mnist classification raises runtime error

[sony]

mnist classification raises runtime error

[TomonobuTsujikawa]

The default return value of grad_depends_{input,output}_data() was changed by:

• sony/nnabla#884
• sony/nnabla-ext-cuda#313

MaxPoolingCudaCudnn should be kept until input/output data is used by backward calculation due to cudnn limitation.
Also, to prevent broken data that are used in backward by inplace, show error messages at graph construction:

Failed `input->allow_modify_data()`: Modifying data is prohibited by the parent function of the 0-th input data of 'ReLUCudaCudnn' (depth=3). (Parent is 'MaxPoolingCudaCudnn').

It is necessary to set False for all places where inplace=True when giving MaxPooling output to ReLU.
So, you can run it by changing following in mnist_lenet_prediction().

    c1 = F.relu(F.max_pooling(c1, (2, 2)), inplace=False)
    c2 = PF.convolution(c1, 16, (5, 5), name='conv2')
    c2 = F.relu(F.max_pooling(c2, (2, 2)), inplace=False)

We will consider other ways such as ignoring inplace.
Sorry for inconvenience.

[TomonobuTsujikawa]

Now we prepared PRs to ignore inplace option from ReLU, LeakyReLU, RandomErase, Add2, BcAdd2, Sub2, AddScalar, MulScalar.
Please be careful if you expect inplace behavior.

• sony/nnabla#950
• sony/nnabla-ext-cuda#336

This change will be released by nnabla v1.21.0.

[TomonobuTsujikawa]

v1.21.0 was released, so close this issue.