megengine¶

import megengine as mge  # NOT as torch - Even if we look like twins.

Note

MGE 或 mge 是 MegEngine 的官方缩写，我们建议相关衍生库将其作为前缀。

Warning

不要尝试 [ import megengine as torch ] 这是行不通的！ ( /ω＼)

MegEngine 的部分 API 设计借鉴了 PyTorch, 我们相信这对开发者而言会更加友好；
但二者的底层设计完全不同，MegEngine is just MegEngine.

Tensor¶

from megengine import Tensor
from megengine import tensor  # tensor is an alias of Tensor

MegEngine 中提供了一种名为 “张量” （Tensor ）的数据结构，区别于物理学中的定义，其概念与 NumPy 中的 ndarray 更加相似，即张量是一类多维数组，其中每个元素的数据类型和空间大小一致，而数据值可以不同。

`Tensor`	A tensor object represents a multidimensional, homogeneous array of fixed-size items.
`Parameter`	A kind of Tensor that is to be considered a module parameter.

上面 Parameter 可以被看作是一种特殊的 Tensor, 通常用来表示神经网络中的参数。

想要了解更多，请参考 Deep understanding of Tensor data structure 。

Serialization¶

`save`	Save an object to disk file.
`load`	Load an object saved with `save` from a file.

Device¶

`device.what_is_xpu`	Return the precise device type like `cpu`, `cuda` and so on.
`is_cuda_available`	Returns whether cuda device is available on this system.
`is_atlas_available`	Returns whether atlas device is available on this system.
`is_cambricon_available`	Returns whether cambricon device is available on this system.
`is_rocm_available`	Returns whether rocm device is available on this system.
`get_device_count`	Gets number of devices installed on this system.
`get_default_device`	Gets default computing node.
`set_default_device`	Sets default computing node.
`get_mem_status_bytes`	Get total and free memory on the computing device in bytes.
`get_cuda_device_property`
`get_cuda_compute_capability`	Gets compute capability of the specified device.
`get_allocated_memory`	Returns the current memory occupied by tensors on the computing device in bytes.
`get_reserved_memory`	Returns the current memory managed by the caching allocator on the computing device in bytes.
`get_max_reserved_memory`	Returns the maximum memory managed by the caching allocator on the computing device in bytes.
`get_max_allocated_memory`	Returns the maximum memory occupied by tensors on the computing device in bytes.
`reset_max_memory_stats`	Resets the maximum stats on the computing device.
`set_prealloc_config`	Specifies how to pre-allocate from raw device allocator.
`coalesce_free_memory`	This function will try it best to free all consecutive free chunks back to operating system, small pieces may not be returned.
`get_cuda_version`	Gets the CUDA version used when compiling MegEngine.
`get_cudnn_version`	Get the Cudnn version used when compiling MegEngine.
`get_tensorrt_version`	Get the TensorRT version used when compiling MegEngine.

Logger¶

`enable_debug_log`	Sets logging level to debug for all components.
`get_logger`	Gets megengine logger with given name.
`set_log_file`	Sets log output file.
`set_log_level`	Sets default logging level.

Version¶

__version__ （MegEngine 版本号）¶

API References

megengine.Tensor