TARGET_LOSS

Description

In the TensorFlow 1.15 training scenario, if the training acceleration function is enabled through the experimental_accelerate_train_mode or accelerate_train_mode parameter, you can use this environment variable to set the target training loss value on the NPU.

• If the training script uses the session configuration, see the description of the experimental_accelerate_train_mode parameter in "Session Configuration > Session Configuration Parameters > Experiment Parameters" in TF Adapter API (1.x).
• If NPURunConfig is used in the training script, see the description of the accelerate_train_mode parameter in "npu_bridge.estimator.npu.npu_config > ExperimentalConfig Constructor" in TF Adapter API (1.x).

The value of this environment variable is of the floating point type. There is no default value.

Example

export TARGET_LOSS=3.0

Restrictions

This environment variable applies only to the scenario where the TensorFlow 1.15 network is trained on the Ascend platform.

Applicability

Atlas Training Series Product