Protobuf: Different behavior with float precision

[twmht]

I found the difference when dealing with float point between different version of python-protobuf (2.6.1 vs 3.5.0.post1)

test.proto

message A {
  optional float a = 1;
}

$ protoc --version # libprotoc 2.6.1
protoc test.proto --python_out=./

python-protobuf 2.6.1

A = test_pb2.A()
A.a = 0.01
print A.__str__() # a: 0.01

python-protobuf 3.5.0.post1

A = test_pb2.A()
A.a = 0.01
print A.__str__() # a: 0.00999999977648

Not sure if it's a bug or not.

What makes the difference?

Thank you

[zchrissirhcz]

In my case, I use pycaffe to read, modify and write 'xxx.prototxt' file. Similar as metioned in this issue https://github.com/google/protobuf/issues/3632, by setting the python protobuf's backend to python instead of the default cpp (I guess, please correct me if I am wrong), the float numbers in saved prototxt has the exactly same precision as expected:

export PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION=python
python my_code.py

@twmht Now I use python2.7.12 with protobuf 3.5.2, and this environment setting also works for the testing code you put in this issue.

[anandolee]

zchrissirhcz@ is right. We have a known issue for float type precision if it is using cpp extension:

Python does not have C-style float , it only has a C-style double. Thus pure python is using double precision for both float and double field, cpp extension is using float precision for float field.

[anandolee]

Closing it for cleaning up. Feel free to reopen if you still have questions