)
reconstruction = self.decoder_bottom(featmap)
return class_logits, reconstruction
class FakeModule(torch.nn.Module):
def forward(self, inputs):
return inputs
class AutoEncoderSkipResNet(ResNet):
"""Autoencoder-U-Net architecture based on ResNet blocks+layers configurations."""
def __init__(self, task, output_pads=None, decoder_dropout=False, dropout_prob=0.1, **kwargs):
assert isinstance(task, thelper.tasks.Segmentation)
super().__init__(task, activation="leaky_relu", **kwargs)
convt = thelper.nn.coordconv.CoordConvTranspose2d if self.coordconv else torch.nn.ConvTranspose2d
pass_through_layer = FakeModule()
self.decoder_top = torch.nn.Sequential(
thelper.nn.coordconv.CoordConv2d(
self.out_features, self.out_features, kernel_size=1, stride=1, padding=0
),
torch.nn.BatchNorm2d(self.out_features),
torch.nn.Dropout2d(p=dropout_prob) if decoder_dropout else pass_through_layer,
torch.nn.LeakyReLU(),
)
self.output_pads = output_pads if not None else [1, 1, 1, 1, 1]
self.ae_decoder_depths = [self.out_features // 2 ** d for d in range(0, 5)]
self.ae_decoder_layers = torch.nn.ModuleList([
torch.nn.Sequential(
convt(depth, depth // 2, kernel_size=3, stride=2, padding=1, output_padding=out_pad),
torch.nn.BatchNorm2d(depth // 2),
torch.nn.Dropout2d(p=dropout_prob) if decoder_dropout else pass_through_layer,
torch.nn.LeakyReLU(), # try w/ regular? @@@@
)
for depth, out_pad in zip(self.ae_decoder_depths, self.output_pads)
])
self.ae_decoder_bottom = torch.nn.Sequential(
thelper.nn.coordconv.CoordConv2d(
self.ae_decoder_depths[-1] // 2, self.input_channels,
kernel_size=3, stride=1, padding=1,
),
torch.nn.Tanh()
)
self.unet_decoder_depths = [
(self.out_features, self.out_features // 2, self.out_features // 2),
(self.out_features, self.out_features // 4, self.out_features // 4),
(self.out_features // 2, self.out_features // 8, self.out_features // 8),
(self.out_features // 4, self.out_features // 8, self.out_features // 8),
(self.out_features // 4, self.out_features // 8, self.out_features // 8),
]
self.unet_decoder_layers = torch.nn.ModuleList([
torch.nn.Sequential(
convt(d_in, d_mid, kernel_size=3, stride=2, padding=1, output_padding=out_pad),
torch.nn.BatchNorm2d(d_mid),
torch.nn.LeakyReLU(), # try w/ regular? @@@@
self._make_conv2d(d_mid, d_out, kernel_size=3, stride=1, padding=1, bias=False),
torch.nn.BatchNorm2d(d_out),
torch.nn.Dropout2d(p=dropout_prob) if decoder_dropout else pass_through_layer,
torch.nn.LeakyReLU(), # try w/ regular? @@@@
)
for (d_in, d_mid, d_out), out_pad in zip(self.unet_decoder_depths, self.output_pads)
])
assert self.unet_decoder_depths[-1][2] > len(self.task.class_names), "woopsie"
self.unet_decoder_bottom = torch.nn.Conv2d(
self.unet_decoder_depths[-1][2], len(self.task.class_names),
kernel_size=1, stride=1, padding=0,
)
self._init_weights(activation="leaky_relu")
# note: cannot rely on pretrained imagenet weights since we reset just above
def forward(self, input):
# forward while keeping refs for latent build? @@@
encoder1 = self.activ(self.bn1(self.conv1(input)))
if self.maxpool is not None:
encoder2 = self.layer1(self.maxpool(encoder1))
else:
encoder2 = self.layer1(encoder1)
encoder3 = self.layer2(encoder2)
encoder4 = self.layer3(encoder3)
encoder5 = self.layer4(encoder4)
assert self.layer5 is None, "missing e5+ impl"
featmap = self.decoder_top(encoder5)
reconstruction = featmap
for decoder_layer in self.ae_decoder_layers:
reconstruction = decoder_layer(reconstruction)
reconstruction = self.ae_decoder_bottom(reconstruction)
encoder_maps = [None, encoder4, encoder3, encoder2, encoder1]
for decoder_layer, encoder_map in zip(self.unet_decoder_layers, encoder_maps):
if encoder_map is not None:
featmap = torch.cat([featmap, encoder_map], dim=1)
featmap = decoder_layer(featmap)
class_logits = self.unet_decoder_bottom(featmap)
return class_logits, reconstruction
