Troubleshooting the deep learning module

Deep learning troubleshooting

Introduction

The deep learning module requires specific libraries to be downloaded, which are dependent on the type of computer you are using. Deep learning models must also be PAMGuard compatible.

Deep learning engine failure

You may receive a “deep learning engine failure” message. This is usually because the required libraries have not downloaded correctly. A more detailed explanation of how these libraries are used is available here. When you first use the module, PAMGuard attempts to download the relevant libraries, and if it cannot, you will receive the “Deep learning engine failure” message. Often, downloads are prevented by security settings or a lack of internet connection. One way to circumvent this is to manually download the libraries and place them in the correct folder.

Notes:

• We do not recommend this, but it is a last resort if your system cannot be modified to allow PAMGuard to download the files.
• You must enable your File browser to show hidden files, otherwise the folder will be invisible when you unzip.

For Windows (x86 only)

Download this zip and place the hidden .djl.ai folder in your user folder, e.g., C:\Users\me.

For MacOS (M series only)

Download this zip and place the hidden .djl.ai folder in your user folder, e.g., /Users/me/.

Model metadata failure

This indicates that the model metadata has failed. Each model contains a .pgtf file, which contains JSON-formatted metadata on how to load the model—i.e., what the model needs to run. More information on how this works can be found in the help files. This error usually indicates that the JSON data is incorrect or the .pgdl file is not present at all.

If you have created your own model file for PAMGuard by zipping a .pgtf file with a deep learning model, this is most likely a user error. Check your JSON using a JSON validator, ensure you are following the JSON schema, and compare it against another .pgtf file. For example, here is the JSON for a right whale model by Shiu et al., 2019:

{
  "framework_info": {"framework": "Bespoke"},
  "model_info": {
    "output_shape": [
      -1,
      2
    ],
    "input_shape": [
      -1,
      40,
      40,
      1
    ]
  },
  "class_info": {
    "name_class": [
      "Noise",
      "Right Whale"
    ],
    "num_class": 2
  },
  "transforms": [
    {
      "name": "load_audio",
      "params": {"sr": 2000}
    },
    {
      "name": "spectrogram",
      "params": {
        "fft": 256,
        "hop": 100
      }
    },
    {
      "name": "freq_compression",
      "params": {
        "bins": 40,
        "fmin": 47,
        "fmax": 357
      }
    },
    {
      "name": "norm_row_sum",
      "params": {}
    }
  ],
  "description": "Metadata for acoustic deep learning",
  "version_info": {"version": 1},
  "seg_size": {"size_ms": 2000}
}

If you have used a Koogu, Ketos or AnimalSpot model or downloaded a model file from the PAMGuard website then this is likely a bug in PAMGuard or support issue. Please contact us with a bug report.

Model load failure

This indicates a more general error loading the model. Please contact us with a bug report.

The input file is null

Most likely you have imported a configuration or moved the model file and PAMGuard can no longer find the file.