TF-Slim uses its own training loop. However, I would like to use a TF-Slim model (Resnet50) while still using my own tensorflow training loop. The TF-Slim model simply outputs the prediction and I calculate my own total loss. I am able to train the model without error and training error seems to converge. I am asking because I experienced issues with batch normalization during evaluation (the error is very high compared to the training error). I found out that this might be due to insufficient training steps. But I want to make sure that I am not using TF-Slim incorrectly.
The TF-Slim training procedure looks like this:
#create_train_op ensures that each time we ask for the loss, the
update_ops
# are run and the gradients being computed are applied too.
train_op = slim.learning.create_train_op(total_loss, optimizer)
logdir = ... # Where checkpoints are stored.
slim.learning.train(
train_op,
logdir,
number_of_steps=1000,
save_summaries_secs=300,
save_interval_secs=600):
I don't want to use the train_opbut something like this
def update_gradients(update_ops, optimizer, total_loss, variables_to_train, global_step, summaries):
for grad, var in gradients:
if grad is not None:
summaries.add(tf.summary.histogram(var.op.name + '/gradients', grad))
grad_updates = optimizer.apply_gradients(gradients,
global_step=global_step)
update_ops.append(grad_updates)
update_op = tf.group(*update_ops)
with tf.control_dependencies([update_op]):
train_tensor = tf.identity(total_loss, name='train_op')
return train_tensor
and then call sess.run(train_tensor)
Does this cause any issues internally? I read here that one should use train_op: github issues
Or is it simply not allowed to pass, for example, train_tensor into the
slim.learning.train() function directly?
