parameter sharing in fluid with simple test cases

.gitignore

+# tilde
+*~
+
 # Byte-compiled / optimized / DLL files
 __pycache__/
 *.py[cod]

README.md

-# RL
-Reinforcement learning framework
+# PPRL
+PaddlePaddle Reinforcement Learning Framework

pprl/__init__.py

+#   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.

pprl/layers/__init__.py

+#   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.
+"""
+This file wraps Fluid layers that have parameters to support parameter sharing.
+For other layers that don't have parameters, we simply copy them to this namespace.
+"""
+from paddle.fluid.layers import *
+from layer_wrappers import *

pprl/layers/tests/test_param_name.py

+#   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 unittest
+import pprl.layers as layers
+
+
+class TestParamName(unittest.TestCase):
+    def test_name_number(self):
+        self.fc1 = layers.fc(100)
+        self.fc2 = layers.fc(100)
+        self.fc3 = layers.fc(100, bias_attr=False)
+        self.fc4 = layers.fc(100, param_attr=False)
+        self.fc5 = layers.fc(100, name="fc", bias_attr=False)
+        self.embedding = layers.embedding((100, 128))
+        self.embedding_custom = layers.embedding(
+            (100, 128), name="embedding_custom")
+        self.conv2d = layers.conv2d(
+            num_filters=64,
+            filter_size=3,
+            name="my_conv2d",
+            set_paras=self.embedding.parameters())
+        self.dynamic_grus = []
+        for i in range(5):
+            self.dynamic_grus.append(layers.dynamic_gru(50))
+
+        ## fc1 and fc2 have different parameters
+        self.assertEqual(self.fc1.param_name, "fc_0.w")
+        self.assertEqual(self.fc2.param_name, "fc_1.w")
+
+        ## fc3 has no bias and fc4 has no param; so the names are None
+        self.assertEqual(self.fc3.bias_name, None)
+        self.assertEqual(self.fc4.param_name, None)
+
+        ## fc5 has a custom name without a bias
+        ## fc5 has a different param name with fc1
+        self.assertEqual(self.fc5.param_name, "fc_0_.w")
+        self.assertEqual(self.fc5.bias_name, None)
+
+        ## embedding layer has no bias
+        self.assertEqual(self.embedding.param_name, "embedding_0.w")
+        self.assertEqual(self.embedding.bias_name, None)
+
+        ## embedding layer with a custom name; the custom id is 1 up to this point
+        self.assertEqual(self.embedding_custom.param_name,
+                         "embedding_custom_1_.w")
+
+        ## conv2d shares param with embedding; has a custom bias name; the custom id is 2 now
+        self.assertEqual(self.conv2d.param_name, "embedding_0.w")
+        self.assertEqual(self.conv2d.bias_name, "my_conv2d_2_.wbias")
+
+        for i, gru in enumerate(self.dynamic_grus):
+            self.assertEqual(gru.param_name, "dynamic_gru_%d.w" % i)
+
+
+if __name__ == '__main__':
+    unittest.main()

pprl/layers/tests/test_param_sharing.py

+#   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 unittest
+import pprl.layers as layers
+import paddle.fluid as fluid
+import numpy as np
+
+
+class TestParamSharing(unittest.TestCase):
+    def __init__(self, *args, **kwargs):
+        super(TestParamSharing, self).__init__(*args, **kwargs)
+        self.fc1 = layers.fc(64, bias_attr=False)
+        self.fc2 = layers.fc(64, bias_attr=False)
+        self.fc3 = layers.fc(64, name="fc")
+        self.fc4 = layers.fc(64, name="fc")
+        ## we bind the paras of self.embedding to those of self.fc1
+        self.embedding = layers.embedding(
+            (100, 64), set_paras=self.fc1.parameters())
+
+    def test_param_sharing(self):
+        """
+        Test case for parameter sharing between layers of the same type
+        """
+        main_program = fluid.Program()
+        startup_program = fluid.Program()
+
+        with fluid.program_guard(main_program, startup_program):
+            x = layers.data(name='x', shape=[100], dtype="float32")
+            y1 = self.fc1(input=x)
+            y11 = self.fc1(input=x)
+            y2 = self.fc2(input=x)
+            y3 = self.fc3(input=x)
+            y4 = self.fc4(input=x)
+
+        place = fluid.CPUPlace()
+        exe = fluid.Executor(place)
+        exe.run(startup_program)
+        batch_size = 10
+        input_x = np.random.uniform(0, 1, [batch_size, 100]).astype("float32")
+        outputs = exe.run(main_program,
+                          feed={"x": input_x},
+                          fetch_list=[y1, y11, y2, y3, y4])
+
+        self.assertEqual(
+            np.sum(outputs[0].flatten()), np.sum(outputs[1].flatten()))
+        self.assertNotEqual(
+            np.sum(outputs[1].flatten()), np.sum(outputs[2].flatten()))
+        self.assertNotEqual(
+            np.sum(outputs[3].flatten()), np.sum(outputs[4].flatten()))
+
+    def test_manual_param_sharing(self):
+        """
+        Test case for parameter sharing between layers of different types
+        """
+        batch_size = 10
+        dict_size = 100
+
+        main_program = fluid.Program()
+        startup_program = fluid.Program()
+        with fluid.program_guard(main_program, startup_program):
+            x = layers.data(name='x', shape=[1], dtype="int")
+            cx = layers.cast(
+                x=layers.one_hot(
+                    input=x, depth=dict_size), dtype="float32")
+            ## remove bias because embedding layer does not have one
+            y1 = self.fc1(input=cx)
+            y2 = self.embedding(input=x)
+
+        place = fluid.CPUPlace()
+        exe = fluid.Executor(place)
+        exe.run(startup_program)
+        input_x = np.random.randint(
+            dict_size, size=(batch_size, 1)).astype("int")
+        outputs = exe.run(main_program,
+                          feed={'x': input_x},
+                          fetch_list=[y1, y2])
+
+        self.assertEqual(
+            np.sum(outputs[0].flatten()), np.sum(outputs[1].flatten()))
+
+
+if __name__ == "__main__":
+    unittest.main()