tensorflow：“试图突变冻结的对象”，巴兹尔

Question

MacOS high sierra, MBP 2016, in terminal.

I'm following the directions here: https://github.com/tensorflow/models/tree/master/research/syntaxnet All options for ./configure chosen as default (and all python directories double-checked.). All steps have completed cleanly until this:

bazel test ...
# On Mac, run the following:
bazel test --linkopt=-headerpad_max_install_names \
  dragnn/... syntaxnet/... util/utf8/...

I assume I'm supposed to run the latter line ("bazel test --linkopt" etc.). But I get the same result either way, interestingly.

This throws about 10 errors, each of the same type "trying to mutate a frozen object", and concludes tests not run, error loading package dragnn/protos, and couldn't start build.

This is the general form of the errors:

syntaxnet>> bazel test --linkopt=-headerpad_max_install_names dragnn/... syntaxnet/... util/utf8/...

.

ERROR: /Users/XXX/Desktop/NLP/syntaxnet/models/research/syntaxnet/dragnn/protos/BUILD:35:1: Traceback (most recent call last): File "/Users/XXX/Desktop/NLP/syntaxnet/models/research/syntaxnet/dragnn/protos/BUILD", line 35 tf_proto_library_py(name = "data_py_pb2", srcs = ["dat..."]) File "/Users/XXX/Desktop/NLP/syntaxnet/models/research/syntaxnet/syntaxnet/syntaxnet.bzl", line 53, in tf_proto_library_py py_proto_library(name = name, srcs = srcs, srcs_versi...", <5 more arguments>) File "/private/var/tmp/_bazel_XXX/f74e5a21c3ad09aeb110d9de15110035/external/protobuf_archive/protobuf.bzl", line 374, in py_proto_library py_libs += [default_runtime] trying to mutate a frozen object ERROR: package contains errors: dragnn/protos

---

... [same error for various 'name = "...pb2"' files] ...

INFO: Elapsed time: 0.709s FAILED: Build did NOT complete successfully (17 packages loaded) ERROR: Couldn't start the build. Unable to run tests

Any idea what could be doing this? Thanks.

Answer 1

This error indicates a bug in the py_proto_library rule implementation.

tf_proto_library_py is defined in syntaxnet.bzl . It is a wrapper around py_proto_library , which is defined by the tf_workspace macro's protobuf_archive rule.

"protobuf_archive" downloads Protobuf 3.3.0, which contains //:protobuf.bzl with the buggy py_proto_library rule implementation: in line #374 it tries to mutate an immutable object py_libs .

Make sure you use the latest Bazel version , currently that's 0.8.1.

If the problem still persists, then:

• I suggest filing a bug with:

  • Protobuf, to fix the py_proto_library rule
  • TensorFlow, to update their Protobuf version in tf_workspace , and
  • Syntaxnet to update their TF submodule reference in //research/syntaxnet to the bugfixed version.

• As a workaround, perhaps you can patch protobuf.bzl .

The patch is to change these lines:

   373    if default_runtime and not default_runtime in py_libs + deps:
   374      py_libs += [default_runtime]
   375
   376    native.py_library(
   377        name=name,
   378        srcs=outs+py_extra_srcs,
   379        deps=py_libs+deps,
   380        imports=includes,
   381        **kargs)

to these:

   373    if default_runtime and not default_runtime in py_libs + deps:
   374      py_libs2 = py_libs + [default_runtime]
   375    else:
   376      py_libs2 = py_libs
   377
   378    native.py_library(
   379        name=name,
   380        srcs=outs+py_extra_srcs,
   381        deps=py_libs2+deps,
   382        imports=includes,
   383        **kargs)

Disclaimer: this is a "blind" fix; I have not tried whether it works.

Answer 2

Tried same pattern patch for cc_libs. 

  if default_runtime and not default_runtime in cc_libs:
    cc_libs2 = cc_libs + [default_runtime]
  else:
    cc_libs2 = cc_libs
  if use_grpc_plugin:
    cc_libs += ["//external:grpc_lib"]

  native.cc_library(
      name=name,
      srcs=gen_srcs,
      hdrs=gen_hdrs,
      deps=cc_libs2 + deps,
      includes=includes,
      **kargs)

Shows new error, but keeps compiling. (Ubuntu 16 on Windows System for Linux--don't ask, native tensorflow 1.4 winx64 works, but not syntaxnet).

greg@FX11:/mnt/c/code/models/research/syntaxnet$ bazel test ... ERROR: /home/greg/.cache/bazel/_bazel_greg/adb8eb0eab8b9680449366fbebe59ec2/external/org_tensorflow/tensorflow/core/kernels/BUILD:451:1: in _transitive_hdrs rule @org_tensorflow//tensorflow/core/kernels:bounds_check_lib_gather: Traceback (most recent call last): File "/home/greg/.cache/bazel/_bazel_greg/adb8eb0eab8b9680449366fbebe59ec2/external/org_tensorflow/tensorflow/core/kernels/BUILD", line 451 _transitive_hdrs(name = 'bounds_check_lib_gather') File "/home/greg/.cache/bazel/_bazel_greg/adb8eb0eab8b9680449366fbebe59ec2/external/org_tensorflow/tensorflow/tensorflow.bzl", line 869, in _transitive_hdrs_impl set()

Just changed set() to depset() and that seems to have avoided the error.

Answer 3

To make a long story short. I was inspired by a sstrasburg 's comment.

Firstly, uninstall a fresh version of bazel.

brew uninstall bazel

Download bazel 0.5.4 from here .

chmod +x bazel-0.5.4-without-jdk-installer-darwin-x86_64.sh
./bazel-0.5.4-without-jdk-installer-darwin-x86_64.sh

After that, again run

bazel test --linkopt=-headerpad_max_install_names dragnn/... syntaxnet/... util/utf8/...

Finally, I got

Executed 57 out of 57 tests: 57 tests pass.