mindspore.nn.Conv1d and mindspore.nn.Conv1dtranspose 's shapes are four-dimensional

[PhyllisJi]

Environment

Hardware Environment(Ascend/GPU/CPU):

CPU and GPU

Software Environment:

• MindSpore version (source or binary): 2.2.14
• Python version (e.g., Python 3.7.5):3.0
• OS platform and distribution (e.g., Linux Ubuntu 16.04):Ubuntu 20.04

Describe the current behavior

The parameter shapes of mindspore.nn.Conv1d and mindspore.nn.Conv1dtranspose are four-dimensional, causing the model to be unable to complete the reading of the parameters. The dimensions described in the official documentation are supposed to be three-dimensional, and other frameworks such as TensorFlow and PyTorch are also three-dimensional.

Describe the expected behavior

Their parameter shapes should be one-dimensional

Related log / screenshot

[VectorSL]

你好。确实存在上述差异，MindSpore的Conv1d底层使用Conv2d算子实现，对于 kernel_size 等参数会在第0维进行自动补0操作，导致初始化的权重weight比PyTorch多一维，正常使用不感知，但涉及网络模型迁移时，通过参数加载的方式调用Conv1d可能会出现权重维度不一致的情况，需要用户手动对weight权重在对应维度上进行升维。 详见 conv1d差异faq