diff options
Diffstat (limited to 'tensorflow/contrib/periodic_resample/ops/array_ops.cc')
| -rw-r--r-- | tensorflow/contrib/periodic_resample/ops/array_ops.cc | 42 |
1 files changed, 28 insertions, 14 deletions
diff --git a/tensorflow/contrib/periodic_resample/ops/array_ops.cc b/tensorflow/contrib/periodic_resample/ops/array_ops.cc index c90fc06c7f..82bd796956 100644 --- a/tensorflow/contrib/periodic_resample/ops/array_ops.cc +++ b/tensorflow/contrib/periodic_resample/ops/array_ops.cc @@ -34,26 +34,40 @@ This function implements a slightly more generic version of the subpixel convolutions found in this [paper](https://arxiv.org/abs/1609.05158). The formula for computing the elements in the `output` tensor is as follows: + `T` = `values` tensor of rank `R` + `S` = desired `shape` of output tensor (vector of length `R`) + `P` = `output` tensor of rank `R` - \((T_1,\ldots,T_R)\) = shape(`T`) - \([S_1,\ldots,S_q,\ldots,S_R]\) = elements of vector `S` - A single element in `S` is left unspecified (denoted \(S_q=-1\)). - Let \(f_i\) denote the (possibly non-integer) factor that relates the original - dimension to the desired dimensions, \(S_i=f_i T_i\), for \(i\neq q\) where - \(f_i>0\). + \\((T_1,\\ldots,T_R)\\) = shape(`T`) + + \\([S_1,\\ldots,S_q,\\ldots,S_R]\\) = elements of vector `S` + + A single element in `S` is left unspecified (denoted \\(S_q=-1\\)). + + Let \\(f_i\\) denote the (possibly non-integer) factor that relates the original + dimension to the desired dimensions, \\(S_i=f_i T_i\\), for \\(i\\neq q\\) where + \\(f_i>0\\). + Define the following: - \(g_i=\lceil f_i\rceil\) - \(t=\prod_i T_i\) - \(s=\prod_{i\neq q} S_i\) - \(S_q\) can then be defined as by \(S_q=\lfloor t/s\rfloor\). + + \\(g_i=\\lceil f_i\\rceil\\) + + \\(t=\\prod_i T_i\\) + + \\(s=\\prod_{i\\neq q} S_i\\) + + \\(S_q\\) can then be defined by \\(S_q=\\lfloor t/s\\rfloor\\). The elements of the resulting tensor are defined as - \(P_{s_1,\ldots,s_R}=T_{h_1,\ldots,h_q,\ldots,h_R}\). - The \(h_i\) (\(i\neq q\)) are defined by \(h_i=\lfloor s_i/g_i\rfloor\). - \(h_q=S_q\sum_{j\neq q}^{q-1}G_j \mathrm{mod}(s_j,g_j) + s_q\), where - \(G_j=\prod_{i}^{j-1}g_i\) (\(G_0=1\)). + + \\(P_{s_1,\\ldots,s_R}=T_{h_1,\\ldots,h_q,\\ldots,h_R}\\). + + The \\(h_i\\) (\\(i\\neq q\\)) are defined by \\(h_i=\\lfloor s_i/g_i\\rfloor\\). + + \\(h_q=S_q\\sum_{j\\neq q}^{q-1}G_j \\mathrm{mod}(s_j,g_j) + s_q\\), where + \\(G_j=\\prod_{i}^{j-1}g_i\\) (\\(G_0=1\\)). One drawback of this method is that whenever the output dimensions are slightly less than integer multiples of the input dimensions, many of the tensor elements |