In a long-running process invoking `nsfw_detector.predict.classify_nd()` often, we observed memory leakage consistent with what [keras-team/keras#13118][13118] suggests can be avoided through predicting via `model()` instead of `model.predict()` when combined with some memory management following invocation:

```python
import gc
import tensorflow.keras as keras
from nsfw_detector import predict
# …
predict.classify_nd(model, image)
gc.collect()
keras.backend.clear_session()
```

An alternative suggestion was to use `model.predict_on_batch(x)` instead of `model()` but we didn't try that because using `model()` solved our memory leak problem adequately.

If this fix limits a use case, an alternative implementation for this fix may be to allow the caller to specify what of the three methods to use or to pass a callable that will receive the `nd_images` parameter.

[13118]: keras-team/keras#13118