ImageBatch Convolution::PoolingBackward(ImageBatch diff, ImageBatch top, ImageBatch bottom, PoolingInfo info) {
CHECK_EQ(diff.Size(), top.Size()) << "inputs sizes mismatch";
CHECK_EQ(diff.GetNumImages(), bottom.GetNumImages()) << "#images mismatch";
CHECK_EQ(diff.GetNumFeatureMaps(), bottom.GetNumFeatureMaps()) << "#channels mismatch";
int pooled_height = (bottom.GetHeight() + 2 * info.pad_height - info.height + info.stride_vertical - 1) / info.stride_vertical + 1;
int pooled_width = (bottom.GetWidth() + 2 * info.pad_width - info.width + info.stride_horizontal - 1) / info.stride_horizontal + 1;
if (0 <= (pooled_height - 1) * info.stride_vertical - bottom.GetHeight() - info.pad_height) {
--pooled_height;
}
if (0 <= (pooled_width - 1) * info.stride_horizontal - bottom.GetWidth() - info.pad_width) {
--pooled_width;
}
CHECK_EQ(top.GetHeight(), pooled_height) << "height mismatch";
CHECK_EQ(top.GetWidth(), pooled_width) << "width mismatch";
PoolingBackwardOp* op = new PoolingBackwardOp();
op->closure = {
info.algorithm,
info.height,
info.width,
info.stride_vertical,
info.stride_horizontal,
info.pad_height,
info.pad_width
};
return NArray::ComputeOne({diff, top, bottom}, bottom.Size(), op);
}
NArray Convolution::ConvBackwardBias(ImageBatch diff) {
Scale new_size {
diff.GetNumFeatureMaps()
};
ConvBackwardBiasOp* op = new ConvBackwardBiasOp();
return NArray::ComputeOne({diff}, new_size, op);
}
ImageBatch Convolution::PoolingForward(ImageBatch src, PoolingInfo info) {
int pooled_height = (src.GetHeight() + 2 * info.pad_height - info.height + info.stride_vertical - 1) / info.stride_vertical + 1;
int pooled_width = (src.GetWidth() + 2 * info.pad_width - info.width + info.stride_horizontal - 1) / info.stride_horizontal + 1;
if (0 <= (pooled_height - 1) * info.stride_vertical - src.GetHeight() - info.pad_height) {
--pooled_height;
}
if (0 <= (pooled_width - 1) * info.stride_horizontal - src.GetWidth() - info.pad_width) {
--pooled_width;
}
Scale new_size {
pooled_height,
pooled_width,
src.GetNumFeatureMaps(),
src.GetNumImages()
};
PoolingForwardOp* op = new PoolingForwardOp();
op->closure = {
info.algorithm,
info.height,
info.width,
info.stride_vertical,
info.stride_horizontal,
info.pad_height,
info.pad_width
};
return NArray::ComputeOne({src}, new_size, op);
}
ImageBatch Convolution::ConvBackwardData(ImageBatch diff, ImageBatch bottom, Filter filter, ConvInfo info) {
CHECK_EQ(diff.GetNumFeatureMaps(), filter.GetNumOutputs()) << "#output channels mismatch";
/*
* We can't get filter size when (top + 2*pad) % stride != 0
Scale new_size {
(diff.GetWidth() - 1) * info.stride_horizontal + filter.GetWidth() - 2 * info.pad_width,
(diff.GetHeight() - 1) * info.stride_vertical + filter.GetHeight() - 2 * info.pad_height,
filter.GetNumInputs(),
diff.GetNumImages()
};
*/
ConvBackwardDataOp* op = new ConvBackwardDataOp();
op->closure = {
info.pad_height,
info.pad_width,
info.stride_vertical,
info.stride_horizontal
};
return NArray::ComputeOne({diff, filter}, bottom.Size(), op);
}
ImageBatch Convolution::ConvForward(ImageBatch src, Filter filter, NArray bias, ConvInfo info) {
CHECK_EQ(src.GetNumFeatureMaps(), filter.GetNumInputs()) << "#input channels mismatch";
CHECK_EQ(bias.Size().NumDims(), 1) << "bias dimension mismatch";
CHECK_EQ(bias.Size()[0], filter.GetNumOutputs()) << "bias size mismatch";
//no such limit
//CHECK_EQ((src.GetHeight() + 2 * info.pad_height - filter.GetHeight()) % info.stride_vertical, 0) << "filter height mismatch";
//CHECK_EQ((src.GetWidth() + 2 * info.pad_width - filter.GetWidth()) % info.stride_horizontal, 0) << "filter width mismatch";
Scale new_size {
(src.GetWidth() + 2 * info.pad_width - filter.GetWidth()) / info.stride_horizontal + 1,
(src.GetHeight() + 2 * info.pad_height - filter.GetHeight()) / info.stride_vertical + 1,
filter.GetNumOutputs(),
src.GetNumImages()
};
ConvForwardOp* op = new ConvForwardOp();
op->closure = {
info.pad_height,
info.pad_width,
info.stride_vertical,
info.stride_horizontal
};
return NArray::ComputeOne({src, filter, bias}, new_size, op);
}