[numpy] Fix unit tests after using numpy-compatible shape definition

src/c_api/c_api_common.h

@@ -91,8 +91,9 @@ struct MXAPIThreadLocalEntry {
     data->resize(shapes.size());
     size_t size = 0;
     for (const auto& s : shapes) {
-      if (s.ndim() > 0);
-      size += s.ndim();
+      if (s.ndim() > 0) {
+        size += s.ndim();
+      }
     }
     buffer->resize(size);
     int *ptr = buffer->data();

src/common/utils.h

@@ -746,6 +746,11 @@ inline void ParallelCopy(DType* dst, const DType* src, index_t size) {
  * 4. -1 dim size means the dimension's size is unknown.
  * so that operator's infer shape function can work in backend.
  * \param shape to be converted.
+ * Note: It is possible that the shape to be converted is already
+ * numpy compatible. For example, when a subgraph operator's infer
+ * shape function is called from the infer shape pass of the whole
+ * graph, its input/output shapes have been converted to numpy
+ * compatible shapes.
  */
 inline void ConvertToNumpyShape(mxnet::TShape* shape) {
   if (shape->ndim() == 0) {  // legacy shape ndim = 0 means unknown

src/operator/leaky_relu-inl.h

@@ -338,10 +338,10 @@ class LeakyReLUProp : public OperatorProperty {
       CHECK_EQ(in_shape->size(), 1U) << "Input:[data]";
     }
     const mxnet::TShape &dshape = in_shape->at(leakyrelu::kData);
-    if (dshape.ndim() == 0) return false;
+    if (!mxnet::ndim_is_known(dshape)) return false;
     if (param_.act_type == leakyrelu::kPReLU) {
       const mxnet::TShape &gshape = in_shape->at(leakyrelu::kGamma);
-      if (gshape.ndim() == 0) {
+      if (!mxnet::ndim_is_known(gshape)) {
         in_shape->at(leakyrelu::kGamma) = mxnet::TShape(Shape1(dshape[1]));
       }
       if (dshape == gshape) {

src/operator/nn/deconvolution-inl.h

@@ -134,11 +134,13 @@ struct DeconvolutionParam : public dmlc::Parameter<DeconvolutionParam> {
       for (size_t i = 0; i < ndim; i++) {
         // input.ndim() can be larger than ndim, in case that the complete input
         // shape was passed and not only the ndim last ones
-        o_pad[i] = stride[i] * (input[(input_ndim - ndim) + i] - 1) + DilatedKernelSize(i);
-        CHECK_GE(o_pad[i], target_shape[i]) << "too big target shape";
-        o_pad[i] -= target_shape[i];
-        o_adj[i] = o_pad[i] % 2;
-        o_pad[i] = (o_pad[i] + 1) / 2;
+        if (mxnet::dim_size_is_known(input, input_ndim - ndim + i)) {
+          o_pad[i] = stride[i] * (input[(input_ndim - ndim) + i] - 1) + DilatedKernelSize(i);
+          CHECK_GE(o_pad[i], target_shape[i]) << "too big target shape";
+          o_pad[i] -= target_shape[i];
+          o_adj[i] = o_pad[i] % 2;
+          o_pad[i] = (o_pad[i] + 1) / 2;
+        }
       }
     } else {
       for (size_t i = 0; i < ndim; i++) {

src/operator/nn/deconvolution.cc

@@ -54,7 +54,7 @@ static bool DeconvolutionShape(const nnvm::NodeAttrs& attrs,
   }
   out_shape->resize(1, mxnet::TShape());
   const mxnet::TShape &dshape = (*in_shape)[deconv::kData];
-  if (!shape_is_known(dshape)) return false;
+  if (!mxnet::ndim_is_known(dshape)) return false;
 
   if (param_.kernel.ndim() == 1) {
     // 1d conv
@@ -90,8 +90,12 @@ static bool DeconvolutionShape(const nnvm::NodeAttrs& attrs,
     Shape<3> oshape;
     oshape[0] = dshape_ncw[0];
     oshape[1] = param_.num_filter;
-    oshape[2] = param_.stride[0] * (dshape_ncw[2] - 1) +
-      dilated_ksize_x - 2 * o_pad[0] + o_adj[0];
+    if (mxnet::dim_size_is_known(dshape_ncw[2])) {
+      oshape[2] = param_.stride[0] * (dshape_ncw[2] - 1) +
+          dilated_ksize_x - 2 * o_pad[0] + o_adj[0];
+    } else {
+      oshape[2] = -1;
+    }
 
     if (param_.target_shape.ndim() > 0) {
       if (param_.target_shape[0] > 0) {
@@ -141,10 +145,18 @@ static bool DeconvolutionShape(const nnvm::NodeAttrs& attrs,
     Shape<4> oshape;
     oshape[0] = dshape_nchw[0];
     oshape[1] = param_.num_filter;
-    oshape[2] = param_.stride[0] * (dshape_nchw[2] - 1) +
-      dilated_ksize_y - 2 * o_pad[0] + o_adj[0];
-    oshape[3] = param_.stride[1] * (dshape_nchw[3] - 1) +
-      dilated_ksize_x - 2 * o_pad[1] + o_adj[1];
+    if (mxnet::dim_size_is_known(dshape_nchw[2])) {
+      oshape[2] = param_.stride[0] * (dshape_nchw[2] - 1) +
+          dilated_ksize_y - 2 * o_pad[0] + o_adj[0];
+    } else {
+      oshape[2] = -1;
+    }
+    if (mxnet::dim_size_is_known(dshape_nchw[3])) {
+      oshape[3] = param_.stride[1] * (dshape_nchw[3] - 1) +
+          dilated_ksize_x - 2 * o_pad[1] + o_adj[1];
+    } else {
+      oshape[3] = -1;
+    }
 
     if (param_.target_shape.ndim() > 1) {
       if (param_.target_shape[0] > 0) {
@@ -203,12 +215,24 @@ static bool DeconvolutionShape(const nnvm::NodeAttrs& attrs,
     Shape<5> oshape;
     oshape[0] = dshape_ncdhw[0];
     oshape[1] = param_.num_filter;
-    oshape[2] = param_.stride[0] * (dshape_ncdhw[2] - 1) +
-      dilated_ksize_d - 2 * o_pad[0] + o_adj[0];
-    oshape[3] = param_.stride[1] * (dshape_ncdhw[3] - 1) +
-      dilated_ksize_y - 2 * o_pad[1] + o_adj[1];
-    oshape[4] = param_.stride[2] * (dshape_ncdhw[4] - 1) +
-      dilated_ksize_x - 2 * o_pad[2] + o_adj[2];
+    if (mxnet::dim_size_is_known(dshape_ncdhw[2])) {
+      oshape[2] = param_.stride[0] * (dshape_ncdhw[2] - 1) +
+          dilated_ksize_d - 2 * o_pad[0] + o_adj[0];
+    } else {
+      oshape[2] = -1;
+    }
+    if (mxnet::dim_size_is_known(dshape_ncdhw[3])) {
+      oshape[3] = param_.stride[1] * (dshape_ncdhw[3] - 1) +
+          dilated_ksize_y - 2 * o_pad[1] + o_adj[1];
+    } else {
+      oshape[3] = -1;
+    }
+    if (mxnet::dim_size_is_known(dshape_ncdhw[4])) {
+      oshape[4] = param_.stride[2] * (dshape_ncdhw[4] - 1) +
+          dilated_ksize_x - 2 * o_pad[2] + o_adj[2];
+    } else {
+      oshape[4] = -1;
+    }
 
     if (param_.target_shape.ndim() > 2) {
       if (param_.target_shape[0] > 0) {

src/operator/tensor/matrix_op-inl.h

@@ -594,7 +594,7 @@ void SliceCsrImpl(const SliceParam &param, const OpContext& ctx,
   mxnet::TShape begin(N, -1), end(N, -1);
   for (int i = 0; i < N; ++i) {
     int s = 0;
-    if (param.begin[i]) {
+    if (i < param.begin.ndim() && param.begin[i]) {
       s = *param.begin[i];
       if (s < 0) s += ishape[i];
     }

tests/python/unittest/test_ndarray.py

@@ -120,7 +120,11 @@ def test_ndarray_setitem():
 
     # numpy assignment for empty axis
     for trivial_shape in [(), (1,), (1, 1), (1, 1, 1)]:
-        x = mx.nd.zeros(trivial_shape)
+        if trivial_shape == tuple():
+            with mx.numpy.enable_np_comp():
+                x = mx.nd.zeros(trivial_shape)
+        else:
+            x = mx.nd.zeros(trivial_shape)
         x[:] = np.ones(trivial_shape)
         x_np = np.ones(trivial_shape, dtype=x.dtype)
         assert x.shape == trivial_shape