bc::copy, copy vector data to cube, don't support, now!

[xinsuinizhuan]

vector<vector<double>> matricsInputSeries;
cube inputs(img_sz, network->m_batch_size, training_sets);
bc::copy(inputs[i][j].get_stream(),matricsInputSeries[m_batch_size + j].begin(), matricsInputSeries[i * m_batch_size + j].end(), inputs[i][j].cw_begin());

now i can't use bc::copy, copy vector data to cube input!

[josephjaspers]

Hi, I've added tests to cover this issue
b035502
6ed9e59

However I was able to successfully compile the above code (Windows VS2019) and have tests run successfully. Without having to change any internal code, so I am unsure why it wouldn't work.

Could you post the error and compilation output?
Thanks!

[josephjaspers]

Additionally I haven't yet documented but there is a new class named

VecList which behaves like a regular bc::Vector but also has the "push_back" method.
Ergo, you might be able to do something like...

bc::VecList<double>  veclist; 
for (int i = 0; i < other_data.size(); ++i) {
     veclist.push_back(other_data[i]); 
}

bc::Cube<double> data(size, batch_sz, other_size); 
data[0][0] = veclist; 

[xinsuinizhuan]

I meet a strange problem, every functions is ok, when i do not call them. But when i call the function, so many compile errors.
do not call fuction:

call function:

error.txt

blackcat_tensors_consoledemo.zip

[josephjaspers]

Hi, I fixed the issue.
I had previously removed utility support (print/copying) from expressions though I have re-added them in the last few commits.

A while back a re-ordered the template arguments for the Allocator class.
So in your code you have to change

	using allocator_type = bc::Allocator<System, value_type>

to

        using allocator_type = bc::Allocator<value_type, System>;

So if you pull the newest commits and make those changes your code should be able to work.
(I was able to successfully compile it on my windows VM).

I added tests to cover this issue.
Let me know if it works!

Thanks!

[xinsuinizhuan]

thanks, it works. Other problem is that, i predict with the trained data, it seems well, but i predict with the untrained data, it seems not good, i try to decrease the epoch, but it seems not work. I don't know why?

[josephjaspers]

Hmmm... I am not sure why, if you send me what you are working on i could take a look?

[xinsuinizhuan]

like this:

auto network = bc::nn::neuralnetwork(
bc::nn::lstm(system_tag, 784/4, 128),
bc::nn::lstm(system_tag, 128, 64),
bc::nn::feedforward(system_tag, 64, 10),
bc::nn::softmax(system_tag, 10),
bc::nn::logging_output_layer(system_tag, 10, bc::nn::RMSE).skip_every(100)
);

bc::print("Neural Network architecture:");
bc::print(network.get_string_architecture());


network.set_learning_rate(0.001);
network.set_batch_size(batch_size);

it break at, network.set_learning_rate(0.001);

mnist_test.txt

[josephjaspers]

I will check!

[josephjaspers]

Fixed!
87e9749

Silly bug (infinite recursion), added set_learning_rate to the mnist_recurrent example to catch this bug in the future!