CI: update to TF 2.10 #1160

albertz · 2022-10-19T19:06:36Z

No description provided.

albertz · 2022-10-19T19:11:32Z

Currently fails:

OSError: /var/tmp/runner/returnn_tf_cache/tf_utils/kernels_registered_for_op/86a95c7ddc/kernels_registered_for_op.so: undefined symbol: _ZN10tensorflow22KernelsRegisteredForOpB5cxx11EN4absl12lts_2022062311string_viewE

Or:

tensorflow.python.framework.errors_impl.NotFoundError: /var/tmp/runner/returnn_tf_cache/ops/NextEditDistanceRowOp/eb12481e22/NextEditDistanceRowOp.so: undefined symbol: _ZN10tensorflow8str_util13StringReplaceB5cxx11EN4absl12lts_2022062311string_viewES3_S3_b

Or:

tensorflow.python.framework.errors_impl.NotFoundError: /var/tmp/runner/returnn_tf_cache/ops/EditDistanceOp/5f70f62376/EditDistanceOp.so: undefined symbol: _ZN10tensorflow8str_util13StringReplaceB5cxx11EN4absl12lts_2022062311string_viewES3_S3_b

Or:

tensorflow.python.framework.errors_impl.NotFoundError: /var/tmp/runner/returnn_tf_cache/ops/LstmGenericBase/30db5737ad/LstmGenericBase.so: undefined symbol: _ZN10tensorflow8str_util13StringReplaceB5cxx11EN4absl12lts_2022062311string_viewES3_S3_b

albertz · 2022-10-19T19:19:53Z

GitHub Actions seems to pick up a reasonable GCC version. From test_TFNativeOp:

TF compiler version: 9.3.1 20200408
GCC for TF: /usr/bin/gcc-9

Other relevant:

TF __file__: /home/runner/.local/lib/python3.7/site-packages/tensorflow/__init__.py
TF version: 2.10.0
TF describe version: 2.10.0 (v2.10.0-rc3-6-g359c3cdfc5f) (<site-package> in /home/runner/.local/lib/python3.7/site-packages/tensorflow)
TF include: /home/runner/.local/lib/python3.7/site-packages/tensorflow/include
TF lib: /home/runner/.local/lib/python3.7/site-packages/tensorflow
TF link flags: ['-L/home/runner/.local/lib/python3.7/site-packages/tensorflow', '-l:libtensorflow_framework.so.2']
TF compile flags: ['-I/home/runner/.local/lib/python3.7/site-packages/tensorflow/include', '-D_GLIBCXX_USE_CXX11_ABI=1', '-DEIGEN_MAX_ALIGN_BYTES=64']
TF cxx11 abi flag: 1
...
TF lib so does not(!) exist: /home/runner/.local/lib/python3.7/site-packages/tensorflow/libtensorflow_framework.so
TF pywrap so exists: /home/runner/.local/lib/python3.7/site-packages/tensorflow/python/_pywrap_tensorflow_internal.so

albertz · 2022-10-19T19:29:07Z

Maybe related problems:
tensorflow/tensorflow#23561
abseil/abseil-cpp#211
https://groups.google.com/a/tensorflow.org/g/build/c/inpbEu4QaEI
tensorflow/serving#1935
FloopCZ/tensorflow_cc#266
FloopCZ/tensorflow_cc#269

albertz · 2022-10-19T20:17:03Z

I don't really understand: Why are some of the seemingly unrelated tests using older TF versions failing now? And also, why with such weird stacktrace? E.g. tf-tests (TEST=TFUtil, 3.7, 1.15.3):

NotImplementedError: Cannot convert a symbolic Tensor (single_strided_slice_3/mod:0) to a numpy array.

__________________________ test_single_strided_slice ___________________________
Traceback (most recent call last):
  File "/home/runner/work/returnn/returnn/tests/test_TFUtil.py", line 1709, in test_single_strided_slice
    line: assert_equal(list(single_strided_slice(x, axis=tf.constant(1), end=3)[0].eval()), [0, 1, 2])
    locals:
      assert_equal = <global> <bound method TestCase.assertEqual of <nose.tools.trivial.Dummy testMethod=nop>>
      list = <builtin> <class 'list'>
      single_strided_slice = <global> <function single_strided_slice at 0x7f406542b830>
      x = <local> <tf.Tensor 'ExpandDims_2:0' shape=(1, 10) dtype=int32>
      axis = <not found>
      tf = <global> <module 'tensorflow' from '/home/runner/.local/lib/python3.7/site-packages/tensorflow/__init__.py'>
      tf.constant = <global> <function constant_v1 at 0x7f406ffd4b90>
      end = <not found>
      eval = <builtin> <built-in function eval>
  File "/home/runner/work/returnn/returnn/returnn/tf/util/basic.py", line 3502, in single_strided_slice
    line: begins = tf.concat([tf.zeros((axis,), tf.int32), (begin,)], axis=0)
    locals:
      begins = <not found>
      tf = <global> <module 'tensorflow' from '/home/runner/.local/lib/python3.7/site-packages/tensorflow/__init__.py'>
      tf.concat = <global> <function concat at 0x7f406fe144d0>
      tf.zeros = <global> <function zeros at 0x7f406fe14dd0>
      axis = <local> <tf.Tensor 'single_strided_slice_3/mod:0' shape=() dtype=int32>
      tf.int32 = <global> tf.int32
      begin = <local> 0
  File "/home/runner/.local/lib/python3.7/site-packages/tensorflow_core/python/ops/array_ops.py", line 2338, in zeros
    line: output = _constant_if_small(zero, shape, dtype, name)
    locals:
      output = <not found>
      _constant_if_small = <global> <function _constant_if_small at 0x7f406fe14a70>
      zero = <local> 0
      shape = <local> (<tf.Tensor 'single_strided_slice_3/mod:0' shape=() dtype=int32>,)
      dtype = <local> tf.int32
      name = <local> 'single_strided_slice_3/zeros/', len = 29
  File "/home/runner/.local/lib/python3.7/site-packages/tensorflow_core/python/ops/array_ops.py", line 2295, in _constant_if_small
    line: if np.prod(shape) < 1000:
    locals:
      np = <global> <module 'numpy' from '/home/runner/.local/lib/python3.7/site-packages/numpy/__init__.py'>
      np.prod = <global> <function prod at 0x7f4089f34dd0>
      shape = <local> (<tf.Tensor 'single_strided_slice_3/mod:0' shape=() dtype=int32>,)
  File "<__array_function__ internals>", line 6, in prod
    -- code not available --
  File "/home/runner/.local/lib/python3.7/site-packages/numpy/core/fromnumeric.py", line 3052, in prod
    line: return _wrapreduction(a, np.multiply, 'prod', axis, dtype, out,
                                keepdims=keepdims, initial=initial, where=where)
    locals:
      _wrapreduction = <global> <function _wrapreduction at 0x7f4089f1e3b0>
      a = <local> (<tf.Tensor 'single_strided_slice_3/mod:0' shape=() dtype=int32>,)
      np = <global> <module 'numpy' from '/home/runner/.local/lib/python3.7/site-packages/numpy/__init__.py'>
      np.multiply = <global> <ufunc 'multiply'>
      axis = <local> None
      dtype = <local> None
      out = <local> None
      keepdims = <local> <no value>
      initial = <local> <no value>
      where = <local> <no value>
  File "/home/runner/.local/lib/python3.7/site-packages/numpy/core/fromnumeric.py", line 86, in _wrapreduction
    line: return ufunc.reduce(obj, axis, dtype, out, **passkwargs)
    locals:
      ufunc = <local> <ufunc 'multiply'>
      ufunc.reduce = <local> <built-in method reduce of numpy.ufunc object at 0x7f408a85d050>
      obj = <local> (<tf.Tensor 'single_strided_slice_3/mod:0' shape=() dtype=int32>,)
      axis = <local> None
      dtype = <local> None
      out = <local> None
      passkwargs = <local> {}
  File "/home/runner/.local/lib/python3.7/site-packages/tensorflow_core/python/framework/ops.py", line 736, in Tensor.__array__
    line: raise NotImplementedError("Cannot convert a symbolic Tensor ({}) to a numpy"
                                    " array.".format(self.name))
    locals:
      NotImplementedError = <builtin> <class 'NotImplementedError'>
      format = <builtin> <built-in function format>
      self = <local> <tf.Tensor 'single_strided_slice_3/mod:0' shape=() dtype=int32>
      self.name = <local> 'single_strided_slice_3/mod:0', len = [28](https://github.com/rwth-i6/returnn/actions/runs/3284624825/jobs/5410808347#step:8:29)
NotImplementedError: Cannot convert a symbolic Tensor (single_strided_slice_3/mod:0) to a numpy array.

And when you look up _constant_if_small, actually it should cover that?

Maybe the GitHub CI cache is somehow messing up different TF versions?

NumPy: 1.21.6
TensorFlow: v1.15.2-30-g4386a66 1.15.3 /home/runner/.local/lib/python3.7/site-packages/tensorflow/__init__.py

In the log I see:

      tf = <global> <module 'tensorflow' from '/home/runner/.local/lib/python3.7/site-packages/tensorflow/__init__.py'>

And then:

  File "/home/runner/.local/lib/python3.7/site-packages/tensorflow_core/python/ops/array_ops.py", line 2338, in zeros

So tensorflow and tensorflow_core. Is that correct?

Or maybe Numpy version?

albertz · 2022-10-19T20:24:27Z

Related issues:
tensorflow/models#9706
https://stackoverflow.com/questions/66139250/tf-data-function-fails-tries-to-convert-to-numpy-array
tensorflow/tensorflow#38836
https://www.reddit.com/r/tensorflow/comments/lgcgby/numpyrelated_error_when_building_model/

albertz · 2022-10-19T20:26:33Z

And when you look up _constant_if_small, actually it should cover that?

Actually no it does not, at least TF 2.3 and probably all earlier versions: https://github.com/tensorflow/tensorflow/blob/v2.3.0/tensorflow/python/ops/array_ops.py#L2730

albertz · 2022-10-19T20:30:21Z

Maybe the GitHub CI cache is somehow messing up different TF versions?
NumPy: 1.21.6
TensorFlow: v1.15.2-30-g4386a66 1.15.3 /home/runner/.local/lib/python3.7/site-packages/tensorflow/__init__.

For comparison, in the current master:

Python env: python is /opt/hostedtoolcache/Python/3.7.14/x64/bin/python Python 3.7.14
NumPy: 1.19.5
TensorFlow: v1.15.2-30-g4386a66 1.15.3 /home/runner/.local/lib/python3.7/site-packages/tensorflow/__init__.py

So interestingly it has an older NumPy version. So the newer NumPy version probably causes the error.

albertz · 2022-10-19T21:01:54Z

Ok, that old Numpy version was too specific. For Python 2.7, it fails:

ERROR: Could not find a version that satisfies the requirement numpy==1.19.5 (from versions: 1.3.0, 1.4.1, 1.5.0, 1.5.1, 1.6.0, 1.6.1, 1.6.2, 1.7.0, 1.7.1, 1.7.2, 1.8.0, 1.8.1, 1.8.2, 1.9.0, 1.9.1, 1.9.2, 1.9.3, 1.10.0, 1.10.0.post2, 1.10.1, 1.10.2, 1.10.3, 1.10.4, 1.11.0, 1.11.1, 1.11.2, 1.11.3, 1.12.0, 1.12.1, 1.13.0rc1, 1.13.0rc2, 1.13.0, 1.13.1, 1.13.3, 1.14.0rc1, 1.14.0, 1.14.1, 1.14.2, 1.14.3, 1.14.4, 1.14.5, 1.14.6, 1.15.0rc1, 1.15.0rc2, 1.15.0, 1.15.1, 1.15.2, 1.15.3, 1.15.4, 1.16.0rc1, 1.16.0rc2, 1.16.0, 1.16.1, 1.16.2, 1.16.3, 1.16.4, 1.16.5, 1.16.6)

albertz · 2022-10-19T21:16:30Z

The PyCharm inspections raise many new warnings for TF 2.10 (see here), for example:

  returnn/datasets/hdf.py:1180: WARNING PyUnresolvedReferencesInspection: Cannot find reference 'object' in '__init__.pyi | __init__.pyi'
  returnn/tf/layers/rec.py:8558: WEAK WARNING PyAbstractClassInspection: Class VanillaLSTMCell must implement all abstract methods
  returnn/tf/updater.py:956: WEAK WARNING PyAbstractClassInspection: Class _KerasOptimizerWrapper must implement all abstract methods
  returnn/tf/util/basic.py:4679: WARNING PyUnresolvedReferencesInspection: Cannot find reference 'float' in '__init__.pyi | __init__.pyi'
  tools/tf_inspect_checkpoint.py:123: WARNING PyTypedDictInspection: TypedDict key must be a string literal; expected one of ('precision', 'threshold', 'edgeitems', 'linewidth', 'suppress', 'nanstr', 'infstr', 'formatter', 'sign', 'floatmode', 'legacy')

I will just keep the old TF version for the PyCharm inspections for now. We can later look at that.

CI: update to TF 2.10

e086ae5

albertz added 2 commits October 19, 2022 21:58

OpCodeCompiler, use C++17 for TF >=2.10

05bd38c

TFNativeUtilCompiler small fix

8cbfcb9

albertz added 2 commits October 19, 2022 22:40

CI: older Numpy for older TF

4a1c809

CI: older Numpy for older TF, only if Python 3

1567cf7

albertz marked this pull request as ready for review October 19, 2022 21:02

albertz requested a review from a team as a code owner October 19, 2022 21:02

CI: PyCharm inspect keeps using 2.4.2 for now

1b8e447

albertz merged commit 89252f1 into master Oct 19, 2022

albertz deleted the albert-tf210 branch October 19, 2022 21:38

This was referenced Oct 20, 2022

OpCodeCompiler/TFNativeUtilCompiler, better C++ standard version detection, use get_compile_flags #1163

Open

CI, update TF to 2.6.2 #929

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

CI: update to TF 2.10 #1160

CI: update to TF 2.10 #1160

albertz commented Oct 19, 2022

albertz commented Oct 19, 2022 •

edited

Loading

albertz commented Oct 19, 2022

albertz commented Oct 19, 2022 •

edited

Loading

albertz commented Oct 19, 2022 •

edited

Loading

albertz commented Oct 19, 2022

albertz commented Oct 19, 2022

albertz commented Oct 19, 2022 •

edited

Loading

albertz commented Oct 19, 2022

albertz commented Oct 19, 2022 •

edited

Loading

CI: update to TF 2.10 #1160

CI: update to TF 2.10 #1160

Conversation

albertz commented Oct 19, 2022

albertz commented Oct 19, 2022 • edited Loading

albertz commented Oct 19, 2022

albertz commented Oct 19, 2022 • edited Loading

albertz commented Oct 19, 2022 • edited Loading

albertz commented Oct 19, 2022

albertz commented Oct 19, 2022

albertz commented Oct 19, 2022 • edited Loading

albertz commented Oct 19, 2022

albertz commented Oct 19, 2022 • edited Loading

albertz commented Oct 19, 2022 •

edited

Loading

albertz commented Oct 19, 2022 •

edited

Loading

albertz commented Oct 19, 2022 •

edited

Loading

albertz commented Oct 19, 2022 •

edited

Loading

albertz commented Oct 19, 2022 •

edited

Loading