Operators Compile
Linux
Windows
Ascend
GPU
CPU
Environment Preparation
Basic
Intermediate
Q: What is the function of the TransData
operator? Can the performance be optimized?
A: The TransData
operator is used in the scenario where the data formats (such as NC1HWC0) used by interconnected operators on the network are inconsistent. In this case, the framework automatically inserts the TransData
operator to convert the data formats into the same format and then performs computation. Huawei Ascend supports 5D format operations, and uses the transdata
operator to convert data from 4D to 5D to improve performance.
Q: An error occurs when the Concat
operator concatenates tuples containing multiple tensors. An error occurs when the number of tensor list
elements entered is greater than or equal to 192. What is a better solution (running in dynamic mode) for Concat
to concatenate tuples containing multiple Tensors?
A: The number of tensors to be concatenated at a time cannot exceed 192 according to the bottom-layer specifications of the Ascend operator. You can try to concatenate them twice.
Q: When Conv2D
is used to define convolution, the group
parameter is used. Is it necessary to ensure that the value of group
can be exactly divided by the input and output dimensions? How is the group parameter transferred?
A: The Conv2d
operator has the following constraint: When the value of group
is greater than 1, the value must be the same as the number of input and output channels. Do not use ops.Conv2D
. Currently, this operator does not support a value of group
that is greater than 1. Currently, only the nn.Conv2d
API of MindSpore supports group
convolution. However, the number of groups must be the same as the number of input and output channels.
Q: Does MindSpore support matrix transposition?
A: Yes. For details, see mindspore.ops.Transpose.
Q: Can MindSpore calculate the variance of any tensor?
A: Currently, mindspore does not have APIs or operators similar to variance which can directly calculate the variance of a tensor
. However, MindSpore has sufficient small operators to support such operations. For details, see class Moments(Cell).
Q: Compared with PyTorch, the nn.Embedding
layer lacks the padding operation. Can other operators implement this operation?
A: In PyTorch, padding_idx
is used to set the word vector in the padding_idx
position in the embedding matrix to 0, and the word vector in the padding_idx
position is not updated during backward propagation.
In MindSpore, you can manually initialize the weight corresponding to the padding_idx
position of embedding to 0. In addition, the loss corresponding to padding_idx
is filtered out through the mask operation during training.
Q: When the Tile
module in operations executes __infer__
, the value
is None
. Why is the value lost?
A: The multiples input
of the Tile
operator must be a constant. (The value cannot directly or indirectly come from the input of the graph.) Otherwise, the None
data will be obtained during graph composition because the graph input is transferred only during graph execution and the input data cannot be obtained during graph composition.
Q: When conv2d is set to (3,10), Tensor[2,2,10,10] and it runs on Ascend on ModelArts, the error message FM_W+pad_left+pad_right-KW>=strideW
is displayed. However, no error message is displayed when it runs on a CPU. What should I do?
A: TBE (Tensor Boost Engine) operator is Huawei’s self-developed Ascend operator development tool, which is extended on the basis of the TVM framework to develop custom operators. The above problem is the limitation of this TBE operator, the width of x must be greater than the width of the kernel. The CPU operator does not have this restriction, so no error is reported.
Q: Has MindSpore implemented the anti-pooling operation similar to nn.MaxUnpool2d
?
A: Currently, MindSpore does not provide anti-pooling APIs but you can customize the operator to implement the operation. For details, refer to Custom Operators.
Q: What can I do if the error message Pynative run op ExpandDims failed
is displayed when the ExpandDims operator is used? The code is as follows:
context.set_context(mode=context.GRAPH_MODE,device_target='Ascend')
input_tensor=Tensor(np.array([[2,2],[2,2]]),mindspore.float32)
expand_dims=ops.ExpandDims()
output=expand_dims(input_tensor,0)
A: The problem is that the Graph mode is selected but the PyNative mode is used. As a result, an error is reported. MindSpore supports the following running modes which are optimized in terms of debugging or running:
PyNative mode: dynamic graph mode. In this mode, operators in the neural network are delivered and executed one by one, facilitating the compilation and debugging of the neural network model.
Graph mode: static graph mode. In this mode, the neural network model is compiled into an entire graph and then delivered for execution. This mode uses technologies such as graph optimization to improve the running performance and facilitates large-scale deployment and cross-platform running.
You can select a proper mode and writing method to complete the training by referring to the official website tutorial.