From 1577c52976f16bc481b6b033d839c6311870830e Mon Sep 17 00:00:00 2001 From: Ross Wightman Date: Fri, 1 Mar 2019 15:29:02 -0800 Subject: [PATCH] Resnext added, changes to bring it and seresnet in line with rest of models --- models/model_factory.py | 6 +- models/resnext.py | 194 ++++++++++++++++++++++++++++++++++++++++ models/senet.py | 70 +++++++++------ scheduler/cosine_lr.py | 21 ++++- scheduler/tanh_lr.py | 9 ++ train.py | 20 +++-- 6 files changed, 277 insertions(+), 43 deletions(-) create mode 100644 models/resnext.py diff --git a/models/model_factory.py b/models/model_factory.py index 806b6ee..37eeac9 100644 --- a/models/model_factory.py +++ b/models/model_factory.py @@ -10,8 +10,8 @@ from .densenet import densenet161, densenet121, densenet169, densenet201 from .resnet import resnet18, resnet34, resnet50, resnet101, resnet152 from .fbresnet200 import fbresnet200 from .dpn import dpn68, dpn68b, dpn92, dpn98, dpn131, dpn107 -from .senet import seresnet18, seresnet34, seresnet50, seresnet101, seresnet152,\ - seresnext50_32x4d, seresnext101_32x4d +from .senet import seresnet18, seresnet34, seresnet50, seresnet101, seresnet152, \ + seresnext26_32x4d, seresnext50_32x4d, seresnext101_32x4d from .resnext import resnext50, resnext101, resnext152 @@ -112,6 +112,8 @@ def create_model( model = seresnet101(num_classes=num_classes, pretrained=pretrained, **kwargs) elif model_name == 'seresnet152': model = seresnet152(num_classes=num_classes, pretrained=pretrained, **kwargs) + elif model_name == 'seresnext26_32x4d': + model = seresnext26_32x4d(num_classes=num_classes, pretrained=pretrained, **kwargs) elif model_name == 'seresnext50_32x4d': model = seresnext50_32x4d(num_classes=num_classes, pretrained=pretrained, **kwargs) elif model_name == 'seresnext101_32x4d': diff --git a/models/resnext.py b/models/resnext.py new file mode 100644 index 0000000..ee4ea51 --- /dev/null +++ b/models/resnext.py @@ -0,0 +1,194 @@ +import torch.nn as nn +import torch.nn.functional as F +import math +import torch.utils.model_zoo as model_zoo +from models.adaptive_avgmax_pool import AdaptiveAvgMaxPool2d + +__all__ = ['ResNeXt', 'resnext50', 'resnext101', 'resnext152'] + + +def conv3x3(in_planes, out_planes, stride=1): + "3x3 convolution with padding" + return nn.Conv2d( + in_planes, out_planes, kernel_size=3, stride=stride, padding=1, bias=False) + + +class ResNeXtBottleneckC(nn.Module): + expansion = 4 + + def __init__(self, inplanes, planes, stride=1, downsample=None, cardinality=32, base_width=4): + super(ResNeXtBottleneckC, self).__init__() + + width = math.floor(planes / 64 * cardinality * base_width) + + self.conv1 = nn.Conv2d(inplanes, width, kernel_size=1, bias=False) + self.bn1 = nn.BatchNorm2d(width) + self.conv2 = nn.Conv2d(width, width, kernel_size=3, stride=stride, + padding=1, bias=False, groups=cardinality) + self.bn2 = nn.BatchNorm2d(width) + self.conv3 = nn.Conv2d(width, planes * 4, kernel_size=1, bias=False) + self.bn3 = nn.BatchNorm2d(planes * 4) + self.relu = nn.ReLU(inplace=True) + self.downsample = downsample + self.stride = stride + + def forward(self, x): + residual = x + + out = self.conv1(x) + out = self.bn1(out) + out = self.relu(out) + + out = self.conv2(out) + out = self.bn2(out) + out = self.relu(out) + + out = self.conv3(out) + out = self.bn3(out) + + if self.downsample is not None: + residual = self.downsample(x) + + out += residual + out = self.relu(out) + + return out + + +class ResNeXt(nn.Module): + + def __init__(self, block, layers, num_classes=1000, cardinality=32, base_width=4, shortcut='C', + drop_rate=0., global_pool='avg'): + self.num_classes = num_classes + self.inplanes = 64 + self.cardinality = cardinality + self.base_width = base_width + self.shortcut = shortcut + self.drop_rate = drop_rate + super(ResNeXt, self).__init__() + self.conv1 = nn.Conv2d(3, 64, kernel_size=7, stride=2, padding=3, bias=False) + self.bn1 = nn.BatchNorm2d(64) + self.relu = nn.ReLU(inplace=True) + self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1) + self.layer1 = self._make_layer(block, 64, layers[0]) + self.layer2 = self._make_layer(block, 128, layers[1], stride=2) + self.layer3 = self._make_layer(block, 256, layers[2], stride=2) + self.layer4 = self._make_layer(block, 512, layers[3], stride=2) + self.avgpool = AdaptiveAvgMaxPool2d(pool_type=global_pool) + self.num_features = 512 * block.expansion + self.fc = nn.Linear(self.num_features, num_classes) + + for m in self.modules(): + if isinstance(m, nn.Conv2d): + n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels + m.weight.data.normal_(0, math.sqrt(2. / n)) + elif isinstance(m, nn.BatchNorm2d): + m.weight.data.fill_(1) + m.bias.data.zero_() + + def _make_layer(self, block, planes, blocks, stride=1): + downsample = None + reshape = 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( + nn.Conv2d( + self.inplanes, planes * block.expansion, kernel_size=1, stride=stride, bias=False), + 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)] + 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): + layers.append(block(self.inplanes, planes, 1, shortcut, self.cardinality, self.base_width)) + + return nn.Sequential(*layers) + + def get_classifier(self): + return self.fc + + def reset_classifier(self, num_classes, global_pool='avg'): + self.avgpool = AdaptiveAvgMaxPool2d(pool_type=global_pool) + self.num_classes = num_classes + del self.fc + if num_classes: + self.fc = nn.Linear(self.num_features, num_classes) + else: + self.fc = None + + def forward_features(self, x, pool=True): + x = self.conv1(x) + x = self.bn1(x) + x = self.relu(x) + x = self.maxpool(x) + + x = self.layer1(x) + x = self.layer2(x) + x = self.layer3(x) + x = self.layer4(x) + if pool: + x = self.avgpool(x) + x = x.view(x.size(0), -1) + return x + + def forward(self, x): + x = self.forward_features(x) + if self.drop_rate > 0.: + x = F.dropout(x, p=self.drop_rate, training=self.training) + x = self.fc(x) + + return x + + +def resnext50(cardinality=32, base_width=4, shortcut='C', pretrained=False, **kwargs): + """Constructs a ResNeXt-50 model. + + Args: + cardinality (int): Cardinality of the aggregated transform + base_width (int): Base width of the grouped convolution + shortcut ('A'|'B'|'C'): 'B' use 1x1 conv to downsample, 'C' use 1x1 conv on every residual connection + """ + model = ResNeXt( + ResNeXtBottleneckC, [3, 4, 6, 3], cardinality=cardinality, + base_width=base_width, shortcut=shortcut, **kwargs) + return model + + +def resnext101(cardinality=32, base_width=4, shortcut='C', pretrained=False, **kwargs): + """Constructs a ResNeXt-101 model. + + Args: + cardinality (int): Cardinality of the aggregated transform + base_width (int): Base width of the grouped convolution + shortcut ('A'|'B'|'C'): 'B' use 1x1 conv to downsample, 'C' use 1x1 conv on every residual connection + """ + model = ResNeXt( + ResNeXtBottleneckC, [3, 4, 23, 3], cardinality=cardinality, + base_width=base_width, shortcut=shortcut, **kwargs) + return model + + +def resnext152(cardinality=32, base_width=4, shortcut='C', pretrained=False, **kwargs): + """Constructs a ResNeXt-152 model. + + Args: + cardinality (int): Cardinality of the aggregated transform + base_width (int): Base width of the grouped convolution + shortcut ('A'|'B'|'C'): 'B' use 1x1 conv to downsample, 'C' use 1x1 conv on every residual connection + """ + model = ResNeXt( + ResNeXtBottleneckC, [3, 8, 36, 3], cardinality=cardinality, + base_width=base_width, shortcut=shortcut, **kwargs) + return model diff --git a/models/senet.py b/models/senet.py index cf169f2..2e88d48 100644 --- a/models/senet.py +++ b/models/senet.py @@ -7,7 +7,9 @@ from collections import OrderedDict import math import torch.nn as nn +import torch.nn.functional as F from torch.utils import model_zoo +from models.adaptive_avgmax_pool import AdaptiveAvgMaxPool2d __all__ = ['SENet', 'senet154', 'seresnet50', 'seresnet101', 'seresnet152', 'seresnext50_32x4d', 'seresnext101_32x4d'] @@ -193,9 +195,9 @@ class SEResNetBlock(nn.Module): class SENet(nn.Module): - def __init__(self, block, layers, groups, reduction, dropout_p=0.2, + def __init__(self, block, layers, groups, reduction, drop_rate=0.2, inchans=3, inplanes=128, input_3x3=True, downsample_kernel_size=3, - downsample_padding=1, num_classes=1000): + downsample_padding=1, num_classes=1000, global_pool='avg'): """ Parameters ---------- @@ -304,8 +306,8 @@ class SENet(nn.Module): downsample_kernel_size=downsample_kernel_size, downsample_padding=downsample_padding ) - self.avg_pool = nn.AdaptiveAvgPool2d(1) - self.dropout = nn.Dropout(dropout_p) if dropout_p is not None else None + self.avg_pool = AdaptiveAvgMaxPool2d(pool_type=global_pool) + self.drop_rate = drop_rate self.num_features = 512 * block.expansion self.last_linear = nn.Linear(self.num_features, num_classes) @@ -354,8 +356,8 @@ class SENet(nn.Module): return x def logits(self, x): - if self.dropout is not None: - x = self.dropout(x) + if self.drop_rate > 0.: + x = F.dropout(x, p=self.drop_rate, training=self.training) x = self.last_linear(x) return x @@ -375,79 +377,89 @@ def _load_pretrained(model, url, inchans=3): model.load_state_dict(state_dict) -def senet154(num_classes=1000, inchans=3, pretrained='imagenet'): +def senet154(num_classes=1000, inchans=3, pretrained='imagenet', **kwargs): model = SENet(SEBottleneck, [3, 8, 36, 3], groups=64, reduction=16, - dropout_p=0.2, num_classes=num_classes) + num_classes=num_classes, **kwargs) if pretrained: _load_pretrained(model, model_urls['senet154'], inchans) return model -def seresnet18(num_classes=1000, inchans=3, pretrained='imagenet'): +def seresnet18(num_classes=1000, inchans=3, pretrained='imagenet', **kwargs): model = SENet(SEResNetBlock, [2, 2, 2, 2], groups=1, reduction=16, - dropout_p=None, inplanes=64, input_3x3=False, + inplanes=64, input_3x3=False, downsample_kernel_size=1, downsample_padding=0, - num_classes=num_classes) + num_classes=num_classes, **kwargs) if pretrained: _load_pretrained(model, model_urls['seresnet18'], inchans) return model -def seresnet34(num_classes=1000, inchans=3, pretrained='imagenet'): +def seresnet34(num_classes=1000, inchans=3, pretrained='imagenet', **kwargs): model = SENet(SEResNetBlock, [3, 4, 6, 3], groups=1, reduction=16, - dropout_p=None, inplanes=64, input_3x3=False, + inplanes=64, input_3x3=False, downsample_kernel_size=1, downsample_padding=0, - num_classes=num_classes) + num_classes=num_classes, **kwargs) if pretrained: _load_pretrained(model, model_urls['seresnet34'], inchans) return model -def seresnet50(num_classes=1000, inchans=3, pretrained='imagenet'): +def seresnet50(num_classes=1000, inchans=3, pretrained='imagenet', **kwargs): model = SENet(SEResNetBottleneck, [3, 4, 6, 3], groups=1, reduction=16, - dropout_p=None, inplanes=64, input_3x3=False, + inplanes=64, input_3x3=False, downsample_kernel_size=1, downsample_padding=0, - num_classes=num_classes) + num_classes=num_classes, **kwargs) if pretrained: _load_pretrained(model, model_urls['seresnet50'], inchans) return model -def seresnet101(num_classes=1000, inchans=3, pretrained='imagenet'): +def seresnet101(num_classes=1000, inchans=3, pretrained='imagenet', **kwargs): model = SENet(SEResNetBottleneck, [3, 4, 23, 3], groups=1, reduction=16, - dropout_p=None, inplanes=64, input_3x3=False, + inplanes=64, input_3x3=False, downsample_kernel_size=1, downsample_padding=0, - num_classes=num_classes) + num_classes=num_classes, **kwargs) if pretrained: _load_pretrained(model, model_urls['seresnet101'], inchans) return model -def seresnet152(num_classes=1000, inchans=3, pretrained='imagenet'): +def seresnet152(num_classes=1000, inchans=3, pretrained='imagenet', **kwargs): model = SENet(SEResNetBottleneck, [3, 8, 36, 3], groups=1, reduction=16, - dropout_p=None, inplanes=64, input_3x3=False, + inplanes=64, input_3x3=False, downsample_kernel_size=1, downsample_padding=0, - num_classes=num_classes) + num_classes=num_classes, **kwargs) if pretrained: _load_pretrained(model, model_urls['seresnet152'], inchans) return model -def seresnext50_32x4d(num_classes=1000, inchans=3, pretrained='imagenet'): +def seresnext26_32x4d(num_classes=1000, inchans=3, pretrained='imagenet', **kwargs): + model = SENet(SEResNeXtBottleneck, [2, 2, 2, 2], groups=32, reduction=16, + inplanes=64, input_3x3=False, + downsample_kernel_size=1, downsample_padding=0, + num_classes=num_classes, **kwargs) + if pretrained: + _load_pretrained(model, model_urls['se_resnext26_32x4d'], inchans) + return model + + +def seresnext50_32x4d(num_classes=1000, inchans=3, pretrained='imagenet', **kwargs): model = SENet(SEResNeXtBottleneck, [3, 4, 6, 3], groups=32, reduction=16, - dropout_p=None, inplanes=64, input_3x3=False, + inplanes=64, input_3x3=False, downsample_kernel_size=1, downsample_padding=0, - num_classes=num_classes) + num_classes=num_classes, **kwargs) if pretrained: _load_pretrained(model, model_urls['seresnext50_32x4d'], inchans) return model -def seresnext101_32x4d(num_classes=1000, inchans=3, pretrained='imagenet'): +def seresnext101_32x4d(num_classes=1000, inchans=3, pretrained='imagenet', **kwargs): model = SENet(SEResNeXtBottleneck, [3, 4, 23, 3], groups=32, reduction=16, - dropout_p=None, inplanes=64, input_3x3=False, + inplanes=64, input_3x3=False, downsample_kernel_size=1, downsample_padding=0, - num_classes=num_classes) + num_classes=num_classes, **kwargs) if pretrained: _load_pretrained(model, model_urls['seresnext101_32x4d'], inchans) return model diff --git a/scheduler/cosine_lr.py b/scheduler/cosine_lr.py index 97aeb9f..36a13ea 100644 --- a/scheduler/cosine_lr.py +++ b/scheduler/cosine_lr.py @@ -24,6 +24,7 @@ class CosineLRScheduler(Scheduler): warmup_t=0, warmup_lr_init=0, warmup_prefix=False, + cycle_limit=0, t_in_epochs=True, initialize=True) -> None: super().__init__(optimizer, param_group_field="lr", initialize=initialize) @@ -37,6 +38,7 @@ class CosineLRScheduler(Scheduler): self.t_mul = t_mul self.lr_min = lr_min self.decay_rate = decay_rate + self.cycle_limit = cycle_limit self.warmup_t = warmup_t self.warmup_lr_init = warmup_lr_init self.warmup_prefix = warmup_prefix @@ -67,9 +69,13 @@ class CosineLRScheduler(Scheduler): lr_min = self.lr_min * gamma lr_max_values = [v * gamma for v in self.base_values] - lrs = [ - lr_min + 0.5 * (lr_max - lr_min) * (1 + math.cos(math.pi * t_curr / t_i)) for lr_max in lr_max_values - ] + if self.cycle_limit == 0 or (self.cycle_limit > 0 and i < self.cycle_limit): + lrs = [ + lr_min + 0.5 * (lr_max - lr_min) * (1 + math.cos(math.pi * t_curr / t_i)) for lr_max in lr_max_values + ] + else: + lrs = [self.lr_min for _ in self.base_values] + return lrs def get_epoch_values(self, epoch: int): @@ -83,3 +89,12 @@ class CosineLRScheduler(Scheduler): return self._get_lr(num_updates) else: return None + + def get_cycle_length(self, cycles=0): + if not cycles: + cycles = self.cycle_limit + assert cycles > 0 + if self.t_mul == 1.0: + return self.t_initial * cycles + else: + return int(math.floor(-self.t_initial * (self.t_mul ** cycles - 1) / (1 - self.t_mul))) diff --git a/scheduler/tanh_lr.py b/scheduler/tanh_lr.py index a942e67..4335e9b 100644 --- a/scheduler/tanh_lr.py +++ b/scheduler/tanh_lr.py @@ -97,3 +97,12 @@ class TanhLRScheduler(Scheduler): return self._get_lr(num_updates) else: return None + + def get_cycle_length(self, cycles=0): + if not cycles: + cycles = self.cycle_limit + assert cycles > 0 + if self.t_mul == 1.0: + return self.t_initial * cycles + else: + return int(math.floor(-self.t_initial * (self.t_mul ** cycles - 1) / (1 - self.t_mul))) diff --git a/train.py b/train.py index 43aeeaf..c8432fa 100644 --- a/train.py +++ b/train.py @@ -91,7 +91,6 @@ def main(): output_dir = get_outdir(output_base, 'train', exp_name) batch_size = args.batch_size - num_epochs = args.epochs torch.manual_seed(args.seed) dataset_train = Dataset( @@ -155,9 +154,7 @@ def main(): else: model.cuda() - train_loss_fn = validate_loss_fn = torch.nn.CrossEntropyLoss() - train_loss_fn = train_loss_fn.cuda() - validate_loss_fn = validate_loss_fn.cuda() + train_loss_fn = validate_loss_fn = torch.nn.CrossEntropyLoss().cuda() if args.opt.lower() == 'sgd': optimizer = optim.SGD( @@ -183,34 +180,39 @@ def main(): #if optimizer_state is not None: # optimizer.load_state_dict(optimizer_state) + num_epochs = args.epochs if args.sched == 'cosine': lr_scheduler = scheduler.CosineLRScheduler( optimizer, - t_initial=130, - t_mul=1.0, - lr_min=0, + t_initial=args.epochs, + t_mul=1.5, + lr_min=1e-5, decay_rate=args.decay_rate, warmup_lr_init=1e-4, warmup_t=3, + cycle_limit=3, t_in_epochs=True, ) + num_epochs = lr_scheduler.get_cycle_length() + 10 elif args.sched == 'tanh': lr_scheduler = scheduler.TanhLRScheduler( optimizer, - t_initial=130, + t_initial=args.epochs, t_mul=1.0, - lr_min=1e-6, + lr_min=1e-5, warmup_lr_init=.001, warmup_t=3, cycle_limit=1, t_in_epochs=True, ) + num_epochs = lr_scheduler.get_cycle_length() + 10 else: lr_scheduler = scheduler.StepLRScheduler( optimizer, decay_t=args.decay_epochs, decay_rate=args.decay_rate, ) + print(num_epochs) saver = CheckpointSaver(checkpoint_dir=output_dir) best_loss = None -- GitLab