Merge pull request #15679 from jacquesqiao/update-lookup_table_grad-padding-index

lookup_table_grad kernel should consider padding_idx test=develop

paddle/fluid/operators/lookup_table_op.h

@@ -129,6 +129,7 @@ class LookupTableGradKernel : public framework::OpKernel<T> {
           "must be either LoDTensor or SelectedRows");
     }
 
+    int64_t padding_idx = context.Attr<int64_t>("padding_idx");
     bool is_sparse = context.Attr<bool>("is_sparse");
     // Since paddings are not trainable and fixed in forward, the gradient of
     // paddings makes no sense and we don't deal with it in backward.
@@ -187,6 +188,10 @@ class LookupTableGradKernel : public framework::OpKernel<T> {
       memset(d_table_data, 0, d_table->numel() * sizeof(T));
 
       for (int64_t i = 0; i < ids->numel(); ++i) {
+        if (padding_idx != kNoPadding && ids_data[i] == padding_idx) {
+          // the gradient of padding_idx should be 0, already done by memset, so
+          // do nothing.
+        } else {
           PADDLE_ENFORCE_LT(ids_data[i], N);
           PADDLE_ENFORCE_GE(ids_data[i], 0);
           for (int j = 0; j < D; ++j) {
@@ -195,6 +200,7 @@ class LookupTableGradKernel : public framework::OpKernel<T> {
         }
       }
     }
+  }
 };
 
 }  // namespace operators