Bug in `data_adapter.deconfigure` on Mac OS with MPS devices using Torch backend

[arrjon]

Bug regarding the new dev-branch: When using the Torch backend on Mac OS with MPS devices, a TypeError is raised during the execution of the deconfigure function in data_adapters/concatenate_keys_data_adapter.py:

TypeError: can't convert mps:0 device type tensor to numpy. Use Tensor.cpu() to copy the tensor to host memory first.

Adding the following code to the deconfigure method resolves the issue for me:

        processed_data = processed_data.cpu()

While this fix works for the Torch backend on MPS devices, I am uncertain whether it is backend-agnostic enough for other libraries (e.g., JAX, TensorFlow). Any ideas how to fix this?

[LarsKue]

Looks like the data adapter deconfigure() is being called in an unintended way: with tensors instead of numpy arrays.

The fix would be to find this call and convert to numpy first, e.g. with

data = keras.tree.map_structure(keras.ops.convert_to_numpy, data)

Can you please share the code that led you to this error?

[arrjon]

Unfortunately, I cannot share a full working example now. But the error was invoked by calling the sample function after model training.

test_data = simulator.sample((10,))

conditions = {
        "obs":  test_data["obs"],
 }

samples = approximator.sample(conditions=conditions, num_samples=100)

However, your proposed fix solves the issue for me.

[arrjon]

I just noticed, that I cannot pass the output of the simulator test_data directly to the sample function, but I have to drop the entries related to the parameters first.

File ~site-packages/bayesflow/approximators/continuous_approximator.py:145, in ContinuousApproximator.sample(self, conditions, num_samples, numpy, batch_shape)
    142 conditions = self.data_adapter.configure(conditions)
    143 conditions = keras.tree.map_structure(keras.ops.convert_to_tensor, conditions)
    144 conditions = {
--> 145     "inference_variables": self._sample(num_samples=num_samples, batch_shape=batch_shape, **conditions)
    146 }
    147 conditions = keras.tree.map_structure(keras.ops.convert_to_numpy, conditions)
    148 conditions = self.data_adapter.deconfigure(conditions)

TypeError: ContinuousApproximator._sample() got an unexpected keyword argument 'inference_variables'

This is somewhat unintuitive and probably also an easy fix.

[paul-buerkner]

Thank you for this point. We will discuss it this week and see what we can do to fix/improve it.