R 中的 mlpack 导致致命错误

Question

我在 R 中使用 mlpack package 进行字典学习。 我已经使用下面的代码进行了学习：

mnist <- keras::dataset_mnist()

N_train <- 300
train_set <- mnist$train$x[sample(1:nrow(mnist$train$x), size = N_train), , ]

n_samples <- dim[1]
i_width <- dim[2]
i_height <- dim[3]

train_resh <- keras::array_reshape(train_set, dim = c(n_samples, i_width * i_height))

dictionary <- readRDS("dictionary.RDS") # a dictionary trained before and stored
dictionary_dim <- dim(dictionary)

N_MODEL_ATOMS <- 50

init_dictionary <- dictionary

# The main function:
sparse_cod_out <- mlpack::sparse_coding(atoms = N_MODEL_ATOMS,
                                          training = train_resh,
                                          verbose = TRUE,
                                          initial_dictionary = init_dictionary)
saveRDS(sparse_cod_out$dictionary, file = "dictionary.RDS")

# Store the model for future purposes as the learning can take a lot of time
mlpack::Serialize(sparse_cod_out$output_model, "model_ptr")

I use mlpack::Serialize to save the output model - the model is externalptr object, so it references memory address.

当我只想 model 新实例时，我会跳过训练步骤并使用以下代码：

test_data <- mnist$test$x[1, ,]
test_data_resh <- keras::array_reshape(test_data, dim = c(1, i_width * i_height))

codes <- sparse_coding(
  input_model = mlpack::Unserialize("model_ptr"), # load the model stroed previously
  test = test_data_resh
)

recovered_signal <- codes$codes %*% codes$dictionary

它工作正常，但如果我想多次运行上述代码 R 会因致命错误而崩溃。 It happens for R 4.0.3 and RStudio 1.3 as well as R 4.0.5, RStudio 1.4, both Windows 10 Pro. 我还在 Ubuntu 上运行了 RStudio docker 容器（摇杆/诗歌图像）中的代码，它是一样的。

序列化/反序列化的许多不同组合 - 无效。

我需要帮助来摆脱这些致命错误。