Hi,dhungle!

In the class visualization part, we want to generate a image to maxmize the loss which includes the score value (we got from CNN model) and regularization part. Right?

But in your implementation, you wrote as following: loss = model.scores[0, target_y] - l2_reg * tf.reduce_sum(model.scores**2), I think it should be loss = model.scores[0, target_y] - l2_reg * tf.reduce_sum(model.image**2) cause the regularization part of loss is for generated image.

You might get a nicer picture after modifying this. my pictures

Did you experienced any slow down after implementing the batch-norm while running a non-batch-norm solver?

I know this question might sound stupid, by I am currently struggling with it.
Thanks