I am confused as to how the model trains on the logit layer. If logits are a result of the sigmoid activation function, then what are the trainable variables? Logits and then softmax, but what part is being trained if the hidden layers are frozen?
output_layer_vars = tf.get_collection(tf.TRAINABLE_VARIABLES, scope='logits')
optimizer = tf.train.AdamOptimizer(learning_rate, name='Adam2')
training_op = optimizer.minimize(loss, var_list=output_layer_vars)```
