Skip to content
  • Guides
  • CAREamist API
  • FAQ

Frequently asked questions#

Training#

NotImplementedError: The operator [...] is not currently implemented for the MPS device#

Unfortunately, not all the operations in PyTorch are implemented for the Metal architecture of recent Apple devices (MPS stands for Metal Performance Shaders). This error occurs when you are trying to run CAREamics on such a device without forcing training on CPU.

NotImplementedError: The operator 'aten::max_pool3d_with_indices' is not currently implemented for the MPS device. If you want this op to be added in priority during the prototype phase of this feature, please comment on https://github.com/pytorch/pytorch/issues/77764. As a temporary fix, you can set the environment variable `PYTORCH_ENABLE_MPS_FALLBACK=1` to use the CPU as a fallback for this op. WARNING: this will be slower than running natively on MPS.

Solution

In your environment (e.g. terminal before you start the jupyter notebook), set the environment variable PYTORCH_ENABLE_MPS_FALLBACK=1. This will force the training to run on the CPU, which is slower but will work.

This can be done for instance by running the following command in your terminal:

export PYTORCH_ENABLE_MPS_FALLBACK=1

Prediction#

Sizes of tensors must match#

If you get an error similar to:

RuntimeError: Sizes of tensors must match except in dimension 1. Expected size 320 but got size 321 for tensor number 1 in the list.

This is likely because the input data size is not compatible with the model (e.g. the model has layers that downsample the data, then upsample it, and this is only compatible with dimensions that are power of 2!).

Solution

By tiling the prediction, you can ensure that every input to the model has the correct shape. Then upon stitching back the tiles, you recover your whole image!

BMZ model export#

pydantic_core._pydantic_core.ValidationError#

The BMZ format is also validated by the pydantic library. If some of the metadata given to the export_to_bmz function is not correct, you might get such errors:

Code
train_data = np.random.randint(0, 255, (256, 256)).
careamist.export_to_bmz( 
    path="n2v_models", 
    name="n2v_model_example.bmz", 
    input_array=train_data, 
    authors=[{ "name": "nobody", "email": "nobody@nobody", }]
)
Error message
pydantic_core._pydantic_core.ValidationError: 2 validation errors for bioimage.io model specification
name
  Value error, 'n2v_model_example.bmz' is not restricted to 'abcdefghijklmnopqrstuvwxyzABCDEFGHIJKLMNOPQRSTUVWXYZ0123456789_- ()' [type=value_error, input_value='n2v_model_example.bmz', input_type=str]
    For further information visit https://errors.pydantic.dev/2.7/v/value_error
authors.0.email
  value is not a valid email address: The part after the @-sign is not valid. It should have a period. [type=value_error, input_value='nobody@nobody', input_type=str]

If we look closely at the error message, Pydantic is telling us that there are actually two errors: - name should only contain letters, numbers, hyphens, underscores, spaces and parentheses. - authors.0.email is not a valid email address, as it is missing a period after the @-sign.

Solution

Look at the input name, it contains a forbidden character: the dot. The email should have something like .com after the @ sign. Fixing these two issues will solve the problem.