fix pruned model save and load, fix bugs of pruning depthwise conv (#422)

* fix pruned model save and load, fix bugs of pruning depthwise conv

* add unit test of pruned model save and load

* temp delete

* add unit test

paddleslim/prune/prune_io.py

@@ -26,7 +26,7 @@ def save_model(exe, graph, dirname):
     assert graph is not None and dirname is not None
     graph = GraphWrapper(graph) if isinstance(graph, Program) else graph
 
-    fluid.io.save_params(
+    fluid.io.save_persistables(
         executor=exe,
         dirname=dirname,
         main_program=graph.program,
@@ -34,8 +34,8 @@ def save_model(exe, graph, dirname):
     weights_file = os.path.join(dirname, _PARAMS_FILE)
     _logger.info("Save model weights into {}".format(weights_file))
     shapes = {}
-    for var in graph.all_parameters():
-        shapes[var.name()] = var.shape()
+    for var in fluid.io.get_program_persistable_vars(graph.program):
+        shapes[var.name] = var.shape
     SHAPES_FILE = os.path.join(dirname, _SHAPES_FILE)
     with open(SHAPES_FILE, "w") as f:
         json.dump(shapes, f)
@@ -62,7 +62,7 @@ def load_model(exe, graph, dirname):
 
     _logger.info("Load shapes of weights from {}".format(SHAPES_FILE))
 
-    fluid.io.load_params(
+    fluid.io.load_persistables(
         executor=exe,
         dirname=dirname,
         main_program=graph.program,

paddleslim/prune/prune_walker.py

@@ -480,9 +480,6 @@ class depthwise_conv2d(PruneWorker):
             self.pruned_params.append((filter_var, 0, pruned_idx))
             self._visit(filter_var, 0)
 
-            new_groups = filter_var.shape()[0] - len(pruned_idx)
-            self.op.set_attr("groups", new_groups)
-
             for op in filter_var.outputs():
                 self._prune_op(op, filter_var, 0, pruned_idx)
 
@@ -499,8 +496,6 @@ class depthwise_conv2d(PruneWorker):
                         (self.op.inputs("Bias"), channel_axis, pruned_idx))
 
                 self.pruned_params.append((var, 0, pruned_idx))
-                new_groups = var.shape()[0] - len(pruned_idx)
-                self.op.set_attr("groups", new_groups)
 
                 for op in var.outputs():
                     self._prune_op(op, var, 0, pruned_idx)
@@ -518,9 +513,6 @@ class depthwise_conv2d(PruneWorker):
             self.pruned_params.append((filter_var, 0, pruned_idx))
             self._visit(filter_var, 0)
 
-            new_groups = filter_var.shape()[0] - len(pruned_idx)
-            self.op.set_attr("groups", new_groups)
-
             for op in filter_var.outputs():
                 self._prune_op(op, filter_var, 0, pruned_idx)
 

tests/test_pruned_model_save_load.py

+# Copyright (c) 2020  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 sys
+sys.path.append("../")
+import unittest
+import paddle.fluid as fluid
+from paddleslim.prune import Pruner, save_model, load_model
+from layers import conv_bn_layer
+import numpy as np
+import numpy
+
+
+class TestSaveAndLoad(unittest.TestCase):
+    def test_prune(self):
+        train_program = fluid.Program()
+        startup_program = fluid.Program()
+        with fluid.program_guard(train_program, startup_program):
+            input = fluid.data(name="image", shape=[None, 3, 16, 16])
+            conv1 = conv_bn_layer(input, 8, 3, "conv1")
+            conv2 = conv_bn_layer(conv1, 8, 3, "conv2")
+            sum1 = conv1 + conv2
+            conv3 = conv_bn_layer(sum1, 8, 3, "conv3")
+            conv4 = conv_bn_layer(conv3, 8, 3, "conv4")
+            sum2 = conv4 + sum1
+            conv5 = conv_bn_layer(sum2, 8, 3, "conv5")
+            conv6 = conv_bn_layer(conv5, 8, 3, "conv6")
+
+        place = fluid.CPUPlace()
+        exe = fluid.Executor(place)
+
+        scope = fluid.global_scope()
+        exe.run(startup_program, scope=scope)
+        criterion = 'bn_scale'
+        pruner = Pruner(criterion)
+        main_program, _, _ = pruner.prune(
+            train_program,
+            scope,
+            params=["conv4_weights"],
+            ratios=[0.5],
+            place=place,
+            lazy=False,
+            only_graph=False,
+            param_backup=None,
+            param_shape_backup=None)
+
+        x = numpy.random.random(size=(10, 3, 16, 16)).astype('float32')
+        loss_data, = exe.run(train_program,
+                             feed={"image": x},
+                             fetch_list=[conv6.name])
+
+        save_model(exe, main_program, 'model_file')
+        pruned_program = fluid.Program()
+        pruned_startup_program = fluid.Program()
+        with fluid.program_guard(pruned_program, pruned_startup_program):
+            input = fluid.data(name="image", shape=[None, 3, 16, 16])
+            conv1 = conv_bn_layer(input, 8, 3, "conv1")
+            conv2 = conv_bn_layer(conv1, 8, 3, "conv2")
+            sum1 = conv1 + conv2
+            conv3 = conv_bn_layer(sum1, 8, 3, "conv3")
+            conv4 = conv_bn_layer(conv3, 8, 3, "conv4")
+            sum2 = conv4 + sum1
+            conv5 = conv_bn_layer(sum2, 8, 3, "conv5")
+            conv6 = conv_bn_layer(conv5, 8, 3, "conv6")
+        exe.run(pruned_startup_program)
+        load_model(exe, pruned_program, 'model_file')
+        shapes = {
+            "conv1_weights": (4, 3, 3, 3),
+            "conv2_weights": (4, 4, 3, 3),
+            "conv3_weights": (8, 4, 3, 3),
+            "conv4_weights": (4, 8, 3, 3),
+            "conv5_weights": (8, 4, 3, 3),
+            "conv6_weights": (8, 8, 3, 3)
+        }
+
+        for param in pruned_program.global_block().all_parameters():
+            if "weights" in param.name:
+                print("param: {}; param shape: {}".format(param.name,
+                                                          param.shape))
+                self.assertTrue(param.shape == shapes[param.name])
+
+
+if __name__ == '__main__':
+    unittest.main()