Calling Blob->Reshape with zero causes a floating point exception

[elezar]

Adding the following test to test_blob.cpp:

TYPED_TEST(BlobSimpleTest, TestReshapeZeroShape) {
  vector<int> shape(4);
  shape[0] = 0; shape[1] = 0; shape[2] = 0; shape[3] = 0;
  this->blob_->Reshape(shape);
  EXPECT_EQ(this->blob_->shape(), shape);
}

Results in the following crash:

[ RUN      ] BlobSimpleTest/0.TestReshapeZeroShape
*** Aborted at 1453388430 (unix time) try "date -d @1453388430" if you are using GNU date ***
PC: @     0x7ff80a5d596d caffe::Blob<>::Reshape()
*** SIGFPE (@0x7ff80a5d596d) received by PID 26958 (TID 0x7ff80aceca40) from PID 173889901; stack trace: ***
    @     0x7ff803ef1d40 (unknown)
    @     0x7ff80a5d596d caffe::Blob<>::Reshape()
    @           0xb4acd4 caffe::BlobSimpleTest_TestReshapeZeroShape_Test<>::TestBody()
    @           0xd7198b testing::internal::HandleSehExceptionsInMethodIfSupported<>()
    @           0xd6d092 testing::internal::HandleExceptionsInMethodIfSupported<>()
    @           0xd5a4a9 testing::Test::Run()
    @           0xd5ac42 testing::TestInfo::Run()
    @           0xd5b230 testing::TestCase::Run()
    @           0xd600ba testing::internal::UnitTestImpl::RunAllTests()
    @           0xd729be testing::internal::HandleSehExceptionsInMethodIfSupported<>()
    @           0xd6dcbd testing::internal::HandleExceptionsInMethodIfSupported<>()
    @           0xd5ee4e testing::UnitTest::Run()
    @           0xa6f1eb main
    @     0x7ff803edcec5 (unknown)
    @           0x97c5c9 (unknown)
    @                0x0 (unknown)
Floating point exception (core dumped)

What is the expected behavior for reshaping to zero?

[seanbell]

I think zero-sized blobs weren't contemplated in the current design. If you reshape to zero you get a user-unfriendly exception.

In general, I think it makes sense to allow zero-sized blobs, and to return null when grabbing the data pointer, instead of throwing the exception during Reshape (which is what happens now).

@shelhamer do you have an opinion on this issue?

[elezar]

I think this may have been addressed in #4101. I will have a look as soon as I get a chance.

[jeffdonahue]

Yes, I'm pretty sure this was the bug that #4101 fixes. (Please comment if you find otherwise though.)