RoundToNearest
, which realizes the lossless compression parameters of Llama2 7B/13B/70B and Baichuan2 13B networks by over 40%.PTQConfig
to configure the post-training quantization algorithm.PTQMode
enumeration class, which can be configured in 'PTQConfig', is used to distinguish between the two phases of the quantization algorithm: the quantization phase and the deployment phase.BackendTarget
enumeration class, which can be configured in PTQConfig
, to indicate the backend to which the quantized network will eventually be deployed. For example, 'BackendTarget.Ascend' indicates that it will eventually be deployed to the Ascend backend of MindSpore.Thanks goes to these wonderful people:
zhuxiaoxiong, hangangqiang
Contributions of any kind are welcome!
Thanks goes to these wonderful people:
hangangqiang, yangruoqi713, kevinkunkun.
Contributions of any kind are welcome!
set_enable_bn_calibration
api to enable BatchNorm calibration. For a network with a BatchNorm layer, the BatchNorm calibration can reduces the decrease in network accuracy caused by the SLB quantization algorithm. (!150)set_enable_bn_calibration
interface to enable or disable BatchNorm calibration.(!117)convert
interface to the algorithm base class, which is configured to convert training network to inferring network. And the network will be exported to MindIR file for Deployment. For details, please refer to Model Deployment.(!176)set_save_mindir
interface to the algorithm base class, which is configured to automatically export MindIR after training. For details, please refer to Model Deployment.(!168)Thanks goes to these wonderful people:
liuzhicheng01, fuzhongqian, hangangqiang, yangruoqi713, kevinkunkun.
Contributions of any kind are welcome!
callbacks
interface to the algorithm base class, which returns the callback logic of the algorithm which will be called during the training process. In order to facilitate different algorithms to implement their own callback logic, this method has variable parameter inputs.(!117)set_epoch_size
interface, which is used to configure the total number of epochs of training, and is used to implement the temperature adjustment callback logic.(!117)set_has_trained_epoch
interface. If a pre-trained checkpoint is used in training, it is used to configure the number of pre-trained epochs corresponding to the pre-trained checkpoint used in the current training, which is used to implement the temperature adjustment callback logic.(!117)set_t_start_val
interface, which is used to configure the initialization value of the temperature in the temperature adjustment mechanism, and is used to implement the temperature adjustment callback logic.(!117)set_t_start_time
interface, which is used to configure the time when the temperature adjustment mechanism start to work, and is used to implement the temperature adjustment callback logic.(!117)set_t_end_time
interface, which is used to configure the time when the temperature adjustment mechanism stop to work, and is used to implement the temperature adjustment callback logic.(!117)set_t_factor
interface, which is used to configure the temperature adjustment factor in the temperature adjustment mechanism, and is used to implement the temperature adjustment callback logic.(!117)Thanks goes to these wonderful people:
ghostnet, liuzhicheng01, fuzhongqian, hangangqiang, cjh9368, yangruoqi713, kevinkunkun.
Contributions of any kind are welcome!
MindSpore Golden Stick is a model compression algorithm set jointly designed and developed by Huawei's Noah team and Huawei's MindSpore team. MindSpore Golden Stick provides an unified user interface allowing users to apply model compression algorithms such as quantization and pruning in a unified and convenient manner. MindSpore Golden Stick also provides front-end network modification capabilities to reduce algorithm development costs. MindSpore Golden Stick provides three algorithms in current version.
Thanks goes to these wonderful people:
ghostnet, liuzhicheng01, fuzhongqian, hangangqiang, cjh9368.
Contributions of any kind are welcome!
此处可能存在不合适展示的内容,页面不予展示。您可通过相关编辑功能自查并修改。
如您确认内容无涉及 不当用语 / 纯广告导流 / 暴力 / 低俗色情 / 侵权 / 盗版 / 虚假 / 无价值内容或违法国家有关法律法规的内容,可点击提交进行申诉,我们将尽快为您处理。