Merge pull request #150 from rwightman/regnet

Add RegNet models and weights

timm/models/__init__.py

@@ -19,6 +19,7 @@ from .hrnet import *
 from .sknet import *
 from .tresnet import *
 from .resnest import *
+from .regnet import *
 
 from .registry import *
 from .factory import create_model

timm/models/layers/se.py

@@ -3,10 +3,10 @@ from torch import nn as nn
 
 class SEModule(nn.Module):
 
-    def __init__(self, channels, reduction=16, act_layer=nn.ReLU):
+    def __init__(self, channels, reduction=16, act_layer=nn.ReLU, min_channels=8, reduction_channels=None):
         super(SEModule, self).__init__()
         self.avg_pool = nn.AdaptiveAvgPool2d(1)
-        reduction_channels = max(channels // reduction, 8)
+        reduction_channels = reduction_channels or max(channels // reduction, min_channels)
         self.fc1 = nn.Conv2d(
             channels, reduction_channels, kernel_size=1, padding=0, bias=True)
         self.act = act_layer(inplace=True)

timm/models/regnet.py

@@ -0,0 +1,485 @@
+"""RegNet
+
+Paper: `Designing Network Design Spaces` - https://arxiv.org/abs/2003.13678
+Original Impl: https://github.com/facebookresearch/pycls/blob/master/pycls/models/regnet.py
+
+Based on original PyTorch impl linked above, but re-wrote to use my own blocks (adapted from ResNet here)
+and cleaned up with more descriptive variable names.
+
+Weights from original impl have been modified
+* first layer from BGR -> RGB as most PyTorch models are
+* removed training specific dict entries from checkpoints and keep model state_dict only
+* remap names to match the ones here
+
+"""
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import numpy as np
+
+from .registry import register_model
+from .helpers import load_pretrained
+from .layers import SelectAdaptivePool2d, AvgPool2dSame, ConvBnAct, SEModule
+from timm.data import IMAGENET_DEFAULT_MEAN, IMAGENET_DEFAULT_STD
+
+
+def _mcfg(**kwargs):
+    cfg = dict(se_ratio=0., bottle_ratio=1., stem_width=32)
+    cfg.update(**kwargs)
+    return cfg
+
+
+# Model FLOPS = three trailing digits * 10^8
+model_cfgs = dict(
+    x_002=_mcfg(w0=24, wa=36.44, wm=2.49, group_w=8, depth=13),
+    x_004=_mcfg(w0=24, wa=24.48, wm=2.54, group_w=16, depth=22),
+    x_006=_mcfg(w0=48, wa=36.97, wm=2.24, group_w=24, depth=16),
+    x_008=_mcfg(w0=56, wa=35.73, wm=2.28, group_w=16, depth=16),
+    x_016=_mcfg(w0=80, wa=34.01, wm=2.25, group_w=24, depth=18),
+    x_032=_mcfg(w0=88, wa=26.31, wm=2.25, group_w=48, depth=25),
+    x_040=_mcfg(w0=96, wa=38.65, wm=2.43, group_w=40, depth=23),
+    x_064=_mcfg(w0=184, wa=60.83, wm=2.07, group_w=56, depth=17),
+    x_080=_mcfg(w0=80, wa=49.56, wm=2.88, group_w=120, depth=23),
+    x_120=_mcfg(w0=168, wa=73.36, wm=2.37, group_w=112, depth=19),
+    x_160=_mcfg(w0=216, wa=55.59, wm=2.1, group_w=128, depth=22),
+    x_320=_mcfg(w0=320, wa=69.86, wm=2.0, group_w=168, depth=23),
+    y_002=_mcfg(w0=24, wa=36.44, wm=2.49, group_w=8, depth=13, se_ratio=0.25),
+    y_004=_mcfg(w0=48, wa=27.89, wm=2.09, group_w=8, depth=16, se_ratio=0.25),
+    y_006=_mcfg(w0=48, wa=32.54, wm=2.32, group_w=16, depth=15, se_ratio=0.25),
+    y_008=_mcfg(w0=56, wa=38.84, wm=2.4, group_w=16, depth=14, se_ratio=0.25),
+    y_016=_mcfg(w0=48, wa=20.71, wm=2.65, group_w=24, depth=27, se_ratio=0.25),
+    y_032=_mcfg(w0=80, wa=42.63, wm=2.66, group_w=24, depth=21, se_ratio=0.25),
+    y_040=_mcfg(w0=96, wa=31.41, wm=2.24, group_w=64, depth=22, se_ratio=0.25),
+    y_064=_mcfg(w0=112, wa=33.22, wm=2.27, group_w=72, depth=25, se_ratio=0.25),
+    y_080=_mcfg(w0=192, wa=76.82, wm=2.19, group_w=56, depth=17, se_ratio=0.25),
+    y_120=_mcfg(w0=168, wa=73.36, wm=2.37, group_w=112, depth=19, se_ratio=0.25),
+    y_160=_mcfg(w0=200, wa=106.23, wm=2.48, group_w=112, depth=18, se_ratio=0.25),
+    y_320=_mcfg(w0=232, wa=115.89, wm=2.53, group_w=232, depth=20, se_ratio=0.25),
+)
+
+
+def _cfg(url=''):
+    return {
+        'url': url, 'num_classes': 1000, 'input_size': (3, 224, 224), 'pool_size': (7, 7),
+        'crop_pct': 0.875, 'interpolation': 'bicubic',
+        'mean': IMAGENET_DEFAULT_MEAN, 'std': IMAGENET_DEFAULT_STD,
+        'first_conv': 'stem.conv', 'classifier': 'head.fc',
+    }
+
+
+default_cfgs = dict(
+    x_002=_cfg(url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-regnet/regnetx_002-e7e85e5c.pth'),
+    x_004=_cfg(url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-regnet/regnetx_004-7d0e9424.pth'),
+    x_006=_cfg(url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-regnet/regnetx_006-85ec1baa.pth'),
+    x_008=_cfg(url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-regnet/regnetx_008-d8b470eb.pth'),
+    x_016=_cfg(url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-regnet/regnetx_016-65ca972a.pth'),
+    x_032=_cfg(url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-regnet/regnetx_032-ed0c7f7e.pth'),
+    x_040=_cfg(url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-regnet/regnetx_040-73c2a654.pth'),
+    x_064=_cfg(url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-regnet/regnetx_064-29278baa.pth'),
+    x_080=_cfg(url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-regnet/regnetx_080-7c7fcab1.pth'),
+    x_120=_cfg(url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-regnet/regnetx_120-65d5521e.pth'),
+    x_160=_cfg(url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-regnet/regnetx_160-c98c4112.pth'),
+    x_320=_cfg(url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-regnet/regnetx_320-8ea38b93.pth'),
+    y_002=_cfg(url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-regnet/regnety_002-e68ca334.pth'),
+    y_004=_cfg(url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-regnet/regnety_004-0db870e6.pth'),
+    y_006=_cfg(url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-regnet/regnety_006-c67e57ec.pth'),
+    y_008=_cfg(url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-regnet/regnety_008-dc900dbe.pth'),
+    y_016=_cfg(url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-regnet/regnety_016-54367f74.pth'),
+    y_032=_cfg(url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-regnet/regnety_032-62b47782.pth'),
+    y_040=_cfg(url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-regnet/regnety_040-f0d569f9.pth'),
+    y_064=_cfg(url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-regnet/regnety_064-0a48325c.pth'),
+    y_080=_cfg(url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-regnet/regnety_080-e7f3eb93.pth'),
+    y_120=_cfg(url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-regnet/regnety_120-721ba79a.pth'),
+    y_160=_cfg(url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-regnet/regnety_160-d64013cd.pth'),
+    y_320=_cfg(url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-regnet/regnety_320-ba464b29.pth'),
+)
+
+
+def quantize_float(f, q):
+    """Converts a float to closest non-zero int divisible by q."""
+    return int(round(f / q) * q)
+
+
+def adjust_widths_groups_comp(widths, bottle_ratios, groups):
+    """Adjusts the compatibility of widths and groups."""
+    bottleneck_widths = [int(w * b) for w, b in zip(widths, bottle_ratios)]
+    groups = [min(g, w_bot) for g, w_bot in zip(groups, bottleneck_widths)]
+    bottleneck_widths = [quantize_float(w_bot, g) for w_bot, g in zip(bottleneck_widths, groups)]
+    widths = [int(w_bot / b) for w_bot, b in zip(bottleneck_widths, bottle_ratios)]
+    return widths, groups
+
+
+def generate_regnet(width_slope, width_initial, width_mult, depth, q=8):
+    """Generates per block widths from RegNet parameters."""
+    assert width_slope >= 0 and width_initial > 0 and width_mult > 1 and width_initial % q == 0
+    widths_cont = np.arange(depth) * width_slope + width_initial
+    width_exps = np.round(np.log(widths_cont / width_initial) / np.log(width_mult))
+    widths = width_initial * np.power(width_mult, width_exps)
+    widths = np.round(np.divide(widths, q)) * q
+    num_stages, max_stage = len(np.unique(widths)), width_exps.max() + 1
+    widths, widths_cont = widths.astype(int).tolist(), widths_cont.tolist()
+    return widths, num_stages, max_stage, widths_cont
+
+
+class Bottleneck(nn.Module):
+    """ RegNet Bottleneck
+
+    This is almost exactly the same as a ResNet Bottlneck. The main difference is the SE block is moved from
+    after conv3 to after conv2. Otherwise, it's just redefining the arguments for groups/bottleneck channels.
+    """
+
+    def __init__(self, in_chs, out_chs, stride=1, bottleneck_ratio=1, group_width=1, se_ratio=0.25,
+                 dilation=1, first_dilation=None, downsample=None, act_layer=nn.ReLU, norm_layer=nn.BatchNorm2d,
+                 aa_layer=None, drop_block=None, drop_path=None):
+        super(Bottleneck, self).__init__()
+        bottleneck_chs = int(round(out_chs * bottleneck_ratio))
+        groups = bottleneck_chs // group_width
+        first_dilation = first_dilation or dilation
+
+        cargs = dict(act_layer=act_layer, norm_layer=norm_layer, aa_layer=aa_layer, drop_block=drop_block)
+        self.conv1 = ConvBnAct(in_chs, bottleneck_chs, kernel_size=1, **cargs)
+        self.conv2 = ConvBnAct(
+            bottleneck_chs, bottleneck_chs, kernel_size=3, stride=stride, dilation=first_dilation,
+            groups=groups, **cargs)
+        if se_ratio:
+            se_channels = int(round(in_chs * se_ratio))
+            self.se = SEModule(bottleneck_chs, reduction_channels=se_channels)
+        else:
+            self.se = None
+        cargs['act_layer'] = None
+        self.conv3 = ConvBnAct(bottleneck_chs, out_chs, kernel_size=1, **cargs)
+        self.act3 = act_layer(inplace=True)
+        self.downsample = downsample
+        self.drop_path = drop_path
+
+    def zero_init_last_bn(self):
+        nn.init.zeros_(self.conv3.bn.weight)
+
+    def forward(self, x):
+        shortcut = x
+        x = self.conv1(x)
+        x = self.conv2(x)
+        if self.se is not None:
+            x = self.se(x)
+        x = self.conv3(x)
+        if self.drop_path is not None:
+            x = self.drop_path(x)
+        if self.downsample is not None:
+            shortcut = self.downsample(shortcut)
+        x += shortcut
+        x = self.act3(x)
+        return x
+
+
+def downsample_conv(
+        in_chs, out_chs, kernel_size, stride=1, dilation=1, first_dilation=None, norm_layer=None):
+    norm_layer = norm_layer or nn.BatchNorm2d
+    kernel_size = 1 if stride == 1 and dilation == 1 else kernel_size
+    first_dilation = (first_dilation or dilation) if kernel_size > 1 else 1
+    return ConvBnAct(
+        in_chs, out_chs, kernel_size, stride=stride, dilation=first_dilation, norm_layer=norm_layer, act_layer=None)
+
+
+def downsample_avg(
+        in_chs, out_chs, kernel_size, stride=1, dilation=1, first_dilation=None, norm_layer=None):
+    """ AvgPool Downsampling as in 'D' ResNet variants. This is not in RegNet space but I might experiment."""
+    norm_layer = norm_layer or nn.BatchNorm2d
+    avg_stride = stride if dilation == 1 else 1
+    pool = nn.Identity()
+    if stride > 1 or dilation > 1:
+        avg_pool_fn = AvgPool2dSame if avg_stride == 1 and dilation > 1 else nn.AvgPool2d
+        pool = avg_pool_fn(2, avg_stride, ceil_mode=True, count_include_pad=False)
+    return nn.Sequential(*[
+        pool, ConvBnAct(in_chs, out_chs, 1, stride=1, norm_layer=norm_layer, act_layer=None)])
+
+
+class RegStage(nn.Module):
+    """Stage (sequence of blocks w/ the same output shape)."""
+
+    def __init__(self, in_chs, out_chs, stride, depth, block_fn, bottle_ratio, group_width, se_ratio):
+        super(RegStage, self).__init__()
+        block_kwargs = {}  # FIXME setup to pass various aa, norm, act layer common args
+        for i in range(depth):
+            block_stride = stride if i == 0 else 1
+            block_in_chs = in_chs if i == 0 else out_chs
+            if (block_in_chs != out_chs) or (block_stride != 1):
+                proj_block = downsample_conv(block_in_chs, out_chs, 1, stride)
+            else:
+                proj_block = None
+
+            name = "b{}".format(i + 1)
+            self.add_module(
+                name, block_fn(
+                    block_in_chs, out_chs, block_stride, bottle_ratio, group_width, se_ratio,
+                    downsample=proj_block, **block_kwargs)
+            )
+
+    def forward(self, x):
+        for block in self.children():
+            x = block(x)
+        return x
+
+
+class ClassifierHead(nn.Module):
+    """Head."""
+
+    def __init__(self, in_chs, num_classes, pool_type='avg', drop_rate=0.):
+        super(ClassifierHead, self).__init__()
+        self.drop_rate = drop_rate
+        self.global_pool = SelectAdaptivePool2d(pool_type=pool_type)
+        if num_classes > 0:
+            self.fc = nn.Linear(in_chs, num_classes, bias=True)
+        else:
+            self.fc = nn.Identity()
+
+    def forward(self, x):
+        x = self.global_pool(x).flatten(1)
+        if self.drop_rate:
+            x = F.dropout(x, p=float(self.drop_rate), training=self.training)
+        x = self.fc(x)
+        return x
+
+
+class RegNet(nn.Module):
+    """RegNet model.
+
+    Paper: https://arxiv.org/abs/2003.13678
+    Original Impl: https://github.com/facebookresearch/pycls/blob/master/pycls/models/regnet.py
+    """
+
+    def __init__(self, cfg, in_chans=3, num_classes=1000, global_pool='avg', drop_rate=0.,
+                 zero_init_last_bn=True):
+        super().__init__()
+        # TODO add drop block, drop path, anti-aliasing, custom bn/act args
+        self.num_classes = num_classes
+        self.drop_rate = drop_rate
+
+        # Construct the stem
+        stem_width = cfg['stem_width']
+        self.stem = ConvBnAct(in_chans, stem_width, 3, stride=2)
+        
+        # Construct the stages
+        block_fn = Bottleneck
+        prev_width = stem_width
+        stage_params = self._get_stage_params(cfg)
+        se_ratio = cfg['se_ratio']
+        for i, (d, w, s, br, gw) in enumerate(stage_params):
+            self.add_module(
+                "s{}".format(i + 1), RegStage(prev_width, w, s, d, block_fn, br, gw, se_ratio))
+            prev_width = w
+
+        # Construct the head
+        self.num_features = prev_width
+        self.head = ClassifierHead(
+            in_chs=prev_width, num_classes=num_classes, pool_type=global_pool, drop_rate=drop_rate)
+
+        for m in self.modules():
+            if isinstance(m, nn.Conv2d):
+                nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
+            elif isinstance(m, nn.BatchNorm2d):
+                nn.init.ones_(m.weight)
+                nn.init.zeros_(m.bias)
+            elif isinstance(m, nn.Linear):
+                nn.init.normal_(m.weight, mean=0.0, std=0.01)
+                nn.init.zeros_(m.bias)
+        if zero_init_last_bn:
+            for m in self.modules():
+                if hasattr(m, 'zero_init_last_bn'):
+                    m.zero_init_last_bn()
+
+    def _get_stage_params(self, cfg, stride=2):
+        # Generate RegNet ws per block
+        w_a, w_0, w_m, d = cfg['wa'], cfg['w0'], cfg['wm'], cfg['depth']
+        widths, num_stages, _, _ = generate_regnet(w_a, w_0, w_m, d)
+
+        # Convert to per stage format
+        stage_widths, stage_depths = np.unique(widths, return_counts=True)
+
+        # Use the same group width, bottleneck mult and stride for each stage
+        stage_groups = [cfg['group_w'] for _ in range(num_stages)]
+        stage_bottle_ratios = [cfg['bottle_ratio'] for _ in range(num_stages)]
+        stage_strides = [stride for _ in range(num_stages)]
+        # FIXME add dilation / output_stride support
+
+        # Adjust the compatibility of ws and gws
+        stage_widths, stage_groups = adjust_widths_groups_comp(stage_widths, stage_bottle_ratios, stage_groups)
+        stage_params = list(zip(stage_depths, stage_widths, stage_strides, stage_bottle_ratios, stage_groups))
+        return stage_params
+
+    def get_classifier(self):
+        return self.head.fc
+
+    def reset_classifier(self, num_classes, global_pool='avg'):
+        self.head = ClassifierHead(self.num_features, num_classes, pool_type=global_pool, drop_rate=self.drop_rate)
+
+    def forward_features(self, x):
+        for block in list(self.children())[:-1]:
+            x = block(x)
+        return x
+
+    def forward(self, x):
+        for block in self.children():
+            x = block(x)
+        return x
+
+
+def _regnet(variant, pretrained, **kwargs):
+    load_strict = True
+    model_class = RegNet
+    if kwargs.pop('features_only', False):
+        assert False, 'Not Implemented'  # TODO
+        load_strict = False
+        kwargs.pop('num_classes', 0)
+    model_cfg = model_cfgs[variant]
+    default_cfg = default_cfgs[variant]
+    model = model_class(model_cfg, **kwargs)
+    model.default_cfg = default_cfg
+    if pretrained:
+        load_pretrained(
+            model, default_cfg,
+            num_classes=kwargs.get('num_classes', 0), in_chans=kwargs.get('in_chans', 3), strict=load_strict)
+    return model
+
+
+@register_model
+def regnetx_002(pretrained=False, **kwargs):
+    """RegNetX-200MF"""
+    return _regnet('x_002', pretrained, **kwargs)
+
+
+@register_model
+def regnetx_004(pretrained=False, **kwargs):
+    """RegNetX-400MF"""
+    return _regnet('x_004', pretrained, **kwargs)
+
+
+@register_model
+def regnetx_006(pretrained=False, **kwargs):
+    """RegNetX-600MF"""
+    return _regnet('x_006', pretrained, **kwargs)
+
+
+@register_model
+def regnetx_008(pretrained=False, **kwargs):
+    """RegNetX-800MF"""
+    return _regnet('x_008', pretrained, **kwargs)
+
+
+@register_model
+def regnetx_016(pretrained=False, **kwargs):
+    """RegNetX-1.6GF"""
+    return _regnet('x_016', pretrained, **kwargs)
+
+
+@register_model
+def regnetx_032(pretrained=False, **kwargs):
+    """RegNetX-3.2GF"""
+    return _regnet('x_032', pretrained, **kwargs)
+
+
+@register_model
+def regnetx_040(pretrained=False, **kwargs):
+    """RegNetX-4.0GF"""
+    return _regnet('x_040', pretrained, **kwargs)
+
+
+@register_model
+def regnetx_064(pretrained=False, **kwargs):
+    """RegNetX-6.4GF"""
+    return _regnet('x_064', pretrained, **kwargs)
+
+
+@register_model
+def regnetx_080(pretrained=False, **kwargs):
+    """RegNetX-8.0GF"""
+    return _regnet('x_080', pretrained, **kwargs)
+
+
+@register_model
+def regnetx_120(pretrained=False, **kwargs):
+    """RegNetX-12GF"""
+    return _regnet('x_120', pretrained, **kwargs)
+
+
+@register_model
+def regnetx_160(pretrained=False, **kwargs):
+    """RegNetX-16GF"""
+    return _regnet('x_160', pretrained, **kwargs)
+
+
+@register_model
+def regnetx_320(pretrained=False, **kwargs):
+    """RegNetX-32GF"""
+    return _regnet('x_320', pretrained, **kwargs)
+
+
+@register_model
+def regnety_002(pretrained=False, **kwargs):
+    """RegNetY-200MF"""
+    return _regnet('y_002', pretrained, **kwargs)
+
+
+@register_model
+def regnety_004(pretrained=False, **kwargs):
+    """RegNetY-400MF"""
+    return _regnet('y_004', pretrained, **kwargs)
+
+
+@register_model
+def regnety_006(pretrained=False, **kwargs):
+    """RegNetY-600MF"""
+    return _regnet('y_006', pretrained, **kwargs)
+
+
+@register_model
+def regnety_008(pretrained=False, **kwargs):
+    """RegNetY-800MF"""
+    return _regnet('y_008', pretrained, **kwargs)
+
+
+@register_model
+def regnety_016(pretrained=False, **kwargs):
+    """RegNetY-1.6GF"""
+    return _regnet('y_016', pretrained, **kwargs)
+
+
+@register_model
+def regnety_032(pretrained=False, **kwargs):
+    """RegNetY-3.2GF"""
+    return _regnet('y_032', pretrained, **kwargs)
+
+
+@register_model
+def regnety_040(pretrained=False, **kwargs):
+    """RegNetY-4.0GF"""
+    return _regnet('y_040', pretrained, **kwargs)
+
+
+@register_model
+def regnety_064(pretrained=False, **kwargs):
+    """RegNetY-6.4GF"""
+    return _regnet('y_064', pretrained, **kwargs)
+
+
+@register_model
+def regnety_080(pretrained=False, **kwargs):
+    """RegNetY-8.0GF"""
+    return _regnet('y_080', pretrained, **kwargs)
+
+
+@register_model
+def regnety_120(pretrained=False, **kwargs):
+    """RegNetY-12GF"""
+    return _regnet('y_120', pretrained, **kwargs)
+
+
+@register_model
+def regnety_160(pretrained=False, **kwargs):
+    """RegNetY-16GF"""
+    return _regnet('y_160', pretrained, **kwargs)
+
+
+@register_model
+def regnety_320(pretrained=False, **kwargs):
+    """RegNetY-32GF"""
+    return _regnet('y_320', pretrained, **kwargs)