I have created a Squeezenet model using Facenet I'd like to convert my Squeezenet tensorflow model using the Qualcomm NPE. The problem is, that Qualcomm's NPE only supports a 3D tensor as an input. Squeezenet uses an image tensor plus a phase_train tensor (4D tensor) as an input. The model can be found here: squeezenet.py
My question is: How can I remove the phase_train tensor from the squeezenet model in order to input only an image? Maybe one can remove the phase_train during freezing of the graph?
I think both inputs tensors are defined in line 46:
with tf.variable_scope('squeezenet', [images], reuse=reuse):
with slim.arg_scope([slim.batch_norm, slim.dropout], is_training=phase_train):