Merge pull request #2202 from huggingface/mnv4_first_weights

First set of MobileNetV4 weights trained in timm
huggingface · Jun 12, 2024 · 5aa49d5 · 5aa49d5
2 parents 2673693 + c63da14
commit 5aa49d5
Showing 1 changed file with 24 additions and 24 deletions.
diff --git a/timm/models/mobilenetv3.py b/timm/models/mobilenetv3.py
@@ -996,28 +996,28 @@ def _cfg(url: str = '', **kwargs):
     ),
     "lcnet_150.untrained": _cfg(),
 
-    'mobilenetv4_conv_small': _cfg(
-        # hf_hub_id='timm/',
-        interpolation='bicubic'),
-    'mobilenetv4_conv_medium.r224': _cfg(
-        # hf_hub_id='timm/',
-        crop_pct=0.95, interpolation='bicubic'),
-    'mobilenetv4_conv_medium.r256': _cfg(
-        # hf_hub_id='timm/',
-        input_size=(3, 256, 256), pool_size=(8, 8), crop_pct=0.95, interpolation='bicubic'),
-    'mobilenetv4_conv_large.r256': _cfg(
+    'mobilenetv4_conv_small.e1200_r224_in1k': _cfg(
+        hf_hub_id='timm/',
+        test_input_size=(3, 256, 256), test_crop_pct=0.95, interpolation='bicubic'),
+    'mobilenetv4_conv_medium.e500_r224_in1k': _cfg(
+        hf_hub_id='timm/',
+        crop_pct=0.95, test_input_size=(3, 256, 256), test_crop_pct=1.0, interpolation='bicubic'),
+    'mobilenetv4_conv_medium.e500_r256_in1k': _cfg(
+        hf_hub_id='timm/',
+        input_size=(3, 256, 256), pool_size=(8, 8),
+        crop_pct=0.95, test_input_size=(3, 320, 320), test_crop_pct=1.0, interpolation='bicubic'),
+    'mobilenetv4_conv_large.e500_r256_in1k': _cfg(
         # hf_hub_id='timm/',
-        input_size=(3, 256, 256), pool_size=(8, 8), crop_pct=0.95, interpolation='bicubic'),
+        file='mobilenetv4_conv_large-r256.pth',
+        input_size=(3, 256, 256), pool_size=(8, 8),
+        crop_pct=0.95, test_input_size=(3, 320, 320), test_crop_pct=1.0, interpolation='bicubic'),
     'mobilenetv4_conv_large.r384': _cfg(
         # hf_hub_id='timm/',
         input_size=(3, 384, 384), pool_size=(12, 12), crop_pct=0.95, interpolation='bicubic'),
 
-    'mobilenetv4_hybrid_small': _cfg(
-        # hf_hub_id='timm/',
-        interpolation='bicubic'),
-    'mobilenetv4_hybrid_medium.r224': _cfg(
-        # hf_hub_id='timm/',
-        crop_pct=0.95, interpolation='bicubic'),
+    'mobilenetv4_hybrid_medium.e500_r224_in1k': _cfg(
+        hf_hub_id='timm/',
+        crop_pct=0.95, test_input_size=(3, 256, 256), test_crop_pct=1.0, interpolation='bicubic'),
     'mobilenetv4_hybrid_medium.r256': _cfg(
         # hf_hub_id='timm/',
         input_size=(3, 256, 256), pool_size=(8, 8), crop_pct=0.95, interpolation='bicubic'),
@@ -1029,16 +1029,16 @@ def _cfg(url: str = '', **kwargs):
         input_size=(3, 384, 384), pool_size=(12, 12), crop_pct=0.95, interpolation='bicubic'),
 
     # experimental
-    'mobilenetv4_conv_aa_medium.r256': _cfg(
-        # hf_hub_id='timm/',
-        input_size=(3, 256, 256), pool_size=(8, 8), crop_pct=0.95, interpolation='bicubic'),
-    'mobilenetv4_conv_blur_medium.r256': _cfg(
+    'mobilenetv4_conv_aa_medium.untrained': _cfg(
         # hf_hub_id='timm/',
         input_size=(3, 256, 256), pool_size=(8, 8), crop_pct=0.95, interpolation='bicubic'),
-    'mobilenetv4_hybrid_medium_075': _cfg(
+    'mobilenetv4_conv_blur_medium.e500_r224_in1k': _cfg(
+        hf_hub_id='timm/',
+        crop_pct=0.95, test_input_size=(3, 256, 256), test_crop_pct=1.0, interpolation='bicubic'),
+    'mobilenetv4_hybrid_medium_075.untrained': _cfg(
         # hf_hub_id='timm/',
         crop_pct=0.95, interpolation='bicubic'),
-    'mobilenetv4_hybrid_large_075.r256': _cfg(
+    'mobilenetv4_hybrid_large_075.untrained': _cfg(
         # hf_hub_id='timm/',
         input_size=(3, 256, 256), pool_size=(8, 8), crop_pct=0.95, interpolation='bicubic'),
 })
@@ -1256,7 +1256,7 @@ def mobilenetv4_hybrid_medium_075(pretrained: bool = False, **kwargs) -> MobileN
 @register_model
 def mobilenetv4_hybrid_large_075(pretrained: bool = False, **kwargs) -> MobileNetV3:
     """ MobileNet V4 Hybrid"""
-    model = _gen_mobilenet_v4('mobilenetv4_hybrid_large', 0.75, pretrained=pretrained, **kwargs)
+    model = _gen_mobilenet_v4('mobilenetv4_hybrid_large_075', 0.75, pretrained=pretrained, **kwargs)
     return model