Edit TensorFlow model for training

Before adding a TensorFlow model to IBM Spectrum Conductor Deep Learning Impact, edit the model to enable distributed training capabilities or deep learning insights.

By default, IBM Spectrum Conductor Deep Learning Impact supports single-node training for TensorFlow models without deep learning insights. To change the training engine or to add deep learning insight capabilities the model must be configured accordingly.

IBM Spectrum Conductor Deep Learning Impact supports the following training engines:
  • Single node training
  • Distributed training with TensorFlow
  • Distributed training with IBM Fabric
IBM Spectrum Conductor Deep Learning Impact supports single-node training for TensorFlow models with no extra configurations needed. For distributed engines, some additional configurations are required.
Each TensorFlow model has the following files:
  • main.py: TensorFlow train model program main entrance
  • inference.py: TensorFlow inference model program main entrance
  • fabricmodel.py: Callback program to convert training model into TensorFlow compute graph
  • ps.conf: Training parameters which is optional.
Both main.py and inference.py are mandatory for a TensorFlow model for any training engine. The fabricmodel.py is required if you want to run distributed training with IBM Fabric. To enable distributed training with IBM Fabric, see Edit a TensorFlow training model for distributed training with IBM Fabric.

To get the deep learning insights feature for TensorFlow models, edit the TensorFlow model to include deep learning insights, see Edit a TensorFlow training model for deep learning insights.