crop: also support splitting double pages

[bertsky]

builds on OCR-D/core#1344

simply run with 2 output file groups, e.g. ocrd-anybaseocr-crop -I OCR-D-IMG -O OCR-D-CROP-L,OCR-D-CROP-R

[bertsky]

builds on OCR-D/core#1344

hence the CI's failure

TypeError: 'OcrdPageResult' object is not iterable

(we have to wait for this to be merged and released on PyPI)

But what has me worried is the failure of the layout-analysis test:

File "/home/circleci/repo/src/ocrd_anybaseocr/cli/ocrd_anybaseocr_layout_analysis.py", line 116, in _predict
    pred = self.model.predict(img_array)
  File "/home/circleci/.pyenv/versions/3.8.20/lib/python3.8/site-packages/keras/engine/training_v1.py", line 1059, in predict
    return func.predict(
  File "/home/circleci/.pyenv/versions/3.8.20/lib/python3.8/site-packages/keras/engine/training_arrays_v1.py", line 801, in predict
    return predict_loop(
  File "/home/circleci/.pyenv/versions/3.8.20/lib/python3.8/site-packages/keras/engine/training_arrays_v1.py", line 192, in model_iteration
    f = _make_execution_function(model, mode)
  File "/home/circleci/.pyenv/versions/3.8.20/lib/python3.8/site-packages/keras/engine/training_arrays_v1.py", line 620, in _make_execution_function
    return model._make_execution_function(mode)
  File "/home/circleci/.pyenv/versions/3.8.20/lib/python3.8/site-packages/keras/engine/training_v1.py", line 2372, in _make_execution_function
    self._make_predict_function()
  File "/home/circleci/.pyenv/versions/3.8.20/lib/python3.8/site-packages/keras/engine/training_v1.py", line 2356, in _make_predict_function
    self.predict_function = backend.function(
  File "/home/circleci/.pyenv/versions/3.8.20/lib/python3.8/site-packages/keras/backend.py", line 4677, in function
    return GraphExecutionFunction(
  File "/home/circleci/.pyenv/versions/3.8.20/lib/python3.8/site-packages/keras/backend.py", line 4448, in __init__
    with tf.control_dependencies([self.outputs[0]]):
  File "/home/circleci/.pyenv/versions/3.8.20/lib/python3.8/site-packages/tensorflow/python/framework/ops.py", line 5698, in control_dependencies
    return get_default_graph().control_dependencies(control_inputs)
  File "/home/circleci/.pyenv/versions/3.8.20/lib/python3.8/site-packages/tensorflow/python/framework/ops.py", line 5107, in control_dependencies
    c = self.as_graph_element(c)
  File "/home/circleci/.pyenv/versions/3.8.20/lib/python3.8/site-packages/tensorflow/python/framework/ops.py", line 4012, in as_graph_element
    return self._as_graph_element_locked(obj, allow_tensor, allow_operation)
  File "/home/circleci/.pyenv/versions/3.8.20/lib/python3.8/site-packages/tensorflow/python/framework/ops.py", line 4091, in _as_graph_element_locked
    raise ValueError("Tensor %s is not an element of this graph." % obj)
ValueError: Tensor Tensor("dense_2/Sigmoid:0", shape=(?, 1), dtype=float32) is not an element of this graph.

We don't do much more than tensorflow.keras.models.load_model and then .predict() – in the same process and thread. The only change to earlier might be the upstream versions (although we already required keras < 3.0 and tensorflow < 2.16 before)...

[bertsky]

The only change to earlier might be the upstream versions (although we already required keras < 3.0 and tensorflow < 2.16 before)...

But all Python versions fail the same (regardless of which TF and Keras is available).

in the same process and thread

Maybe not the same thread – since OCR-D/core@3cc4780 we use a ThreadPool for our page timeout...