Fix dygraph unique name bug (#17592)

* fix unique_name growth bug in dygraph mode,test=develop

* change generate_tmp to generate_with_ignorable_key,test=develop

python/paddle/fluid/layer_helper_base.py

@@ -85,19 +85,19 @@ class LayerHelperBase(object):
                       block=self.startup_program.global_block()):
             if out is None:
                 out = block.create_var(
-                    name=unique_name.generate(".".join(
+                    name=unique_name.generate_with_ignorable_key(".".join(
                         [self.name, 'weight_norm_norm'])),
                     dtype=dtype,
                     persistable=False)
             abs_out = block.create_var(
-                name=unique_name.generate(".".join(
+                name=unique_name.generate_with_ignorable_key(".".join(
                     [self.name, 'weight_norm_abs'])),
                 dtype=dtype,
                 persistable=False)
             block.append_op(
                 type='abs', inputs={'X': x}, outputs={'Out': abs_out})
             pow_out = block.create_var(
-                name=unique_name.generate(".".join(
+                name=unique_name.generate_with_ignorable_key(".".join(
                     [self.name, 'weight_norm_pow'])),
                 dtype=dtype,
                 persistable=False)
@@ -107,7 +107,7 @@ class LayerHelperBase(object):
                 outputs={'Out': pow_out},
                 attrs={'factor': float(p)})
             sum_out = block.create_var(
-                name=unique_name.generate(".".join(
+                name=unique_name.generate_with_ignorable_key(".".join(
                     [self.name, 'weight_norm_sum'])),
                 dtype=dtype,
                 persistable=False)
@@ -133,7 +133,7 @@ class LayerHelperBase(object):
                          block=self.startup_program.global_block()):
             if out is None:
                 out = block.create_var(
-                    name=unique_name.generate(".".join(
+                    name=unique_name.generate_with_ignorable_key(".".join(
                         [self.name, 'weight_norm_reshape'])),
                     dtype=dtype,
                     persistable=False)
@@ -150,7 +150,7 @@ class LayerHelperBase(object):
                            block=self.startup_program.global_block()):
             if out is None:
                 out = block.create_var(
-                    name=unique_name.generate(".".join(
+                    name=unique_name.generate_with_ignorable_key(".".join(
                         [self.name, 'weight_norm_transpose'])),
                     dtype=dtype,
                     persistable=False)
@@ -168,7 +168,7 @@ class LayerHelperBase(object):
             """Computes the norm over all dimensions except dim"""
             if out is None:
                 out = block.create_var(
-                    name=unique_name.generate(".".join(
+                    name=unique_name.generate_with_ignorable_key(".".join(
                         [self.name, 'weight_norm_norm'])),
                     dtype=dtype,
                     persistable=False)
@@ -327,7 +327,8 @@ class LayerHelperBase(object):
             infer_var_type.
         """
         return self.main_program.current_block().create_var(
-            name=unique_name.generate(".".join([self.name, 'tmp'])),
+            name=unique_name.generate_with_ignorable_key(".".join(
+                [self.name, 'tmp'])),
             dtype=dtype,
             type=core.VarDesc.VarType.LOD_TENSOR,
             persistable=False,

python/paddle/fluid/layers/collective.py

@@ -33,7 +33,8 @@ def _allreduce(x, out=None, reduce_type="sum", sync_mode=False):
 
     if out is None:
         out = helper.create_variable(
-            name=unique_name.generate(".".join([x.name, 'tmp'])),
+            name=unique_name.generate_with_ignorable_key(".".join(
+                [x.name, 'tmp'])),
             shape=x.shape,
             dtype=x.dtype,
             type=x.type,

python/paddle/fluid/layers/control_flow.py

@@ -395,7 +395,7 @@ class StaticRNN(object):
                 raise ValueError(
                     "if init is None, memory at least need shape and batch_ref")
             parent_block = self._parent_block()
-            var_name = unique_name.generate("@".join(
+            var_name = unique_name.generate_with_ignorable_key("@".join(
                 [self.helper.name, "memory_boot"]))
             boot_var = parent_block.create_var(
                 name=var_name,
@@ -418,7 +418,8 @@ class StaticRNN(object):
             return self.memory(init=boot_var)
         else:
             pre_mem = self.helper.create_variable(
-                name=unique_name.generate("@".join([self.helper.name, "mem"])),
+                name=unique_name.generate_with_ignorable_key("@".join(
+                    [self.helper.name, "mem"])),
                 dtype=init.dtype,
                 shape=init.shape)
             self.memories[pre_mem.name] = StaticRNNMemoryLink(
@@ -1563,11 +1564,13 @@ class IfElse(object):
         if id(x) not in self.input_table:
             parent_block = self._parent_block()
             out_true = parent_block.create_var(
-                name=unique_name.generate('ifelse_input' + self.helper.name),
+                name=unique_name.generate_with_ignorable_key('ifelse_input' +
+                                                             self.helper.name),
                 dtype=x.dtype)
 
             out_false = parent_block.create_var(
-                name=unique_name.generate('ifelse_input' + self.helper.name),
+                name=unique_name.generate_with_ignorable_key('ifelse_input' +
+                                                             self.helper.name),
                 dtype=x.dtype)
             parent_block.append_op(
                 type='split_lod_tensor',
@@ -1609,7 +1612,7 @@ class IfElse(object):
                 raise TypeError("Each output should be a variable")
             # create outside tensor
             outside_out = parent_block.create_var(
-                name=unique_name.generate("_".join(
+                name=unique_name.generate_with_ignorable_key("_".join(
                     [self.helper.name, 'output'])),
                 dtype=each_out.dtype)
             out_table.append(outside_out)
@@ -2027,7 +2030,7 @@ class DynamicRNN(object):
         parent_block = self._parent_block_()
         for each in outputs:
             outside_array = parent_block.create_var(
-                name=unique_name.generate("_".join(
+                name=unique_name.generate_with_ignorable_key("_".join(
                     [self.helper.name, "output_array", each.name])),
                 type=core.VarDesc.VarType.LOD_TENSOR_ARRAY,
                 dtype=each.dtype)

python/paddle/fluid/layers/device.py

@@ -30,7 +30,7 @@ __all__ = []
 def get_places(device_count=None, device_type=None):
     helper = LayerHelper('get_places', **locals())
     out_places = helper.create_variable(
-        name=unique_name.generate(helper.name + ".out"))
+        name=unique_name.generate_with_ignorable_key(helper.name + ".out"))
     attrs = dict()
     if device_count is not None:
         attrs['device_count'] = int(device_count)

python/paddle/fluid/layers/nn.py

@@ -9844,7 +9844,8 @@ def clip(x, min, max, name=None):
     helper = LayerHelper("clip", **locals())
 
     if name is None:
-        name = unique_name.generate(".".join([helper.name, 'tmp']))
+        name = unique_name.generate_with_ignorable_key(".".join(
+            [helper.name, 'tmp']))
 
     out = helper.create_variable(
         type=x.type, name=name, dtype=x.dtype, persistable=False)
@@ -9883,7 +9884,8 @@ def clip_by_norm(x, max_norm, name=None):
     helper = LayerHelper("clip_by_norm", **locals())
 
     if name is None:
-        name = unique_name.generate(".".join([helper.name, 'tmp']))
+        name = unique_name.generate_with_ignorable_key(".".join(
+            [helper.name, 'tmp']))
 
     out = helper.create_variable(
         type=x.type, name=name, dtype=x.dtype, persistable=False)

python/paddle/fluid/optimizer.py

@@ -839,7 +839,8 @@ class DGCMomentumOptimizer(MomentumOptimizer):
         helper = LayerHelper("dgc_clip_by_norm_op", **args)
 
         if name is None:
-            name = unique_name.generate(".".join([helper.name, 'tmp']))
+            name = unique_name.generate_with_ignorable_key(".".join(
+                [helper.name, 'tmp']))
 
         out = helper.create_variable(
             type=x.type, name=name, dtype=x.dtype, persistable=False)
@@ -2057,7 +2058,8 @@ class ModelAverage(Optimizer):
         ).all_parameters():
             if param.do_model_average != False:
                 grad = param.block.create_var(
-                    name=unique_name.generate(".".join([param.name, 'tmp'])),
+                    name=unique_name.generate_with_ignorable_key(".".join(
+                        [param.name, 'tmp'])),
                     dtype=param.dtype,
                     persistable=False,
                     stop_gradient=True)

python/paddle/fluid/unique_name.py

@@ -58,6 +58,29 @@ def generate(key):
     return generator(key)
 
 
+# FIXME(zjl): The previous naming rule in static graph would
+# cause memory leak in dygraph mode. It is because the previous
+# nameing rule would use `conv_0.tmp` as the key, and in dygraph
+# mode, `conv_i` increases as batch increases. Thus, keys would
+# increase in a way like `conv_0.tmp`, `conv_1.tmp`, .... 
+# Not find a better way to fix this bug in dygraph mode. In TF,
+# variable name is meaningless in eager execution mode, and in
+# PyTorch, there is no variable name at all. Maybe we should
+# discard variable name in dygraph mode.
+#
+# Another concern is that save/load inference. Usually, user
+# would save model in static graph mode, and load it in dygraph
+# mode. Therefore, we keep the variable name of Parameter currently.
+# 
+# Please fix me if a better method is found.        
+def generate_with_ignorable_key(key):
+    from .framework import in_dygraph_mode
+    if in_dygraph_mode():
+        key = "tmp"
+
+    return generator(key)
+
+
 def switch(new_generator=None):
     global generator
     old = generator