Concatenates several layers into one, by concatenating each in the x-dimension or y-dimension, based on a user-specified flag. For x-concatenation, layers are aligned at matching y and depth coordinates, and for y-concatenation, they are aligned at matching x and depth coordinates. The inputs must all have the same depth and batch size. For x-concatenation, the inputs must have the same height (y-
| 2176 | // concat_direction: either dnn:SpaceConcatenateMode::XDirection or |
| 2177 | // dnn::SpaceConcatenateMode::YDirection. |
| 2178 | virtual bool DoSpaceConcatenate( |
| 2179 | Stream* stream, port::ArraySlice<dnn::BatchDescriptor> input_dimensions, |
| 2180 | port::ArraySlice<const DeviceMemory<float>*> input_data, |
| 2181 | DeviceMemory<float>* output_data, |
| 2182 | dnn::SpaceConcatenateMode concat_direction) { |
| 2183 | return false; |
| 2184 | } |
| 2185 | |
| 2186 | // Change the layout of the data by shrinking one dimension (or set of |
| 2187 | // dimensions) and growing another dimension (or set of dimensions), while |