something wrong inupdate_window_and_padding about computelibrary HOT 5 CLOSED

When you call allocate before configuring the function the tensor by default will be allocated with no padding. The ActivationLayer function tries to process 16 elements at once (in X direction) but since your tensor is smaller the end of the window is automatically reduced to prevent invalid memory accesses. In your case it is reduced all the way down to zero which later triggers the assertion. We will add a better error message to indicate the problem.

To use the ActivationLayer function your options are:

• to set the padding manually before allocating the tensors (TensorInfo::extend_padding(PaddingSize)),
• to use auto padding which will be large enough to run any function but will allocate unnecessarily much memory for many functions (TensorInfo::auto_padding(), not recommended) or
• to allocate the tensors after configuring the functions. In that case the function will update the padding according to their needs. But this can only be done if the tensors are not already allocated.

from computelibrary.

In your piece of code it's unclear where and how you allocate your buffers, could you please give us more details ?

from computelibrary.

We've tried to run the following code and it seems to work:

    Tensor input;
    Tensor output;

    const TensorShape shape(14U, 14U, 192U);

    input.allocator()->init(TensorInfo(shape, 1, DataType::F32));
    output.allocator()->init(TensorInfo(shape, 1, DataType::F32));

    NEActivationLayer activation;
    activation.configure(&input, &output, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));

    input.allocator()->allocate();
    //TODO: Fill the input's content
    output.allocator()->allocate();

    activation.run();

from computelibrary.

Hi @AnthonyARM ,
I found that if I call the allocate() function before configure(), it will crash.

The following code crashed.

    input->allocator()->allocate();
    output->allocator()->allocate();
    activation.configure(input, output, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));

    activation.run();

The following code worked.

    activation.configure(input, output, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
    input->allocator()->allocate();
    output->allocator()->allocate();

    activation.run();

I have not got how to use Tensor properly. ^_^!

This is the complete test code

    const TensorShape shape(14U, 14U, 192U);

    Tensor *input = new Tensor();
    input->allocator()->init(TensorInfo(shape, Format::F32));

    Tensor *output = new Tensor();
    output->allocator()->init(TensorInfo(shape, Format::F32));

    NEActivationLayer activation;
    activation.configure(input, output, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
    input->allocator()->allocate();
    output->allocator()->allocate();

    activation.run();

from computelibrary.

Cool! Thank you all.

from computelibrary.