Fix

apache · Feb 14, 2020 · a245021 · a245021
1 parent cc236ae
commit a245021
Show file tree

Hide file tree

Showing 2 changed files with 14 additions and 15 deletions.
diff --git a/src/operator/numpy/random/np_weibull_op.h b/src/operator/numpy/random/np_weibull_op.h
@@ -61,9 +61,9 @@ struct NumpyWeibullParam : public dmlc::Parameter<NumpyWeibullParam> {
 
 template <typename DType>
 struct scalar_weibull_kernel {
-  MSHADOW_XINLINE static void Map(index_t i, float a, float *threshold,
+  MSHADOW_XINLINE static void Map(index_t i, float a, float *noise,
                                   DType *out) {
-    out[i] = powf(-log(threshold[i]), DType(1.0/a));
+    out[i] = powf(-log(noise[i]), DType(1.0/a));
   }
 };
 
@@ -84,11 +84,13 @@ struct weibull_kernel {
   MSHADOW_XINLINE static void Map(index_t i,
                                   const Shape<ndim> &stride,
                                   const Shape<ndim> &oshape,
-                                  IType *aparams, float* threshold, OType *out) {
+                                  IType *aparams, float *noise, OType *out) {
     Shape<ndim> coord = unravel(i, oshape);
     auto idx = static_cast<index_t>(dot(coord, stride));
-    threshold[i] = -log(threshold[i]);
-    out[i] =  powf(threshold[i], IType(1.0/aparams[idx]));
+    noise[i] = -log(noise[i]);
+    out[i] =  powf(noise[i], IType(1.0/aparams[idx]));
+    // get grad
+    noise[i] = -log(noise[i]) * out[i] * (1.0/(aparams[idx] * aparams[idx]));
   }
 };
 
@@ -167,10 +169,8 @@ inline void ScalarWeibullReparamBackwardImpl(const OpContext& ctx,
       ReduceWorkspaceSize<ndim, DType>(s, igrad.shape_, req[0], ograd.shape_);
   Tensor<xpu, 1, char> workspace =
       ctx.requested[0].get_space_typed<xpu, 1, char>(Shape1(workspace_size), s);
-  // igrad=sum(ograd*samples*log(noise)*reciprocal_grad_a)
-  // temp = sum(ograd*samples*noise) 
-  Reduce<red::sum, ndim, DType, op::mshadow_op::mul, op::mshadow_op::mul>(
-      s, igrad, req[0], workspace, ograd, noise, samples);
+  Reduce<red::sum, ndim, DType, op::mshadow_op::mul, op::mshadow_op::left>(
+      s, igrad, req[0], workspace, ograd, noise, noise);
   }
 
 template<typename xpu>

diff --git a/tests/python/unittest/test_numpy_op.py b/tests/python/unittest/test_numpy_op.py
@@ -3941,8 +3941,8 @@ def hybrid_forward(self, F, a):
 
     output_shapes = [
         (3, 2),
-        # (4, 3, 2, 2),
-        # (3, 4, 5)
+        (4, 3, 2, 2),
+        (3, 4, 5)
     ]
     for hybridize in [False, True]: 
         for out_shape in output_shapes:
@@ -3954,12 +3954,11 @@ def hybrid_forward(self, F, a):
             with mx.autograd.record():
                 mx_out = test_w_grad(a)
             mx_out.backward()
-            
+
             # gradient formula calculus (a=1)
-            # formula_grad = - np.sum(mx_out * np.log(mx_out))
+            formula_grad = - mx_out * np.log(mx_out)
             assert a.grad.shape == out_shape
-            # assert_almost_equal(a.grad.asnumpy().sum(), formula_grad.asnumpy().sum(), rtol=1e-3, atol=1e-5)
-            # assert_almost_equal(a.grad.asnumpy().sum(), (mx_out*mx_out).asnumpy().sum(), rtol=1e-3, atol=1e-5)
+            assert_almost_equal(a.grad.asnumpy().sum(), formula_grad.asnumpy().sum(), rtol=1e-3, atol=1e-5)
 
 
 @with_seed()