Program crashes with constant tensors

[gbenfis]

My program crashes when loading the model from the json file. Upon debugging, it crashes when looking for a keras_history in a node, and not finding any attribute with that name:

Assertion failed: it != m_data.m_value.object->end(), file C:\Users\e1081782\dependency\json\single_include\nlohmann\json.hpp, line 21449

This is most probably linked to the fact that my model normalizes the input tensor by subtracting a constant tensor (which has no history). Is there any workaround to make it execute?

    {
      "module": "keras.src.ops.numpy",
      "class_name": "Subtract",
      "config": { "name": "subtract_2" },
      "registered_name": "Subtract",
      "name": "subtract_2",
      "inbound_nodes": [
        {
          "args": [
            {
              "class_name": "__keras_tensor__",
              "config": {
                "shape": [ null, 11 ],
                "dtype": "float32",
                "keras_history": [ "input_1", 0, 0 ]
              }
            },
            {
              "class_name": "__tensor__",
              "config": {
                "value": [ [ 0.014991031028330326, -5.773593159119628e-08, 0.3049706220626831, 0.9997414946556091, 0.04997607693076134, 1.0001437664031982, 0.9999415874481201, 1.0001988410949707, 1.0000635385513306, 2.4990406036376953, 0.4999502897262573 ] ],
                "dtype": "float32"
              }
            }
          ],
          "kwargs": {}
        }
      ]
    },

[Dobiasd]

Thanks for reporting!

Could you upload your model (.keras format), so I can reproduce the error locally and debug it?

[gbenfis]

fdeep_model.json

It doesn't let me upload .keras files here. I've uploaded the json file in case it can help!

[Dobiasd]

When testing with this, I get a different error message.

It would be great if you could find a way to upload the .keras model too, because I'd like to check the convert_model step too. Maybe you can put it on Google Drive or Dropbox and post the link here?

[gbenfis]

Sure, here it is: https://drive.google.com/file/d/1bZPpGPowOxmEW1qH16-7hV5Xgkh8ty0L/view?usp=drivesdk

Thanks!

[Dobiasd]

Thanks. With this, I can reproduce the problem locally too. 👍

Can you create a minimal Python source code that builds a model running into that problem? (no history, likely from the constant tensor)

[gbenfis]

Sure, here it is. I think the problem boils down to the normalization layer.

    import tensorflow as tf
    import numpy as np
    
    x = np.array([[0.1, 0.2]])
    inputs = tf.keras.Input(shape=(2,), name="input_1")
    input_norm_layer = tf.keras.layers.Normalization()
    input_norm_layer.adapt(x)
    outputs = (inputs - input_norm_layer.mean) / tf.sqrt(input_norm_layer.variance)
    model = tf.keras.Model(inputs=inputs, outputs=outputs)
    model.save('test.keras')

[Dobiasd]

Thanks a lot. This works nicely to reproduce the problem. I'll see if I can fix it and get back to you here. 🧑‍🔧

[gbenfis]

Hi,

So I found a way around this problem by avoiding doing the normalization manually as I did above, but letting the normalization layer handle it:

    import tensorflow as tf
    import numpy as np
    
    x = np.array([[0.1, 0.2], [0.5, 0.7]])
    y = np.array([[0.1, 0.15], [0.2, 0.35]])
    inputs = tf.keras.Input(shape=(2,), name="input_1")
    input_norm_layer = tf.keras.layers.Normalization()
    input_norm_layer.adapt(x)
    outputs = input_norm_layer(inputs)
    output_norm_layer = tf.keras.layers.Normalization(invert=True)
    output_norm_layer.adapt(y)
    outputs = output_norm_layer(outputs)
    model = tf.keras.Model(inputs=inputs, outputs=outputs)
    model.save('test.keras')

Now the previous error disappears but I get this new one (which seems linked to the inverse normalization layer):

FAILURE test failed: output=0 pos=0,0,0 value=-45.8373 target=0.0429067 

Any idea why?

[Dobiasd]

Awesome that you found a workaround! Implementing support for constant-value layers in frugally-deep is non-trivial. Having you unblocked by this solution takes away the pressure. 😌

Your current problem is way simpler. Frugally-deep does not yet support invert=True for Normalization layers. I'll add this support and get back to you.

[gbenfis]

Sounds great, thank you!

[Dobiasd]

Ok, support for Normalization layers with invert=True is implemented now and available in the new version (0.18.1). ✅
Please let me know if it works for your particular problem too.

And thanks for these great reports, helping to make frugally-deep a bit better. ❤

I'll keep this issue open nonetheless, to have a public todo item for implementing support for constant-value layers. 🧑‍💼