Merge pull request huggingface#413 from rwightman/eca-weights

Add new weights for ecaresnet26t/50t/269d models.

README.md

@@ -2,6 +2,14 @@
 
 ## What's New
 
+### Feb 8, 2021
+* Add several ResNet weights with ECA attention. 26t & 50t trained @ 256, test @ 320. 269d train @ 256, fine-tune @320, test @ 352.
+  * `ecaresnet26t` - 79.88 top-1 @ 320x320, 79.08 @ 256x256
+  * `ecaresnet50t` - 82.35 top-1 @ 320x320, 81.52 @ 256x256
+  * `ecaresnet269d` - 84.93 top-1 @ 352x352, 84.87 @ 320x320
+* Remove separate tiered (`t`) vs tiered_narrow (`tn`) ResNet model defs, all `tn` changed to `t` and `t` models removed (`seresnext26t_32x4d` only model w/ weights that was removed).
+* Support model default_cfgs with separate train vs test resolution `test_input_size`
+
 ### Jan 30, 2021
 * Add initial "Normalization Free" NF-RegNet-B* and NF-ResNet model definitions based on [paper](https://arxiv.org/abs/2101.08692)
 

sotabench.py

@@ -122,8 +122,6 @@ def _entry(model_name, paper_model_name, paper_arxiv_id, batch_size=BATCH_SIZE,
            model_desc='Block cfg of SE-ResNeXt-34 w/ Bottleneck, deep stem, and avg-pool in downsample layers.'),
     _entry('seresnext26t_32x4d', 'SE-ResNeXt-26-T 32x4d', '1812.01187',
            model_desc='Block cfg of SE-ResNeXt-34 w/ Bottleneck, deep tiered stem, and avg-pool in downsample layers.'),
-    _entry('seresnext26tn_32x4d', 'SE-ResNeXt-26-TN 32x4d', '1812.01187',
-           model_desc='Block cfg of SE-ResNeXt-34 w/ Bottleneck, deep tiered narrow stem, and avg-pool in downsample layers.'),
     _entry('seresnext50_32x4d', 'SE-ResNeXt-50 32x4d', '1709.01507'),
 
     _entry('skresnet18', 'SK-ResNet-18', '1903.06586'),

timm/data/config.py

@@ -5,7 +5,7 @@
 _logger = logging.getLogger(__name__)
 
 
-def resolve_data_config(args, default_cfg={}, model=None, verbose=True):
+def resolve_data_config(args, default_cfg={}, model=None, use_test_size=False, verbose=True):
     new_config = {}
     default_cfg = default_cfg
     if not default_cfg and model is not None and hasattr(model, 'default_cfg'):
@@ -25,8 +25,11 @@ def resolve_data_config(args, default_cfg={}, model=None, verbose=True):
     elif 'img_size' in args and args['img_size'] is not None:
         assert isinstance(args['img_size'], int)
         input_size = (in_chans, args['img_size'], args['img_size'])
-    elif 'input_size' in default_cfg:
-        input_size = default_cfg['input_size']
+    else:
+        if use_test_size and 'test_input_size' in default_cfg:
+            input_size = default_cfg['test_input_size']
+        elif 'input_size' in default_cfg:
+            input_size = default_cfg['input_size']
     new_config['input_size'] = input_size
 
     # resolve interpolation method

timm/models/efficientnet_blocks.py

@@ -350,7 +350,8 @@ def __init__(self, in_chs, out_chs, exp_kernel_size=3, exp_ratio=1.0, fake_in_ch
         self.drop_path_rate = drop_path_rate
 
         # Expansion convolution
-        self.conv_exp = create_conv2d(in_chs, mid_chs, exp_kernel_size, padding=pad_type)
+        self.conv_exp = create_conv2d(
+            in_chs, mid_chs, exp_kernel_size, stride=stride, dilation=dilation, padding=pad_type)
         self.bn1 = norm_layer(mid_chs, **norm_kwargs)
         self.act1 = act_layer(inplace=True)
 
@@ -362,8 +363,7 @@ def __init__(self, in_chs, out_chs, exp_kernel_size=3, exp_ratio=1.0, fake_in_ch
             self.se = None
 
         # Point-wise linear projection
-        self.conv_pwl = create_conv2d(
-            mid_chs, out_chs, pw_kernel_size, stride=stride, dilation=dilation, padding=pad_type)
+        self.conv_pwl = create_conv2d(mid_chs, out_chs, pw_kernel_size, padding=pad_type)
         self.bn2 = norm_layer(out_chs, **norm_kwargs)
 
     def feature_info(self, location):

timm/models/layers/test_time_pool.py

@@ -36,14 +36,17 @@ def forward(self, x):
         return x.view(x.size(0), -1)
 
 
-def apply_test_time_pool(model, config):
+def apply_test_time_pool(model, config, use_test_size=True):
     test_time_pool = False
     if not hasattr(model, 'default_cfg') or not model.default_cfg:
         return model, False
-    if (config['input_size'][-1] > model.default_cfg['input_size'][-1] and
-            config['input_size'][-2] > model.default_cfg['input_size'][-2]):
+    if use_test_size and 'test_input_size' in model.default_cfg:
+        df_input_size = model.default_cfg['test_input_size']
+    else:
+        df_input_size = model.default_cfg['input_size']
+    if config['input_size'][-1] > df_input_size[-1] and config['input_size'][-2] > df_input_size[-2]:
         _logger.info('Target input size %s > pretrained default %s, using test time pooling' %
-                     (str(config['input_size'][-2:]), str(model.default_cfg['input_size'][-2:])))
+                     (str(config['input_size'][-2:]), str(df_input_size[-2:])))
         model = TestTimePoolHead(model, original_pool=model.default_cfg['pool_size'])
         test_time_pool = True
     return model, test_time_pool