Add IntermediateLayerGetter

python/paddle/fluid/tests/unittests/CMakeLists.txt

@@ -1229,6 +1229,9 @@ if(WITH_GLOO)
   set_tests_properties(test_parallel_dygraph_sparse_embedding_over_height_gloo
                        PROPERTIES TIMEOUT 120)
 endif()
+if(NOT WITH_GPU)
+  set_tests_properties(test_IntermediateLayerGetter PROPERTIES TIMEOUT 900)
+endif()
 
 if($ENV{USE_STANDALONE_EXECUTOR})
   # these test will fail in some server due to PR#42149, temporarily set it use old executor.

python/paddle/fluid/tests/unittests/test_IntermediateLayerGetter.py

@@ -0,0 +1,96 @@
+#   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import paddle
+import unittest
+import random
+from paddle.vision.models.utils import IntermediateLayerGetter
+
+
+class TestBase:
+    def setUp(self):
+
+        self.init_model()
+        self.model.eval()
+
+        self.layer_names = [
+            (order, name)
+            for order, (name, _) in enumerate(self.model.named_children())
+        ]
+        # choose two layer children of model randomly
+        self.start, self.end = sorted(
+            random.sample(self.layer_names, 2), key=lambda x: x[0]
+        )
+
+        self.return_layers_dic = {self.start[1]: "feat1", self.end[1]: "feat2"}
+        self.new_model = IntermediateLayerGetter(
+            self.model, self.return_layers_dic
+        )
+
+    def init_model(self):
+        self.model = paddle.vision.models.resnet50(pretrained=False)
+
+    @paddle.no_grad()
+    def test_inter_result(self):
+
+        inp = paddle.randn([32, 3, 224, 224])
+        inter_oup = self.new_model(inp)
+
+        for layer_name, layer in self.model.named_children():
+
+            if (isinstance(layer, paddle.nn.Linear) and inp.ndim == 4) or (
+                len(layer.sublayers()) > 0
+                and isinstance(layer.sublayers()[0], paddle.nn.Linear)
+                and inp.ndim == 4
+            ):
+                inp = paddle.flatten(inp, 1)
+
+            inp = layer(inp)
+            if layer_name in self.return_layers_dic:
+                feat_name = self.return_layers_dic[layer_name]
+                self.assertTrue((inter_oup[feat_name] == inp).all())
+
+
+class TestIntermediateLayerGetterResNet50(TestBase, unittest.TestCase):
+    def init_model(self):
+        self.model = paddle.vision.models.resnet50(pretrained=False)
+
+
+class TestIntermediateLayerGetterDenseNet201(TestBase, unittest.TestCase):
+    def init_model(self):
+        self.model = paddle.vision.models.densenet201(pretrained=False)
+
+
+class TestIntermediateLayerGetterVGG19(TestBase, unittest.TestCase):
+    def init_model(self):
+        self.model = paddle.vision.models.vgg19(pretrained=False)
+
+
+class TestIntermediateLayerGetterMobileNetV3Large(TestBase, unittest.TestCase):
+    def init_model(self):
+        self.model = paddle.vision.models.MobileNetV3Large()
+
+
+class TestIntermediateLayerGetterShuffleNetV2(TestBase, unittest.TestCase):
+    def init_model(self):
+        self.model = paddle.vision.models.ShuffleNetV2()
+
+
+class TestIntermediateLayerGetterAlexNet(TestBase, unittest.TestCase):
+    def init_model(self):
+        self.model = paddle.vision.models.AlexNet()
+
+
+if __name__ == "__main__":
+    unittest.main()

python/paddle/vision/__init__.py

@@ -79,6 +79,7 @@
 from .models import shufflenet_v2_x1_5  # noqa: F401
 from .models import shufflenet_v2_x2_0  # noqa: F401
 from .models import shufflenet_v2_swish  # noqa: F401
+from .models import IntermediateLayerGetter
 from .transforms import BaseTransform  # noqa: F401
 from .transforms import Compose  # noqa: F401
 from .transforms import Resize  # noqa: F401

python/paddle/vision/models/__init__.py

@@ -63,6 +63,7 @@
 from .shufflenetv2 import shufflenet_v2_x1_5  # noqa: F401
 from .shufflenetv2 import shufflenet_v2_x2_0  # noqa: F401
 from .shufflenetv2 import shufflenet_v2_swish  # noqa: F401
+from .utils import IntermediateLayerGetter
 
 __all__ = [  # noqa
     'ResNet',

python/paddle/vision/models/utils.py

@@ -12,6 +12,13 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+import paddle
+import paddle.nn as nn
+from collections import OrderedDict
+from typing import Dict
+
+__all__ = ["IntermediateLayerGetter"]
+
 
 def _make_divisible(v, divisor=8, min_value=None):
     """
@@ -30,3 +37,74 @@ def _make_divisible(v, divisor=8, min_value=None):
     if new_v < 0.9 * v:
         new_v += divisor
     return new_v
+
+
+class IntermediateLayerGetter(nn.LayerDict):
+    """
+    Layer wrapper that returns intermediate layers from a model.
+
+    It has a strong assumption that the layers have been registered into the model in the
+    same order as they are used. This means that one should **not** reuse the same nn.Layer
+    twice in the forward if you want this to work.
+
+    Additionally, it is only able to query sublayer that are directly assigned to the model.
+    So if `model` is passed, `model.feature1` can be returned, but not `model.feature1.layer2`.
+
+    Args:
+        model (nn.Layer): model on which we will extract the features
+        return_layers (Dict[name, new_name]): a dict containing the names of the layers for
+        which the activations will be returned as the key of the dict, and the value of the
+        dict is the name of the returned activation (which the user can specify).
+
+    Examples:
+        .. code-block:: python
+
+        >>> import paddle
+        >>> m = paddle.vision.models.resnet18(pretrained=True)
+        >>> # extract layer1 and layer3, giving as names `feat1` and feat2`
+        >>> new_m = paddle.vision.models.utils.IntermediateLayerGetter(m,
+        >>>     {'layer1': 'feat1', 'layer3': 'feat2'})
+        >>> out = new_m(paddle.rand([1, 3, 224, 224]))
+        >>> print([(k, v.shape) for k, v in out.items()])
+        >>>     [('feat1', [1, 64, 56, 56]),
+        >>>      ('feat2', [1, 256, 14, 14])]
+    """
+
+    __annotations__ = {
+        "return_layers": Dict[str, str],
+    }
+
+    def __init__(self, model: nn.Layer, return_layers: Dict[str, str]) -> None:
+        if not set(return_layers).issubset(
+            [name for name, _ in model.named_children()]
+        ):
+            raise ValueError("return_layers are not present in model")
+        orig_return_layers = return_layers
+        return_layers = {str(k): str(v) for k, v in return_layers.items()}
+        layers = OrderedDict()
+        for name, module in model.named_children():
+            layers[name] = module
+            if name in return_layers:
+                del return_layers[name]
+            if not return_layers:
+                break
+
+        super(IntermediateLayerGetter, self).__init__(layers)
+        self.return_layers = orig_return_layers
+
+    def forward(self, x):
+        out = OrderedDict()
+        for name, module in self.items():
+
+            if (isinstance(module, nn.Linear) and x.ndim == 4) or (
+                len(module.sublayers()) > 0
+                and isinstance(module.sublayers()[0], nn.Linear)
+                and x.ndim == 4
+            ):
+                x = paddle.flatten(x, 1)
+
+            x = module(x)
+            if name in self.return_layers:
+                out_name = self.return_layers[name]
+                out[out_name] = x
+        return out