Merge pull request #436 from guoruoqian/replication_padXd_fix_bug

fix bug of replication_pad when input is dynamic

Author: narendasan

core/conversion/converters/impl/replication_pad.cpp

@@ -67,6 +67,9 @@ bool replication_padXd(ConversionCtx* ctx, const torch::jit::Node* n, args& args
         at::Tensor dimValue = torch::tensor({axis}, torch::kInt32);
         auto dimTensor = tensor_to_const(ctx, dimValue);
         indicesTensor = ctx->net->addGather(*shapeTensor, *dimTensor, 0)->getOutput(0);
+        auto oneTensor = tensor_to_const(ctx, torch::tensor({1}, torch::kInt32));
+        indicesTensor =
+            ctx->net->addElementWise(*indicesTensor, *oneTensor, nvinfer1::ElementWiseOperation::kSUB)->getOutput(0);
       } else {
         auto indices = torch::tensor({inDims.d[axis] - 1}, torch::kInt32);
         indicesTensor = tensor_to_const(ctx, indices);