Enhance the unit test framework to explicitly test whether the operator correctly handles gradients for multiple inputs.

Fix conv3d DeConv3d (bias shape)

Invoke check_grad many times for no_grad_set

add missing protostr for sub_nested_seq_layer.

package a new USE_NO_KERNEL_OP for USE_OP_ITSELF

Add a graph building example

Revert "Remove `grad_op_type` in `REGISTER_OP`"

fix ctc edit distance evaluator can not print information in v2 API.

cuDNN Wrapper.

rename mse_cost into square_error_cost.

add op() to InferShapeContext

Make operator Input/Output can return nullptr

Design of if else op

Lod Tensor design doc