Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into fix-sendrec-ci

paddle/fluid/framework/block_desc.cc

@@ -147,15 +147,52 @@ void BlockDesc::RemoveOp(size_t s, size_t e) {
   if (ops_.begin() + s == ops_.end() || ops_.begin() + e == ops_.end()) {
     return;
   }
+  auto get_vars = [](std::deque<std::unique_ptr<OpDesc>>::iterator &op,
+                     std::vector<std::string> &v) {
+    auto in_names = (*op)->InputArgumentNames();
+    v.insert(v.end(), in_names.begin(), in_names.end());
+    auto out_names = (*op)->OutputArgumentNames();
+    v.insert(v.end(), out_names.begin(), out_names.end());
+    std::sort(v.begin(), v.end());
+    auto last = std::unique(v.begin(), v.end());
+    v.erase(last, v.end());
+  };
   need_update_ = true;
-  for (auto it = ops_.begin() + s; it != ops_.begin() + e; it++) {
-    auto names = (*it)->InputArgumentNames();
-    for (auto n : names) {
-      // TODO(typhoonzero): delete vars if no other op use it.
-      VLOG(3) << "deleting var " << n;
+
+  for (size_t i = s; i < e; i++) {
+    // since remove op one by one, every time remove the first op.
+    auto op = ops_.begin() + s;
+
+    // collect input and output variables from current delete op
+    std::vector<std::string> cur_vars;
+    get_vars(op, cur_vars);
+
+    // remove current op
+    ops_.erase(ops_.begin() + s);
+
+    // collect input and output variables from other ops
+    std::vector<std::string> other_vars;
+    for (auto it = ops_.begin(); it != ops_.end(); it++) {
+      get_vars(it, other_vars);
+    }
+
+    // variables should be deleted
+    std::vector<std::string> delete_vars;
+    // delete_vars = cur_vars -  cur_vars ^ other_input_vars
+    std::set_difference(cur_vars.begin(), cur_vars.end(), other_vars.begin(),
+                        other_vars.end(),
+                        std::inserter(delete_vars, delete_vars.end()));
+    // remove variables
+    for (size_t i = 0; i < delete_vars.size(); i++) {
+      auto name = delete_vars[i];
+      auto it = vars_.find(name);
+      PADDLE_ENFORCE(it != vars_.end(),
+                     "%s is not in variable list, it should not be deleted",
+                     name);
+      vars_.erase(it);
+      VLOG(3) << "deleting variable " << name;
     }
   }
-  ops_.erase(ops_.begin() + s, ops_.begin() + e);
 }
 
 std::vector<OpDesc *> BlockDesc::AllOps() const {

paddle/fluid/framework/block_desc.h

@@ -89,6 +89,11 @@ class BlockDesc {
 
   OpDesc *InsertOp(size_t index);
 
+  /*
+   * Remove Op and its input/output variables.
+   * Note that for either input or ouput variable, if it is also an input or
+   * output variable of other ops, we should remain it.
+   */
   void RemoveOp(size_t s, size_t e);
 
   std::vector<OpDesc *> AllOps() const;

paddle/scripts/submit_local.sh.in

@@ -153,9 +153,15 @@ if [ $? -ne 0 ]; then
     exit 1
 fi
 
-INSTALLED_VERSION=`pip freeze 2>/dev/null | grep '^paddle' | sed 's/.*==//g'`
+if [ "@WITH_GPU@" == "ON" ]; then
+    PADDLE_NAME="paddlepaddle-gpu"
+else 
+    PADDLE_NAME="paddlepaddle"
+fi
+
+INSTALLED_VERSION=`pip freeze 2>/dev/null | grep "^${PADDLE_NAME}==" | sed 's/.*==//g'`
 
-if [ -z ${INSTALLED_VERSION} ]; then
+if [ -z "${INSTALLED_VERSION}" ]; then
    INSTALLED_VERSION="0.0.0"  # not installed
 fi
 cat <<EOF | python -

python/paddle/fluid/layers/nn.py

@@ -1483,6 +1483,7 @@ def batch_norm(input,
                param_attr=None,
                bias_attr=None,
                data_layout='NCHW',
+               in_place=False,
                name=None,
                moving_mean_name=None,
                moving_variance_name=None):
@@ -1538,7 +1539,7 @@ def batch_norm(input,
     saved_mean = helper.create_tmp_variable(dtype=dtype, stop_gradient=True)
     saved_variance = helper.create_tmp_variable(dtype=dtype, stop_gradient=True)
 
-    batch_norm_out = helper.create_tmp_variable(dtype)
+    batch_norm_out = input if in_place else helper.create_tmp_variable(dtype)
 
     helper.append_op(
         type="batch_norm",

python/paddle/fluid/nets.py

@@ -98,7 +98,7 @@ def img_conv_group(input,
             use_mkldnn=use_mkldnn)
 
         if conv_with_batchnorm[i]:
-            tmp = layers.batch_norm(input=tmp, act=conv_act)
+            tmp = layers.batch_norm(input=tmp, act=conv_act, in_place=True)
             drop_rate = conv_batchnorm_drop_rate[i]
             if abs(drop_rate) > 1e-5:
                 tmp = layers.dropout(x=tmp, dropout_prob=drop_rate)

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

@@ -186,6 +186,34 @@ class TestBlockDesc(unittest.TestCase):
             all_ops.append(block.op(idx))
         self.assertEqual(all_ops, [op0, op1, op2])
 
+    def test_remove_op(self):
+        prog = core.ProgramDesc()
+        self.assertIsNotNone(prog)
+        block = prog.block(0)
+        self.assertIsNotNone(block)
+        op1 = block.append_op()
+        op2 = block.append_op()
+        var1 = block.var("var1")
+        var2 = block.var("var2")
+        var3 = block.var("var3")
+        var4 = block.var("var4")
+        var5 = block.var("var5")
+        op1.set_input("X", ["var1", "var2"])
+        op1.set_output("Y", ["var3", "var4"])
+        op2.set_input("X", ["var1"])
+        op2.set_output("Y", ["var4", "var5"])
+
+        # remove op1, its input var2 and output var3 will be removed at the same time,
+        # but its input var1 and output var4 will not be removed since they are used for op2.
+        block.remove_op(0, 1)
+
+        all_ops = []
+        for idx in xrange(0, block.op_size()):
+            all_ops.append(block.op(idx))
+        self.assertEqual(all_ops, [op2])
+        all_vars = block.all_vars()
+        self.assertEqual(set(all_vars), {var1, var4, var5})
+
 
 if __name__ == '__main__':
     unittest.main()