Forums - Getting segmentation fault with my converted dlc model

I'm trying to get the input and output tensor names on my converted DLC model on an Android phone (just to make sure my model is working correctly), but I keep getting 'Segmentation fault' errors no matter what. I tried using other models already converted by another party that made the platform I'm working on and it works fine, it's just my converted model that's not working. I converted my keras TensorFlow model to a frozen graph (.pb), then converted that to DLC using snpe-tensorflow-to-dlc successfully. But when I load it with loadContainerFromFile() and run this code on it:

std::unique_ptr<zdl::DlContainer::IDlContainer> container = loadContainerFromFile("/data/openpilot/selfdrive/df/new_dlc.dlc");
std::unique_ptr<zdl::SNPE::SNPE> snpe = setBuilderOptions(container, runtime);
const auto &strListi_opt = snpe->getInputTensorNames();
const auto &strListo_opt = snpe->getOutputTensorNames();

const auto &strListi = *strListi_opt;
//assert(strListi.size() == 1);
const char *input_tensor_name = strListi.at(0);

const auto &strListo = *strListo_opt;
assert(strListo.size() == 1);
const char *output_tensor_name = strListo.at(0);

I get a segmentation fault at line getInputTensorNames. Again, this all works with other models I've tested made by other people, but not mine. Is there anything I could have done wrong converting it to a dlc file? To go from keras h5 to pb, I used the freeze_session function found online along with:

frozen_graph = freeze_session(K.get_session(),

output_names=[out.op.name for out in model.outputs])

 

tf.train.write_graph(frozen_graph, "pb_models", model_name+".pb", as_text=False)

to convert to a pb file. Then I used this to get the input and output node names to use in the conversion command for SNPE:

model = load_model("h5_models/"+model_name+".h5")

print([out.op.name for out in model.inputs])

print([out.op.name for out in model.outputs])

My keras TF model takes an array of 5 floats and returns one single float from -1 to 1. That's it, it doesn't need to process any images or anything, just that one 5 number array. The first layer is a dense layer with 5 nodes, and the last layer is also a dense layer with 1 node. For the input shape on the snpe conversion command, I used "1,5"

Here's the full command I'm using:

python snpe-tensorflow-to-dlc --graph /home/shane/model.pb --input_dim dense_1_input "1,5" --out_node "dense_5/BiasAdd" --dlc /home/shane/test.dlc

The SNPE version I used to convert to a dlc file was 1.25.1, and the version I'm running the model on is 1.19.2.0.

If you notice anything out of the ordinary that might be causing SNPE to not work, please let me know. I've been working on this for a week and a half and for the life of me can't figure anything out. Thank you very much.