Merge branch 'master' into norm_norm_norm

.github/workflows/tests.yml

@@ -48,4 +48,5 @@ jobs:
       env:
         LD_PRELOAD: /usr/lib/x86_64-linux-gnu/libtcmalloc.so.4
       run: |
+        export PYTHONDONTWRITEBYTECODE=1
         pytest -vv --forked --durations=0 ./tests

timm/data/distributed_sampler.py

@@ -103,15 +103,16 @@ class RepeatAugSampler(Sampler):
         g = torch.Generator()
         g.manual_seed(self.epoch)
         if self.shuffle:
-            indices = torch.randperm(len(self.dataset), generator=g).tolist()
+            indices = torch.randperm(len(self.dataset), generator=g)
         else:
-            indices = list(range(len(self.dataset)))
+            indices = torch.arange(start=0, end=len(self.dataset))
 
         # produce repeats e.g. [0, 0, 0, 1, 1, 1, 2, 2, 2....]
-        indices = [x for x in indices for _ in range(self.num_repeats)]
+        indices = torch.repeat_interleave(indices, repeats=self.num_repeats, dim=0)
         # add extra samples to make it evenly divisible
         padding_size = self.total_size - len(indices)
-        indices += indices[:padding_size]
+        if padding_size > 0:
+            indices = torch.cat([indices, indices[:padding_size]], dim=0)
         assert len(indices) == self.total_size
 
         # subsample per rank

timm/models/efficientnet.py

@@ -23,6 +23,10 @@ An implementation of EfficienNet that covers variety of related models with effi
 * Single-Path NAS Pixel1
   - Single-Path NAS: Designing Hardware-Efficient ConvNets - https://arxiv.org/abs/1904.02877
 
+* TinyNet
+    - Model Rubik's Cube: Twisting Resolution, Depth and Width for TinyNets - https://arxiv.org/abs/2010.14819
+    - Definitions & weights borrowed from https://github.com/huawei-noah/CV-Backbones/tree/master/tinynet_pytorch
+
 * And likely more...
 
 The majority of the above models (EfficientNet*, MixNet, MnasNet) and original weights were made available
@@ -427,11 +431,27 @@ default_cfgs = {
         url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-weights/tf_mixnet_m-0f4d8805.pth'),
     'tf_mixnet_l': _cfg(
         url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-weights/tf_mixnet_l-6c92e0c8.pth'),
+
+    "tinynet_a": _cfg(
+        input_size=(3, 192, 192), pool_size=(6, 6),  # int(224 * 0.86)
+        url='https://github.com/huawei-noah/CV-Backbones/releases/download/v1.2.0/tinynet_a.pth'),
+    "tinynet_b": _cfg(
+        input_size=(3, 188, 188), pool_size=(6, 6),  # int(224 * 0.84)
+        url='https://github.com/huawei-noah/CV-Backbones/releases/download/v1.2.0/tinynet_b.pth'),
+    "tinynet_c": _cfg(
+        input_size=(3, 184, 184), pool_size=(6, 6),  # int(224 * 0.825)
+        url='https://github.com/huawei-noah/CV-Backbones/releases/download/v1.2.0/tinynet_c.pth'),
+    "tinynet_d": _cfg(
+        input_size=(3, 152, 152), pool_size=(5, 5),  # int(224 * 0.68)
+        url='https://github.com/huawei-noah/CV-Backbones/releases/download/v1.2.0/tinynet_d.pth'),
+    "tinynet_e": _cfg(
+        input_size=(3, 106, 106), pool_size=(4, 4),  # int(224 * 0.475)
+        url='https://github.com/huawei-noah/CV-Backbones/releases/download/v1.2.0/tinynet_e.pth'),
 }
 
 
 class EfficientNet(nn.Module):
-    """ (Generic) EfficientNet
+    """ EfficientNet
 
     A flexible and performant PyTorch implementation of efficient network architectures, including:
       * EfficientNet-V2 Small, Medium, Large, XL & B0-B3
@@ -443,7 +463,7 @@ class EfficientNet(nn.Module):
       * MobileNet-V2
       * FBNet C
       * Single-Path NAS Pixel1
-
+      * TinyNet
     """
 
     def __init__(self, block_args, num_classes=1000, num_features=1280, in_chans=3, stem_size=32, fix_stem=False,
@@ -1160,6 +1180,31 @@ def _gen_mixnet_m(variant, channel_multiplier=1.0, depth_multiplier=1.0, pretrai
     return model
 
 
+def _gen_tinynet(
+    variant, model_width=1.0, depth_multiplier=1.0, pretrained=False, **kwargs
+):
+    """Creates a TinyNet model.
+    """
+    arch_def = [
+        ['ds_r1_k3_s1_e1_c16_se0.25'], ['ir_r2_k3_s2_e6_c24_se0.25'],
+        ['ir_r2_k5_s2_e6_c40_se0.25'], ['ir_r3_k3_s2_e6_c80_se0.25'],
+        ['ir_r3_k5_s1_e6_c112_se0.25'], ['ir_r4_k5_s2_e6_c192_se0.25'],
+        ['ir_r1_k3_s1_e6_c320_se0.25'],
+    ]
+    model_kwargs = dict(
+        block_args=decode_arch_def(arch_def, depth_multiplier, depth_trunc='round'),
+        num_features=max(1280, round_channels(1280, model_width, 8, None)),
+        stem_size=32,
+        fix_stem=True,
+        round_chs_fn=partial(round_channels, multiplier=model_width),
+        act_layer=resolve_act_layer(kwargs, 'swish'),
+        norm_layer=kwargs.pop('norm_layer', None) or partial(nn.BatchNorm2d, **resolve_bn_args(kwargs)),
+        **kwargs,
+    )
+    model = _create_effnet(variant, pretrained, **model_kwargs)
+    return model
+
+
 @register_model
 def mnasnet_050(pretrained=False, **kwargs):
     """ MNASNet B1, depth multiplier of 0.5. """
@@ -2298,3 +2343,33 @@ def tf_mixnet_l(pretrained=False, **kwargs):
     model = _gen_mixnet_m(
         'tf_mixnet_l', channel_multiplier=1.3, pretrained=pretrained, **kwargs)
     return model
+
+
+@register_model
+def tinynet_a(pretrained=False, **kwargs):
+    model = _gen_tinynet('tinynet_a', 1.0, 1.2, pretrained=pretrained, **kwargs)
+    return model
+
+
+@register_model
+def tinynet_b(pretrained=False, **kwargs):
+    model = _gen_tinynet('tinynet_b', 0.75, 1.1, pretrained=pretrained, **kwargs)
+    return model
+
+
+@register_model
+def tinynet_c(pretrained=False, **kwargs):
+    model = _gen_tinynet('tinynet_c', 0.54, 0.85, pretrained=pretrained, **kwargs)
+    return model
+
+
+@register_model
+def tinynet_d(pretrained=False, **kwargs):
+    model = _gen_tinynet('tinynet_d', 0.54, 0.695, pretrained=pretrained, **kwargs)
+    return model
+
+
+@register_model
+def tinynet_e(pretrained=False, **kwargs):
+    model = _gen_tinynet('tinynet_e', 0.51, 0.6, pretrained=pretrained, **kwargs)
+    return model

timm/models/visformer.py

@@ -167,14 +167,14 @@ class Visformer(nn.Module):
             self.patch_embed1 = PatchEmbed(
                 img_size=img_size, patch_size=patch_size, in_chans=in_chans,
                 embed_dim=embed_dim, norm_layer=embed_norm, flatten=False)
-            img_size = [x // 16 for x in img_size]
+            img_size = [x // patch_size for x in img_size]
         else:
             if self.init_channels is None:
                 self.stem = None
                 self.patch_embed1 = PatchEmbed(
                     img_size=img_size, patch_size=patch_size // 2, in_chans=in_chans,
                     embed_dim=embed_dim // 2, norm_layer=embed_norm, flatten=False)
-                img_size = [x // 8 for x in img_size]
+                img_size = [x // (patch_size // 2) for x in img_size]
             else:
                 self.stem = nn.Sequential(
                     nn.Conv2d(in_chans, self.init_channels, 7, stride=2, padding=3, bias=False),
@@ -185,7 +185,7 @@ class Visformer(nn.Module):
                 self.patch_embed1 = PatchEmbed(
                     img_size=img_size, patch_size=patch_size // 4, in_chans=self.init_channels,
                     embed_dim=embed_dim // 2, norm_layer=embed_norm, flatten=False)
-                img_size = [x // 4 for x in img_size]
+                img_size = [x // (patch_size // 4) for x in img_size]
 
         if self.pos_embed:
             if self.vit_stem:
@@ -207,7 +207,7 @@ class Visformer(nn.Module):
             self.patch_embed2 = PatchEmbed(
                 img_size=img_size, patch_size=patch_size // 8, in_chans=embed_dim // 2,
                 embed_dim=embed_dim, norm_layer=embed_norm, flatten=False)
-            img_size = [x // 2 for x in img_size]
+            img_size = [x // (patch_size // 8) for x in img_size]
             if self.pos_embed:
                 self.pos_embed2 = nn.Parameter(torch.zeros(1, embed_dim, *img_size))
         self.stage2 = nn.ModuleList([
@@ -224,7 +224,7 @@ class Visformer(nn.Module):
             self.patch_embed3 = PatchEmbed(
                 img_size=img_size, patch_size=patch_size // 8, in_chans=embed_dim,
                 embed_dim=embed_dim * 2, norm_layer=embed_norm, flatten=False)
-            img_size = [x // 2 for x in img_size]
+            img_size = [x // (patch_size // 8) for x in img_size]
             if self.pos_embed:
                 self.pos_embed3 = nn.Parameter(torch.zeros(1, embed_dim*2, *img_size))
         self.stage3 = nn.ModuleList([