[Embedding] Initialize the layout of ValuePtr in InitializeOp and ImportOp. (#405)

Author: lixy9474

tensorflow/core/kernels/kv_variable_ops.cc

@@ -138,7 +138,6 @@ class InitializeKvVariableOp : public OpKernel {
           &false_positive_probability_));
     OP_REQUIRES_OK(c, c->GetAttr("l2_weight_threshold",
           &l2_weight_threshold_));
-    OP_REQUIRES_OK(c, c->GetAttr("layout", &layout_));
     OP_REQUIRES_OK(c, c->GetAttr("default_value_dim", &default_value_dim_));
     OP_REQUIRES_OK(c, c->GetAttr("slot_num", &slot_num_));
     OP_REQUIRES_OK(c, c->GetAttr("record_freq", &record_freq_));
@@ -156,6 +155,20 @@ class InitializeKvVariableOp : public OpKernel {
       filter_freq_ = 0;
     }
 
+    if ((filter_freq_ != 0 && max_element_size_ == 0)
+         || steps_to_live_ != 0 || record_freq_
+         || record_version_ || storage_type > 5) {
+      if (block_num_ > 1 || (filter_freq_ != 0 && storage_type <= 5)) {
+        layout_ = "normal";
+      } else {
+        layout_ = "normal_contiguous";
+      }
+    } else {
+      layout_ = "light";
+    }
+
+    CHECK(block_num_ == 1 || layout_ != "normal_contiguous");
+
     if (steps_to_live_ == kEmbeddingVarUseDB ||
         steps_to_live_ == kInitializableEmbeddingVarUseDB) {
       LOG(INFO) << "hashmap use db";
@@ -199,7 +212,6 @@ class InitializeKvVariableOp : public OpKernel {
     std::string opname = handle_self.name();
 
     EmbeddingVar<TKey, TValue>* ev = nullptr;
-    CHECK(block_num_ == 1 || layout_ != "normal_contiguous");
 
     if (handle_self.name() == handle_primary.name() &&
         handle_self.container() == handle_primary.container()) {
@@ -642,7 +654,6 @@ class KvResourceImportV2Op: public AsyncOpKernel {
           &false_positive_probability_));
     OP_REQUIRES_OK(c, c->GetAttr("l2_weight_threshold",
           &l2_weight_threshold_));
-    OP_REQUIRES_OK(c, c->GetAttr("layout", &layout_));
     OP_REQUIRES_OK(c, c->GetAttr("max_freq", &max_freq_));
     OP_REQUIRES_OK(c, c->GetAttr("default_value_dim",
           &default_value_dim_));
@@ -656,6 +667,20 @@ class KvResourceImportV2Op: public AsyncOpKernel {
     OP_REQUIRES_OK(c, c->GetAttr("record_freq", &record_freq_));
     OP_REQUIRES_OK(c, c->GetAttr("record_version", &record_version_));
 
+    if ((filter_freq_ != 0 && max_element_size_ == 0)
+         || steps_to_live_ != -1 || record_freq_
+         || record_version_ || storage_type > 5) {
+      if (block_num_ > 1 || (filter_freq_ != 0 && storage_type <= 5)) {
+        layout_ = "normal";
+      } else {
+        layout_ = "normal_contiguous";
+      }
+    } else {
+      layout_ = "light";
+    }
+
+    CHECK(block_num_ == 1 || layout_ != "normal_contiguous");
+
     TF_CHECK_OK(ReadBoolFromEnvVar("TF_ENABLE_EV_ASYNC_RESTORE", true,
                                    &ev_async_restore_));
   }

tensorflow/python/ops/kv_variable_ops.py

@@ -293,18 +293,6 @@ def _init_from_args(self,
     self._storage_path = evconfig.storage_path
     self._storage_size = evconfig.storage_size
     self._default_value_dim = evconfig.default_value_dim
-    if (isinstance(evconfig.filter_strategy, variables.CounterFilter)  and self._filter_freq != 0) or \
-       self._steps_to_live not in [0, None] or self._record_version or \
-       self._storage_type in multi_level_list or self._record_freq:
-      if self._block_num not in [1, None] and self._storage_type in multi_level_list:
-        raise ValueError("Dynamic-dimension Embedding and Multi-level EV can't be enabled together") 
-      if self._block_num not in [1, None] or \
-          (self._filter_freq != 0 and self._storage_type not in multi_level_list):
-        self._layout = "normal"
-      else:
-        self._layout = "normal_contiguous"
-    else:
-      self._layout = "light"
 
     if self._primary is None:
       self._is_primary = True
@@ -409,7 +397,7 @@ def _init_from_args(self,
                     false_positive_probability = self._false_positive_probability,
                     counter_type = self._counter_type,
                     max_freq = 99999,
-                    layout = self._layout,
+                    layout = "",
                     storage_type = self._storage_type,
                     storage_path = self._storage_path,
                     storage_size = self._storage_size,

tensorflow/python/training/saving/saveable_object_util.py

@@ -208,7 +208,7 @@ def restore(self, restored_tensors, unused_restored_shapes):
             max_element_size = self.var._max_element_size,
             false_positive_probability = self.var._false_positive_probability,
             counter_type = self.var._counter_type,
-            layout = self.var._layout,
+            layout = "",
             storage_type=self.var._storage_type,
             storage_path=self.var._storage_path,
             storage_size=self.var._storage_size,

tensorflow/python/training/slot_creator.py

@@ -123,6 +123,8 @@ def _create_slot_var(primary, val, scope, validate_shape, shape, dtype, slot_con
             primary=primary._primary,
             slot_num=slot_config.slot_num,
             storage_type=primary.storage_type,
+            storage_path=primary._storage_path,
+            storage_size=primary._storage_size,
             l2_weight_threshold=primary._l2_weight_threshold,
             filter_strategy=filter_strategy)
         )