void convert_to_binary_row(mxnet::NDArray& array) {
CHECK(array.shape().ndim() >= 2); // second dimension is input depth from prev. layer, needed for next line
std::cout << "array shape: " << array.shape() << std::endl;
//if(array.shape()[1] < BITS_PER_BINARY_WORD) return;
CHECK(array.shape()[1] % BITS_PER_BINARY_WORD == 0); // depth from input has to be divisible by 32 (or 64)
nnvm::TShape binarized_shape(1);
size_t size = array.shape().Size();
binarized_shape[0] = size / BITS_PER_BINARY_WORD;
mxnet::NDArray temp(binarized_shape, mxnet::Context::CPU(), false, mxnet::op::xnor_cpu::corresponding_dtype());
mxnet::op::xnor_cpu::get_binary_row((float*) array.data().dptr_, (BINARY_WORD*) temp.data().dptr_, size);
array = temp;
}
void transpose(mxnet::NDArray& array) {
CHECK(array.shape().ndim() == 2);
nnvm::TShape tansposed_shape(2);
int rows = array.shape()[0];
int cols = array.shape()[1];
tansposed_shape[0] = cols;
tansposed_shape[1] = rows;
mxnet::NDArray temp(tansposed_shape, mxnet::Context::CPU(), false, array.dtype());
MSHADOW_REAL_TYPE_SWITCH(array.dtype(), DType, {
for (int row = 0; row < rows; row++) {
for (int col = 0; col < cols; col++) {
((DType*)temp.data().dptr_)[col * rows + row] = ((DType*)array.data().dptr_)[row * cols + col];
}
}
})
void Train(mxnet::NDArray data_array, mxnet::NDArray label_array,
int max_epoches, int val_fold, float start_learning_rate, std::vector<mxnet::NDArray> &argsl) {
/*prepare ndarray*/
learning_rate = start_learning_rate;
size_t data_count = data_array.shape()[0];
size_t val_data_count = data_count * val_fold / 10;
size_t train_data_count = data_count - val_data_count;
train_data = data_array.Slice(0, train_data_count);
train_label = label_array.Slice(0, train_data_count);
val_data = data_array.Slice(train_data_count, data_count);
val_label = label_array.Slice(train_data_count, data_count);
size_t batch_size = in_args[0].shape()[0];
/*start the training*/
for (int iter = 0; iter < max_epoches; ++iter) {
CHECK(optimizer);
size_t start_index = 0;
in_args[0] =
train_data.Slice(start_index, start_index + batch_size).Copy(ctx_dev);
in_args[in_args.size() - 1] =
train_label.Slice(start_index, start_index + batch_size)
.Copy(ctx_dev);
in_args[0].WaitToRead();
in_args[in_args.size() - 1].WaitToRead();
while (start_index < train_data_count) {
/*rebind the excutor*/
delete exe;
exe = mxnet::Executor::Bind(net, ctx_dev, g2c, in_args, arg_grad_store,
grad_req_type, aux_states);
CHECK(exe);
exe->Forward(true);
exe->Backward(std::vector<mxnet::NDArray>());
start_index += batch_size;
if (start_index < train_data_count) {
if (start_index + batch_size >= train_data_count)
start_index = train_data_count - batch_size;
in_args[0] = train_data.Slice(start_index, start_index + batch_size)
.Copy(ctx_dev);
in_args[in_args.size() - 1] =
train_label.Slice(start_index, start_index + batch_size)
.Copy(ctx_dev);
}
for (size_t i = 1; i < in_args.size() - 1; ++i) {
optimizer->Update(i, &in_args[i], &arg_grad_store[i], learning_rate);
}
for (size_t i = 1; i < in_args.size() - 1; ++i) {
in_args[i].WaitToRead();
}
in_args[0].WaitToRead();
in_args[in_args.size() - 1].WaitToRead();
}
/*call every iter*/
TrainingCallBack(iter, exe);
}
argsl.assign(in_args.begin(), in_args.end());
}
