# Converting TensorFlow FaceNet Models¶

Public pretrained FaceNet models contain both training and inference part of graph. Switch between this two states is manageable with placeholder value. Intermediate Representation (IR) models are intended for inference, which means that train part is redundant.

There are two inputs in this network: boolean phase_train which manages state of the graph (train/infer) and batch_size which is a part of batch joining pattern.

## Converting a TensorFlow FaceNet Model to the IR¶

To generate a FaceNet OpenVINO model, feed a TensorFlow FaceNet model to Model Optimizer with the following parameters:

 mo
--input_model path_to_model/model_name.pb       \
--freeze_placeholder_with_value "phase_train->False"

The batch joining pattern transforms to a placeholder with the model default shape if --input_shape or --batch */* -b are not provided. Otherwise, the placeholder shape has custom parameters.

• --freeze_placeholder_with_value "phase_train->False" to switch graph to inference mode

• --batch */* -b is applicable to override original network batch

• --input_shape is applicable with or without --input

• other options are applicable