Merge pull request #2195 from huggingface/refactor_pre_logits

Fix consistency, testing for forward_head w/ pre_logits, reset_classifier, models with pre_logits size != unpooled feature size

Author: rwightman

tests/test_models.py

@@ -169,6 +169,18 @@ def test_model_backward(model_name, batch_size):
     assert not torch.isnan(outputs).any(), 'Output included NaNs'
 
 
+# models with extra conv/linear layers after pooling
+EARLY_POOL_MODELS = (
+    timm.models.EfficientVit,
+    timm.models.EfficientVitLarge,
+    timm.models.HighPerfGpuNet,
+    timm.models.GhostNet,
+    timm.models.MetaNeXt, # InceptionNeXt
+    timm.models.MobileNetV3,
+    timm.models.RepGhostNet,
+    timm.models.VGG,
+)
+
 @pytest.mark.cfg
 @pytest.mark.timeout(timeout300)
 @pytest.mark.parametrize('model_name', list_models(
@@ -179,6 +191,9 @@ def test_model_default_cfgs(model_name, batch_size):
     model = create_model(model_name, pretrained=False)
     model.eval()
     model.to(torch_device)
+    assert getattr(model, 'num_classes') >= 0
+    assert getattr(model, 'num_features') > 0
+    assert getattr(model, 'head_hidden_size') > 0
     state_dict = model.state_dict()
     cfg = model.default_cfg
 
@@ -195,37 +210,37 @@ def test_model_default_cfgs(model_name, batch_size):
         input_size = tuple([min(x, MAX_FWD_OUT_SIZE) for x in input_size])
         input_tensor = torch.randn((batch_size, *input_size), device=torch_device)
 
-        # test forward_features (always unpooled)
+        # test forward_features (always unpooled) & forward_head w/ pre_logits
         outputs = model.forward_features(input_tensor)
-        assert outputs.shape[spatial_axis[0]] == pool_size[0], 'unpooled feature shape != config'
-        assert outputs.shape[spatial_axis[1]] == pool_size[1], 'unpooled feature shape != config'
-        if not isinstance(model, (timm.models.MobileNetV3, timm.models.GhostNet, timm.models.RepGhostNet, timm.models.VGG)):
-            assert outputs.shape[feat_axis] == model.num_features
+        outputs_pre = model.forward_head(outputs, pre_logits=True)
+        assert outputs.shape[spatial_axis[0]] == pool_size[0], f'unpooled feature shape {outputs.shape} != config'
+        assert outputs.shape[spatial_axis[1]] == pool_size[1], f'unpooled feature shape {outputs.shape} != config'
+        assert outputs.shape[feat_axis] == model.num_features, f'unpooled feature dim {outputs.shape[feat_axis]} != model.num_features {model.num_features}'
+        assert outputs_pre.shape[1] == model.head_hidden_size, f'pre_logits feature dim {outputs_pre.shape[1]} != model.head_hidden_size {model.head_hidden_size}'
 
         # test forward after deleting the classifier, output should be poooled, size(-1) == model.num_features
         model.reset_classifier(0)
         model.to(torch_device)
         outputs = model.forward(input_tensor)
         assert len(outputs.shape) == 2
-        assert outputs.shape[1] == model.num_features
+        assert outputs.shape[1] == model.head_hidden_size, f'feature dim w/ removed classifier {outputs.shape[1]} != model.head_hidden_size {model.head_hidden_size}'
+        assert outputs.shape == outputs_pre.shape, f'output shape of pre_logits {outputs_pre.shape} does not match reset_head(0) {outputs.shape}'
 
-        # test model forward without pooling and classifier
-        model.reset_classifier(0, '')  # reset classifier and set global pooling to pass-through
-        model.to(torch_device)
-        outputs = model.forward(input_tensor)
-        assert len(outputs.shape) == 4
-        if not isinstance(model, (timm.models.MobileNetV3, timm.models.GhostNet, timm.models.RepGhostNet, timm.models.VGG)):
-            # mobilenetv3/ghostnet/repghostnet/vgg forward_features vs removed pooling differ due to location or lack of GAP
+        # test model forward after removing pooling and classifier
+        if not isinstance(model, EARLY_POOL_MODELS):
+            model.reset_classifier(0, '')  # reset classifier and disable global pooling
+            model.to(torch_device)
+            outputs = model.forward(input_tensor)
+            assert len(outputs.shape) == 4
             assert outputs.shape[spatial_axis[0]] == pool_size[0] and outputs.shape[spatial_axis[1]] == pool_size[1]
 
-        if 'pruned' not in model_name:  # FIXME better pruned model handling
-            # test classifier + global pool deletion via __init__
+        # test classifier + global pool deletion via __init__
+        if 'pruned' not in model_name and not isinstance(model, EARLY_POOL_MODELS):
             model = create_model(model_name, pretrained=False, num_classes=0, global_pool='').eval()
             model.to(torch_device)
             outputs = model.forward(input_tensor)
             assert len(outputs.shape) == 4
-            if not isinstance(model, (timm.models.MobileNetV3, timm.models.GhostNet, timm.models.RepGhostNet, timm.models.VGG)):
-                assert outputs.shape[spatial_axis[0]] == pool_size[0] and outputs.shape[spatial_axis[1]] == pool_size[1]
+            assert outputs.shape[spatial_axis[0]] == pool_size[0] and outputs.shape[spatial_axis[1]] == pool_size[1]
 
     # check classifier name matches default_cfg
     if cfg.get('num_classes', None):
@@ -253,6 +268,9 @@ def test_model_default_cfgs_non_std(model_name, batch_size):
     model = create_model(model_name, pretrained=False)
     model.eval()
     model.to(torch_device)
+    assert getattr(model, 'num_classes') >= 0
+    assert getattr(model, 'num_features') > 0
+    assert getattr(model, 'head_hidden_size') > 0
     state_dict = model.state_dict()
     cfg = model.default_cfg
 
@@ -264,13 +282,15 @@ def test_model_default_cfgs_non_std(model_name, batch_size):
     feat_dim = getattr(model, 'feature_dim', None)
 
     outputs = model.forward_features(input_tensor)
+    outputs_pre = model.forward_head(outputs, pre_logits=True)
     if isinstance(outputs, (tuple, list)):
         # cannot currently verify multi-tensor output.
         pass
     else:
         if feat_dim is None:
             feat_dim = -1 if outputs.ndim == 3 else 1
         assert outputs.shape[feat_dim] == model.num_features
+        assert outputs_pre.shape[1] == model.head_hidden_size
 
     # test forward after deleting the classifier, output should be poooled, size(-1) == model.num_features
     model.reset_classifier(0)
@@ -280,7 +300,8 @@ def test_model_default_cfgs_non_std(model_name, batch_size):
         outputs = outputs[0]
     if feat_dim is None:
         feat_dim = -1 if outputs.ndim == 3 else 1
-    assert outputs.shape[feat_dim] == model.num_features, 'pooled num_features != config'
+    assert outputs.shape[feat_dim] == model.head_hidden_size, 'pooled num_features != config'
+    assert outputs.shape == outputs_pre.shape
 
     model = create_model(model_name, pretrained=False, num_classes=0).eval()
     model.to(torch_device)

timm/models/_prune.py

@@ -101,7 +101,10 @@ def adapt_model_from_string(parent_module, model_string):
                 in_features=num_features, out_features=old_module.out_features, bias=old_module.bias is not None)
             set_layer(new_module, n, new_fc)
             if hasattr(new_module, 'num_features'):
+                if getattr(new_module, 'head_hidden_size', 0) == new_module.num_features:
+                    new_module.head_hidden_size = num_features
                 new_module.num_features = num_features
+
     new_module.eval()
     parent_module.eval()
 

timm/models/beit.py

@@ -291,7 +291,7 @@ def __init__(
         super().__init__()
         self.num_classes = num_classes
         self.global_pool = global_pool
-        self.num_features = self.embed_dim = embed_dim  # num_features for consistency with other models
+        self.num_features = self.head_hidden_size = self.embed_dim = embed_dim  # for consistency with other models
         self.num_prefix_tokens = 1
         self.grad_checkpointing = False
 
@@ -392,7 +392,7 @@ def group_matcher(self, coarse=False):
         return matcher
 
     @torch.jit.ignore
-    def get_classifier(self):
+    def get_classifier(self) -> nn.Module:
         return self.head
 
     def reset_classifier(self, num_classes: int, global_pool: Optional[str] = None):

timm/models/byobnet.py

@@ -1224,6 +1224,7 @@ def __init__(
             dict(num_chs=self.num_features, reduction=reduction, module='final_conv', stage=len(self.stages))]
         self.stage_ends = [f['stage'] for f in self.feature_info]
 
+        self.head_hidden_size = self.num_features
         self.head = ClassifierHead(
             self.num_features,
             num_classes,
@@ -1250,10 +1251,11 @@ def set_grad_checkpointing(self, enable=True):
         self.grad_checkpointing = enable
 
     @torch.jit.ignore
-    def get_classifier(self):
+    def get_classifier(self) -> nn.Module:
         return self.head.fc
 
-    def reset_classifier(self, num_classes, global_pool='avg'):
+    def reset_classifier(self, num_classes: int, global_pool: Optional[str] = None):
+        self.num_classes = num_classes
         self.head.reset(num_classes, global_pool)
 
     def forward_intermediates(

timm/models/cait.py

@@ -239,7 +239,7 @@ def __init__(
 
         self.num_classes = num_classes
         self.global_pool = global_pool
-        self.num_features = self.embed_dim = embed_dim
+        self.num_features = self.head_hidden_size = self.embed_dim = embed_dim
         self.grad_checkpointing = False
 
         self.patch_embed = patch_layer(
@@ -328,7 +328,7 @@ def _matcher(name):
         return _matcher
 
     @torch.jit.ignore
-    def get_classifier(self):
+    def get_classifier(self) -> nn.Module:
         return self.head
 
     def reset_classifier(self, num_classes: int, global_pool: Optional[str] = None):

timm/models/coat.py

@@ -7,7 +7,7 @@
 
 Modified from timm/models/vision_transformer.py
 """
-from typing import List, Optional, Union, Tuple
+from typing import List, Optional, Tuple, Union
 
 import torch
 import torch.nn as nn
@@ -380,7 +380,7 @@ def __init__(
         self.return_interm_layers = return_interm_layers
         self.out_features = out_features
         self.embed_dims = embed_dims
-        self.num_features = embed_dims[-1]
+        self.num_features = self.head_hidden_size = embed_dims[-1]
         self.num_classes = num_classes
         self.global_pool = global_pool
 
@@ -556,7 +556,7 @@ def group_matcher(self, coarse=False):
         return matcher
 
     @torch.jit.ignore
-    def get_classifier(self):
+    def get_classifier(self) -> nn.Module:
         return self.head
 
     def reset_classifier(self, num_classes: int, global_pool: Optional[str] = None):

timm/models/convit.py

@@ -269,7 +269,7 @@ def __init__(
         self.num_classes = num_classes
         self.global_pool = global_pool
         self.local_up_to_layer = local_up_to_layer
-        self.num_features = self.embed_dim = embed_dim  # num_features for consistency with other models
+        self.num_features = self.head_hidden_size = self.embed_dim = embed_dim  # for consistency with other models
         self.locality_strength = locality_strength
         self.use_pos_embed = use_pos_embed
 
@@ -345,7 +345,7 @@ def set_grad_checkpointing(self, enable=True):
         assert not enable, 'gradient checkpointing not supported'
 
     @torch.jit.ignore
-    def get_classifier(self):
+    def get_classifier(self) -> nn.Module:
         return self.head
 
     def reset_classifier(self, num_classes: int, global_pool: Optional[str] = None):

timm/models/convmixer.py

@@ -40,7 +40,7 @@ def __init__(
     ):
         super().__init__()
         self.num_classes = num_classes
-        self.num_features = dim
+        self.num_features = self.head_hidden_size = dim
         self.grad_checkpointing = False
 
         self.stem = nn.Sequential(
@@ -74,7 +74,7 @@ def set_grad_checkpointing(self, enable=True):
         self.grad_checkpointing = enable
 
     @torch.jit.ignore
-    def get_classifier(self):
+    def get_classifier(self) -> nn.Module:
         return self.head
 
     def reset_classifier(self, num_classes: int, global_pool: Optional[str] = None):

timm/models/convnext.py

@@ -358,7 +358,7 @@ def __init__(
             # NOTE feature_info use currently assumes stage 0 == stride 1, rest are stride 2
             self.feature_info += [dict(num_chs=prev_chs, reduction=curr_stride, module=f'stages.{i}')]
         self.stages = nn.Sequential(*stages)
-        self.num_features = prev_chs
+        self.num_features = self.head_hidden_size = prev_chs
 
         # if head_norm_first == true, norm -> global pool -> fc ordering, like most other nets
         # otherwise pool -> norm -> fc, the default ConvNeXt ordering (pretrained FB weights)
@@ -382,6 +382,7 @@ def __init__(
                 norm_layer=norm_layer,
                 act_layer='gelu',
             )
+            self.head_hidden_size = self.head.num_features
         named_apply(partial(_init_weights, head_init_scale=head_init_scale), self)
 
     @torch.jit.ignore
@@ -401,10 +402,11 @@ def set_grad_checkpointing(self, enable=True):
             s.grad_checkpointing = enable
 
     @torch.jit.ignore
-    def get_classifier(self):
+    def get_classifier(self) -> nn.Module:
         return self.head.fc
 
-    def reset_classifier(self, num_classes=0, global_pool=None):
+    def reset_classifier(self, num_classes: int, global_pool: Optional[str] = None):
+        self.num_classes = num_classes
         self.head.reset(num_classes, global_pool)
 
     def forward_intermediates(

timm/models/crossvit.py

@@ -330,7 +330,7 @@ def __init__(
         num_patches = _compute_num_patches(self.img_size_scaled, patch_size)
         self.num_branches = len(patch_size)
         self.embed_dim = embed_dim
-        self.num_features = sum(embed_dim)
+        self.num_features = self.head_hidden_size = sum(embed_dim)
         self.patch_embed = nn.ModuleList()
 
         # hard-coded for torch jit script
@@ -415,17 +415,18 @@ def set_grad_checkpointing(self, enable=True):
         assert not enable, 'gradient checkpointing not supported'
 
     @torch.jit.ignore
-    def get_classifier(self):
+    def get_classifier(self) -> nn.Module:
         return self.head
 
     def reset_classifier(self, num_classes: int, global_pool: Optional[str] = None):
         self.num_classes = num_classes
         if global_pool is not None:
             assert global_pool in ('token', 'avg')
             self.global_pool = global_pool
-        self.head = nn.ModuleList(
-            [nn.Linear(self.embed_dim[i], num_classes) if num_classes > 0 else nn.Identity() for i in
-             range(self.num_branches)])
+        self.head = nn.ModuleList([
+            nn.Linear(self.embed_dim[i], num_classes) if num_classes > 0 else nn.Identity()
+            for i in range(self.num_branches)
+        ])
 
     def forward_features(self, x) -> List[torch.Tensor]:
         B = x.shape[0]