Add missing reset for extra_loss_arg tensor list to preserve reference to user data.

src/csrc/training.cpp

@@ -178,6 +178,8 @@ void train_multiarg(const char* name, torchfort_tensor_list_t inputs_in, torchfo
 
   inputs->reset();
   labels->reset();
+  if (extra_loss_args)
+    extra_loss_args->reset();
 
   torchfort::nvtx::rangePop();
 }

tests/supervised/scripts/setup_tests.py

@@ -36,22 +36,22 @@ def __init__(self):
     super(Loss1, self).__init__()
 
   def forward(self, prediction, label):
-    return (torch.sum(prediction) + torch.sum(label)) / (2 * prediction.numel())
+    return (torch.sum(prediction) + torch.sum(label))
 
 class Loss2(torch.nn.Module):
   def __init__(self):
     super(Loss2, self).__init__()
 
   def forward(self, prediction1, prediction2, label1, label2):
-    return (torch.sum(prediction1) + torch.sum(prediction2) + torch.sum(label1) + torch.sum(label2)) / (4 * prediction1.numel())
+    return (torch.sum(prediction1) + torch.sum(prediction2) + torch.sum(label1) + torch.sum(label2))
 
 class Loss2Extra(torch.nn.Module):
   def __init__(self):
     super(Loss2Extra, self).__init__()
 
   def forward(self, prediction1, prediction2, label1, label2, extra_args1, extra_args2):
     return (torch.sum(prediction1) + torch.sum(prediction2) + torch.sum(label1) + torch.sum(label2) +
-            torch.sum(extra_args1) + torch.sum(extra_args2)) / (6 * prediction1.numel())
+            torch.sum(extra_args1) + torch.sum(extra_args2))
 
 def main():
   model1 = Net1()

tests/supervised/test_training.cpp

@@ -220,26 +220,31 @@ void training_test_multiarg(const std::string& model_config, int dev_model, int
     }
 
     // Check that external data changes reflect in tensor list
+    float expected_loss_val = 0.0;
     for (int i = 0; i < 2; ++i) {
-      auto tmp = generate_random<float>(shape);
+      auto tmp = generate_constant<float>(shape, 1);
       inputs[i].assign(tmp.begin(), tmp.end());
+      labels[i].assign(tmp.begin(), tmp.end());
+      expected_loss_val += 2 * 1.0f * tmp.size();
+      if (use_extra_args) {
+        extra_args[i].assign(tmp.begin(), tmp.end());
+        expected_loss_val += 1.0f * tmp.size();
+      }
 #ifdef ENABLE_GPU
       if (dev_input != TORCHFORT_DEVICE_CPU) {
         copy_from_host_vector(input_ptrs[i], inputs[i]);
+        copy_from_host_vector(label_ptrs[i], labels[i]);
+        if (use_extra_args) {
+          copy_from_host_vector(extra_args_ptrs[i], extra_args[i]);
+        }
       }
 #endif
     }
 
-    CHECK_TORCHFORT(torchfort_inference_multiarg(model_name.c_str(), inputs_tl, outputs_tl, 0));
+    CHECK_TORCHFORT(torchfort_train_multiarg(model_name.c_str(), inputs_tl, labels_tl, &loss_val,
+                                             (use_extra_args) ? extra_args_tl : nullptr, 0));
 
-    for (int i = 0; i < 2; ++i) {
-#ifdef ENABLE_GPU
-      if (dev_input != TORCHFORT_DEVICE_CPU) {
-        copy_to_host_vector(outputs[i], output_ptrs[i]);
-      }
-#endif
-      EXPECT_EQ(inputs[i], outputs[i]);
-    }
+    EXPECT_EQ(loss_val, expected_loss_val);
   }
 
   for (int i = 0; i < 2; ++i) {

tests/test_utils.h

@@ -45,6 +45,15 @@ template <typename T> std::vector<T> generate_random(const std::vector<int64_t>&
   return data;
 }
 
+// Generate constant vector data for testing
+template <typename T> std::vector<T> generate_constant(const std::vector<int64_t>& shape, T value) {
+
+  int64_t num_values = std::accumulate(shape.begin(), shape.end(), 1, std::multiplies<int64_t>());
+  std::vector<T> data(num_values, value);
+
+  return data;
+}
+
 // Generate random names to use as model keys to avoid conflicts between tests
 std::string generate_random_name(int length) {