Merge pull request #1099 from pytorch/max_fix

fix: Fix keep_dims functionality for aten::max

Author: peri044

core/conversion/converters/impl/max.cpp

@@ -18,17 +18,36 @@ auto max_registrations TORCHTRT_UNUSED = RegisterNodeConversionPatterns().patter
      [](ConversionCtx* ctx, const torch::jit::Node* n, args& args) -> bool {
        auto self = args[0].ITensorOrFreeze(ctx);
        auto dim = args[1].unwrapToInt();
+       auto keep_dims = args[2].unwrapToBool();
        auto selfDim = util::toVec(self->getDimensions());
        if (dim < 0) {
          dim = selfDim.size() + dim;
        }
        uint32_t shiftDim = 1 << dim;
        auto TopKOperation = nvinfer1::TopKOperation::kMAX;
-       auto new_layer = ctx->net->addTopK(*self, TopKOperation, 1, shiftDim);
-       TORCHTRT_CHECK(new_layer, "Unable to create max layer from node: " << *n);
+       auto topk_layer = ctx->net->addTopK(*self, TopKOperation, 1, shiftDim);
+       TORCHTRT_CHECK(topk_layer, "Unable to create max layer from node: " << *n);
+       auto topk_dims = util::toVec(topk_layer->getOutput(0)->getDimensions());
 
-       auto out0 = ctx->AssociateValueAndTensor(n->outputs()[0], new_layer->getOutput(0));
-       auto out1 = ctx->AssociateValueAndTensor(n->outputs()[1], new_layer->getOutput(1));
+       nvinfer1::ITensor* out0;
+       nvinfer1::ITensor* out1;
+       if (!keep_dims) {
+         if (topk_dims[dim] == 1) {
+           auto squeeze_layer = ctx->net->addShuffle(*topk_layer->getOutput(0));
+           squeeze_layer->setReshapeDimensions(util::squeezeDims(topk_layer->getOutput(0)->getDimensions(), dim));
+           TORCHTRT_CHECK(squeeze_layer, "Unable to create squeeze_layer layer from node: " << *n);
+           out0 = ctx->AssociateValueAndTensor(n->outputs()[0], squeeze_layer->getOutput(0));
+
+           auto squeeze_layer_indices = ctx->net->addShuffle(*topk_layer->getOutput(1));
+           squeeze_layer_indices->setReshapeDimensions(
+               util::squeezeDims(topk_layer->getOutput(1)->getDimensions(), dim));
+           TORCHTRT_CHECK(squeeze_layer_indices, "Unable to create squeeze_layer_indices layer from node: " << *n);
+           out1 = ctx->AssociateValueAndTensor(n->outputs()[1], squeeze_layer_indices->getOutput(0));
+         }
+       } else {
+         out0 = ctx->AssociateValueAndTensor(n->outputs()[0], topk_layer->getOutput(0));
+         out1 = ctx->AssociateValueAndTensor(n->outputs()[1], topk_layer->getOutput(1));
+       }
 
        LOG_DEBUG("Output tensor(0) shape: " << out0->getDimensions());
        LOG_DEBUG("Output tensor(1) shape: " << out1->getDimensions());

tests/core/conversion/converters/test_reduce.cpp

@@ -212,6 +212,18 @@ TEST(Converters, ATenProdKeepDimsConvertsCorrectly) {
   test_body(graph, in);
 }
 
+TEST(Converters, ATenMaxKeepDimsConvertsCorrectly) {
+  const auto graph = R"IR(
+    graph(%x : Tensor):
+          %2 : int = prim::Constant[value=-1]()
+          %3 : bool = prim::Constant[value=1]()
+          %keep.1 : Tensor, %6 : Tensor = aten::max(%x, %2, %3)
+          return (%keep.1, %6))IR";
+
+  auto in = at::randint(-5, 5, {4, 4}, at::kCUDA);
+  test_body(graph, in);
+}
+
 TEST(Converters, ATenMeanDimNegOneIndexConvertsCorrectly) {
   const auto graph = R"IR(
     graph(%0 : Tensor):