Merge pull request #16891 from tink2123/cherry-pick

[Cherry pick] modified infer shape

paddle/fluid/operators/affine_channel_op.cc

@@ -79,9 +79,13 @@ class AffineChannelOp : public framework::OperatorWithKernel {
                            : x_dims[x_dims.size() - 1]);
 
     PADDLE_ENFORCE_EQ(scale_dims.size(), 1UL);
-    PADDLE_ENFORCE_EQ(scale_dims[0], C);
     PADDLE_ENFORCE_EQ(b_dims.size(), 1UL);
-    PADDLE_ENFORCE_EQ(b_dims[0], C);
+    if (ctx->IsRuntime() || scale_dims[0] > 0) {
+      PADDLE_ENFORCE_EQ(scale_dims[0], C);
+    }
+    if (ctx->IsRuntime() || b_dims[0] > 0) {
+      PADDLE_ENFORCE_EQ(b_dims[0], C);
+    }
 
     ctx->SetOutputDim("Out", ctx->GetInputDim("X"));
     ctx->ShareLoD("X", "Out");

paddle/fluid/operators/conv_op.cc

@@ -68,9 +68,14 @@ void ConvOp::InferShape(framework::InferShapeContext* ctx) const {
 
   std::vector<int64_t> output_shape({in_dims[0], filter_dims[0]});
   for (size_t i = 0; i < strides.size(); ++i) {
-    output_shape.push_back(ConvOutputSize(in_dims[i + 2], filter_dims[i + 2],
-                                          dilations[i], paddings[i],
-                                          strides[i]));
+    if ((!ctx->IsRuntime()) &&
+        (in_dims[i + 2] <= 0 || filter_dims[i + 2] <= 0)) {
+      output_shape.push_back(-1);
+    } else {
+      output_shape.push_back(ConvOutputSize(in_dims[i + 2], filter_dims[i + 2],
+                                            dilations[i], paddings[i],
+                                            strides[i]));
+    }
   }
   ctx->SetOutputDim("Output", framework::make_ddim(output_shape));
   ctx->ShareLoD("Input", "Output");

paddle/fluid/operators/detection_map_op.cc

@@ -51,8 +51,10 @@ class DetectionMAPOp : public framework::OperatorWithKernel {
     PADDLE_ENFORCE_EQ(label_dims.size(), 2,
                       "The rank of Input(Label) must be 2, "
                       "the shape is [N, 6].");
-    PADDLE_ENFORCE(label_dims[1] == 6 || label_dims[1] == 5,
-                   "The shape of Input(Label) is [N, 6] or [N, 5].");
+    if (ctx->IsRuntime() || label_dims[1] > 0) {
+      PADDLE_ENFORCE(label_dims[1] == 6 || label_dims[1] == 5,
+                     "The shape of Input(Label) is [N, 6] or [N, 5].");
+    }
 
     if (ctx->HasInput("PosCount")) {
       PADDLE_ENFORCE(ctx->HasInput("TruePos"),

paddle/fluid/operators/row_conv_op.cc

@@ -41,9 +41,12 @@ class RowConvOp : public framework::OperatorWithKernel {
     auto filter_dims = ctx->GetInputDim("Filter");
     PADDLE_ENFORCE_EQ(x_dims.size(), 2, "Input(X)'s rank should be 2.");
     PADDLE_ENFORCE_EQ(filter_dims.size(), 2, "Input(Y)'s rank should be 2.");
-    PADDLE_ENFORCE_EQ(
-        x_dims[1], filter_dims[1],
-        "The 2nd dimension of Input(X) and Input(Filter) should be same.");
+    if (ctx->IsRuntime() || (x_dims[1] > 0 && filter_dims[1] > 0)) {
+      PADDLE_ENFORCE_EQ(
+          x_dims[1], filter_dims[1],
+          "The 2nd dimension of Input(X) and Input(Filter) should be same.");
+    }
+
     ctx->SetOutputDim("Out", x_dims);
     ctx->ShareLoD("X", "Out");
   }

paddle/fluid/operators/unpool_op.cc

@@ -99,10 +99,15 @@ class UnpoolOp : public framework::OperatorWithKernel {
     PADDLE_ENFORCE(in_x_dims.size() == 4,
                    "Unpooling intput must be of 4-dimensional.");
     PADDLE_ENFORCE_EQ(in_x_dims, in_y_dims);
+
     std::vector<int64_t> output_shape({in_x_dims[0], in_x_dims[1]});
     for (size_t i = 0; i < ksize.size(); ++i) {
-      output_shape.push_back(UnpoolOutputSize(in_x_dims[i + 2], ksize[i],
-                                              paddings[i], strides[i]));
+      if (!ctx->IsRuntime() && in_x_dims[i + 2] <= 0) {
+        output_shape.push_back(-1);
+      } else {
+        output_shape.push_back(UnpoolOutputSize(in_x_dims[i + 2], ksize[i],
+                                                paddings[i], strides[i]));
+      }
     }
     ctx->SetOutputDim("Out", framework::make_ddim(output_shape));
   }