Why validate on LFW with phase_train_placeholder:True?

This is relative to #56, #72. I check TensorFlow-Slim image classification library Evaluation Code and found setting is_training is to False in the eval_image_classifier.py. ` #################### # Select the model # #################### network_fn = nets_factory.get_network_fn( FLAGS.model_name, num_classes=(dataset.num_classes - FLAGS.labels_offset), is_training=False)

##############################################################
# Create a dataset provider that loads data from the dataset #
##############################################################
provider = slim.dataset_data_provider.DatasetDataProvider(
    dataset,
    shuffle=False,
    common_queue_capacity=2 * FLAGS.batch_size,
    common_queue_min=FLAGS.batch_size)
[image, label] = provider.get(['image', 'label'])
label -= FLAGS.labels_offset

#####################################
# Select the preprocessing function #
#####################################
preprocessing_name = FLAGS.preprocessing_name or FLAGS.model_name
image_preprocessing_fn = preprocessing_factory.get_preprocessing(
    preprocessing_name,
    is_training=False)

`

So I think setting is_training with False is true when embedding for new face. If validate on LFW with is_training is False, accuracy is low 67%. Is it mean this model is not good trainable?

[LiuzcEECS]

I think the decay in moving_average was too high (0.9997) (seems to be fixed yesterday) to record the variance and as a result, the moving_variance kept unchanged during the training. When we validate the model on LFW with phase_train_placeholder:True, tensorflow doesn’t use the moving_variance we saved. Instead, it just calculates mean and variance of the mini-batch by tf.nn.moments. So my solution is to just set the phase_train_placeholder:True or you can write another batch_norm module which use tf.nn.moments when phase_train_placeholder is set to False instead of using the moving_variance we saved as tf.Slim does.

[davidsandberg]

Closing due to inactivity. Open if needed.