Add efficientnetv2_rw_t defs w/ weights, and gc variant, as well as gcresnet26ts for experiments. Version 0.4.13

README.md

@@ -23,11 +23,12 @@ I'm fortunate to be able to dedicate significant time and money of my own suppor
 
 ## What's New
 
-### July 8, 2021
+### July 5-9, 2021
+* Add `efficientnetv2_rw_t` weights, a custom 'tiny' 13.6M param variant that is a bit better than (non NoisyStudent) B3 models. Both faster and better accuracy (at same or lower res)
+  * top-1 82.34 @ 288x288 and 82.54 @ 320x320
 * Add [SAM pretrained](https://arxiv.org/abs/2106.01548) in1k weight for ViT B/16 (`vit_base_patch16_sam_224`) and B/32 (`vit_base_patch32_sam_224`)  models.
-
-### July 5, 2021
 * Add 'Aggregating Nested Transformer' (NesT) w/ weights converted from official [Flax impl](https://github.com/google-research/nested-transformer). Contributed by [Alexander Soare](https://github.com/alexander-soare).
+  * `jx_nest_base` - 83.534, `jx_nest_small` - 83.120, `jx_nest_tiny` - 81.426
 
 ### June 23, 2021
 * Reproduce gMLP model training, `gmlp_s16_224` trained to 79.6 top-1, matching [paper](https://arxiv.org/abs/2105.08050). Hparams for this and other recent MLP training [here](https://gist.github.com/rwightman/d6c264a9001f9167e06c209f630b2cc6)

timm/models/byobnet.py

@@ -101,6 +101,8 @@ default_cfgs = {
 
     'gcresnext26ts': _cfg(
         first_conv='stem.conv1.conv', input_size=(3, 256, 256), pool_size=(8, 8), interpolation='bicubic'),
+    'gcresnet26ts': _cfg(
+        first_conv='stem.conv1.conv', input_size=(3, 256, 256), pool_size=(8, 8), interpolation='bicubic'),
     'bat_resnext26ts': _cfg(
         first_conv='stem.conv1.conv', input_size=(3, 256, 256), pool_size=(8, 8), interpolation='bicubic',
         min_input_size=(3, 256, 256)),
@@ -329,6 +331,21 @@ model_cfgs = dict(
         attn_layer='gc',
     ),
 
+    gcresnet26ts=ByoModelCfg(
+        blocks=(
+            ByoBlockCfg(type='bottle', d=2, c=256, s=1, gs=0, br=0.25),
+            ByoBlockCfg(type='bottle', d=3, c=512, s=2, gs=0, br=0.25),
+            ByoBlockCfg(type='bottle', d=3, c=1536, s=2, gs=0, br=0.25),
+            ByoBlockCfg(type='bottle', d=2, c=1536, s=2, gs=0, br=0.25),
+        ),
+        stem_chs=64,
+        stem_type='tiered',
+        stem_pool='',
+        num_features=1280,
+        act_layer='silu',
+        attn_layer='gc',
+    ),
+
     bat_resnext26ts=ByoModelCfg(
         blocks=(
             ByoBlockCfg(type='bottle', d=2, c=256, s=1, gs=32, br=0.25),
@@ -470,6 +487,13 @@ def gcresnext26ts(pretrained=False, **kwargs):
     return _create_byobnet('gcresnext26ts', pretrained=pretrained, **kwargs)
 
 
+@register_model
+def gcresnet26ts(pretrained=False, **kwargs):
+    """
+    """
+    return _create_byobnet('gcresnet26ts', pretrained=pretrained, **kwargs)
+
+
 @register_model
 def bat_resnext26ts(pretrained=False, **kwargs):
     """

timm/models/efficientnet.py

@@ -165,6 +165,12 @@ default_cfgs = {
         url='https://imvl-automl-sh.oss-cn-shanghai.aliyuncs.com/darts/hyperml/hyperml/job_45403/outputs/effnetb3_pruned_5abcc29f.pth',
         input_size=(3, 300, 300), pool_size=(10, 10), crop_pct=0.904, mean=IMAGENET_INCEPTION_MEAN, std=IMAGENET_INCEPTION_STD),
 
+    'efficientnetv2_rw_t': _cfg(
+        url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-weights/efficientnetv2_t_agc-3620981a.pth',
+        input_size=(3, 224, 224), test_input_size=(3, 288, 288), pool_size=(7, 7), crop_pct=1.0),
+    'gc_efficientnetv2_rw_t': _cfg(
+        url='',
+        input_size=(3, 224, 224), test_input_size=(3, 288, 288), pool_size=(7, 7), crop_pct=1.0),
     'efficientnetv2_rw_s': _cfg(
         url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-weights/efficientnet_v2s_ra2_288-a6477665.pth',
         input_size=(3, 288, 288), test_input_size=(3, 384, 384), pool_size=(9, 9), crop_pct=1.0),
@@ -1487,9 +1493,26 @@ def efficientnet_b3_pruned(pretrained=False, **kwargs):
     return model
 
 
+@register_model
+def efficientnetv2_rw_t(pretrained=False, **kwargs):
+    """ EfficientNet-V2 Tiny (Custom variant, tiny not in paper). """
+    model = _gen_efficientnetv2_s(
+        'efficientnetv2_rw_t', channel_multiplier=0.8, depth_multiplier=0.9, rw=False, pretrained=pretrained, **kwargs)
+    return model
+
+
+@register_model
+def gc_efficientnetv2_rw_t(pretrained=False, **kwargs):
+    """ EfficientNet-V2 Tiny w/ Global Context Attn (Custom variant, tiny not in paper). """
+    model = _gen_efficientnetv2_s(
+        'gc_efficientnetv2_t', channel_multiplier=0.8, depth_multiplier=0.9,
+        rw=False, se_layer='gc', pretrained=pretrained, **kwargs)
+    return model
+
+
 @register_model
 def efficientnetv2_rw_s(pretrained=False, **kwargs):
-    """ EfficientNet-V2 Small RW variant.
+    """ EfficientNet-V2 Small (RW variant).
     NOTE: This is my initial (pre official code release) w/ some differences.
     See efficientnetv2_s and tf_efficientnetv2_s for versions that match the official w/ PyTorch vs TF padding
     """
@@ -1499,7 +1522,7 @@ def efficientnetv2_rw_s(pretrained=False, **kwargs):
 
 @register_model
 def efficientnetv2_rw_m(pretrained=False, **kwargs):
-    """ EfficientNet-V2 Medium RW variant.
+    """ EfficientNet-V2 Medium (RW variant).
     """
     model = _gen_efficientnetv2_s(
         'efficientnetv2_rw_m', channel_multiplier=1.2, depth_multiplier=(1.2,) * 4 + (1.6,) * 2, rw=True,

timm/version.py

@@ -1 +1 @@
-__version__ = '0.4.12'
+__version__ = '0.4.13'