auto_decomposition

Applicability

Product	Supported
Atlas 350 Accelerator Card	√
Atlas A3 training product/Atlas A3 inference product	√
Atlas A2 training product/Atlas A2 inference product	√
Atlas 200I/500 A2 inference product	√
Atlas inference product	√
Atlas training product	√

Description

Performs tensor decomposition on the original TensorFlow model, generating a new model file.

Prototype

add_ops = auto_decomposition(meta_path, ckpt_path, save_path)

Parameters

Parameter	Input/Output	Description
meta_path	Input	Path of the TensorFlow model definition file (.meta) to be decomposed. A string. NOTE: Ensure that the .meta file can be properly loaded by tf.compat.v1.train.import_meta_graph. For example, for a .meta file obtained through Horovod training, you must run the import horovod.tensorflow command for the file to be successfully called by the auto_decomposition API.
ckpt_path	Input	Path of the original TensorFlow model weight file. Set this parameter to the common path prefix (including the file name without file name extensions) of the .data-XXXXX-of-XXXXX file and .index file. For example, the two files are path/model-200.data-00000-of-00001 and path/model-200.index. Then ckpt_path should be set to path/model-200. A string.
save_path	Input	Save path, for storing the tensor decomposition resultant files such as .data-XXXXX-of-XXXXX. The resultant files are named after this value. For example, if this parameter is set to path/model, the generated .data-XXXXX-of-XXXXX and .index files are named as follows: path/model.data-XXXXX-of-XXXXX and path/model.index. A string.

Parameter

Input/Output

Description

meta_path

Input

Path of the TensorFlow model definition file (.meta) to be decomposed.

A string.

NOTE:

Ensure that the .meta file can be properly loaded by tf.compat.v1.train.import_meta_graph.

For example, for a .meta file obtained through Horovod training, you must run the import horovod.tensorflow command for the file to be successfully called by the auto_decomposition API.

ckpt_path

Input

Path of the original TensorFlow model weight file. Set this parameter to the common path prefix (including the file name without file name extensions) of the .data-XXXXX-of-XXXXX file and .index file.

For example, the two files are path/model-200.data-00000-of-00001 and path/model-200.index.

Then ckpt_path should be set to path/model-200.

A string.

save_path

Input

Save path, for storing the tensor decomposition resultant files such as .data-XXXXX-of-XXXXX. The resultant files are named after this value.

For example, if this parameter is set to path/model, the generated .data-XXXXX-of-XXXXX and .index files are named as follows:

path/model.data-XXXXX-of-XXXXX and path/model.index.

A string.

Returns

Returns a list of names of new convolutional layers after tensor decomposition. The data type is list.

Restrictions

Ensure the input TensorFlow model files match each other: model definition file (.meta), model weight file (.data-XXXXX-of-XXXXX), and weight index file (.index).
Pass the path of the original model and the output path to the tensor decomposition API. This API automatically decomposes the convolutional layers that meet the decomposition conditions. For details about the decomposition conditions, see Restrictions.

Example

from amct_tensorflow.tensor_decompose import auto_decomposition
auto_decomposition(meta_path='src_path/model.meta',
                   ckpt_path='src_path/model',
                   save_path='decomposed_path/model'
)

Flush files:

A checkpoint list file
A model definition file (.meta)
A model weight file (.data-00000-of-00001)
A weight index file (.index)
A graph change history file (.pkl), which is used by the decompose_graph API to modify the graph defined by the source training code.

Parent topic: Tensor Decomposition APIs