Merge pull request #16930 from phlrain/pick_infer_shape_many

Merge pull request #16840 from phlrain/fix_shape_check_many

paddle/fluid/operators/controlflow/tensor_array_read_write_op.cc

@@ -81,8 +81,10 @@ class WriteToArrayInferShape : public framework::InferShapeBase {
  public:
   void operator()(framework::InferShapeContext *context) const override {
     PADDLE_ENFORCE(context->HasInput("I"), "Must set the subscript index");
-    PADDLE_ENFORCE_EQ(framework::product(context->GetInputDim("I")), 1,
-                      "The number of element of subscript index must be 1");
+    if (context->IsRuntime()) {
+      PADDLE_ENFORCE_EQ(framework::product(context->GetInputDim("I")), 1,
+                        "The number of element of subscript index must be 1");
+    }
     if (!context->HasInput("X")) {
       return;
     }

paddle/fluid/operators/data_norm_op.cc

@@ -13,6 +13,7 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/fluid/operators/data_norm_op.h"
+#include <memory>
 #include <string>
 #include "paddle/fluid/framework/data_layout.h"
 #ifdef PADDLE_WITH_MKLDNN
@@ -65,9 +66,11 @@ class DataNormOp : public framework::OperatorWithKernel {
     PADDLE_ENFORCE_EQ(ctx->GetInputDim("BatchSize").size(), 1UL);
     PADDLE_ENFORCE_EQ(ctx->GetInputDim("BatchSum").size(), 1UL);
     PADDLE_ENFORCE_EQ(ctx->GetInputDim("BatchSquareSum").size(), 1UL);
-    PADDLE_ENFORCE_EQ(ctx->GetInputDim("BatchSize")[0], C);
-    PADDLE_ENFORCE_EQ(ctx->GetInputDim("BatchSum")[0], C);
-    PADDLE_ENFORCE_EQ(ctx->GetInputDim("BatchSquareSum")[0], C);
+    if (ctx->IsRuntime()) {
+      PADDLE_ENFORCE_EQ(ctx->GetInputDim("BatchSize")[0], C);
+      PADDLE_ENFORCE_EQ(ctx->GetInputDim("BatchSum")[0], C);
+      PADDLE_ENFORCE_EQ(ctx->GetInputDim("BatchSquareSum")[0], C);
+    }
 
     ctx->SetOutputDim("Y", x_dims);
     ctx->SetOutputDim("Means", {C});

paddle/fluid/operators/huber_loss_op.cc

@@ -28,13 +28,18 @@ class HuberLossOp : public framework::OperatorWithKernel {
     auto x_dims = ctx->GetInputDim("X");
     auto y_dims = ctx->GetInputDim("Y");
 
-    PADDLE_ENFORCE_EQ(x_dims, y_dims);
     PADDLE_ENFORCE_EQ(x_dims.size(), 2,
                       "The rank of Input(X) must be 2 and the shape is "
                       "[batch_size, 1].");
-    PADDLE_ENFORCE_EQ(x_dims[1], 1,
-                      "Each row of Input(X) contains a real value, "
-                      "so the 2nd dimension of Input(X) must be 1.");
+    if (ctx->IsRuntime() ||
+        (framework::product(x_dims) > 0 && framework::product(y_dims) > 0)) {
+      PADDLE_ENFORCE_EQ(x_dims, y_dims, "Shape of X and Y should be same");
+    }
+    if (ctx->IsRuntime()) {
+      PADDLE_ENFORCE_EQ(x_dims[1], 1,
+                        "Each row of Input(X) contains a real value, "
+                        "so the 2nd dimension of Input(X) must be 1.");
+    }
 
     ctx->SetOutputDim("Residual", x_dims);
     ctx->SetOutputDim("Out", {x_dims[0], 1});

paddle/fluid/operators/layer_norm_op.cc

@@ -46,11 +46,18 @@ class LayerNormOp : public framework::OperatorWithKernel {
     int right = static_cast<int>(matrix_dim[1]);
     if (ctx->HasInput("Scale")) {
       PADDLE_ENFORCE_EQ(ctx->GetInputDim("Scale").size(), 1);
-      PADDLE_ENFORCE_EQ(ctx->GetInputDim("Scale")[0], right);
+
+      if (ctx->IsRuntime()) {
+        PADDLE_ENFORCE_EQ(ctx->GetInputDim("Scale")[0], right,
+                          "scale should with right");
+      }
     }
     if (ctx->HasInput("Bias")) {
       PADDLE_ENFORCE_EQ(ctx->GetInputDim("Bias").size(), 1);
-      PADDLE_ENFORCE_EQ(ctx->GetInputDim("Bias")[0], right);
+      if (ctx->IsRuntime()) {
+        PADDLE_ENFORCE_EQ(ctx->GetInputDim("Bias")[0], right,
+                          "bias should with right");
+      }
     }
 
     ctx->SetOutputDim("Y", ctx->GetInputDim("X"));

paddle/fluid/operators/metrics/precision_recall_op.cc

@@ -40,30 +40,40 @@ class PrecisionRecallOp : public framework::OperatorWithKernel {
     auto max_probs_dims = ctx->GetInputDim("MaxProbs");
     auto labels_dims = ctx->GetInputDim("Labels");
 
-    PADDLE_ENFORCE_EQ(max_probs_dims[1], 1,
-                      "Each instance contains one max probability, so the "
-                      "shape of Input(MaxProbs) should be [batch_size, 1].");
-    PADDLE_ENFORCE_EQ(ctx->GetInputDim("Indices"), max_probs_dims,
-                      "The shape of Input(Indices) should be [batch_size, 1].");
-    PADDLE_ENFORCE_EQ(max_probs_dims[0], labels_dims[0],
-                      "The 1st dimension of Input(MaxProbs) and "
-                      "Input(Labels) both are batch_size and the shape should "
-                      "be the same.");
-    PADDLE_ENFORCE_EQ(labels_dims[1], 1,
-                      "The 2nd dimension of Input(Labels) contains instance "
-                      "label and the shape should be equal to 1.");
+    if (ctx->IsRuntime()) {
+      PADDLE_ENFORCE_EQ(max_probs_dims[1], 1,
+                        "Each instance contains one max probability, so the "
+                        "shape of Input(MaxProbs) should be [batch_size, 1].");
+      PADDLE_ENFORCE_EQ(
+          ctx->GetInputDim("Indices"), max_probs_dims,
+          "The shape of Input(Indices) should bes same with max_probs_dims");
+      PADDLE_ENFORCE_EQ(
+          max_probs_dims[0], labels_dims[0],
+          "The 1st dimension of Input(MaxProbs) and "
+          "Input(Labels) both are batch_size and the shape should "
+          "be the same.");
+      PADDLE_ENFORCE_EQ(labels_dims[1], 1,
+                        "The 2nd dimension of Input(Labels) contains instance "
+                        "label and the shape should be equal to 1.");
+    }
     if (ctx->HasInput("Weights")) {
       auto weights_dims = ctx->GetInputDim("Weights");
-      PADDLE_ENFORCE_EQ(weights_dims,
-                        framework::make_ddim({max_probs_dims[0], 1}),
-                        "The shape of Input(Weights) should be "
-                        "[batch_size, 1].");
+
+      if (ctx->IsRuntime()) {
+        PADDLE_ENFORCE_EQ(weights_dims,
+                          framework::make_ddim({max_probs_dims[0], 1}),
+                          "The shape of Input(Weights) should be "
+                          "[batch_size, 1].");
+      }
     }
     if (ctx->HasInput("StatesInfo")) {
       auto states_dims = ctx->GetInputDim("StatesInfo");
-      PADDLE_ENFORCE_EQ(states_dims, framework::make_ddim({cls_num, 4}),
-                        "The shape of Input(StatesInfo) should be "
-                        "[class_number, 4].");
+
+      if (ctx->IsRuntime()) {
+        PADDLE_ENFORCE_EQ(states_dims, framework::make_ddim({cls_num, 4}),
+                          "The shape of Input(StatesInfo) should be "
+                          "[class_number, 4].");
+      }
     }
 
     // Layouts of BatchMetrics and AccumMetrics both are:

paddle/fluid/operators/minus_op.cc

@@ -14,6 +14,7 @@ limitations under the License. */
 
 #include "paddle/fluid/operators/minus_op.h"
 
+#include <memory>
 #include <string>
 #include <vector>
 
@@ -38,9 +39,12 @@ class MinusOp : public framework::OperatorWithKernel {
     auto x_dims = ctx->GetInputDim("X");
     auto y_dims = ctx->GetInputDim("Y");
 
-    PADDLE_ENFORCE_EQ(
-        x_dims, y_dims,
-        "Minus operator must take two tensor with same num of elements");
+    if (ctx->IsRuntime() ||
+        (framework::product(x_dims) > 0 && framework::product(y_dims) > 0)) {
+      PADDLE_ENFORCE_EQ(
+          x_dims, y_dims,
+          "Minus operator must take two tensor with same num of elements");
+    }
     ctx->SetOutputDim("Out", x_dims);
     ctx->ShareLoD("X", /*->*/ "Out");
   }

paddle/fluid/operators/modified_huber_loss_op.cc

@@ -28,9 +28,16 @@ class ModifiedHuberLossOp : public framework::OperatorWithKernel {
     auto x_dims = ctx->GetInputDim("X");
     auto y_dims = ctx->GetInputDim("Y");
 
-    PADDLE_ENFORCE_EQ(x_dims, y_dims, "The shape of X and Y must be the same.");
     PADDLE_ENFORCE_EQ(x_dims.size(), 2, "The tensor rank of X must be 2.");
-    PADDLE_ENFORCE_EQ(x_dims[1], 1, "The 2nd dimension of X must be 1.");
+    if (ctx->IsRuntime() ||
+        (framework::product(x_dims) > 0 && framework::product(y_dims) > 0)) {
+      PADDLE_ENFORCE_EQ(x_dims, y_dims,
+                        "The shape of X and Y must be the same.");
+    }
+
+    if (ctx->IsRuntime()) {
+      PADDLE_ENFORCE_EQ(x_dims[1], 1, "The 2nd dimension of X must be 1.");
+    }
 
     ctx->SetOutputDim("IntermediateVal", x_dims);
     ctx->SetOutputDim("Out", {x_dims[0], 1});
@@ -90,11 +97,13 @@ class ModifiedHuberLossGradOp : public framework::OperatorWithKernel {
     auto intermediate_dims = ctx->GetInputDim("IntermediateVal");
     auto out_grad_dims = ctx->GetInputDim(framework::GradVarName("Out"));
 
-    PADDLE_ENFORCE_EQ(
-        intermediate_dims, x_dims,
-        "The shape of X and intermediate value must be the same.");
-    PADDLE_ENFORCE_EQ(out_grad_dims, x_dims,
-                      "The shape of Input(Out@Grad) and X must be the same.");
+    if (ctx->IsRuntime()) {
+      PADDLE_ENFORCE_EQ(
+          intermediate_dims, x_dims,
+          "The shape of X and intermediate value must be the same.");
+      PADDLE_ENFORCE_EQ(out_grad_dims, x_dims,
+                        "The shape of Input(Out@Grad) and X must be the same.");
+    }
 
     if (ctx->HasOutput(framework::GradVarName("X"))) {
       ctx->SetOutputDim(framework::GradVarName("X"), x_dims);

paddle/fluid/operators/space_to_depth_op.cc

@@ -34,19 +34,44 @@ class SpaceToDepthOp : public framework::OperatorWithKernel {
     auto blocksize = ctx->Attrs().Get<int64_t>("blocksize");
 
     PADDLE_ENFORCE_GT(blocksize, 1, "The blocksize should be Greater than 1");
-    PADDLE_ENFORCE_GT(x_dims[1], 0, "input channel should be Greater than 0");
-    PADDLE_ENFORCE_GT(x_dims[2], 0, "input Height should be Greater than 0");
-    PADDLE_ENFORCE_GT(x_dims[3], 0, "input Width should be Greater than 0");
-
-    PADDLE_ENFORCE_EQ(x_dims[1] % (blocksize * blocksize), 0,
-                      "input channel should be divisible of the square of "
-                      "SpaceToDepthOp blocksize");
-    PADDLE_ENFORCE_EQ(x_dims[2] % (blocksize), 0,
-                      "input Height should be divisible of the square of "
-                      "SpaceToDepthOp blocksize");
-    PADDLE_ENFORCE_EQ(x_dims[3] % (blocksize), 0,
-                      "input Width should be divisible of the square of "
-                      "SpaceToDepthOp blocksize");
+    if (ctx->IsRuntime()) {
+      PADDLE_ENFORCE_GT(x_dims[1], 0, "input channel should be Greater than 0");
+      PADDLE_ENFORCE_GT(x_dims[2], 0, "input Height should be Greater than 0");
+      PADDLE_ENFORCE_GT(x_dims[3], 0, "input Width should be Greater than 0");
+
+      PADDLE_ENFORCE_EQ(x_dims[1] % (blocksize * blocksize), 0,
+                        "input channel should be divisible of the square of "
+                        "SpaceToDepthOp blocksize");
+      PADDLE_ENFORCE_EQ(x_dims[2] % (blocksize), 0,
+                        "input Height should be divisible of the square of "
+                        "SpaceToDepthOp blocksize");
+      PADDLE_ENFORCE_EQ(x_dims[3] % (blocksize), 0,
+                        "input Width should be divisible of the square of "
+                        "SpaceToDepthOp blocksize");
+    } else {
+      if (x_dims[1] != -1) {
+        PADDLE_ENFORCE_GT(x_dims[1], 0,
+                          "input channel should be Greater than 0");
+        PADDLE_ENFORCE_EQ(x_dims[1] % (blocksize * blocksize), 0,
+                          "input channel should be divisible of the square of "
+                          "SpaceToDepthOp blocksize");
+      }
+      if (x_dims[2] != -1) {
+        PADDLE_ENFORCE_GT(x_dims[2], 0,
+                          "input Height should be Greater than 0");
+        PADDLE_ENFORCE_EQ(x_dims[2] % (blocksize), 0,
+                          "input Height should be divisible of the square of "
+                          "SpaceToDepthOp blocksize");
+      }
+
+      if (x_dims[3] != -1) {
+        PADDLE_ENFORCE_GT(x_dims[3], 0, "input Width should be Greater than 0");
+
+        PADDLE_ENFORCE_EQ(x_dims[3] % (blocksize), 0,
+                          "input Width should be divisible of the square of "
+                          "SpaceToDepthOp blocksize");
+      }
+    }
 
     VLOG(3) << "SpaceToDepthOp operator x.shape=" << x_dims
             << "Attribute blocksize" << blocksize << std::endl;

paddle/fluid/operators/tree_conv_op.cc

@@ -62,17 +62,38 @@ class TreeConvOp : public framework::OperatorWithKernel {
     auto edge_dims = ctx->GetInputDim("EdgeSet");
     auto vector_dims = ctx->GetInputDim("NodesVector");
     auto filter_dims = ctx->GetInputDim("Filter");
-    PADDLE_ENFORCE_EQ(edge_dims[2], 2, "Input(EdgeSet) dim[2] should be 2");
+
+    if (ctx->IsRuntime()) {
+      PADDLE_ENFORCE_EQ(edge_dims[2], 2, "Input(EdgeSet) dim[2] should be 2");
+    } else {
+      if (edge_dims[2] != -1) {
+        PADDLE_ENFORCE_EQ(edge_dims[2], 2, "Input(EdgeSet) dim[2] should be 2");
+      }
+    }
     PADDLE_ENFORCE_EQ(edge_dims.size(), 3,
                       "The dimension of EdgeSet Tensor should be 3");
     PADDLE_ENFORCE_EQ(vector_dims.size(), 3,
                       "The dimension of NodesVector Tensor should be 3");
     PADDLE_ENFORCE_EQ(filter_dims.size(), 4,
                       "The dimension of Filter Tensor should be 4");
-    PADDLE_ENFORCE_EQ(filter_dims[1], 3, "Input(Filter) dim[1] should be 3");
-    PADDLE_ENFORCE_EQ(
-        filter_dims[0], vector_dims[2],
-        "Input(Filter) dim[0] must equal to Input(NodesVector) dim[2]");
+
+    if (ctx->IsRuntime()) {
+      PADDLE_ENFORCE_EQ(filter_dims[1], 3, "Input(Filter) dim[1] should be 3");
+      PADDLE_ENFORCE_EQ(
+          filter_dims[0], vector_dims[2],
+          "Input(Filter) dim[0] must equal to Input(NodesVector) dim[2]");
+    } else {
+      if (filter_dims[1] != -1) {
+        PADDLE_ENFORCE_EQ(filter_dims[1], 3,
+                          "Input(Filter) dim[1] should be 3");
+      }
+
+      if (filter_dims[0] != -1 && vector_dims[2] != -1) {
+        PADDLE_ENFORCE_EQ(
+            filter_dims[0], vector_dims[2],
+            "Input(Filter) dim[0] must equal to Input(NodesVector) dim[2]");
+      }
+    }
     auto output_dims = framework::make_ddim(
         {vector_dims[0], vector_dims[1], filter_dims[2], filter_dims[3]});
     ctx->SetOutputDim("Out", output_dims);