fix save inference model conditional op (PaddlePaddle#37579)

paddle/fluid/framework/prune.cc

@@ -145,6 +145,23 @@ int FindMapByValue(const std::map<int, int>& m, int val) {
   return -1;
 }
 
+// In other two cases，the op that has feed vars as output vars is dependent:
+// 1. op has subblock, like while/for/ifelse/recurrent
+// 2. op is in subblock
+bool IsSubBlockDependent(const proto::OpDesc& op_desc,
+                         const std::set<std::string>& feed_vars,
+                         int parent_block_id) {
+  for (auto& var : op_desc.outputs()) {
+    for (auto& argu : var.arguments()) {
+      if ((HasSubBlock(op_desc) || parent_block_id != -1) &&
+          feed_vars.count(argu) != 0) {
+        return true;
+      }
+    }
+  }
+  return false;
+}
+
 // block_id is the idx of the current block in the input desc
 // parent_block_id is the idx of the parent of the current block
 // in the output desc, -1 means the current block is global block
@@ -210,7 +227,8 @@ void prune_impl(const proto::ProgramDesc& input, proto::ProgramDesc* output,
     // }
 
     if (IsTarget(op_desc) ||
-        (HasDependentOutputVar(op_desc, *dependent_vars) &&
+        ((HasDependentOutputVar(op_desc, *dependent_vars) ||
+          (IsSubBlockDependent(op_desc, feed_var_names, parent_block_id))) &&
          (GetOpRole(op_desc) & static_cast<int>(OpRole::kOptimize)) == 0)) {
       // NOTE(zhiqiu): since optimize op takes the trainable parameters as
       // inputs and output, it may introduce wrong dependency graph.
@@ -227,30 +245,6 @@ void prune_impl(const proto::ProgramDesc& input, proto::ProgramDesc* output,
       should_run.push_back(true);
     } else {
       should_run.push_back(false);
-      // If the output of an op modifies feed vars, the op should not clip.
-      // For example, in the transformer structure, the third parameter returned
-      // by beam_search op is generally assigned to a feed var. Cutting the
-      // assign op will cause an error.
-      if (parent_block_id != -1) {
-        bool flag = false;
-        for (auto& var : op_desc.outputs()) {
-          for (auto& argu : var.arguments()) {
-            if (feed_var_names.count(argu)) {
-              flag = true;
-            }
-          }
-        }
-        if (flag) {
-          should_run.back() = true;
-
-          // If any op should run, then there inputs are dependent_vars
-          for (auto& var : op_desc.inputs()) {
-            for (auto& argu : var.arguments()) {
-              dependent_vars->insert(argu);
-            }
-          }
-        }
-      }
     }
   }
 

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

@@ -0,0 +1,148 @@
+#   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.
+
+from __future__ import print_function
+
+import os
+import unittest
+import numpy as np
+
+import paddle
+import paddle.fluid as fluid
+import paddle.nn.functional as F
+
+
+def getModelOp(model_path):
+    model_bytes = paddle.static.load_from_file(model_path)
+    pg = paddle.static.deserialize_program(model_bytes)
+    main_block = pg.desc.block(0)
+    size = main_block.op_size()
+
+    result = set()
+    for i in range(0, size):
+        #print(main_block.op(i).type())    
+        result.add(main_block.op(i).type())
+
+    return result
+
+
+class WhileNet(paddle.nn.Layer):
+    def __init__(self):
+        super(WhileNet, self).__init__()
+
+    def forward(self, x):
+        y = paddle.rand(shape=[1, 3, 4, 4])
+
+        w1 = paddle.shape(y)[0]
+        w2 = paddle.shape(x)[0]
+
+        while w2 != w1:
+            x = F.avg_pool2d(x, kernel_size=3, padding=1, stride=2)
+            w2 = paddle.shape(x)[0]
+
+        return x + y
+
+
+class ForNet(paddle.nn.Layer):
+    def __init__(self):
+        super(ForNet, self).__init__()
+
+    def forward(self, x):
+        y = paddle.randint(low=0, high=5, shape=[1], dtype='int32')
+        z = paddle.randint(low=0, high=5, shape=[1], dtype='int32')
+        for i in range(0, z):
+            x = x + i
+
+        return x + y
+
+
+class IfElseNet(paddle.nn.Layer):
+    def __init__(self):
+        super(IfElseNet, self).__init__()
+
+    def forward(self, x):
+        y = paddle.to_tensor([5])
+        if x > y:
+            x = x + 1
+        else:
+            x = x - 1
+        return x
+
+
+class TestConditionalOp(unittest.TestCase):
+    def test_while_op(self):
+        paddle.disable_static()
+        net = WhileNet()
+        net = paddle.jit.to_static(
+            net,
+            input_spec=[
+                paddle.static.InputSpec(
+                    shape=[1, 3, 8, 8], dtype='float32')
+            ])
+        paddle.jit.save(net, './while_net')
+
+        right_pdmodel = set([
+            "uniform_random", "shape", "slice", "not_equal", "while",
+            "elementwise_add"
+        ])
+        paddle.enable_static()
+        pdmodel = getModelOp("while_net.pdmodel")
+        #print(len(right_pdmodel.difference(pdmodel)))
+        self.assertTrue(
+            len(right_pdmodel.difference(pdmodel)) == 0,
+            "The while op is pruned by mistake.")
+
+    def test_for_op(self):
+        paddle.disable_static()
+        net = ForNet()
+        net = paddle.jit.to_static(
+            net,
+            input_spec=[paddle.static.InputSpec(
+                shape=[1], dtype='int32')])
+        paddle.jit.save(net, './for_net')
+
+        right_pdmodel = set([
+            "randint", "fill_constant", "cast", "less_than", "while",
+            "elementwise_add"
+        ])
+        paddle.enable_static()
+        pdmodel = getModelOp("for_net.pdmodel")
+        #print(len(right_pdmodel.difference(pdmodel)))
+        self.assertTrue(
+            len(right_pdmodel.difference(pdmodel)) == 0,
+            "The for op is pruned by mistake.")
+
+    def test_if_op(self):
+        paddle.disable_static()
+        net = IfElseNet()
+        net = paddle.jit.to_static(
+            net,
+            input_spec=[paddle.static.InputSpec(
+                shape=[1], dtype='int32')])
+        paddle.jit.save(net, './if_net')
+
+        right_pdmodel = set([
+            "assign_value", "greater_than", "cast", "conditional_block",
+            "logical_not", "select_input"
+        ])
+        paddle.enable_static()
+        pdmodel = getModelOp("if_net.pdmodel")
+        #print(len(right_pdmodel.difference(pdmodel)))
+        self.assertTrue(
+            len(right_pdmodel.difference(pdmodel)) == 0,
+            "The if op is pruned by mistake.")
+
+
+if __name__ == '__main__':
+    unittest.main()