Add upper bound for numpy version in dllib dependency

[hkvision]

According to @hamham223 python 3.8 will default to install numpy 1.24, which will cause issues.
@hamham223 Paste the issues here and check what's the upper bound of the numpy we support.

[hamham223]

Previous Failed Job: [Orca-Python-Ray-Py38-Spark3]

Typical Error Message: test/bigdl/orca/learn/ray/mxnet/test_mxnet_gluon.py - AttributeError: module 'numpy' has no attribute 'bool'

This is due to mxnet does not support numpy 1.24, although the installation requirement for mxnet 1.9.1 is numpy<2.0.0,>1.16.0. Hopefully they could fix this issue soon.

[hamham223]

Reference:

• Numpy 1.20 Deprecate the attributes np.int, np.bool, np.float etc.
• A long email about the deprecation
• Deprecation is expired in numpy 1.24

[hamham223]

Two possible solutions?

1. add setup requirement: numpy <= 1.23.5
2. replace all the np.int with np.int32/64, np.float with np.float32/64:

Here are all the src codes that have those deprecations

• dllib:

  • https://github.com/intel-analytics/BigDL/blob/main/python/dllib/src/bigdl/dllib/models/rnn/rnnexample.py#L104
  • https://github.com/intel-analytics/BigDL/blob/main/python/dllib/src/bigdl/dllib/utils/utils.py#L227

• orca:

  • https://github.com/intel-analytics/BigDL/blob/main/python/orca/src/bigdl/orca/learn/openvino/estimator.py#L145

• chronos:

  • https://github.com/intel-analytics/BigDL/blob/main/python/chronos/src/bigdl/chronos/data/utils/roll_dataset.py#L34
  • https://github.com/intel-analytics/BigDL/blob/main/python/chronos/src/bigdl/chronos/model/tcmf/time.py#L62
  • https://github.com/intel-analytics/BigDL/blob/main/python/chronos/src/bigdl/chronos/model/tf1/VanillaLSTM_keras.py#L74

[hkvision]

https://github.com/intel-analytics/BigDL/blob/main/python/dllib/src/bigdl/dllib/models/rnn/rnnexample.py#L104

can change to np.int32

https://github.com/intel-analytics/BigDL/blob/main/python/dllib/src/bigdl/dllib/utils/utils.py#L227

can change to str

>>> np.array(["x"])
array(['x'], dtype='<U1')
>>> np.array(["x"]).astype(str)
array(['x'], dtype='<U1')

[hkvision]

By default for Python 3.8, pip install numpy>=3.19.5 will install numpy 1.24.2.

[hamham223]

Seems numpy 1.24 requires np.array must have homogeneous shape. e.g. [[1,2], [3,4,5]] is not acceptable. Previously numpy 1.21 does not have this problem.

https://stackoverflow.com/questions/67183501/setting-an-array-element-with-a-sequence-requested-array-has-an-inhomogeneous-sh

Failed job: https://github.com/intel-analytics/BigDL/actions/runs/4192715495/jobs/7268720402

[hkvision]

Fixed. Orca code/unit tests have now supported numpy 1.24

[sharan21]

I think this issue still persists. Are there any updates on this?

[hkvision]

Oh really? I think we have supported at least for numpy 1.24.

Can you provide us with more information? Including your numpy version, and what piece of code you get the issue. We will then look into this.