Merge pull request #16853 from SunGaofeng/affine_modify

modify infer shape pad_op, pad_constant_like_op, no need in psroi_pool_op, crop_op

paddle/fluid/operators/pad_constant_like_op.cc

@@ -13,6 +13,7 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/fluid/operators/pad_constant_like_op.h"
+#include <memory>
 
 namespace paddle {
 namespace operators {
@@ -38,8 +39,16 @@ class PadConstantLikeOp : public framework::OperatorWithKernel {
                       "The dimention of X and Y should be the same.");
 
     for (int i = 0; i < x_dim.size(); ++i) {
-      PADDLE_ENFORCE_GE(x_dim[i], y_dim[i]);
+      if ((!ctx->IsRuntime()) && ((x_dim[i] == -1) || (y_dim[i] == -1))) {
+        continue;
+      } else {
+        PADDLE_ENFORCE_GE(
+            x_dim[i], y_dim[i],
+            "expected X_dim[i] >= Y_dim[i], but received %d < %d for dim %d",
+            x_dim[i], y_dim[i], i);
+      }
     }
+
     ctx->SetOutputDim("Out", x_dim);
     ctx->ShareLoD("X", /*->*/ "Out");
   }
@@ -162,7 +171,14 @@ class PadConstantLikeOpGrad : public framework::OperatorWithKernel {
       ctx->ShareLoD("Y", /*->*/ y_grad_name);
 
       for (int i = 0; i < y_dim.size(); ++i) {
-        PADDLE_ENFORCE_GE(dout_dim[i], y_dim[i]);
+        if ((!ctx->IsRuntime()) && ((dout_dim[i] == -1) || (y_dim[i] == -1))) {
+          continue;
+        } else {
+          PADDLE_ENFORCE_GE(dout_dim[i], y_dim[i],
+                            "expected Out_dim[i] >= Y_dim[i], but received %d "
+                            "< %d for dim %d",
+                            dout_dim[i], y_dim[i], i);
+        }
       }
     }
   }

paddle/fluid/operators/pad_op.cc

@@ -34,9 +34,16 @@ class PadOp : public framework::OperatorWithKernel {
     PADDLE_ENFORCE_EQ(x_dim.size() * 2, int64_t(paddings.size()),
                       "Size of paddings should be equal to 2 * dimension size "
                       "of input tensor.");
+    for (size_t i = 0; i < paddings.size(); ++i) {
+      PADDLE_ENFORCE_GE(paddings[i], 0, "paddings should >= 0.");
+    }
     std::vector<int64_t> out_dims(x_dim.size());
     for (int i = 0; i < x_dim.size(); ++i) {
-      out_dims[i] = x_dim[i] + paddings[i * 2] + paddings[i * 2 + 1];
+      if ((!ctx->IsRuntime()) && (x_dim[i] == -1)) {
+        out_dims[i] = -1;
+      } else {
+        out_dims[i] = x_dim[i] + paddings[i * 2] + paddings[i * 2 + 1];
+      }
     }
     ctx->SetOutputDim("Out", framework::make_ddim(out_dims));
     if (out_dims[0] == x_dim[0]) {
@@ -100,18 +107,14 @@ class PadOpGrad : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
   void InferShape(framework::InferShapeContext* ctx) const override {
-    auto dout_dims = ctx->GetInputDim(framework::GradVarName("Out"));
-    auto& paddings = ctx->Attrs().Get<std::vector<int>>("paddings");
-    for (int i = 0; i < dout_dims.size(); ++i) {
-      dout_dims[i] -= (paddings[i * 2] + paddings[i * 2 + 1]);
-    }
-
     auto x_grad_name = framework::GradVarName("X");
     if (ctx->HasOutput(x_grad_name)) {
       auto dout_dims = ctx->GetInputDim(framework::GradVarName("Out"));
       auto& paddings = ctx->Attrs().Get<std::vector<int>>("paddings");
       for (int i = 0; i < dout_dims.size(); ++i) {
-        dout_dims[i] -= (paddings[i * 2] + paddings[i * 2 + 1]);
+        if (ctx->IsRuntime() || (dout_dims[i] != -1)) {
+          dout_dims[i] -= (paddings[i * 2] + paddings[i * 2 + 1]);
+        }
       }
       ctx->SetOutputDim(x_grad_name, dout_dims);
     }