fix alpha compositing

Summary:
Fix division by zero when alpha is 1.0
In this case, the nominator is already 0 and we need to make sure division with 0 does not occur which would produce nans

Reviewed By: nikhilaravi

Differential Revision: D21650478

fbshipit-source-id: bc457105b3050fef1c8bd4e58e7d6d15c0c81ffd

pytorch3d/csrc/compositing/alpha_composite.cu

@@ -11,6 +11,8 @@
 #include <stdio.h>
 #include <vector>
 
+__constant__ const float kEpsilon = 1e-9;
+
 // TODO(gkioxari) support all data types once AtomicAdd supports doubles.
 // Currently, support is for floats only.
 __global__ void alphaCompositeCudaForwardKernel(
@@ -126,7 +128,7 @@ __global__ void alphaCompositeCudaBackwardKernel(
         atomicAdd(
             &grad_alphas[batch][t][j][i],
             -grad_outputs[batch][ch][j][i] * features[ch][n_idx] * cum_alpha *
-                alpha / (1 - alpha_tvalue));
+                alpha / (1 - alpha_tvalue + kEpsilon));
       }
 
       cum_alpha = cum_alpha * (1 - alphas[batch][k][j][i]);

pytorch3d/csrc/compositing/alpha_composite_cpu.cpp

@@ -5,6 +5,9 @@
 #include <cmath>
 #include <vector>
 
+// Epsilon float
+const float kEps = 1e-9;
+
 torch::Tensor alphaCompositeCpuForward(
     const torch::Tensor& features,
     const torch::Tensor& alphas,
@@ -101,7 +104,8 @@ std::tuple<torch::Tensor, torch::Tensor> alphaCompositeCpuBackward(
               }
               float alpha_tvalue = alphas_a[b][t][j][i];
               grad_alphas_a[b][t][j][i] -= grad_outputs_a[b][c][j][i] *
-                  features_a[c][n_idx] * cum_alpha * alpha / (1 - alpha_tvalue);
+                  features_a[c][n_idx] * cum_alpha * alpha /
+                  (1 - alpha_tvalue + kEps);
             }
 
             cum_alpha = cum_alpha * (1 - alpha);