Merge pull request #13941 from panyx0718/fix2

handle var type inference

Author: panyx0718

paddle/fluid/framework/op_desc.cc

@@ -515,20 +515,14 @@ void OpDesc::InferShape(const BlockDesc &block) const {
 }
 
 void OpDesc::InferVarType(BlockDesc *block) const {
+  // There are a few places that var type can be set.
+  // When VarDesc is created, default set to LOD_TENSOR.
+  // When output variable is created, default is defaut set to LOD_TENSOR.
+  // We limit here to be the only place that operator defines its customized
+  // var type inference. Hence, we don't do any "default" setting here.
   auto &info = OpInfoMap::Instance().Get(this->Type());
   if (info.infer_var_type_) {
     info.infer_var_type_(*this, block);
-  } else {
-    // all output type is LoDTensor by default
-    VLOG(10) << this->Type()
-             << " has not registered InferVarType. Set output variables to "
-                "LOD_TENSOR";
-    for (auto &out_pair : this->outputs_) {
-      for (auto &out_var_name : out_pair.second) {
-        block->FindRecursiveOrCreateVar(out_var_name)
-            .SetType(proto::VarType::LOD_TENSOR);
-      }
-    }
   }
 }
 

python/paddle/fluid/layer_helper.py

@@ -324,10 +324,19 @@ def get_parameter(self, name):
             raise ValueError("no Parameter name %s found" % name)
         return param
 
-    def create_tmp_variable(self, dtype, stop_gradient=False):
+    def create_variable_for_type_inference(self, dtype, stop_gradient=False):
+        """Create a temporary variable that should be type inferred layer.
+
+        Note:
+            The default type will be set to LOD_TENSOR. However, when
+            the var is used as operator output, its type will be updated
+            based on operator's `VarTypeInference` implementation in
+            infer_var_type.
+        """
         return self.main_program.current_block().create_var(
             name=unique_name.generate(".".join([self.name, 'tmp'])),
             dtype=dtype,
+            type=core.VarDesc.VarType.LOD_TENSOR,
             persistable=False,
             stop_gradient=stop_gradient)
 
@@ -388,7 +397,7 @@ def append_bias_op(self, input_var, dim_start=1, dim_end=None):
 
         b = self.create_parameter(
             attr=bias_attr, shape=size, dtype=input_var.dtype, is_bias=True)
-        tmp = self.create_tmp_variable(dtype=input_var.dtype)
+        tmp = self.create_variable_for_type_inference(dtype=input_var.dtype)
         self.append_op(
             type='elementwise_add',
             inputs={'X': [input_var],
@@ -414,7 +423,7 @@ def append_activation(self, input_var):
         tmp = input_var
         # NOTE(dzhwinter): some activation support inplace compution.
         if not core.IsInplace(act_type):
-            tmp = self.create_tmp_variable(dtype=input_var.dtype)
+            tmp = self.create_variable_for_type_inference(dtype=input_var.dtype)
         self.append_op(
             type=act_type,
             inputs={"X": [input_var]},

python/paddle/fluid/layers/control_flow.py

@@ -80,8 +80,8 @@ def split_lod_tensor(input, mask, level=0):
 
     """
     helper = LayerHelper('split_lod_tensor', **locals())
-    out_true = helper.create_tmp_variable(dtype=input.dtype)
-    out_false = helper.create_tmp_variable(dtype=input.dtype)
+    out_true = helper.create_variable_for_type_inference(dtype=input.dtype)
+    out_false = helper.create_variable_for_type_inference(dtype=input.dtype)
     helper.append_op(
         type='split_lod_tensor',
         inputs={
@@ -131,7 +131,7 @@ def merge_lod_tensor(in_true, in_false, x, mask, level=0):
                 in_true=out_true, in_false=out_false, mask=y, x=x, level=level)
     """
     helper = LayerHelper('merge_lod_tensor', **locals())
-    out = helper.create_tmp_variable(dtype=in_true.dtype)
+    out = helper.create_variable_for_type_inference(dtype=in_true.dtype)
     helper.append_op(
         type='merge_lod_tensor',
         inputs={'X': x,
@@ -524,7 +524,7 @@ def step_output(self, o):
         if not isinstance(o, Variable):
             raise TypeError("step output takes a Variable")
 
-        tmp_o = self.helper.create_tmp_variable(dtype=o.dtype)
+        tmp_o = self.helper.create_variable_for_type_inference(dtype=o.dtype)
         self.helper.append_op(
             type='rnn_memory_helper',
             inputs={'X': [o]},
@@ -606,7 +606,8 @@ def _complete_op(self):
             pre_memories.append(mem.pre_mem.name)
             mem_var = rnn_block.var(mem.mem.name)
             assert isinstance(mem_var, Variable)
-            new_mem = self.helper.create_tmp_variable(dtype=mem_var.dtype)
+            new_mem = self.helper.create_variable_for_type_inference(
+                dtype=mem_var.dtype)
 
             rnn_block.append_op(
                 type='rnn_memory_helper',
@@ -813,7 +814,7 @@ def max_sequence_len(rank_table):
         ${out_comment}.
     """
     helper = LayerHelper("max_seqence_len", **locals())
-    res = helper.create_tmp_variable(dtype="int64")
+    res = helper.create_variable_for_type_inference(dtype="int64")
     helper.append_op(
         type="max_sequence_len",
         inputs={"RankTable": rank_table},
@@ -884,7 +885,7 @@ def array_to_lod_tensor(x, table):
           lod_tensor = fluid.layers.array_to_lod_tensor(array, table)
     """
     helper = LayerHelper("array_to_lod_tensor", **locals())
-    tmp = helper.create_tmp_variable(dtype=x.dtype)
+    tmp = helper.create_variable_for_type_inference(dtype=x.dtype)
     helper.append_op(
         type="array_to_lod_tensor",
         inputs={'X': x,
@@ -915,7 +916,7 @@ def increment(x, value=1.0, in_place=True):
     """
     helper = LayerHelper("increment", **locals())
     if not in_place:
-        out = helper.create_tmp_variable(dtype=x.dtype)
+        out = helper.create_variable_for_type_inference(dtype=x.dtype)
     else:
         out = x
     helper.append_op(
@@ -1012,7 +1013,7 @@ def less_than(x, y, force_cpu=None, cond=None, **ignored):
     """
     helper = LayerHelper("less_than", **locals())
     if cond is None:
-        cond = helper.create_tmp_variable(dtype='bool')
+        cond = helper.create_variable_for_type_inference(dtype='bool')
         cond.stop_gradient = True
 
     attrs = dict()
@@ -1051,7 +1052,7 @@ def equal(x, y, cond=None, **ignored):
     """
     helper = LayerHelper("equal", **locals())
     if cond is None:
-        cond = helper.create_tmp_variable(dtype='bool')
+        cond = helper.create_variable_for_type_inference(dtype='bool')
         cond.stop_gradient = True
 
     helper.append_op(
@@ -1098,7 +1099,7 @@ def array_read(array, i):
             array,
             Variable) or array.type != core.VarDesc.VarType.LOD_TENSOR_ARRAY:
         raise TypeError("array should be tensor array vairable")
-    out = helper.create_tmp_variable(dtype=array.dtype)
+    out = helper.create_variable_for_type_inference(dtype=array.dtype)
     helper.append_op(
         type='read_from_array',
         inputs={'X': [array],
@@ -1133,7 +1134,7 @@ def shrink_memory(x, i, table):
         usage.
     """
     helper = LayerHelper('shrink_memory', **locals())
-    out = helper.create_tmp_variable(dtype=x.dtype)
+    out = helper.create_variable_for_type_inference(dtype=x.dtype)
     helper.append_op(
         type='shrink_rnn_memory',
         inputs={'X': [x],
@@ -1170,7 +1171,7 @@ def array_length(array):
 
     """
     helper = LayerHelper('array_length', **locals())
-    tmp = helper.create_tmp_variable(dtype='int64')
+    tmp = helper.create_variable_for_type_inference(dtype='int64')
     tmp.stop_gradient = True
     helper.append_op(
         type='lod_array_length', inputs={'X': [array]}, outputs={'Out': [tmp]})
@@ -1590,7 +1591,7 @@ def __init__(self, name=None):
         self.mem_dict = dict()
         self.output_array = []
         self.outputs = []
-        self.cond = self.helper.create_tmp_variable(dtype='bool')
+        self.cond = self.helper.create_variable_for_type_inference(dtype='bool')
         self.cond.stop_gradient = False
         self.while_op = While(self.cond)
         self.input_array = []
@@ -1924,7 +1925,7 @@ def reorder_lod_tensor_by_rank(x, rank_table):
     helper.is_instance('x', Variable)
     helper.is_instance('rank_table', Variable)
 
-    out = helper.create_tmp_variable(dtype=x.dtype)
+    out = helper.create_variable_for_type_inference(dtype=x.dtype)
     helper.append_op(
         type='reorder_lod_tensor_by_rank',
         inputs={'X': [x],
@@ -1958,7 +1959,7 @@ def is_empty(x, cond=None, **ignored):
     """
     helper = LayerHelper("is_empty", **locals())
     if cond is None:
-        cond = helper.create_tmp_variable(dtype='bool')
+        cond = helper.create_variable_for_type_inference(dtype='bool')
         cond.stop_gradient = True
     elif not isinstance(cond, Variable):
         raise TypeError("cond takes a variable")

python/paddle/fluid/layers/detection.py

@@ -147,10 +147,11 @@ def rpn_target_assign(bbox_pred,
 
     helper = LayerHelper('rpn_target_assign', **locals())
     # Assign target label to anchors
-    loc_index = helper.create_tmp_variable(dtype='int32')
-    score_index = helper.create_tmp_variable(dtype='int32')
-    target_label = helper.create_tmp_variable(dtype='int32')
-    target_bbox = helper.create_tmp_variable(dtype=anchor_box.dtype)
+    loc_index = helper.create_variable_for_type_inference(dtype='int32')
+    score_index = helper.create_variable_for_type_inference(dtype='int32')
+    target_label = helper.create_variable_for_type_inference(dtype='int32')
+    target_bbox = helper.create_variable_for_type_inference(
+        dtype=anchor_box.dtype)
     helper.append_op(
         type="rpn_target_assign",
         inputs={
@@ -282,7 +283,8 @@ class number, M is number of bounding boxes. For each category
     scores = nn.reshape(x=scores, shape=compile_shape, actual_shape=run_shape)
     scores = nn.transpose(scores, perm=[0, 2, 1])
     scores.stop_gradient = True
-    nmsed_outs = helper.create_tmp_variable(dtype=decoded_box.dtype)
+    nmsed_outs = helper.create_variable_for_type_inference(
+        dtype=decoded_box.dtype)
     helper.append_op(
         type="multiclass_nms",
         inputs={'Scores': scores,
@@ -314,7 +316,7 @@ def iou_similarity(x, y, name=None):
     """
     helper = LayerHelper("iou_similarity", **locals())
     if name is None:
-        out = helper.create_tmp_variable(dtype=x.dtype)
+        out = helper.create_variable_for_type_inference(dtype=x.dtype)
     else:
         out = helper.create_variable(
             name=name, dtype=x.dtype, persistable=False)
@@ -351,7 +353,8 @@ def box_coder(prior_box,
     helper = LayerHelper("box_coder", **locals())
 
     if name is None:
-        output_box = helper.create_tmp_variable(dtype=prior_box.dtype)
+        output_box = helper.create_variable_for_type_inference(
+            dtype=prior_box.dtype)
     else:
         output_box = helper.create_variable(
             name=name, dtype=prior_box.dtype, persistable=False)
@@ -382,7 +385,7 @@ def polygon_box_transform(input, name=None):
     """
     helper = LayerHelper("polygon_box_transform", **locals())
     if name is None:
-        output = helper.create_tmp_variable(dtype=input.dtype)
+        output = helper.create_variable_for_type_inference(dtype=input.dtype)
     else:
         output = helper.create_variable(
             name=name, dtype=prior_box.input, persistable=False)
@@ -450,7 +453,7 @@ def detection_map(detect_res,
     helper = LayerHelper("detection_map", **locals())
 
     def __create_var(type):
-        return helper.create_tmp_variable(dtype=type)
+        return helper.create_variable_for_type_inference(dtype=type)
 
     map_out = __create_var('float32')
     accum_pos_count_out = out_states[0] if out_states else __create_var('int32')
@@ -557,8 +560,9 @@ def bipartite_match(dist_matrix,
         >>> matched_indices, matched_dist = fluid.layers.bipartite_match(iou)
     """
     helper = LayerHelper('bipartite_match', **locals())
-    match_indices = helper.create_tmp_variable(dtype='int32')
-    match_distance = helper.create_tmp_variable(dtype=dist_matrix.dtype)
+    match_indices = helper.create_variable_for_type_inference(dtype='int32')
+    match_distance = helper.create_variable_for_type_inference(
+        dtype=dist_matrix.dtype)
     helper.append_op(
         type='bipartite_match',
         inputs={'DistMat': dist_matrix},
@@ -644,8 +648,8 @@ def target_assign(input,
                             gt, matched_indices, mismatch_value=0)
     """
     helper = LayerHelper('target_assign', **locals())
-    out = helper.create_tmp_variable(dtype=input.dtype)
-    out_weight = helper.create_tmp_variable(dtype='float32')
+    out = helper.create_variable_for_type_inference(dtype=input.dtype)
+    out_weight = helper.create_variable_for_type_inference(dtype='float32')
     helper.append_op(
         type='target_assign',
         inputs={
@@ -816,9 +820,10 @@ def __reshape_to_2d(var):
     conf_loss = nn.reshape(
         x=conf_loss, shape=(num, num_prior), actual_shape=actual_shape)
     conf_loss.stop_gradient = True
-    neg_indices = helper.create_tmp_variable(dtype='int32')
+    neg_indices = helper.create_variable_for_type_inference(dtype='int32')
     dtype = matched_indices.dtype
-    updated_matched_indices = helper.create_tmp_variable(dtype=dtype)
+    updated_matched_indices = helper.create_variable_for_type_inference(
+        dtype=dtype)
     helper.append_op(
         type='mine_hard_examples',
         inputs={
@@ -998,8 +1003,8 @@ def _is_list_or_tuple_(data):
             max_sizes = [max_sizes]
         attrs['max_sizes'] = max_sizes
 
-    box = helper.create_tmp_variable(dtype)
-    var = helper.create_tmp_variable(dtype)
+    box = helper.create_variable_for_type_inference(dtype)
+    var = helper.create_variable_for_type_inference(dtype)
     helper.append_op(
         type="prior_box",
         inputs={"Input": input,
@@ -1337,8 +1342,8 @@ def _is_list_or_tuple_(data):
         'offset': offset
     }
 
-    anchor = helper.create_tmp_variable(dtype)
-    var = helper.create_tmp_variable(dtype)
+    anchor = helper.create_variable_for_type_inference(dtype)
+    var = helper.create_variable_for_type_inference(dtype)
     helper.append_op(
         type="anchor_generator",
         inputs={"Input": input},
@@ -1384,7 +1389,7 @@ def roi_perspective_transform(input,
     """
     helper = LayerHelper('roi_perspective_transform', **locals())
     dtype = helper.input_dtype()
-    out = helper.create_tmp_variable(dtype)
+    out = helper.create_variable_for_type_inference(dtype)
     helper.append_op(
         type="roi_perspective_transform",
         inputs={"X": input,
@@ -1418,11 +1423,15 @@ def generate_proposal_labels(rpn_rois,
 
     helper = LayerHelper('generate_proposal_labels', **locals())
 
-    rois = helper.create_tmp_variable(dtype=rpn_rois.dtype)
-    labels_int32 = helper.create_tmp_variable(dtype=gt_classes.dtype)
-    bbox_targets = helper.create_tmp_variable(dtype=rpn_rois.dtype)
-    bbox_inside_weights = helper.create_tmp_variable(dtype=rpn_rois.dtype)
-    bbox_outside_weights = helper.create_tmp_variable(dtype=rpn_rois.dtype)
+    rois = helper.create_variable_for_type_inference(dtype=rpn_rois.dtype)
+    labels_int32 = helper.create_variable_for_type_inference(
+        dtype=gt_classes.dtype)
+    bbox_targets = helper.create_variable_for_type_inference(
+        dtype=rpn_rois.dtype)
+    bbox_inside_weights = helper.create_variable_for_type_inference(
+        dtype=rpn_rois.dtype)
+    bbox_outside_weights = helper.create_variable_for_type_inference(
+        dtype=rpn_rois.dtype)
 
     helper.append_op(
         type="generate_proposal_labels",
@@ -1504,8 +1513,10 @@ def generate_proposals(scores,
     """
     helper = LayerHelper('generate_proposals', **locals())
 
-    rpn_rois = helper.create_tmp_variable(dtype=bbox_deltas.dtype)
-    rpn_roi_probs = helper.create_tmp_variable(dtype=scores.dtype)
+    rpn_rois = helper.create_variable_for_type_inference(
+        dtype=bbox_deltas.dtype)
+    rpn_roi_probs = helper.create_variable_for_type_inference(
+        dtype=scores.dtype)
     helper.append_op(
         type="generate_proposals",
         inputs={

python/paddle/fluid/layers/io.py

@@ -954,7 +954,7 @@ def read_file(reader):
     """
     helper = LayerHelper('read_file')
     out = [
-        helper.create_tmp_variable(
+        helper.create_variable_for_type_inference(
             stop_gradient=True, dtype='float32')
         for _ in range(len(reader.desc.shapes()))
     ]

python/paddle/fluid/layers/layer_function_generator.py

@@ -202,10 +202,12 @@ def func(*args, **kwargs):
             out_var = out[0] if (isinstance(out, list) or
                                  isinstance(out, tuple)) else out
         else:
-            out_var = helper.create_tmp_variable(dtype=dtype)
+            out_var = helper.create_variable_for_type_inference(dtype=dtype)
         outputs[o_name] = [out_var]
         for name in intermediate_output_names:
-            outputs[name] = [helper.create_tmp_variable(dtype=dtype)]
+            outputs[name] = [
+                helper.create_variable_for_type_inference(dtype=dtype)
+            ]
         helper.append_op(
             type=op_type, inputs=inputs, outputs=outputs, attrs=kwargs)
         return helper.append_activation(out_var)
@@ -229,7 +231,7 @@ def generate_layer_fn_noattr(op_type):
 
     def func(x, name=None):
         helper = LayerHelper(op_type, **locals())
-        output = helper.create_tmp_variable(dtype=x.dtype)
+        output = helper.create_variable_for_type_inference(dtype=x.dtype)
         helper.append_op(type=op_type, inputs={"X": x}, outputs={"Out": output})
         return output