ERR_MMEORY_OVERLAP

[glebss]

Hello.

I am trying to compile and run my custom model. I use custom-model-onnx.ipynb notebook as a basis.
It compiles fine to onnx format. After several modifications, all layers are now supported by TIDL. Nevertheless, when I do

img = np.zeros((1, 3, 608, 960), dtype='float32')
output = list(sess.run(['det_out', 'laneaf_out'],
                           input_feed={'image': img}))

in the error log I have

Error : Error Code = <ERR_MMEORY_OVERLAP>
Segmentation fault (core dumped)

By commenting in and out some parts of my code I figured out that something illegal happens in this code snippet:

xs_out = torch.narrow(y, 2, 12, 4) + torch.cat([torch.narrow(grid, 2, 0, 1)] * 4, dim=2)
res_out = torch.cat([bboxes_out, xs_out], 2)

where y is a torch.Tensor of size (1, 3, 24, 9120), grid is a torch.Tensor of shape (1, 3, 2, 9120). Therefore, here two tensors of the exact same shape (1, 3, 4, 9120) are added together.

However, when I skip addition and leave only

xs_out = torch.narrow(y, 2, 12, 4)
res_out = torch.cat([bboxes_out, xs_out], 2)

it compiles and runs without errors.

My compilation parameters are the following

compile_options = {
        'tidl_tools_path' : os.environ['TIDL_TOOLS_PATH'],
        'artifacts_folder' : output_dir,
        'tensor_bits' :  8,
        'accuracy_level' : 0,
        'advanced_options:calibration_frames' : len(calib_images), 
        'advanced_options:calibration_iterations' : 0, # used if accuracy_level = 1
        'debug_level' :  2
    }

Have no idea how to fix it myself. Is there any way to debug it?

Thanks.