1

I have a requirement to implement the forward computing of deconv layer in the 3D filter manner.

Here, by '3D filter manner', I mean convolution like the Gaussian filter in CV. In the contrast, the caffe implements the deconv in the gemm + col2im manner.

I find a similar question here. The guy wrote the code according the introduction in tranposed conv.

Image

He/She does not open the source code. So I finished my own one:

template <typename DataType> int deconv_cpu(
  DataType *src, DataType *dst, DataType *para, DataType *bias,
  int in_width, int in_height, int in_channel,
  int out_width, int out_height, int out_channel,
  int ks, int padding = 0, int step = 1)  { // step indicates the stride

  int col, row, ch_o, ch_i, x, y;
  int r = (ks - 1) / 2; //radius;

  DataType result;
  DataType *output;
  DataType *filter;
  DataType *input;

  int sim_width, sim_height, sim_pad, width_border, height_border;
  sim_width = in_width * step - step + 1;
  sim_height = in_height * step - step + 1;
  sim_pad = ks - padding - 1;
  width_border = sim_pad == 0 ? r : 0;
  height_border = sim_pad == 0 ? r : 0;
  for (row = height_border; row < (sim_height - height_border); row++)
    for (col = width_border; col < (sim_width - width_border); col++)
    {
        for (ch_o = 0; ch_o < out_channel; ch_o++)
        {
            output = dst + ch_o * out_width * out_height;
            result = 0;
            for (ch_i = 0; ch_i < in_channel; ch_i++)
            {
                filter = para + ks * ks * (in_channel * ch_o + ch_i);
                //filter = para + ks*ks * (out_channel * ch_i + ch_o);
                input = src + ch_i * in_width * in_height;
                for (x = -r; x <= r; x++)
                {
                    for (y = -r; y <= r; y++)
                    {
                        if ((row + x) >= 0 && (col + y) >= 0 && (row + x) < sim_height && (col + y) < sim_width)
                        {
                            if ( (row + x) % step != 0 || (col + y) % step != 0) continue;
                            result += input[(row + x) / step * in_width + (col + y) / step] * filter[(x + r) * ks + (y + r)];
                        }
                    }
                }
            }

            if (bias != NULL) result = result + bias[ch_o];
            output[(row - height_border) * out_width + (col - width_border)] = result;
        }
    }
  return 0;
}

I compare the result with the caffe's one:

const caffe::vector<caffe::shared_ptr<caffe::Blob<float> > > blobs = layers[i]->blobs();
float *filter = blobs[0]->mutable_cpu_data();
float *bias = blobs[1]->mutable_cpu_data();

caffe::shared_ptr<caffe::Blob<float> > blob;
blob = caffe_net->blob_by_name(np.bottom(0));
deconv_cpu(blob->mutable_cpu_data(), dst, filter, bias, width1, 
height1, c1, width2, height2, c2, ks, pad, stride);

blob = caffe_net->blob_by_name(np.top(0));
if(compare(dst, blob->mutable_cpu_data()) == 0) printf("match\n");
else printf("do not match\n");

However, the code does not give the same result with the caffe's implementation.

Do anyone know what is wrong? Or any advises or comment on the code?

4

1 回答 1

0

这个问题最终通过更改过滤器索引来解决：filter[(rx) * ks + (ry)]

于 2017-09-04T09:09:23.297 回答